Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Response to Amendment
The amendments filed on 05/08/2026 have been considered. Claims 1, 3, 5-10, 12, and 14-19 are amended. Claims 2, 4, 11, 13, and 20 are canceled. Thus, claims 1, 3, 5-10, 12, and 14-19 are pending and presented for examination.
Applicant's arguments filled on 05/08/2026 with respect to the 35 U.S.C. 101 rejections have
been fully considered and are persuasive. Thus, the 35 U.S.C. 101 rejection is withdrawn.
Applicant's amendments filled on 04/27/2026 with respect to the claim objections have been
fully considered and are persuasive. Thus, the claim objections have been withdrawn.
Applicant's arguments filled on 04/27/2026 with respect to the 35 U.S.C. 102 rejections have
been fully considered and are persuasive. Thus, the 35 U.S.C. 102 rejection is withdrawn.
Applicant's arguments filled on 04/27/2026 with respect to the 35 U.S.C. 103 rejections have
been fully considered but are moot because of the new ground of rejection. With the following argument being the sole exception. The response to this argument is presented below.
Response to Arguments
Applicant argues (pages 17-18):
"For example, amended dependent claim 6, which is representative of amended dependent claims 15 with regard to similarly recited subject matter, recites "creating, by the number of processor units, a project that comprises multiple steps for creating the artifact models; and running, by the number of processor units, an experiment that estimates workload needed to perform the multiple steps of the project to create the artifact models." The Examiner cites Capelo, paragraphs [0086], [0152], and [0088] as teaching the features of original claim 6. Office Action dated April 28, 2026, page 40. However, it appears that the Examiner is using the term "experiments" as taught by Capelo to teach both the terms "project" and "experiment" as recited in amended claim 6.
According to Becton, Dickinson and Co. v. Tyco Healthcare Gr., LP, 616 F.3d 1249 (Fed. Cir. 2010) "[w]here a claim lists elements separately, 'the clear implication of the claim language is that those elements are 'distinct component[s]' of the [claimed] invention." This is particularly true where the specification "confirms" that the separated claim elements are in fact separate elements. In addition, Engel Indus., Inc. v. Lockformer Co., 96 F.3d 1298 (Fed. Cir. 1996) also concludes that where a claim provides for two separate elements (a "second portion" and a "return portion" as recited in Engel), these two elements "logically cannot be one and the same." Further, in Hopkins Manufacturing Corp. v. Cequent Performance Products, Inc. (IPR 2015-00613, Paper 9, August 7, 2015), the PTAB held that the prior art disclosure of an accelerometer with an output signal could not be applied as teaching both a "deceleration signal" and an "inclination signal" as claimed. As a result, Applicant respectfully submits that Capelo, paragraphs [0086], [0152], and [0088] do not teach or suggest the above-recited features of amended dependent claims 6 and 15.”
Response: It seems the Applicant is arguing that the Examiner is relying on Capelo's "experiments" to teach both the "project" and the "experiment" recited as separate elements in claims 6 and 15, which is improper under Becton Dickinson, Engel, and Hopkins because separately listed claim elements are distinct and a single prior art element cannot be applied as teaching two separately claimed elements. The Examiner agrees with the cited legal principle but respectfully disagrees that the current rejection violates it, because the current rejection does not read the two claim elements on a single element of Capelo. The claimed project is mapped to Capelo's run, which is expressly identified by a "run name (e.g., project name)" (Para 72) and comprises multiple steps for creating the trained models, determining a run specification (S100), determining a set of experiments (S200), determining computing resources (S300), provisioning a machine set (S500), scheduling (S600), and running the experiments (S700) (Para 109). The claimed experiment is separately mapped to the individual experiments contained within the run, which train the models and generate the trained-model artifacts (Paras 86, 102, 152; Para 88). Capelo discloses the run and its experiments as distinct elements: "Each run can include one or more experiments" (Para 86), and the experiments within a run are generated from the run specification (Para 102). The rejection maps the two separately recited claim elements to two distinct elements disclosed by Capelo (the run, and the experiments within the run).
Claim Rejections - 35 USC § 112
The following is a quotation of 35 U.S.C. 112(b):
(b) CONCLUSION.—The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the inventor or a joint inventor regards as the invention.
The following is a quotation of 35 U.S.C. 112 (pre-AIA ), second paragraph:
The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the applicant regards as his invention.
Claim 3, 5, 9, 12, 14, 18 rejected under 35 U.S.C. 112(b) or 35 U.S.C. 112 (pre-AIA ), second paragraph, as being indefinite for failing to particularly point out and distinctly claim the subject matter which the inventor or a joint inventor (or for applications subject to pre-AIA 35 U.S.C. 112, the applicant), regards as the invention.
Claims 3, 5, 12, 14 recites the limitation "the custom model recipe". There is insufficient antecedent basis for this limitation in the claim.
Claims 9, 18 recites the limitation "the specific type of assets". There is insufficient antecedent basis for this limitation in the claim.
Claim Objections
Claim 1, 10, 19 objected to because of the following informalities:
The phrase “…using the customized model recipe and input data, the customized model recipe uses a data dictionary…” insert “wherein the customized model recipe uses…”
Appropriate correction is required.
Claim 8 and 17 objected to because of the following informalities:
The phrase “a learning parameter of a single task … or a multitask” is grammatically unclear regarding what the learning parameter specifies. Correction: recite “a learning parameter indicating either single-task anomaly detection … or multitask training …”
Appropriate correction is required.
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
This application currently names joint inventors. In considering patentability of the claims the examiner presumes that the subject matter of the various claims was commonly owned as of the effective filing date of the claimed invention(s) absent any evidence to the contrary. Applicant is advised of the obligation under 37 CFR 1.56 to point out the inventor and effective filing dates of each claim that was not commonly owned as of the effective filing date of the later invention in order for the examiner to consider the applicability of 35 U.S.C. 102(b)(2)(C) for any potential 35 U.S.C. 102(a)(2) prior art against the later invention.
Claims 1, 7, 9-10, and 16, 18-19 are rejected under 35 U.S.C. 103 as being unpatentable over Song et al. (US 11443237 B1), hereinafter “Song”, in view of Boven et al. (US 20210034581 A1), hereinafter “Boven”, and further in view of Zhang et al. (US 11720068 B2), hereinafter “Zhang”.
Claim 1
Song teaches,
A computer implemented method for generating an artificial intelligence system, the computer implemented method comprising: (Abstract, “Systems and techniques are disclosed for a centralized platform for enhanced automated machine learning using disparate datasets. An example method includes receiving user specification of one or more data sources to be integrated with the system, the data sources storing datasets to be utilized to train one or more machine learning models by the system ...” Col 19, lines 4-10, “At block 506, the system trains a machine learning model according to a selected recipe. The system may train the model based on the training dataset. Optionally, the system may train multitudes of models, for example with varying hyperparameters. The trained machine learning model may be utilized to determine forecasts, for example as illustrated in FIG. 6A.”)
selecting, by a number of processor units, a (Col 18, lines 53-62, “At block 502, the system receives selection of a particular machine learning model recipe. As described in FIG. 4, a user (e.g., associated with an entity) may select from amongst presented machine learning model recipes. Optionally, the system may receive an API call associated with initiating training. The machine learning model recipe may inform a particular type of machine learning model to be trained. Thus, for certain users with knowledge of such models, the user may prefer that a particular model be trained.” - EN: this denotes that Song’s machine learning model recipe, which the user selects to inform the type of model to be trained, is the claimed model recipe for generating the artificial intelligence models.)
identifying, by the number of processor units, recipe parameters specified in the (Col 5, line 63 to Col 6, line 2, “A recipe, however, may additionally be fine-grained. For example, the recipe described above may refer to a recurrent neural network with gated recurrent units or long short-term memory cells. As another example, a recipe may further indicate that certain hyperparameters are to be utilized or that respective ranges of hyperparameters are to be utilized.” - EN: this denotes that the model type and the hyperparameters indicated in the recipe are the recipe parameters specified in the model recipe.)
creating, by the number of processor units, a training dataset using the (Col 18, line 63 to Col 19, line 3, “At block 504, the system accesses stored datasets. As described in FIGS. 2A-3B, the system may store, or have access to, datasets associated with the entity. The system may obtain these datasets and generate a training dataset, a validation dataset, and optionally a test dataset. For example, the system may assign a threshold quantity of data (e.g., 85%, 90%, and so on) to be a training dataset. The system may then assign the remaining data to be the validation dataset.” Col 20, lines 9-15, “Optionally, each machine learning model which is being trained may have a different subset of the datasets utilized as training data and validation data. For example, the system may randomly select 90% of the data as being training data, with the random selection occurring for each machine learning model.” Col 17, lines 27-33, “These templates may inform the machine learning model recipes which are utilized by the system to train machine learning models. For example, the system may utilize the template to indicate that a first type of machine learning model (e.g., a neural network) will likely produce less error than a second type of machine learning model (e.g., a clustering model).” Also see FIGS. 5A-5B. - EN: this denotes that after the recipe is selected (block 502) the system assembles the training dataset (block 504), and because the recipe determines which type of model is trained and each model being trained receives its own data subset, the training dataset is created using the model recipe and the stored input data.)
training, by the number of processor units, the artificial intelligence models (Col 19, lines 4-10, “At block 506, the system trains a machine learning model according to a selected recipe. The system may train the model based on the training dataset. Optionally, the system may train multitudes of models, for example with varying hyperparameters. The trained machine learning model may be utilized to determine forecasts, for example as illustrated in FIG. 6A.” Also see FIGS. 5A-5B. - EN: this denotes that the trained models resulting from training according to the recipe and its hyperparameters are the artifact models, consistent with paragraph 77 of the instant application (“artifact models 238 are the artificial intelligence models resulting from training artificial intelligence models 232.”).)
evaluating, by the number of processor units, the artifact models resulting from training the artificial intelligence models (Col 19, lines 19-25, “At block 508, the system generates error metrics associated with the trained machine learning model. For example, the system may generate multitudes of models with varying hyperparameters. The system may thus compute error metrics, for example based on the validation dataset described above. Without being constrained by theory, it should be appreciated that there may be multitudes of error metrics.” Also see FIGS. 5A-5B. - EN: this denotes that the error metrics generated for the trained models are the evaluation of the artifact models.)
selecting, by the number of processor units, a set of the artifact models for the artificial intelligence system using the evaluation of the artifact models; (Col 20, lines 20-25, “At block 520 the system selects one or more machine learning models. The system may select a machine learning model which provides superior performance to that of other machine learning models. Optionally, the system may determine that two or more machine learning models are to be implemented.” Also see FIGS. 5A-5B. - EN: this denotes selecting the trained models with superior performance, i.e., using the error metrics that form the evaluation.)
deploying, by the number of processor units, the artificial intelligence system to a set of target platforms (Col 20, lines 39-40, “Once the machine learning model(s) is/are trained, the system may implement (e.g., host) the model(s).” Col 8, lines 22-33, “To access the trained machine learning model, API 26 may be utilized. This example API 26 may enable an entity to provide queries to the automated machine learning system 100 for processing. Results associated with the queries may optionally be provided to the entity as a visual response (e.g., presented via a user interface). Results may also be provided as a CSV file, for example as a time-series forecast. Optionally, the results may be provided in a particular file format (e.g., JavaScript Object Notation). Using the API 26, the results may be automatically ingested via an enterprise system, a cloud-storage system, a database, and so on. In this way, the entity may easily obtain the results.” - EN: this denotes that once trained, the selected models are implemented (hosted), i.e., the artificial intelligence system is deployed, and the enterprise system, cloud-storage system, and database that automatically ingest the selected trained models’ results are the set of target platforms to which the artificial intelligence system is deployed.)
Song does not explicitly teach:
a customized model recipe (as struck through in the selecting, identifying, creating, and training steps above);
to detect anomalies in a specific type of industrial assets (as struck through in the selecting step, and correspondingly recited as “to detect anomalies in the specific type of industrial assets” in the identifying, training, evaluating, monitoring, and deploying steps);
the customized model recipe uses a data dictionary to process and organize the input data to generate a data model specific to the specific type of industrial assets for training the artificial intelligence models
monitoring, by the number of processor units, a number of performance metrics of the artificial intelligence models after deployment of the artificial intelligence system based on evaluation artifacts generated during the training of the artificial intelligence models to detect anomalies in the specific type of industrial assets; and
retraining, by the number of processor units, the artificial intelligence models in the artificial intelligence system based on the number of performance metrics monitored during use of the artificial intelligence models after the deployment of the artificial intelligence system.
However, Boven teaches:
a customized model recipe (as struck through in the selecting, identifying, creating, and training steps above); (Para 190, “the data analytics module 310 may establish certain of these model parameters based on predefined parameters data associated with the type of data science model being created” Para 187, “if an anomaly detection model is being created, the initial set of model parameters for the anomaly detection model may include a name and/or description of the model, an identification of the type(s) of assets for which anomalies are to be detected, and an identification of the set of data channels that are to serve as inputs for the anomaly detection model, a “sensitivity” level of the model” Para 126, “predefined data science models that are specifically designed for performing certain types of data analytics operations on asset-related data” - EN: this denotes Boven’s predefined set of model parameters, which under the broadest reasonable interpretation is a form of customization relating to the particular analytics operation and the particular asset type for which the model is being created, i.e., the parameters are tailored to anomaly detection for the identified types of assets and to the data channels of those assets (Para 187), rather than being generic to all data science models.)
to detect anomalies in a specific type of industrial assets (as struck through in the selecting step, and correspondingly recited as “to detect anomalies in the specific type of industrial assets” in the identifying, training, evaluating, monitoring, and deploying steps); (Para 5, “disclosed herein is a data science platform that is built with a specific focus on monitoring and analyzing the operation of industrial assets, such as trucking assets, rail assets, construction assets, mining assets, wind assets, thermal assets, oil and gas assets, and manufacturing assets... a determination that an asset is behaving abnormally, etc.), which may involve data science models that have been specifically designed to analyze asset-related data” Para 172, “Another type of data science model that may be created and/or deployed by the data analytics module 310 may take the form of a “anomaly detection model”... which is a data science model that is configured to evaluate whether multivariate data from a set of related data channels, or univariate data from a single data channel, is anomalous relative to the “normal” behavior for that set of related data channels” Para 187, “if an anomaly detection model is being created, the initial set of model parameters for the anomaly detection model may include... an identification of the type(s) of assets for which anomalies are to be detected” - EN: this denotes creating and deploying anomaly detection models for an identified type of industrial assets (e.g., wind assets), which supplies the struck-through recitation at each of the selecting, identifying, training, evaluating, monitoring, and deploying steps of the combined method.)
the customized model recipe uses a data dictionary to process and organize the input data to generate a data model specific to the specific type of industrial assets for training the artificial intelligence models; (Para 134, “schema data in general describes how many and what type of assets the data ingestion module 302 will ingest data from, what kinds of data the data ingestion module 302 will ingest, what format this data is collected in, and how this data is collected, among other possibilities. By knowing this information, data ingestion module 302, and the platform 300 generally, can efficiently ingest asset-related data, store it in an appropriate location, apply appropriate transformations to this data, and make the data available to other modules in the platform 300 in formats that these modules expect.” Para 138, “Objects may have a hierarchal relationship with one another... a “child” object may inherent all the data fields of its “parent” object and may include additional data fields that are relevant to the specific “child” object... an “asset” object may be a parent to a “truck asset” object or a “turbine asset” object... and/or a “channels” field for describing the type of data collected by the truck asset, which may be ingested into the platform 300 by data ingestion module 302.” Para 119, “the disclosed data science platform is built to evolve and extend a common industrial data model that has already been built and includes both asset-based contextual data (e.g. asset hierarchy, units of measure, fault events, etc.) and high-volume telemetry data.” Para 193, “if an anomaly detection model is being created, the set of training data may include historical data for the set of related data channels selected to serve as input for the anomaly detection model” Para 47, “an asset’s operating data may include sensor data that comprises time-series measurements for certain operating parameters of the asset” - EN: this denotes that the schema data is a data dictionary under the broadest reasonable interpretation because it includes metadata defining the asset types, data kinds, and formats (Para 134), used to process and organize the ingested sensor input data (Para 134, Para 47) into asset-type-specific objects such as the "turbine asset" object (Para 138), i.e., a data model specific to the specific type of industrial assets (Para 119), and the data science models are specifically designed to consume data in that schema (Para 111), with the training data for an anomaly detection model being the historical data of the data channels that are organized (Para 193).)
Before the effective filing date of the claimed invention, it would have been obvious to one of ordinary skill in the art to combine the recipe-driven automated machine learning pipeline of Song with the asset-specific schema data (data dictionary) and the predefined, asset-type-specific anomaly detection model parameters of Boven during the recipe selection and training dataset creation steps, such that the model recipe selected in Song is a customized model recipe, established for anomaly detection for an identified type of industrial assets, that uses the schema data to organize the input data for training. The motivation for doing so would be to organize the ingested asset data into a common, asset-type-specific structure that predefined models can take in, so that anomaly detection models for the identified type of industrial assets doesn’t have to be custom built for each new deployment. As Boven elaborates regarding the benefit of this schema-based technique in paragraph 111, "when ingested data is transformed into a schema that is specifically designed for asset-related data, this in turn enables the platform to make use of predefined data science models and applications that are specifically designed to consume data in that schema, which provides advantages over platforms that may require all data science models and applications to be custom built." Further, before the effective filing date of the claimed invention, it would have been obvious to one of ordinary skill in the art to apply Song's recipe-driven generation of machine learning models to the field of industrial assets for anomaly detection, as taught by Boven. The motivation for doing so would be to obtain an early indication of potential problems with an asset from its sensor data, meeting the recognized need to monitor and analyze the operation of industrial assets. As Boven elaborates regarding this benefit in paragraph 173, "the anomaly detection model's output may provide an indication of whether there is a potential problem with that particular component or subsystem of the asset."
Song in view of Boven does not explicitly teach:
monitoring, by the number of processor units, a number of performance metrics of the artificial intelligence models after deployment of the artificial intelligence system based on evaluation artifacts generated during the training of the artificial intelligence models to detect anomalies in the specific type of industrial assets; and
retraining, by the number of processor units, the artificial intelligence models in the artificial intelligence system based on the number of performance metrics monitored during use of the artificial intelligence models after the deployment of the artificial intelligence system.
However, Zhang teaches:
monitoring, by the number of processor units, a number of performance metrics of the artificial intelligence models after deployment of the artificial intelligence system based on evaluation artifacts generated during the training of the artificial intelligence models (Col 2, lines 33-38, “(2) deploying a first version of the forecast model trained using the set of model training configurations to predict optimal suggestion for controlling the industrial process, (3) monitoring the performance of the first version of the forecast model on current industrial process data of the industrial process” Col 9, line 61 to Col 10, line 10, “In some implementations, performance metrics (e.g., mean squared prediction error) of the candidate version of forecast model on current industrial process data (e.g., current validation data set) are generated, and a relative and/or absolute performance evaluation may be performed... For relative evaluation, the performance metrics of the candidate version of forecast model is compared to the performance metrics of an approved version of the forecast model. If the difference is within a threshold range, the new version of forecast model passes the relative evaluation. If the difference is outside of the threshold range (e.g., indicating forecast model has degraded beyond certain extend), the candidate version of forecast model does not pass the relative evaluation” Col 5, lines 38-41, “The model training controller evaluates the performance of versions of forecast models trained until an optimal version of forecast model with satisfactory performance metrics is found and approved for deployment.” Also see FIG. 5 (training results 466 including performance metrics 467 and checkpoints 468). - EN: this denotes monitoring the deployed model’s performance metrics on current data and evaluating them against the performance metrics generated when the model versions were trained and approved, which are the evaluation artifacts generated during the training (corresponding to the error metrics of Song, Col 19, lines 19-25, in the combination); Boven, not Zhang, is relied upon for the monitored models being trained to detect anomalies in the specific type of industrial assets, as set forth above.)
retraining, by the number of processor units, the artificial intelligence models in the artificial intelligence system based on the number of performance metrics monitored during use of the artificial intelligence models after the deployment of the artificial intelligence system. (Col 2, lines 38-45, “(4) if the performance of the first version of the forecast model becomes unsatisfactory, retraining a second version of the forecast model on current industrial process data with the set of model training configurations as a single point of truth for guiding the forecast model training using the model training algorithm, and (5) deploying the second version of the forecast model to replace the first version of the forecast model ...” Col 2, lines 16-19, “The retraining and re-deployment of the forecast model may be initiated automatically when the forecast model performance degrades, or when a triggering event has occurred (e.g., equipment has been replaced).” - EN: this denotes automatically retraining the deployed model when its monitored performance metrics degrade during use after deployment, and deploying the retrained version in its place.)
Before the effective filing date of the claimed invention, it would have been obvious to one of ordinary skill in the art to combine the automated machine learning pipeline of Song, as modified by Boven, with the automated performance monitoring and configuration-guided retraining of deployed models of Zhang after the deployment step. The motivation for doing so would be to keep the deployed models accurate when sensor behavior or system dynamics drift over time, without requiring human intervention. Zhang elaborates regarding the benefit of this monitoring and retraining methodology in column 6, lines 32-41, “the control system addresses faulty forecast model that can no longer accurately capture system dynamics due to for example sensor and/or system dynamics drift by continuously monitoring its performance, automatically triggering its training and retraining using a set of model training configurations as a single point of truth for training and retraining of the forecast model, and automatically deploying and redeploying the trained forecast model guided by a single set of forecast model deployment configurations.”
Claim 7
Song further teaches:
The computer implemented method of claim 1, wherein the (Col 19, lines 4-10, "At block 506, the system trains a machine learning model according to a selected recipe. The system may train the model based on the training dataset." Col 17, lines 27-33, "These templates may inform the machine learning model recipes which are utilized by the system to train machine learning models." - EN: this denotes that Song's model recipe is what the system uses to train the models, i.e., the model recipe is for training the artificial intelligence models.)
Song does not explicitly teach: that the model recipe is a customized model recipe, and that the training is for a specific type of industrial asset, as struck through above.
However, Boven teaches:
customized (as struck through above); (Para 190, "the data analytics module 310 may establish certain of these model parameters based on pre-defined parameters data associated with the type of data science model being created" Para 126, "predefined data science models that are specifically designed for performing certain types of data analytics operations on asset-related data" - EN: this denotes that the recipe's parameters are pre-defined per the particular model type and analytics operation, which is the customization of the model recipe, consistent with the instant specification's description of a customized model recipe as one for a specific type of asset (instant application, Para 61).)
a specific type of industrial asset (as struck through above); (Para 252, "The platform 300 may also provide sets of pre-established objects, which are organized into packages based on industry type. As one example, platform 300 may provide a "wind-turbine" package comprising a set of objects that may be specific for tenants that operate in the wind-turbine industry." Para 187, "if an anomaly detection model is being created, the initial set of model parameters for the anomaly detection model may include... an identification of the type(s) of assets for which anomalies are to be detected" Para 118, "the disclosed data science platform offers specialized data science functionality such as an asset fuel optimization engine, wind turbine power curve optimization" - EN: this denotes that the predefined packages and model parameters are provided per an identified asset type, such as wind turbines, so in the combination Song's recipe trains the models for that specific type of industrial asset. )
Before the effective filing date of the claimed invention, it would have been obvious to one of ordinary skill in the art to combine the model recipe used for training the machine learning models of Song with the pre-established, industry-type-specific packages and asset-type-specific model parameters of Boven, such that the customized model recipe is for training the artificial intelligence models for a specific type of industrial asset. The motivation for doing so would be to make setting up models for a given type of industrial asset faster and simpler, because pre-defined, asset-type-specific configurations can be reused rather than building each model from scratch. As Boven elaborates regarding the benefit of the pre-established industry packages in paragraph 252, “Providing pre-defined packages of objects may reduce the complexity of establishing a tenant’s schema. In this way, the platform 300 may facilitate a more efficient setup process for new tenants, which, among other advantages, may decrease the time required before a new tenant can begin to onboard its data to the platform 300 and to receive insights.”
Claim 9
Song further teaches:
The computer implemented method of claim 1, wherein the artificial intelligence system comprises at least one of a single artificial intelligence model for a single asset in the (Col 6, lines 17-23, “the system may determine that a first machine learning model is to be utilized for a first category of items. Thus, for a second category of items the system may use a second, different, machine learning model. Optionally, the system may combine outputs associated with different machine learning models when responding to a query.” - EN: this denotes that a first model utilized for a first category of items is a single artificial intelligence model for a group of assets; the claim recites the configurations in the alternative, so a showing on one suffices.)
Song does not explicitly teach: the struck-through text above, namely that the assets are in the specific type of industrial assets (first and second alternatives) or the “specific type of” assets (third and fourth alternatives). However, Boven teaches the specific type of industrial assets (Para 5, Para 187), as set forth in claim 1. The rationale to combine set forth in claim 1 applies.
Claim 10
Song teaches,
A computer system for generating an artificial intelligence system, comprising: (Col 19, lines 4-10, “At block 506, the system trains a machine learning model according to a selected recipe. The system may train the model based on the training dataset. Optionally, the system may train multitudes of models, for example with varying hyperparameters. The trained machine learning model may be utilized to determine forecasts, for example as illustrated in FIG. 6A.”)
one or more processor units; (Col 24, lines 6-10, “Each of the processes, methods, and algorithms described in the preceding sections may be embodied in, and fully or partially automated by, code modules executed by one or more computer systems or computer processors comprising computer hardware.”)
one or more computer readable storage devices; and (Col 24, lines 10-14, “The code modules (or “engines”) may be stored on any type of, one or more, non-transitory computer-readable media (e.g., a computer storage product) or computer storage devices, such as hard drives, solid state memory, optical disc, and/or the like.”)
computer program instructions, the computer program instructions being stored on the one or more computer readable storage devices for execution by the one or more processor units to perform one or more operations to: (Col 23, lines 38-48, “The memory 710 may include computer program instructions that the processing unit 704 executes in order to implement one or more embodiments. The memory 710 generally includes RAM, ROM, and/or other persistent or non-transitory memory. The memory 710 may store an operating system 714 that provides computer program instructions for use by the processing unit 704 in the general administration and operation of the server 700. The memory 710 may further include computer program instructions and other information for implementing aspects of the present disclosure.”)
The remaining limitations of claim 10 are substantially the same as method claim 1, therefore claim 10 is rejected under the same rationale as claim 1.
Claims 16, 18 recite substantially the same limitations as method claims 7, 9 respectively. Therefore, claims 16, 18 are rejected under the same rationale as claims 7, 9.
Claim 19
Song teaches,
A computer program product for generating an artificial intelligence system, the computer program product comprising a computer readable storage device having computer program instructions embodied therewith, the computer program instructions executable by a computer system to cause the computer system to: (Col 24, lines 6-19, “Each of the processes, methods, and algorithms described in the preceding sections may be embodied in, and fully or partially automated by, code modules executed by one or more computer systems or computer processors comprising computer hardware. The code modules (or "engines") may be stored on any type of, one or more, non-transitory computer-readable media (e.g., a computer storage product) or computer storage devices, such as hard drives, solid state memory, optical disc, and/or the like. The processes and algorithms may be implemented partially or wholly in application-specific circuitry. The results of the disclosed processes and process steps may be stored, persistently or otherwise, in any type of non-transitory computer storage such as, for example, volatile or non-volatile storage.” Col 19, lines 4-10, “At block 506, the system trains a machine learning model according to a selected recipe. The system may train the model based on the training dataset. Optionally, the system may train multitudes of models, for example with varying hyperparameters. The trained machine learning model may be utilized to determine forecasts, for example as illustrated in FIG. 6A.”)
The remaining limitations of claim 19 are substantially the same as method claim 1, therefore claim 19 is rejected under the same rationale as claim 1.
Claims 3 and 12 are rejected under 35 U.S.C. 103 as being unpatentable over Song in view of Boven and Zhang, as applied to claims 1 and 10 above, and further in view of Montanari et al. (US 20210350280 A1), hereinafter “Montanari”.
Claim 3
Song further teaches:
The computer implemented method of claim 1, wherein deploying, by the number of processor units, the artificial intelligence system comprises:
“using the (Col 17, lines 36-47, “The user interface 400 further indicates machine learning model recipes 26. As described above, a machine learning model recipe may indicate a type of machine learning model optionally along with certain hyperparameters or ranges of hyperparameters. The user may optionally select a particular machine learning model recipe to utilize. For example, the user may select from among the presented recipes 30, such as neural network type A, neural network type B, factorization machine, and so on. An example neural network may be a sequence model with or without attention, DeepAR, and so on. The presented recipes 30 may optionally be selected according to the template specified by selection option 24.”)
Song does not explicitly teach:
a custom model recipe (as struck through in the limitations below);
However, Boven teaches:the custom model recipe (as struck through above) (as mapped in claim 1, Para 190, “the data analytics module 310 may establish certain of these model parameters based on pre-defined parameters data associated with the type of data science model being created” Para 187, “if an anomaly detection model is being created, the initial set of model parameters for the anomaly detection model may include a name and/or description of the model, an identification of the type(s) of assets for which anomalies are to be detected, and an identification of the set of data channels that are to serve as inputs for the anomaly detection model, a “sensitivity” level of the model” Para 126, “predefined data science models that are specifically designed for performing certain types of data analytics operations on asset-related data” - EN: this denotes Boven’s predefined set of model parameters, which under the broadest reasonable interpretation is a form of customization relating to the particular analytics operation and the particular asset type for which the model is being created, i.e., the parameters are tailored to anomaly detection for the identified types of assets and to the data channels of those assets (Para 187), rather than being generic to all data science models.)
The rationale to combine set forth in claim 1 applies with respect to the custom model recipe.
Song in view of Boven does not explicitly teach:
identifying, by the number of processor units, the set of target platforms for the artificial intelligence system;
creating, by the number of processor units, a set of production artificial intelligence models to form the artificial intelligence system to run on the set of target platforms using (...) and the set of artifact models; and
deploying, by the number of processor units, the artificial intelligence system comprising the set of production artificial intelligence models to the set of target platforms (...).
However, Montanari teaches:
identifying, by the number of processor units, the set of target platforms for the artificial intelligence system; (Para 78, “The resource discoverer module 72 may read low-level hardware description data of the first device.” Para 82, “The algorithm 80 starts at operation 81, where the requirements of the first device are received, wherein the requirements are based at least in part on hardware of the first device.” Para 47, “FIG. 1 is a block diagram of an example system, indicated generally by the reference numeral 10. The example system 10 shows a pre-trained model 11 being deployed at a plurality of edge devices 14a, 14b, and 14c.” - EN: this denotes the system identifying the platform by having a resource discoverer read its low-level hardware description.)
creating, by the number of processor units, a set of production artificial intelligence models to form the artificial intelligence system to run on the set of target platforms using (...) and the set of artifact models; and (Para 47, “In order to be deployed at each of the edge devices 14, the pre-trained model 11 may need to undergo manual modification 13 in order to be optimized according, for example, to the hardware capabilities of a specific edge device. For example, the pre-trained model 11 may be modified differently at modification steps 12a, 12b, and 12c, in order to be compatible with the hardware requirements of the edge devices 14a, 14b, and 14c respectively.” Para 64, “As such, when running a plurality of models, the plurality of models may be compiled in a way that maximizes the use of the on-board memory, thus obtaining the lowest compilation latency.” - EN: this denotes that the pre-trained models (the set of artifact models) modified and compiled to be compatible with each device’s hardware are the production artificial intelligence models; the text elided as “(...)” in the creating and deploying steps, “using the custom model recipe,” is taught by Song in view of Boven as set forth above; the production models are formed from the artifact models trained according to the custom model recipe (Song, Col 19, lines 4-10; Boven as set forth in claim 1), such that the creating and the deploying use the custom model recipe.)
deploying, by the number of processor units, the artificial intelligence system comprising the set of production artificial intelligence models to the set of target platforms (...). (Para 47, “FIG. 1 is a block diagram of an example system, indicated generally by the reference numeral 10. The example system 10 shows a pre-trained model 11 being deployed at a plurality of edge devices 14a, 14b, and 14c.” - EN: this denotes that the edge devices 14a, 14b, and 14c are the set of target platforms, and in the combination the deploying of claim 1 (Song, Col 8, lines 22-33) is performed to the set of target platforms identified per Montanari.)
Before the effective filing date of the claimed invention, it would have been obvious to one of ordinary skill in the art to combine the automated machine learning architecture of Song, as modified by Boven and Zhang, that includes model recipes with the dynamic modification of the models to accommodate for target platforms of Montanari in order to create and deploy models that are the right fit for the specific platform. The motivation for doing so would be to allow the models to be properly set up for deployment at various target platforms. As Montanari elaborates regarding this benefit in paragraph 48, “Machine learning models (e.g. deep learning models, neural networks etc.) may be designed for a variety of sensing tasks, including speech, vision and motion sensing. These models may be trained on GPU servers to benefit from the computational capabilities and parallelization power of such servers. However, once developed, it may be a lengthy and manual process to deploy these models on edge devices and accelerators as different edge devices (e.g. smartphones, smartwatches, laptops etc.) and accelerators may have different hardware architecture and resource constraints than the servers on which the model was trained.”
Claim 12 recites substantially the same limitations as method claim 3, therefore claim 12 is rejected under the same rationale as claim 3.
Claims 5, 6, 14, and 15 are rejected under 35 U.S.C. 103 as being unpatentable over Song in view of Boven and Zhang, as applied to claims 1 and 10 above, and further in view of Capelo et al. (US 20220107744 A1), hereinafter “Capelo”.
Claim 5
Song further teaches:
The computer implemented method of claim 1, wherein training, by the number of processor units, the artificial intelligence models comprises:
“using the (Col 17, lines 36-47, “The user interface 400 further indicates machine learning model recipes 26. As described above, a machine learning model recipe may indicate a type of machine learning model optionally along with certain hyperparameters or ranges of hyperparameters. The user may optionally select a particular machine learning model recipe to utilize. For example, the user may select from among the presented recipes 30, such as neural network type A, neural network type B, factorization machine, and so on. An example neural network may be a sequence model with or without attention, DeepAR, and so on. The presented recipes 30 may optionally be selected according to the template specified by selection option 24.” Col 19, lines 4-10, “At block 506, the system trains a machine learning model according to a selected recipe. The system may train the model based on the training dataset.”)
Song does not explicitly teach the dynamic steps of:
identifying, by the number of processor units, resources for training the artificial intelligence models (...);
creating, by the number of processor units, an execution cluster based on the resources identified; and
training, by the number of processor units, the artificial intelligence models in the execution cluster (...).
As set forth in claim 1, Song does not explicitly teach that the model recipe is a custom model recipe (as struck through above); Boven teaches the custom model recipe (Para 190, Para 187, Para 126), and the rationale to combine set forth in claim 1 applies.
However, Capelo teaches:
identifying, by the number of processor units, resources for training the artificial intelligence models (...); (Para 131, “Determining the computing resources for the run functions to determine the machine parameters required to execute one or more experiments (e.g., for the run, for an experiment, etc.). Examples of machine parameters can include: the number of machines, the type of machines, and/or any other suitable set of machine parameters.”)
creating, by the number of processor units, an execution cluster based on the resources identified; and (Para 140 and 141, “Provisioning a machine set S500 functions to prepare the computing environment within the cloud computing system for experiment execution. S500 is preferably performed after S200, more preferably after the user approval of the run metrics but alternatively at any other time. The machine set (e.g., cluster) preferably includes the types of machines associated with the run (e.g., specified by the run specification, specified by the experiment specification, etc.), but can alternatively include a subset of the machine types. The machine set preferably includes the number of machines associated with the run (e.g., total number of machines for concurrent execution, number of machines per experiment epoch, etc.), but can alternatively include a subset of the number of machines. However, the machine set can be otherwise constructed. [0141] S500 can include: optionally accessing the user’s cloud account using the user’s access credentials; optionally initializing a cluster orchestrator instance (e.g., to initialize a cluster); provisioning the machine set (e.g., by loading and running a set of container images on each machine); provisioning a dataset storage volume; provisioning an object storage volume; loading agents onto the cluster (e.g., wherein the agents can be specific to the cluster orchestrator and/or machine provider; wherein the agents can be included within the container image, etc.); loading the model onto each machine; and/or otherwise provisioning the machine set.”)
training, by the number of processor units, the artificial intelligence models in the execution cluster (...). (Para 152, “Running the experiments S700 functions to execute the experiments determined in S200... S700 preferably includes training each experiment’s instance of the model using the dataset”)
Before the effective filing date of the claimed invention, it would have been obvious to one of ordinary skill in the art to combine the automated machine learning platform of Song, as modified by Boven and Zhang, that includes model recipes and training datasets with the automated provisioning of computing clusters for model training of Capelo. The motivation for doing so would be to reduce the user overhead of provisioning and managing the machines used for training. As Capelo elaborates regarding the benefit of this automated provisioning methodology in paragraph 32, “In addition to automatically provisioning and controlling the requested machines, the technology can further reduce user overhead by automatically configuring, scheduling, and deploying the experiments. This frees users from manually testing different experiments to determine the best or optimal hyperparameter value permutation, from manually debugging experiments, and from waiting around for each experiment to complete before executing the next one (or from having to manage multiple experiments in parallel).” In the combination, Song’s custom model recipe corresponds to the model and hyperparameter specification of Capelo’s run specification (Para 110), from which the required computing resources are determined (Para 113, Para 131), such that the resources for training are identified, and the training in the execution cluster is performed, using the training dataset and the custom model recipe.
Claim 6
Song further teaches:
The computer implemented method of claim 1, wherein training, by the number of processor units, the artificial intelligence models comprises: (Col 19, lines 4-10, “At block 506, the system trains a machine learning model according to a selected recipe. The system may train the model based on the training dataset.”)
Song does not explicitly teach:
creating, by the number of processor units, a project that comprises multiple steps for creating the artifact models; and
running, by the number of processor units, an experiment that estimates workload needed to perform the multiple steps of the project to create the artifact models.
However, Capelo teaches:
creating, by the number of processor units, a project that comprises multiple steps for creating the artifact models; and (Para 86, “The system can function to run a set of experiments (e.g., a “run”), wherein each experiment trains (or attempts to train) one or more models... Each run can include one or more experiments. Additionally or alternatively, a run can be a process that generates one or more experiments for execution, and/or execute the experiments.” Para 72, “The run metrics can include: the run name (e.g., project name), the total number of experiments in the run” Para 109, “The method for training orchestration includes: determining a run specification S100; determining a set of experiments from the run parameters S200; optionally determining computing resources S300; and executing the set experiments S400. In variants, S400 can include: provisioning a machine set S500; and iteratively: scheduling experiments to machines within the machine set S600 and running the experiments S700.” Para 102, “In a first variation, the experiments within an experiment set (e.g., “run”) are generated from the same run specification.” - EN: this denotes that the run, expressly identified by a “project name” and comprising the steps of determining experiments, determining resources, provisioning machines, scheduling, and running the experiments that train the models, is the claimed project, distinct from the individual experiments within it.)
running, by the number of processor units, an experiment that estimates workload needed to perform the multiple steps of the project to create the artifact models. (Para 152, “Running the experiments S700 functions to execute the experiments determined in S200... S700 preferably includes training each experiment’s instance of the model using the dataset” Para 88, “Each experiment can generate one or more experiment outputs... Artifacts can include: the trained model (e.g., weights, equations, etc.); model checkpoints (e.g., for resuming training later); files created during the training process” Para 113, “Determining a set of experiments from the run parameters S200 functions to determine the resources required to run the experiments and to generate the scripts for individual experiments themselves.” Para 128, “The method can optionally include estimating the run metrics for the run. The run metrics can be estimated: before the run is executed, during run execution (e.g., iteratively), and/or at any other suitable time... The run metrics can be estimated based on information extracted from the run specification, machine provider data (e.g., cost data, availability, latency, processing speed, etc.), prior run metrics... experiment metrics (e.g., from the same cluster, similar clusters, the same run, etc.)” Para 129, “Run metrics can include: cost, runtime, failure percentage, success percentage, and/or any other suitable metric... For example, the run’s estimated runtime can be determined as a function of the number of experiments within the run and the estimated time to run each experiment.” Para 33, “The technology can also estimate and give users control of run metrics, such as cost or total runtime, before deploying the experiments to the machine set.” - EN: this denotes that the experiments run within the project that create the artifact models (the trained models), and their experiment metrics are used, iteratively during run execution, to estimate the workload (runtime, cost, machines) needed to perform the run’s steps, so running the experiments estimates the workload needed to perform the multiple steps of the project.)
Before the effective filing date of the claimed invention, it would have been obvious to one of ordinary skill in the art to combine the automated machine learning platform of Song, as modified by Boven and Zhang, that includes model recipes and artifact models with the structured runs (projects) and experiments with workload estimation of Capelo. The motivation for doing so would be to create the artifact models through an organized, automated framework that reduces user overhead. As Capelo elaborates regarding the benefit of this orchestration methodology in paragraph 23, “The method functions to automatically orchestrate large scale model training across different training parameter combinations and multiple machines (e.g., concurrently or asynchronously), which enables a user to train the same model at both a small scale and a large scale without manual model modifications, result monitoring, or active cloud computing management.”
Claims 14 and 15 recite substantially the same limitations as method claims 5 and 6 respectively. Therefore, claims 14 and 15 are rejected under the same rationale as claims 5 and 6.
Claims 8 and 17 are rejected under 35 U.S.C. 103 as being unpatentable over Song in view of Boven and Zhang, as applied to claims 1 and 10 above, and further in view of Tang et al. (US-20210350636-A1), hereinafter “Tang”.
Claim 8
Song further teaches:
The computer implemented method of claim 1,
wherein the recipe parameters are (…) (Col 17, lines 36-47, “The user interface 400 further indicates machine learning model recipes 26. As described above, a machine learning model recipe may indicate a type of machine learning model optionally along with certain hyperparameters or ranges of hyperparameters.)
Song in view of Boven and Zhang does not explicitly teach:
“an application specific parameter of a prediction window size denoting a forecast horizon, a learning parameter of a single task of anomaly detection for training a different artificial intelligence model for each respective industrial asset or a multitask for training one artificial intelligence model for multiple industrial assets,and a training level of basic, advanced, or comprehensive for training a given artificial intelligence model.”
However tang teaches:
“an application specific parameter of a prediction window size denoting a forecast horizon” (Paragraph 16, “faults … may be predicted for a given time period” and “σ.sub.t is the duration of the prediction window” Paragraph 17, “σ.sub.t is the duration of the prediction window” – EN: this denotes a prediction-window size establishing the forecast horizon)
a learning parameter of a single task (Paragraph 22, the model may be “specific to a particular model of vehicle” or “general and effective across multiple different vehicle models”, Paragraph 44, “During the training of block 304, fault score losses and anomaly score losses may be considered jointly.”; Paragraph 65, “This fault detection model 614 may be transmitted to the different cyber-physical systems for implementation. For example, the network interface 606 may transmit the model 614 to a fleet of vehicles 102, for use during operation to identify and predict faults.” -- EN: this denotes the claimed multitask/shared-model alternative).
and a training level of basic, advanced, or comprehensive for training a given artificial intelligence model. (Paragraph 67,”Referring now to FIG. 7, a generalized diagram of a neural network is shown. Although a specific structure of an ANN is shown, having three layers and a set number of fully connected neurons, it should be understood that this is intended solely for the purpose of illustration. In practice, the present embodiments may take any appropriate form, including any number of layers and any pattern or patterns of connections therebetween.” -- EN: this denotes selectable degrees of neural network complexity which falls under “comprehensive” in the claim because ang permits an unrestricted number of neural-network layers and connection patterns, thereby providing a comparatively extensive model configuration for training the AI model.)
Song does not explicitly teach: the struck-through text above, namely that the single task is anomaly detection and that the assets are industrial assets. However, Boven teaches anomaly detection models for specific types of industrial assets (Para 5, Para 172, Para 187), as set forth in claim 1. The rationale to combine set forth in claim 1 applies.
Before the effective filing date, it would have been obvious to a person of ordinary skill in the art to combine Song in view of Boven and Zhang where it discloses machine-learning model recipe, which may specify a model type and associated hyperparameters or ranges of hyperparameters that’s relating to anomaly detection in industrial assets, with Tang’s prediction-window duration, shared/general model configuration, and adjustable multi-component neural-network training configuration corresponding to a comprehensive training level. Song teaches that a recipe may be fine-grained and may specify particular hyperparameters or ranges of hyperparameters. The motivation for doing so would have been to tailor Song’s automated recipe-based training system for time-series fault detection so that the trained model predicts faults over an appropriate time period, operates effectively across multiple vehicle models, and is sufficiently trained and tested to generalize without overfitting. See tang para 16, “Faults in the vehicle's systems may be predicted and prevented based on data from the ECUs, which can prevent damage to the vehicle and loss of life. Using time series information generated by the ECUs, faults in the vehicle may be predicted for a given time period, with the fault being labeled as to a likely vehicle sub-system that is responsible.” Para 22, “the trained model 108 may be general and effective across multiple different vehicles models” Para 71, “the ANN may be tested against the testing set, to ensure that the training has not resulted in overfitting. If the ANN can generalize to new inputs, beyond those which it was already trained on, then it is ready for use.”
Claims 17 recite substantially the same limitations as method claims 8 respectively. Therefore, claims 17 is rejected under the same rationale as claims 8.
Conclusion
Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to NAYMUR RAHMAN ALI whose telephone number is (571)272-0007. The examiner can normally be reached Mon-Fri. 9:30-6:30 pm.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Alexey Shmatov can be reached at (571)270-3428. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/NAYMUR RAHMAN ALI/Examiner, Art Unit 2123
/ALEXEY SHMATOV/Supervisory Patent Examiner, Art Unit 2123
Prosecution Insights