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 .
Continued Examination Under 37 CFR 1.114
A request for continued examination under 37 CFR 1.114, including the fee set forth in 37 CFR 1.17(e), was filed in this application after final rejection. Since this application is eligible for continued examination under 37 CFR 1.114, and the fee set forth in 37 CFR 1.17(e) has been timely paid, the finality of the previous Office action has been withdrawn pursuant to 37 CFR 1.114. Applicant's submission filed on 03/18/2026 has been entered.
Priority
Acknowledgment is made of applicant's claim for foreign priority based on an application filed in Republic of Korea on 04/27/2022. It is noted, however, that applicant has not filed a certified copy of the KR1020220052223 application as required by 37 CFR 1.55.
Response to Arguments
Applicant’s argument filed 03/18/2026 regarding 103 rejections have been fully considered but they are not persuasive.
Applicant’s Argument: On pages 7-9 of Applicant’s response to rejections under 35 U.S.C. 103, applicant states that Zhou fails to disclose the use of a separate “proxy model”, transmitting training and validation data divided into a plurality of data, and updating a parameter of the proxy model based on accuracy of different architectures. Zhou also fails to disclose a termination condition that is based on both accuracy and depth of the model with a specific convergence condition on the change rate of accuracy.
Examiner’s Response: Applicant’s argument is not persuasive. The claim as a whole does not clearly define how the proxy model is different from the AI diagnostic model and it is not clear how the proxy model is structured in the system. In the Specification (par. 40 and 41; Figure 3) of the claimed invention, the proxy model is created in the environment. Zhou (pg. 3, Section A, par. 1) teaches a similar framework that consists of a controller and an environment during the process of searching for the optimal diagnosis model based on reinforcement learning. The object that the controller interacts with is the environment and Zhou (Figure 1) discloses the interaction between the controller and the candidate model, which is a part of the environment. Under the broadest reasonable interpretation, the candidate model discloses in Zhou teaches on the claimed proxy model.
Claim 1 discloses the training data and the validation data is divided into a plurality of data and does not further disclose how the data is specifically divided. Zhou (pg. 7, Section D, par. 1) discloses the training process with a batch size of 32, which indicates a few training samples used during each iteration. Under the broadest reasonable interpretation, batch size teaches on dividing a dataset into a plurality of data.
As explained above, the candidate model disclosed in Zhou teaches the claimed proxy model. During training, the candidate model parameters are updated. Applicant (Remarks from 03/18/2026, pg. 8) states “in Zhou, parameters begin updated belong either to the controller’s policy or the candidate model itself”. Thus, Applicant agrees the reference disclose the updating of parameters of the candidate model.
Zhou discloses a search operation that is based on both accuracy and depth of the model. Zhou (pg. 4, Section B, par. 5-8) discloses the process of adding a layer on top of the candidate model and comparing the accuracy with a candidate model having a different architecture. Zhou also explicitly states “An expected accuracy is set as the final termination condition of the entire search process, together with two other termination conditions that are set for each episode”. Thus, Zhou teaches a convergence criteria that is based on both accuracy and depth of the model.
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.
Claims 18-19 are 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.
Claim 18 is a dependent claim and recites “a proxy model”. It is not clear whether the proxy model recited in claim 18 is the same or different proxy model as recited in claim 13. Examiner interprets the proxy model to be the same model in the claims. Examiner suggests correcting claim 18 to recite “the proxy model”.
Claim 19 is dependent on claim 18 and is rejected on the same basis as its parent claim.
The following is a quotation of 35 U.S.C. 112(d):
(d) REFERENCE IN DEPENDENT FORMS.—Subject to subsection (e), a claim in dependent form shall contain a reference to a claim previously set forth and then specify a further limitation of the subject matter claimed. A claim in dependent form shall be construed to incorporate by reference all the limitations of the claim to which it refers.
The following is a quotation of pre-AIA 35 U.S.C. 112, fourth paragraph:
Subject to the following paragraph [i.e., the fifth paragraph of pre-AIA 35 U.S.C. 112], a claim in dependent form shall contain a reference to a claim previously set forth and then specify a further limitation of the subject matter claimed. A claim in dependent form shall be construed to incorporate by reference all the limitations of the claim to which it refers.
Claims 11-12 are rejected under 35 U.S.C. 112(d) or pre-AIA 35 U.S.C. 112, 4th paragraph, as being of improper dependent form for failing to further limit the subject matter of the claim upon which it depends, or for failing to include all the limitations of the claim upon which it depends. Claim 11 recites “wherein, based on (i) the accuracy of the Al diagnostic model being greater than the predefined level and (ii) the number of layers greater than or equal to the predefined number being added to the Al diagnostic model according to the change in the depth of the Al diagnostic model, the training process is terminated when the change rate of the accuracy converges to the level equal to or less than the predetermined level” and the claim limitation is the same as presented in independent claim 1 without additional elements to further limit the subject matter of the claim. Applicant may cancel the claim, amend the claim to place the claim in proper dependent form, rewrite the claim in independent form, or present a sufficient showing that the dependent claim complies with the statutory requirements.
Claim 12 is dependent on claim 11 and is rejected on the same basis as its parent claim.
Claim Rejections - 35 USC § 103
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.
Claims 1-2, 4-9, 11-14, and 16-20 are rejected under 35 U.S.C. 103 as being unpatentable over Zhou, “Automated Model Generation for Machinery Fault Diagnosis Based on Reinforcement Learning and Neural Architecture Search” in view of Liu, “Progressive Neural Architecture Search”, Wang, “An Engine-Fault-Diagnosis System Based On Sound Intensity Analysis And Wavelet Packet Pre-Processing Neural Network” and Rho, “NAS-VAD: Neural Architecture Search for Voice Activity Detection”. Zhou and Wang references are disclosed in the IDS dated 07/16/2025. Rho reference is disclosed in the IDS dated 12/18/2025.
Regarding claim 1, Zhou teaches:
“A method of automatically creating an artificial intelligence (Al)diagnostic model for diagnosing an abnormal state of a vehicle, the method comprising” ([abstract; pg. 1, section 1, par. 1], A method of automatically designing high accuracy models for fault diagnosis of machinery. Fault diagnostics of rotating machine components in transmission systems prevents unnecessary failures. Vehicles inherently consists of a transmission system and the experiments conducted by Zhou using different gearbox and bearing dataset can be applied to detect faulty conditions in vehicles.)
“acquiring ” ([pg. 7, section 4, par. 1-7, Figure 7], The five datasets consist of vibration data and the data is sampled using a sensor, such as an acceleration sensor shown in Figure 7. The different dataset is used to validate the model in machinery fault diagnostics and the results of the experiments may be transferrable to transmission systems of vehicles.)
“processing the input data by applying one or more algorithms selected from harmonic/percussive sound separation (HPSS), normalization, and outlier detection” ([pg. 7, section 4.A, par. 1-2; pg. 9, section 4.B, par. 1; pg. 8, Table 3], For the CWRU dataset, experiments are conducted on deep groove ball bearings and the collected data is divided into different health states, which includes “normal” and “ball fault”. Testing for bearing fault is an example of outlier detection in machine health monitoring.)
“extracting one or more features from the processed input data using at least one technique ” ([pg. 5-6, section A, par. 1; pg. 6, Figure 4], The action space contains 3 convolutional layers, one maxpooling layer, one dropout layer, and one fully connected layer. The 3 convolutional layer differ in the number of filters and the kernel size. The vibration signal is used as the input of the diagnosis model. Thus, the vibration data is processed by the convolutional layers to extract features.)
“selecting a combination of features for the Al diagnostic model from the extracted one or more features” ([pg. 5-6, section A, par. 1; pg. 6, section C, par. 1; pg. 6, Figure 4], The convolutional layers extract features from the vibration data and all the extracted features are used to train the diagnostic model. Under the broadest reasonable interpretation, selecting all the extract features for a model is within the scope of the claim limitation.)
“searching and selecting an architecture of the Al diagnostic model based on the selected combination of features, the architecture including a proxy model ” ([pg. 3, section 2, par. 1; pg. 6-7, section C, par. 1-4; pg. 9, section C, par. 1; pg. 9, Figure 9], The controller receives a set of possible states that consists of the information that describes the environment at a particular period of time. The controller selects an action based on the state and the action consists of selecting a network layer to create a new candidate model. The state may be the health state of machinery for fault diagnostics. The created candidate model is trained and tested using a weight sharing mechanism that shares the same value of the weight parameters during the training process. Under the broadest reasonable interpretation, the proxy model is included in the architecture and there are no further limiting definitions on how the proxy model is structured in the system. Therefore, the building and creating of candidate model teaches on the proxy model.)
“transmitting, to the proxy model, training data among the processed input data and validation data among the input data, the training data and the validation data each being divided into a plurality of data” ([pg. 6, section C, par. 1; pg. 7, section D, par. 1], The created candidate model is trained and tested on the training and validation data, which has been randomly divided into a 4:1 ratio. In the training process of the candidate model, the batch size is 32. Each training iteration, the candidate model process one single batch of training data.)
“updating a parameter in the proxy model based on a result of measuring accuracy in the proxy model with different architectures of the Al diagnostic model” ([pg. 9-10, section D, par. 1-4; pg. 10, Figure 11], The weight sharing mechanism significantly improves the search efficiency as well as the accuracy of the candidate model. The proposed framework searches for the candidate model with the best accuracy and selects the architecture with the best performance for each dataset. During each training iteration, the candidate model is searched based on accuracy. For the Ball Screw and PHM Gearbox dataset, it took about 30 minutes for the search operation to achieve over 80% accuracy.)
“validating the Al diagnostic model that is configured to, based on (i) accuracy of the Al diagnostic model being greater than a predefined level and (ii) a number of layers greater than or equal to a predefined number being added to the Al diagnostic model according to a change in a depth of the Al diagnostic model, terminate a training process when ” ([pg. 4, section B, par. 8; pg. 4, section C, par. 1 & 6; pg. 6, section B, par. 1-4; pg. 9-10, section D, par. 1-4], The proposed framework has been repeatedly executed five times to quantify its efficiency and evaluate its reproducibility. For all five runs, candidate models with 100% diagnostic accuracy have been searched out. The training of the candidate model is terminated when the expected accuracy is reached. Additionally, the time for the search operation to achieve over 80% accuracy is determined. Another termination condition during the search operation is that the candidate model has reached a depth level of 12. Thus, the search operation consists of increasing the depth of the candidate model and improving the accuracy of the model.)
“providing the Al diagnostic model to diagnose the abnormal state of the vehicle, wherein an efficient neural architecture search (ENAS) is applied to update the Al diagnostic model and the parameter configuring the Al diagnostic model, the ENAS sharing the parameter with the updated Al diagnostic model” ([pg. 9, section 4.B, par. 1-2; pg. 9, section 4.C, par. 1-2], The framework generates 5 different CNN models with different structures for fault diagnosis of the sample dataset. The framework is configured to update the model to provide a diagnostic model with high accuracy.)
Zhou does not explicitly disclose an implementation of “acquiring noise and vibration data”, “extracting one or more features from the processed input data using at least one technique selected from Fast Fourier Transform (FFT), Mel-spectrogram, or HPSS”, “searching and selecting an architecture ... the architecture including a proxy model and comprising a normal cell and a reduction cell”, and “terminate a training process when a change rate of the accuracy converges to a level equal to or less than a predetermined level”.
However, Liu discloses in the same field of endeavor:
“searching and selecting an architecture of the Al diagnostic model based on the selected combination of features, the architecture including a proxy model and comprising a normal cell and a reduction cell, the searching and selecting including” ([pg. 1-2, section 1, par. 3-4; pg. 5, section 3.2, par 1-3; pg. 6-7, section 4.2, par. 1-5], The proposed method uses heuristic search to search the space of cell structures, starting with simple models and progressing to complex ones. A surrogate model is learned to predict the performance of a structure without needing to train it. The best architecture is found by stacking a predefined number of copies of the basic cell. A reduction cell is emulated by using a normal cell with stride 2.)
“validating the Al diagnostic model that is configured to, … , terminate a training process when a change rate of the accuracy converges to a level equal to or less than a predetermined level” ([pg. 18-20, section C, par. 1-2; pg. 19, Table 2], The test set error rate decreases as the number of blocks of the model increases. The best model is found when the model is trained with the number of blocks equal to 5 because it has a lower error rate when compared to the other intermediate levels.)
It would be obvious to one of the ordinary skill in the art before the effective filing date of the claimed invention to incorporate the teaching of “searching and selecting an architecture ... the architecture including a proxy model and comprising a normal cell and a reduction cell”, and “terminate a training process when a change rate of the accuracy converges to a level equal to or less than a predetermined level” from Liu into the teaching of Zhou. Doing so can improve reinforcement learning based neural architecture search by implementing a surrogate model to guide the search through structure space in order of increasing complexity (Liu, abstract).
Zhou in view of Liu does not explicitly disclose an implementation of “acquiring noise and vibration data” and “extracting one or more features from the processed input data using at least one technique selected from Fast Fourier Transform (FFT), Mel-spectrogram, or HPSS”.
However, Wang discloses in the same field of endeavor:
“acquiring noise and vibration data measured by a sensor of the vehicle as input data” ([pg. 1, section 1, par. 2; pg. 4, section 3, par. 1], A number of publications are provided to show the effective of vibration and noise signals in fault diagnosis of rotating machineries, such as gearbox and bearings. The experiments use a microphone to record the sound signals of an engine.)
It would be obvious to one of the ordinary skill in the art before the effective filing date of the claimed invention to incorporate the teaching of “acquiring noise and vibration data” from Wang into the teaching of Zhou in view of Liu. Doing so can improve engine fault diagnosis by implementing sound data into the classification model for failure detection (Wang, abstract).
Zhou in view of Liu and Wang does not explicitly disclose an implementation of “extracting one or more features from the processed input data using at least one technique selected from Fast Fourier Transform (FFT), Mel-spectrogram, or HPSS”.
However, Rho discloses in the same field of endeavor:
“extracting one or more features from the processed input data using at least one technique selected from Fast Fourier Transform (FFT), Mel-spectrogram, or HPSS” ([pg. 3-4, section 4.1.1, par. 1-2; pg. 2, Figure 1], Speech dataset is processed by adding noise to create noise-added speech datasets. Log-melspectrogram is applied to the dataset to extract features.)
It would be obvious to one of the ordinary skill in the art before the effective filing date of the claimed invention to incorporate the teaching of “extracting one or more features from the processed input data using at least one technique selected from Fast Fourier Transform (FFT), Mel-spectrogram, or HPSS” from Rho into the teaching of Zhou in view of Liu and Wang. Doing so can improve designing and optimizing network architecture by implementing the neural architecture search approaches to detection tasks in audio datasets (Rho, abstract).
Regarding claim 13, the claim recites a method that performs a similar process as described in Claim 1. Therefore claim 13 is rejected under the same reasons mentioned for claim 1. The additional elements of claim 13 are addressed below using Zhou and Liu reference.
Zhou teaches:
“extracting one or more features from the processed input data” ([pg. 5-6, section A, par. 1; pg. 6, Figure 4], The action space contains 3 convolutional layers, one maxpooling layer, one dropout layer, and one fully connected layer. The 3 convolutional layer differ in the number of filters and the kernel size. The vibration signal is used as the input of the diagnosis model. Thus, the vibration data is processed by the convolutional layers to extract features.)
“searching and selecting an architecture of the Al diagnostic model based on the extracted features, …” ([pg. 3, section 1, par. 15-16; pg. 3, section 2, par. 1; pg. 4-5, section 3, par. 1; pg. 9, section 4.C, par. 1; Figure 9], The controller receives a set of possible states that consists of the information that describes the environment at a particular period of time. The controller selects an action based on the state and the action consists of selecting a network layer to create a new candidate model. The state may be the health state of machinery for fault diagnostics. The framework automatically creates CNN models for specific fault diagnosis tasks by processing datasets of bearings and gears to determine the state of the environment. The different layers of the CNN model extracts features from the vibration data and processes those features when searching for the best model architecture.)
“creating a sampled architecture string to transmit, to the proxy model, the created architecture string by a ” ([pg. 3, section 2, par. 1; pg. 4, section B, par. 5-7], The controller selects an action (string) based on the state and the action consists of selecting a network layer (sampled architecture) to create a new candidate model. The controller is trained to optimize its decision making based on the rewards. Liu (pg. 1, section 1, par. 1; pg. 2, section 2, par. 1) further teaches a method a RNN controller is used to generate the sample architecture.)
“optimizing the architecture of the Al diagnostic model based on a parameter” ([pg. 4, section 2.C, par. 1], A greedy algorithm is used to adjust what actions the controller takes in selecting a specific network layer for the candidate model. The greedy algorithm consists of parameters such as greedy degree, greedy decay rate, and minimum greedy degree.)
Regarding claims 2 and 14, Zhou teaches:
“wherein searching and selecting the architecture of the Al diagnostic model includes parameter tuning and controller training” ([pg. 2, section 1, par. 10; pg. 4, section 2.C, par. 2], The candidate models are CNNs whose structure and hyperparameters are designed by the controller. The framework trains the controller.)
Regarding claims 4 and 16, Zhou teaches:
“wherein the parameter tuning includes: transmitting, to the proxy model, the training data among the processed input data” ([pg. 5-6, section A, par. 1], The candidate model receives the vibration signal as input.)
Regarding claims 5 and 17, Zhou teaches:
“wherein the parameter tuning includes: updating the parameter in the proxy model” ([pg. 2, section 1, par. 9-10; Figure 3], The candidate models are CNNs whose structure and hyperparameters are designed by the controller. The controller search and create candidate model at each time step. The candidate model is updated based on the action from the controller.)
Regarding claim 6, Zhou in view of Liu, Wang, and Rho teaches:
“wherein the controller training includes: creating the sampled architecture string by a RNN controller” ([Liu, pg. 1, section 1, par. 1; pg. 2, section 2, par. 1], The proposed method is based on the neural architecture search with reinforcement learning framework outlined in the Zoph reference. An RNN controller is used to generate the sample architecture.)
Regarding claim 7, Zhou teaches:
“wherein the controller training further includes: transmitting, to the proxy model, the validation data among the input data” ([pg. 6, section C, par. 1, Figure 3], The candidate model is tested on validation data, which has been randomly divided into a 4:1 ratio. The controller is trained on the states of the candidate model and the target action return.)
Regarding claim 8, Zhou teaches:
“wherein the controller training further includes: measuring accuracy in the proxy model with a different architecture of the Al diagnostic model” ([pg. 4, section B, par. 7; pg. 4, section C, par. 1, Figure 3], The accuracy of the new candidate model is compared with the accuracy from the previous candidate model. A positive reward is recorded when the new candidate model has higher accuracy.)
Regarding claim 9, Zhou teaches:
“wherein the controller training further includes: updating a value of the parameter using a reinforced training that increases the measured accuracy by performing reinforcement leaning for a reward” ([pg. 4, section C, par. 2-5, Figure 3], The controller is trained using reinforcement learning to maximize the return of the decision-making process and increase the accuracy of the candidate model. The controller updates its policy, which consists of weights to improve the rewards.)
“training the RNN controller by the updated value of the parameter” ([pg. 4, section C, par. 2-5, Figure 3], The state and its current return value pairs are used to train the controller.)
Regarding claim 11, Zhou in view of Liu, Wang, and Rho teaches:
“wherein, based on (i) the accuracy of the Al diagnostic model being greater than the predefined level and (ii) the number of layers greater than or equal to the predefined number being added to the Al diagnostic model according to the change in the depth of the Al diagnostic model, the training process is terminated when ” ([Zhou, pg. 4, section B, par. 8; pg. 4, section C, par. 1 & 6; pg. 6, section B, par. 1-4; pg. 9-10, section D, par. 1-4], The proposed framework has been repeatedly executed five times to quantify its efficiency and evaluate its reproducibility. For all five runs, candidate models with 100% diagnostic accuracy have been searched out. The training of the candidate model is terminated when the expected accuracy is reached. Additionally, the time for the search operation to achieve over 80% accuracy is determined. Another termination condition during the search operation is that the candidate model has reached a depth level of 12. Thus, the search operation consists of increasing the depth of the candidate model and improving the accuracy of the model. [Liu, pg. 18-20, section C, par. 1-2; pg. 19, Table 2], The test set error rate decreases as the number of blocks of the model increases. The best model is found when the model is trained with the number of blocks equal to 5 because it has a lower error rate when compared to the other intermediate levels.)
Regarding claims 12 and 20, Zhou teaches:
“wherein the Al diagnostic model is (i) provided as an API in a server or (ii) stored in a file as a user device environment” ([pg. 9, section B, par. 1-3], The RL-NAS farmwork generated different CNN models based on the 5 dataset. It is implied that the framework is performed on a computer, consisting of a processor and memory. The memory stores instructions to execute the framework on the user computer.)
Regarding claim 18, Zhou in view of Liu, Wang, and Rho teaches:
“creating a sampled architecture string by the RNN controller to transmit, to a proxy model, the created architecture string by the RNN controller” ([Zhou, pg. 3, section 2, par. 1; pg. 4, section B, par. 5-7], The controller selects an action (string) based on the state and the action consists of selecting a network layer (sampled architecture) to create a new candidate model. The controller is trained to optimize its decision making based on the rewards. Liu (pg. 1, section 1, par. 1; pg. 2, section 2, par. 1) further teaches a method a RNN controller is used to generate the sample architecture.)
Regarding claim 19, Zhou teaches:
“wherein the controller training further includes: transmitting, to the proxy model, the validation data among the input data, the validation data divided into a plurality of data” ([pg. 6, section C, par. 1, Figure 3], The candidate model is tested on validation data, which has been randomly divided into a 4:1 ratio. The controller is trained on the states of the candidate model and the target action return.)
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to GARY MAC whose telephone number is (703)756-1517. The examiner can normally be reached Monday - Friday 8:00 AM - 5:00 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, Abdullah Kawsar can be reached at (571) 270-3169. 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.
/GARY MAC/Examiner, Art Unit 2127
/ABDULLAH AL KAWSAR/Supervisory Patent Examiner, Art Unit 2127
Prosecution Insights