DETAILED ACTION
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 .
Claim Rejections - 35 USC § 101
Claims 1-39 are rejected under 35 U.S.C. 101 because the claimed invention is directed to abstract idea without significantly more. The claim(s) recite(s) significantly more. The subject matter eligibility test for products and process is describe below for claim 1 in view of dependent claims.
Regarding claim 1:
Step 1: Is the claim to a process machine manufacture or composition of matter?
Yes – Claim 1 recites a method, which a method falls under the statutory categories.
Step 2A Prong 1: Does the claim recite an abstract idea, law of nature, or natural phenomenon?
Yes – The claim recites the following:
“determining M sets of feature relevance values including a set of feature relevance values for each of the M predicted values, a particular set of feature relevance values associated with a particular predicted value, wherein each feature relevance value of the particular set of feature relevance values represents an estimate of a contribution of a respective one of the N input values to the particular predicted value;” - The limitations recites a mental process of determining M sets of feature relevance values (see MPEP 2106.04(a)(2)III).
“characterizing the operating state of the monitored asset based at least in part on the N aggregate feature relevance values.” - The limitations recites a mental process of characterizing the operating state (see MPEP 2106.04(a)(2)III).
“wherein N is an integer greater than or equal to two and M is an integer greater than or equal to two;” – The limitation defines a mathematical relationship where n and m are greater than or equal to two. (See MPEP 2106.04(a)(2)I.
Step 2 Prong 2: Does the claim recite additional elements that integrate the judicial exception into a particular application? No –
The claim includes the additional element(s):
“A method comprising: providing input data to one or more machine-learning models to generate output data, the input data including an input value for each of N features associated with an operating state of a monitored asset and the output data including a predicted value of each of M features,”
The additional elements fall under Insignificant Extra-Solution Activity as mere data gathering by providing input to one or more machine learning models. See MPEP 2106.5(g). The additional elements fall under “apply it” as using a generic computer to implement machine learning model to generate output. See Mere Instructions to Apply an Exemption (see MPEP 2106.05(f)).
“aggregating, across the M sets of feature relevance values, feature relevance values for each of the N features to generate N aggregate feature relevance values; and”
The additional elements fall under “apply it” as using a generic computer to aggregate M sets of feature relevance values. See Mere Instructions to Apply an Exemption (see MPEP 2106.05(f)).
Step 2B: Does the claim recite additional elements that amount to significantly more than the judicial exception?
No - The claim does not include additional elements that are sufficient to amount to a significantly more than the judicial exemption. As an order whole, the claim is directed to a mental process of determining relevance values to characterize an operating state. As discussed above with respect to integration of the abstract idea into a practical application, the additional elements of providing, outputting, and aggregating fall under using generic computer to apply an exemption and mere data gathering. The method does not improve on the function of a computer, transforms an article into another article, nor is it applied by a particular machine, making the claim not patent eligible.
Regarding claim 2:
Step 2A Prong 2, Step 2B: The additional element(s):
“The method of claim 1, wherein the characterizing the operating state of the monitored asset is further based at least in part on the output data.”
The additional elements fall under Insignificant Extra-Solution Activity. See MPEP 2106.5(g).
No additional elements. The judicial exemptions do not integrate into a practical application nor provide an improvement. The process does not provide an inventive concept nor provides a practical application.
Regarding claim 3:
Step 2A Prong 2, Step 2B: The additional element(s):
“The method of claim 1, wherein the characterizing the operating state of the monitored asset comprises providing input based at least in part on the N aggregate feature relevance values to an operating state model to generate an operating state output.”
The additional elements fall under Insignificant Extra-Solution Activity as mere data gathering by providing input to the operating state model. See MPEP 2106.5(g). The additional elements fall under “apply it” as using a generic computer to implement operating state model to generate output. See Mere Instructions to Apply an Exemption (see MPEP 2106.05(f)).
No additional elements. The judicial exemptions do not integrate into a practical application nor provide an improvement. The process does not provide an inventive concept nor provides a practical application
Regarding claim 4:
Step 2A Prong 2, Step 2B: The additional element(s):
“The method of claim 3, wherein the operating state output indicates whether the operating state of the monitored asset is an anomalous operating state.”
The additional elements fall under “apply it” as using a generic computer to output if the operating sates is anomalous. See Mere Instructions to Apply an Exemption (see MPEP 2106.05(f)).
Regarding claim 5:
Step 2A Prong 1:
“The method of claim 3, further comprising determining one or more residual data values based on comparison of each of the M predicted values to an actual value of a corresponding feature of the M features,” - The limitations recites a mental process to determine one or more residual data values based on a comparison (see MPEP 2106.04(a)(2)III).
Step 2A Prong 2, Step 2B: The additional element(s):
“wherein the input to the operating state model is further based on the one or more residual data values.”
The additional elements fall under Insignificant Extra-Solution Activity as mere data gathering by providing input that further residual data values. See MPEP 2106.5(g).
Regarding claim 6:
Step 2A Prong 2, Step 2B: The additional element(s):
“The method of claim 5, wherein the one or more residual data values include M residual data values.”
The additional elements fall under Insignificant Extra-Solution Activity. See MPEP 2106.5(g).
Regarding claim 7:
Step 2A Prong 2, Step 2B: The additional element(s):
“The method of claim 1, wherein the one or more machine-learning models include one or more autoencoders.”
The additional elements fall under Insignificant Extra-Solution Activity. See MPEP 2106.5(g). The judicial exemptions do not integrate into a practical application nor provide an improvement. The process does not provide an inventive concept nor provides a practical application.
Regarding claim 8:
Step 2A Prong 1:
“The method of claim 1, wherein N is equal to M.” – The limitation defines a mathematical relationship where n is equal to m. (See MPEP 2106.04(a)(2)I.
Step 2A Prong 2, Step 2B: The additional element(s):
No additional elements. The judicial exemptions do not integrate into a practical application nor provide an improvement. The process does not provide an inventive concept nor provides a practical application
Regarding claim 9:
Step 2A Prong 1:
“The method of claim 1, wherein N is less than M.” – The limitation defines a mathematical relationship where n is less to m. (See MPEP 2106.04(a)(2)I.
Step 2A Prong 2, Step 2B: The additional element(s):
No additional elements. The judicial exemptions do not integrate into a practical application nor provide an improvement. The process does not provide an inventive concept nor provides a practical application
Regarding claim 10:
Step 2A Prong 1:
“The method of claim 1, wherein N is greater than M.” – The limitation defines a mathematical relationship where n is greater to m. (See MPEP 2106.04(a)(2)I.
Step 2A Prong 2, Step 2B: The additional element(s):
No additional elements. The judicial exemptions do not integrate into a practical application nor provide an improvement. The process does not provide an inventive concept nor provides a practical application.
Regarding claim 11:
Step 2A Prong 2, Step 2B: The additional element(s):
“The method of claim 1, wherein one or more of the N features represents sensor data values from one or more sensors associated with the monitored asset.”
The additional elements fall under Insignificant Extra-Solution Activity. See MPEP 2106.5(g). The judicial exemptions do not integrate into a practical application nor provide an improvement. The process does not provide an inventive concept nor provides a practical application.
Regarding claim 12:
Step 2A Prong 1:
“The method of claim 1, wherein the determining the M sets of feature relevance values comprises performing layer-wise relevance propagation for each of the M predicted values.” – The limitation defines a mathematical method of performing layer-wise relevance propagation. (See MPEP 2106.04(a)(2)I.
Step 2A Prong 2, Step 2B: The additional element(s):
No additional elements. The judicial exemptions do not integrate into a practical application nor provide an improvement. The process does not provide an inventive concept nor provides a practical application.
Regarding claim 13:
Step 2A Prong 2, Step 2B: The additional element(s):
“The method of claim 1, further comprising generating one or more control signals based on characterization of the operating state of the monitored asset.”
The additional elements fall under “apply it” as using a generic computer to generate one or more control signals based on the characterization. See Mere Instructions to Apply an Exemption (see MPEP 2106.05(f)).
Claims 14-26 recite a system and are analogous to the method of claims 1-13. Therefore, the rejections of claim 1-13 above applies to claims 14-26.
Claims 27-39 recite a CRM and are analogous to the method of claims 1-13. Therefore, the rejections of claim 1-13 above applies to claims 27-39.
Claim Rejections - 35 USC § 102
The following is a quotation of the appropriate paragraphs of 35 U.S.C. 102 that form the basis for the rejections under this section made in this Office action:
A person shall be entitled to a patent unless –
(a)(1) the claimed invention was patented, described in a printed publication, or in public use, on sale, or otherwise available to the public before the effective filing date of the claimed invention.
(a)(2) the claimed invention was described in a patent issued under section 151, or in an application for patent published or deemed published under section 122(b), in which the patent or application, as the case may be, names another inventor and was effectively filed before the effective filing date of the claimed invention.
Claim(s) 1, 2, 10, 11, 12, 15, 23, 24, 25, 27, 28, 36, 37, 38 are rejected under 35 U.S.C. 102(a)(1) as being anticipated by J. Grezmak, J. Zhang, P. Wang, K. A. Loparo and R. X. Gao, "Interpretable Convolutional Neural Network Through Layer-wise Relevance Propagation for Machine Fault Diagnosis," in IEEE Sensors Journal, vol. 20, no. 6, pp. 3172-3181, 15 March15, 2020, doi: 10.1109/JSEN.2019.2958787 (“Grezmak”).
Regarding claim 1 and analogous claims 14, 27, Dalli teaches a method comprising: providing input data to one or more machine-learning models to generate output data,
the input data including an input value for each of N features associated with an operating state of a monitored asset and the output data including a predicted value of each of M features, wherein N is an integer greater than or equal to two and M is an integer greater than or equal to two (Grezmak Page 3172,
PNG
media_image1.png
174
366
media_image1.png
Greyscale
[a method comprising: providing input data to one or more machine-learning models to generate output data]
Page 3176 B. CNN Architectures and Training,
2) CNN Architecture: A CNN to classify the gray-scale 128×128 images is constructed with two convolutional layers with 12 7 × 7 kernels and 15 7 × 7 kernels, respectively, each followed by a max pooling layer with 2 × 2 pooling size. The output of the final pooling layer is flattened, and a fully connected layer consisting of 100 neurons is included before the 4 × 1 softmax classifier layer where each neuron corresponds to a specific fault type [the input data including an input value for each of N features associated with an operating state of a monitored asset and the output data including a predicted value of each of M features, wherein N is an integer greater than or equal to two and M is an integer greater than or equal to two].);
determining M sets of feature relevance values including a set of feature relevance values for each of the M predicted values, a particular set of feature relevance values associated with a particular predicted value, wherein each feature relevance value of the particular set of feature relevance values represents an estimate of a contribution of a respective one of the N input values to the particular predicted value (Grezmak page 3174 III. LRP FOR ANALYSIS OF LEARNED FAULT FEATURES BY CNN CLASSIFIERS para 3, In the LRP method, the attribution scores, known as relevance scores, are computed in a top-down manner with respect to the classifier’s structure. For NN classifiers such as CNNs, relevance scores are propagated starting from the output (i.e. classification) layer. The output layer relevance score is typically taken as the output layer pre-activation value corresponding to the class for which relevance scores are desired, and are obtained from the forward pass of an input through the CNN. Relevance scores are propagated to preceding layers such that the sum of the relevance scores in each layer is constant … We assume these assumptions are satisfied to justify the use of the z-rule. For pooling layers, the relevance scores are up-sampled by the pooling size to match the output dimensions of the previous layer and scaled by the pooling size to satisfy the conservation property of (4). It should be noted that the relevance calculation by (5) works for both 1-D and 2-D convolution, which depends on the network inputs.
page 3175 III. LRP FOR ANALYSIS OF LEARNED FAULT FEATURES BY CNN CLASSIFIERS para 4, The z-rule for relevance propagation as in (5) implies that relevance scores can take on both positive and negative values. Because the probability of an input belonging to a certain class is ultimately measured by the value of its corresponding output layer neuron, relevance scores can represent evidence for (positive values) and against (negative values) the classification decision [a particular set of feature relevance values associated with a particular predicted value]. By analyzing relevance scores in a global manner with respect to the input, regions or patterns in the inputs with high positive relevance scores can be taken as an indication of the patterns that most contribute to a classification decision [determining M sets of feature relevance values including a set of feature relevance values for each of the M predicted values,]
Page 3176 Fig. 2.
PNG
media_image2.png
530
1130
media_image2.png
Greyscale
[wherein each feature relevance value of the particular set of feature relevance values represents an estimate of a contribution of a respective one of the N input values to the particular predicted value;]);
aggregating, across the M sets of feature relevance values, feature relevance values for each of the N features to generate N aggregate feature relevance values (Grezmak page 3176 Fig. 2 Result Interpretation
PNG
media_image3.png
243
1122
media_image3.png
Greyscale
(Examiner Note: The relevance values are aggregating across the sets of relevance values for the convolution.)
Page 3177 Fig 3,
PNG
media_image4.png
363
1109
media_image4.png
Greyscale
V. Results/Discussion para 3 line1-10, The learned fault patterns for distinguishing between fault types are first analyzed. Relevance score heatmaps corresponding to different fault types classified by the same CNN model are illustrated in Fig. 3. Several qualitative observations can be made about the apparent patterns in the heatmaps. First, relevance scores appear to follow band-like patterns consisting of positive or negative relevance scores that span across the time domain of the input at specific locations along the frequency axis. Second, their specific locations along the frequency axis are unique for each fault type. [feature relevance values for each of the N features to generate N aggregate feature relevance values]);
and characterizing the operating state of the monitored asset based at least in part on the N aggregate feature relevance values (page 3177 V. Results/Discussion para 4, Next, the fault patterns for different sample inputs of the same fault type are analyzed. Heatmaps for four randomly chosen sample inputs (from four signal segments) corresponding to broken rotor bar fault and classified by the same CNN are shown in Fig. 4. The heatmaps for each sample show that the relevance scores follow the band-like patterns observed in Fig. 3 [based at least in part on the N aggregate feature relevance values], but the band locations along the frequency axis are essentially the same for each sample. For example, a negative band in the range of approximately 250 Hz to 355 Hz is seen in each sample, indicating that the discriminatory information for the CNN to diagnose broken rotor bar faults has remained stable for each sample [and characterizing the operating state of the monitored asset].).
Regarding claim 2 and analogous claims 15 and 28, Grezmak teaches the method of claim 1.
Grezmak teaches wherein the characterizing the operating state of the monitored asset is further based at least in part on the output data (Grezmak Page 3173 I Introduction para 5 line 6-11, We use Layer-wise Relevance Propagation (LRP) [20], a method for quantifying the contributions of individual values in inputs to the output (i.e. classification decision) of non-linear classifiers such as CNNs [based at least in part on the output data], to determine what information in the images is used by the CNN to distinguish between motor fault types [wherein the characterizing the operating state of the monitored asset]).
Regarding claim 10 and analogous claims 23 and 36, Grezmak teaches the method of claim 1.
Grezmak teaches wherein N is greater than M (Grezmak page 3176, CNN Architecture: A CNN to classify the gray-scale 128×128 images is constructed with two convolutional layers with 12 7 × 7 kernels and 15 7 × 7 kernels, respectively, each followed by a max pooling layer with 2 × 2 pooling size. The output of the final pooling layer is flattened, and a fully connected layer consisting of 100 neurons is included before the 4X1 softmax classifier layer where each neuron corresponds to a specific fault type. (i.e. the N number of features is greater than the M or last layer of features)).
Regarding claim 11 and analogous claims 24 and 37, Grezmak teaches the method of claim 1.
Grezmak teaches wherein one or more of the N features represents sensor data values from one or more sensors associated with the monitored asset (Grezmak page 3175 A. Experimental Setup, The experimental setup is the machinery fault simulator (MFS) shown in Fig. 1 that includes a 3-phase, 373 W induction motor. The system is capable of producing experiments under different motor operating conditions and fault scenarios; three rotor-related faults are included in this study, see Table I for details. To measure the vibration of the motor casing, a Kistler 8702B25 single-axis piezoelectric accelerometer is mounted on the motor casing in the vertical direction via a magnetic mounting base. Signals from this sensor are acquired using an NI 9230 C Series Sound and Vibration Input Module with a sampling rate of 10,240 Hz (614,000 data points per minute). The motor operates at a line frequency of 50 Hz, and the load supplied by the magnetic brake is set as a constant value of 5 lb-in [one or more sensors associated with the monitored asset].
Page 3175 -3176 B. CNN Architectures and Training, The details for pre-processing the vibration signals to create the inputs, as well as the corresponding CNN architectures to process these inputs, are described below: 1) Signal Pre-Processing: Collected vibration data is windowed into one-second segments (10, 240 data points), resulting in a total of 240 segments (4-minutes) of data. To create time-frequency image inputs, wavelet coefficients for each segment are computed by the CWT using a Morlet wavelet, chosen based on an energy-to-Shannon entropy ratio [28]. The wavelet coefficients are converted to time-frequency plots, and the region of a plot corresponding to the frequency range of 120 to 1, 000 Hz is converted to a gray-scale image of size 128×128 pixels. This specific frequency range is chosen such that information related to the vibration (fundamental and harmonics) of the motor rotational frequency, which are known to be affected by rotor-related faults, are preserved in the 128 × 128 pixel inputs [wherein one or more of the N features represents sensor data values].).
Regarding claim 12 and analogous claims 25 and 38, Grezmak teaches the method of claim 1.
Grezmak teaches wherein the determining the M sets of feature relevance values comprises performing layer-wise relevance propagation for each of the M predicted values (Grezmak page 3176,
PNG
media_image5.png
203
835
media_image5.png
Greyscale
Grezmak Page 3173 I Introduction para 5 line 6-11, We use Layer-wise Relevance Propagation (LRP) [20], a method for quantifying the contributions of individual values in inputs to the output (i.e. classification decision) of non-linear classifiers such as CNNs, to determine what information in the images is used by the CNN to distinguish between motor fault types.).
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.
Claim(s) 3, 4, 7, 9, 16, 17, 20, 22, 29, 30, 33, 35 are rejected under 35 U.S.C. 103 as being unpatentable over Grezmak in view of Agarwal, Piyush, Melih Tamer, and Hector Budman. "Explainability: Relevance based dynamic deep learning algorithm for fault detection and diagnosis in chemical processes." Computers & Chemical Engineering 154 (2021) (“Agarwal”).
Regarding claim 3 and analogous claims 16 and 29, Grezmak teaches the method of claim 1.
Gremak does not explicitly teach wherein the characterizing the operating state of the monitored asset comprises providing input based at least in part on the N aggregate feature relevance values to an operating state model to generate an operating state output.
However Agarwal teaches wherein the characterizing the operating state of the monitored asset comprises providing input based at least in part on the N aggregate feature relevance values to an operating state model to generate an operating state output (Agarwal page 6 Fig 4
PNG
media_image6.png
701
1179
media_image6.png
Greyscale
[wherein the characterizing the operating state of the monitored asset comprises providing input based at least in part on the N aggregate feature relevance values],
Page 6 3.2. Fault diagnosis methodology Para 1 line 5-24 First, the static DSAE is used to extract deep features and predict the type of fault in a process plant. For this task one-hot encoded outputs are utilized as the labels for training the model. Initially, the DSAE-NN is trained on the training dataset X l , y l using all the input-variables. The best performing model is chosen using a validation dataset X v , y v . LRP is subsequently implemented to explain the predictions of the selected DSAE-NN by computing the relevance of each input variable. The irrelevant features are removed by comparing the relevances to a threshold. Then an xDSAE NN is trained using the reduced training dataset X l r , y l r by successive iterations of pruning of irrelevant inputs and model re-training until the relevance of all the remaining input variables are above the threshold. Since data collected from chemical processes have strong dynamic/temporal correlations, the in- put data matrix X is augmented with observations at lprevious time steps for each input feature dimension (refer Eq. 10 ) and a DDSAE-NN is trained. The iterative procedure of discarding input variables from the reduced dynamic matrix X l D r is implemented and pruning and re-training is applied as long as the validation accuracy continues to increase after discarding features [an operating state model to generate an operating state output]).
Grezmak and Agarwal are considered to be analogous to the claim invention because they are in the same field of machine learning. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filling date of the claimed invention to have modified Grezmak to incorporate the teachings of Agarwal to input the relevance values into the operating state module. Doing so to prune irrelevant inputs and features from the fault detection model (Agarwal Page 6 3.2. Fault diagnosis methodology Para 1 line 5-24 First, the static DSAE is used to extract deep features and predict the type of fault in a process plant. For this task one-hot encoded outputs are utilized as the labels for training the model. Initially, the DSAE-NN is trained on the training dataset X l , y l using all the input-variables. The best performing model is chosen using a validation dataset X v , y v . LRP is subsequently implemented to explain the predictions of the selected DSAE-NN by computing the relevance of each input variable. The irrelevant features are removed by comparing the relevances to a threshold. Then an xDSAE NN is trained using the reduced training dataset X l r , y l r by successive iterations of pruning of irrelevant inputs and model re-training until the relevance of all the remaining input variables are above the threshold. Since data collected from chemical processes have strong dynamic/temporal correlations, the in- put data matrix X is augmented with observations at previous time steps for each input feature dimension (refer Eq. 10 ) and a DDSAE-NN is trained. The iterative procedure of discarding input variables from the reduced dynamic matrix X l D r is implemented and pruning and re-training is applied as long as the validation accuracy continues to increase after discarding features).
Regarding claim 4 and analogous claims 17 and 30, Grezmak and Agarwal teaches the method of claim 3.
Grezmak teaches wherein the operating state output indicates whether the operating state of the monitored asset is an anomalous operating state (Grezmak Page 3173 I Introduction para 5 line 6-11, We use Layer-wise Relevance Propagation (LRP) [20], a method for quantifying the contributions of individual values in inputs to the output (i.e. classification decision) of non-linear classifiers such as CNNs to determine what information in the images is used by the CNN to distinguish between motor fault types.
Page 3175 Table 1.
PNG
media_image7.png
186
463
media_image7.png
Greyscale
[whether the operating state of the monitored asset is an anomalous operating state].).
Regarding claim 7 and analogous claims 20 and 33, Grezmak teaches the method of claim 1.
Grezmak and Agarwal are combine in the same rational as set forth above with respect to claim 3 and analogous claims 16 and 29.
Grezmak does not explicitly teach wherein the one or more machine-learning models include one or more autoencoders.
However Agarwal teaches wherein the one or more machine-learning models include one or more autoencoders (Both the Deep Supervised Autoencoder NN (DSAE-NN)and Dy- namic Deep Supervised Autoencoder NN (DDSAE-NN) are used for FDD and are the basis for the explainable-pruning based method- ology presented in the previous section. The proposed fault detection algorithm is first used to extract deep features to detect if the process is operating in a normal or faulty region. Then, a fault diagnosis algorithm is applied in case the sample indicates faulty operation to identify the particular fault and possible root-cause of the occurring fault in the process using an DDSAE-NN. Since the latter is iteratively trained by using the LRP based pruning procedure that provides explainability of input variables the resulting DDSAE- NN model will be referred to as xDDSAE-NN [one or more autoencoders]).
Regarding claim 9 and analogous claims 22 and 35, Grezmak teaches the method of claim 1.
Grezmak and Agarwal are combine in the same rational as set forth above with respect to claim 3 and analogous claims 16 and 29.
Agarwal teaches wherein N is less than M (Agarwal Page 3 2. Preliminaries 2.1 Autoencoder neural networks (AE-NNs) line 16-24, The decoder reconstructs back the input variables from the feature or latent space z i ∈ R d z as per the following operation follows: ˆ x i = f d (W d z i + b d ) (2) where f d is a chosen activation function for the decoder, W d ∈ R d x ×d z and b d ∈ R d x is a decoder weight matrix and a bias vec- tor respectively. The ‘ tanh ’ function is used for both transforming the inputs into the latent variables and for reconstructing back the inputs from the latent variables as in the example presented here. The AE-NN is trained based on the following minimization problem
PNG
media_image8.png
412
418
media_image8.png
Greyscale
[wherein N is less than M]
page 6, Both the Deep Supervised Autoencoder NN (DSAE-NN)and Dynamic Deep Supervised Autoencoder NN (DDSAE-NN) are used for FDD and are the basis for the explainable-pruning based methodology presented in the previous section. The proposed fault detection algorithm is first used to extract deep features to detect if the process is operating in a normal or faulty region. (Examiner Note: The decoder starts with less features N that it produces in its output layer as M).
Claim(s) 5, 6, 18, 19, 31, and 32 are rejected under 35 U.S.C. 103 as being unpatentable over Grezmak in view of Agarwal and further in view of Wang, H.; Wang, H.; Jiang, G.; Li, J.; Wang, Y. Early Fault Detection of Wind Turbines Based on Operational Condition Clustering and Optimized Deep Belief Network Modeling. Energies 2019, 12, 984. (“Wang”).
Regarding claim 5 and analogous claims 18 and 31, Grezmak and Agarwal teaches the method of claim 3.
Grezmak and Agarwal are combine in the same rational as set forth above with respect to claim 3 and analogous claims 16 and 29.
Grezmak does not explicitly teach further comprising determining one or more residual data values based on comparison of each of the M predicted values to an actual value of a corresponding feature of the M features, wherein the input to the operating state model is further based on the one or more residual data values.
However Wang teaches further comprising determining one or more residual data values based on comparison of each of the M predicted values to an actual value of a corresponding feature of the M features, wherein the input to the operating state model is further based on the one or more residual data values (Wang Page 4 3. Proposed Health Monitoring Framework, In this study, a novel health monitoring framework for wind turbines under varying operating conditions is proposed, and its flowchart is shown in Figure 2. It is general and can be used for fault detection of different wind turbine subsystems and components. The main idea of the proposed framework is to build normal behavior models relying on only historical normal SCADA data from wind turbines and then perform fault detection based on the evaluation results of residuals between the predicted values and actual measured values. The changes of the residuals will give an indication of possible faults [further comprising determining one or more residual data values based on comparison of each of the M predicted values to an actual value of a corresponding feature of the M features]. Usually, normal test samples will produce a low residual value since they can well satisfy the learned normal model, whereas faulty test samples will produce high residual values and therefore be identified as faults. Generally, the proposed framework mainly consists of four sequential parts: operation condition partition, variable selection, model development and anomaly detection. The detailed procedures are summarized as follows:
Page 5 Figure 2.
PNG
media_image9.png
268
679
media_image9.png
Greyscale
[wherein the input to the operating state model is further based on the one or more residual data values.]).
Grezmak and Wang are considered to be analogous to the claim invention because they are in the same field of machine learning. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filling date of the claimed invention to have modified Grezmak to incorporate the teachings of Wang to determine residual data values. Doing so to determine fault detection of different systems using historical data (Wang Page 4 para 1 line 1-6 3. Proposed Health Monitoring Framework, In this study, a novel health monitoring framework for wind turbines under varying operating conditions is proposed, and its flowchart is shown in Figure 2. It is general and can be used for fault detection of different wind turbine subsystems and components. The main idea of the proposed framework is to build normal behavior models relying on only historical normal SCADA data from wind turbines and then perform fault detection based on the evaluation results of residuals between the predicted values and actual measured values.).
Regarding claim 6 and analogous claims 19 and 32, Grezmak in view of Agarwal and Wang teaches the method of claim 5 and analogous claims 18 and 31.
Grezmak and Agarwal are combine in the same rational as set forth above with respect to claim 3 and analogous claims 16 and 29.
Grezmak and Wang are combine in the same rational as set forth above with respect to claim 5 and analogous claims 18 and 31.
Wang further teaches wherein the one or more residual data values include M residual data values (Wang et al. (Early Fault Detection)
The changes of the residuals will give an indication of possible faults. Usually, normal test samples will produce a low residual value since they can well satisfy the learned normal model, whereas faulty test samples will produce high residual values and therefore be identified as faults. Generally, the proposed framework mainly consists of four sequential parts: operation condition partition, variable selection, model development and anomaly detection. The detailed procedures are summarized as follows:
Page 5 Figure 2.
PNG
media_image9.png
268
679
media_image9.png
Greyscale
Page 4-5 3. Propose Health Monitoring Frame work Collect normal SCADA data from multiple wind turbines on a wind farm.
(2) Choose operation parameters that characterize the complex operating conditions of wind turbines and segment the operation parameter data into K clusters using the k-means method and silhouette index. The obtained K clusters represent the corresponding K operating conditions, i.e., [𝐂1, 𝐂2, ⋯, 𝐂𝐾]. Then, divide the normal state data into corresponding K parts based on the partitioned operating conditions.
(3) Select appropriate modeling variables for each operating condition by combining three variable selection techniques, and the final selected variables for different operating clusters can be represented as [𝐕1, 𝐕2, ⋯, 𝐕𝐾].
(4) Build a normal behavior model under each operating condition using ODBNs to explore the sophisticated nonlinear characteristics among modeling variables, resulting in multiple DBN models, denoted as [DBN1, DBN2, ⋯, DBN𝐾] for K operating clusters.
(5) Calculate the threshold for abnormal detection under different operating conditions using the Mahalanobis distance (MD) measure to automatically identify the anomalies that occur in the operation of the wind turbines, i.e., [MD1, MD2, ⋯, MD𝐾].
(6) For the new incoming SCADA data, first recognize the operating condition 𝐂𝑖 that it belongs to, then select the corresponding modeling input variable 𝐕𝑖 and predict the output using the constructed DBNi. Next, compute the MD value and compare it with the threshold MDi under condition 𝐂𝑖, and then output the real-time online health monitoring results [wherein the one or more residual data values include M residual data values.]).
Claim(s) 8, 21, and 34 are rejected under 35 U.S.C. 103 as being unpatentable over Grezmak in view of Kursun et al. (US20210174562Al) (“Kursun”).
Regarding claim 8 and analogous claims 21 and 34, Grezmak teaches the method of claim 1.
Grezmak does not exility teach wherein N is equal to M.
However Kursun teaches wherein N is equal to M (Fig. 9
PNG
media_image10.png
436
816
media_image10.png
Greyscale
Para 0084 line 1-10, FIG. 7 provides an exemplary visualization output of grouped features, wherein a number data features 702 are processed and arranged in the visualization 704. The features may be organized into a number of subsets within the visualization 704. In some embodiments, the input features are ranked and/or highlighted within the visualization based on their importance and/or relevance to the decisioning event (i.e., in a heat map). FIG. 8 provides a block diagram of an exemplary relevance visualization heat map, in accordance with one embodiment of the invention.
Para 0085, Feature relevance calculations are an important step in explainability of AI solutions. The present invention utilizes a backpropagation of a feed-forward score to deter mine a relevance of the one or more features used to determine the score and provide explainability of the feedforward result to a human user. For example, a feed-forward score may identify an input interaction having one or more features as being associated with the determined misappropriation decision output. FIG. 9 provides a block diagram of a neural network for feed-forward scoring, in accordance with one embodiment of the invention. As illustrated in FIG. 9, the neural network comprises multiple layers of nonlinear functional structure, wherein output from a first layer of nodes is input to the next layer until a final decision is output (i.e. the system takes in N features and the number of features remains the same, thus N and M are equal)).
Grezmak and Kursun are considered to be analogous to the claim invention because they are in the same field of machine learning. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filling date of the claimed invention to have modified Grezmak to incorporate the teachings of Kursun to use an N number of inputs features that equal M number of features of the last layer. Doing so to provide an exemplary visualization of the output groped features (Kursun Para 0084 line 1-10, FIG. 7 provides an exemplary visualization output of grouped features, wherein a number data features 702 are processed and arranged in the visualization 704. The features may be organized into a number of subsets within the visualization 704. In some embodiments, the input features are ranked and/or highlighted within the visualization based on their importance and/or relevance to the decisioning event (i.e., in a heat map). FIG. 8 provides a block diagram of an exemplary relevance visualization heat map, in accordance with one embodiment of the invention).
Claim(s) 13, 26 and 39 are rejected under 35 U.S.C. 103 as being unpatentable over Grezmak in view of Trinh et al. (US11307570B2) (“Trinh”).
Regarding claim 13 and analogous claims 26 and 39, Grezmak teaches the method of claim 1.
Grezmak does not explicitly teach further comprising generating one or more control signals based on characterization of the operating state of the monitored asset.
However Trinh teaches further comprising generating one or more control signals based on characterization of the operating state of the monitored asset (Trinh Col 10 line 1-10, The maintenance recommendation engine 270 may provide one or more alerts (e.g., in the form of recommendations) for inspecting or repairing of pieces of equipment 150. For example, for a particular equipment 150 that newly generates a set of sensor data, the predictive maintenance server 110 may retrieve one or more machine learning models stored in the anomaly detection model store 250 and/or in the failure classification and prediction model store 260. One or more anomaly scores may be determined for the 10 particular equipment 150 [based on characterization of the operating state of the monitored asset].
Col 14 line 13-14, FIG. 4C is a block diagram illustrating a process for generating alerts and reporting anomalies, according to an embodiment. The predictive maintenance server 110 performs various steps, which may include (1) aggregating atomic scores into per-equipment health score (risk score) (2) raising alerts from atomic (e.g., minutely) anomaly) scores, subject to moving average smoothing (e.g., simple weighted moving average or exponentially weighted moving average) and consecutive-day constraints, and (3) ranking/ displaying via a user interface, alerts for an IoT fleet aggregated by various dimensions, for example, by region or by facility site [further comprising generating one or more control signals]).
Grezmak and Trinh are considered to be analogous to the claim invention because they are in the same field of machine learning. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filling date of the claimed invention to have modified Grezmak to incorporate the teachings of Trinh to generate a control signals based on the characterization of the monitored asset. Doing so to provide recommendations for inspecting or repairing of pieces of equipment (Trinh Col 10 line 1-3, Trinh Col 10 line 1-10, The maintenance recommendation engine 270 may provide one or more alerts (e.g., in the form of recommendations) for inspecting or repairing of pieces of equipment 150.).
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to ALFREDO CAMPOS whose telephone number is (571)272-4504. The examiner can normally be reached 7:00 - 4:00 pm M - F.
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, Michael J. Huntley can be reached at (303) 297-4307. 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.
/ALFREDO CAMPOS/Examiner, Art Unit 2129
/MICHAEL J HUNTLEY/Supervisory Patent Examiner, Art Unit 2129