System and method for detecting at least one anomaly in industrial process

Abstract

Systems and methods of detecting at least one anomaly in an industrial process (110) are disclosed. The method includes: receiving at least one signal associated with a computer-controlled machine (112, 114, and 220), the at least one signal indicating a runtime operation performed by the computer-controlled machine (112, 114, and 220); classifying the signal (X1, X2, X3) as valid or invalid by a trained discriminator, where the trained discriminator is generated using a generative adversarial network (GAN) architecture, by which operation performed by the computer-controlled machines (112, 114, 220) is predicted in response to the signal (X1, X2, X3); detecting an anomaly when the signal (X1, X2, X3) is classified as invalid or when the predicted operation is different from a runtime operation performed by the computer-controlled machines (112, 114, and 220); and determining a fault in the computer-controlled machine (112, 114 and 220) on the basis of the detected anomalies, the fault being determined by anomalies in the isolation spectrum and / or data points.

Description

Technical Field

[0001] The present invention relates to detecting anomalies in industrial processes. Specifically, the present invention relates to detecting anomalies using a combination of supervised and unsupervised learning techniques. Background Art

[0002] Industrial processes are typically implemented using one or more computer-controlled machines. Examples of computer-controlled machines include the automatic control of tools such as drills, lathes, milling machines, grinders, slot planers, and 3D printers. Other examples of computer-controlled machines include automated guided vehicles (AGVs), robotic arms, etc. Failures of computer-controlled machines negatively impact industrial processes. For example, the breakage of a drill bit during a machining process in a machine tool causes an interruption of the industrial process and results in unplanned downtime. Delays caused by machine downtime also represent a further cost factor. Strictly speaking, the breakage of the drill bit is an anomaly in the drilling process, which is evident in various signals in and on the machine tool.

[0003] Regardless of how the anomaly itself manifests, the anomaly can also be detected in different ways. Common methods for detecting anomalies include rule-based determination, supervised learning, and unsupervised learning. Rule-based determination is feasible when the standard or expected signals and anomalies are well defined. In particular, when the manifestation of the anomaly is unknown or the anomaly is not necessarily definable, rule-based determination quickly reaches its limits. In addition, it is difficult to define more complex signals, where the actual state depends especially on the operating conditions of the computer-controlled machine. In supervised learning, a trained neural network can be supervised (monitored) based on, for example, anomalies. The presence or absence of an anomaly is described by the labels (ground truth) required for supervised training. However, due to the nature of anomalies, few anomalies have the same characteristics in their manifestations. In addition, this method can only detect anomalies that already exist in the training data and are therefore known. In unsupervised learning, the neural network learns what the normal state of the signal looks like and detects anomalies therein. However, different from the other two methods, it cannot classify what kind of anomalies occur.

[0004] Therefore, there is a need to improve methods for detecting deviations in machine behavior and accordingly detecting anomalies in industrial processes. Summary of the Invention

[0005] In one example, the objective is achieved by a method for detecting at least one anomaly in an industrial process, wherein the industrial process is at least partially performed by a computer-controlled machine, and wherein the method comprises: receiving at least one signal associated with the computer-controlled machine, the at least one signal indicating a runtime operation performed by the computer-controlled machine, wherein the signal comprises at least one of one or more spectra and one or more data points, the one or more data points being generated by a sensor associated with the computer-controlled machine, and wherein the spectrum is an output of a frequency analysis of the data points received from the sensor; classifying the signal as valid or invalid by a trained discriminator, wherein the trained discriminator is generated using a generative adversarial network (GAN) architecture, and wherein the trained discriminator is generated based on a standard signal and a simulated signal generated by a generator of the GAN architecture, and wherein the standard signal comprises at least one of a default signal expected from the computer-controlled machine and a previous signal generated from the computer-controlled machine; predicting, in response to the signal, an operation performed by the computer-controlled machine by the trained discriminator; detecting an anomaly when the signal is classified as invalid or when the predicted operation is different from the runtime operation performed by the computer-controlled machine; and determining a fault in the computer-controlled machine based on the detected anomaly, wherein the fault is determined by isolating the anomaly in the spectrum and / or data points.

[0006] In another example, the objective is achieved by a computer-readable medium having machine-readable instructions stored therein, which when executed by a processing unit cause the processing unit to perform the steps of the method disclosed herein.

[0007] In yet another example, the objective is achieved by a system for detecting anomalies in an industrial process, the system comprising a server including at least one processing unit configured to coordinate the execution of one or more method steps disclosed herein that are executable in a distributed computing environment.

[0008] In yet another alternative example, the object of the present invention is achieved by a method for detecting at least one anomaly in the operation of a computer-controlled machine, wherein an industrial process is at least partially performed by the computer-controlled machine, and wherein the method comprises: receiving at least one signal associated with the computer-controlled machine, the at least one signal indicating a runtime operation performed by the computer-controlled machine, wherein the signal comprises at least one of one or more spectra and one or more data points, the one or more data points being generated by a sensor associated with the computer-controlled machine; classifying the signal as valid or invalid by a trained discriminator, wherein the trained discriminator is generated using a generative adversarial network (GAN) architecture, wherein the trained discriminator is generated based on a standard signal and a simulated signal generated by a generator of the GAN architecture, wherein the standard signal comprises at least one of a default signal expected from the computer-controlled machine and a previous signal generated from the computer-controlled machine; predicting, in response to the signal, an operation performed by the computer-controlled machine by the trained discriminator; detecting an anomaly when the signal is classified as invalid or when the predicted operation is different from the runtime operation performed by the computer-controlled machine; and determining a fault in the computer-controlled machine based on the detected anomaly, wherein the fault is determined by isolating the anomaly in the spectrum and / or data points.

[0009] The present invention advantageously combines supervised and unsupervised learning methods to detect anomalies in industrial processes. Further, the present invention provides a method for isolating one or more faults based on the detected anomalies. The present invention detects an anomaly when a trained discriminator classifies a signal from a computer-controlled machine as invalid or predicts a different operation than the operation performed by the machine. This method of classifying signals and operations enables accurate anomaly detection. By training the discriminator with previous signals generated from the machine, the present invention adapts to the operating conditions of the machine, which results in accurate anomaly detection.

[0010] Furthermore, the present invention is not limited to sensor signals. In operation, process signals as well as sensor signals are analyzed. These signals can be measured internally or externally and can be used without knowing what the signal looks like. Thus, the cost of using industry experts for anomaly detection and analysis can be reduced. Additionally, during runtime, only the trained discriminator is executed, so the total computational effort is significantly lower, and thus signals from machines with high data throughput can also be analyzed. The runtime efficiency enables the present invention to be implemented on devices with low computational power. Thus, the present invention saves hardware costs (since several applications can run in parallel on one edge device) and does not require any changes to the machine.

[0011] Accurate fault isolation is achieved by improving the accuracy of anomaly detection. In addition, by locating faults and quantitatively analyzing one or more anomalies, remote condition monitoring is made more effective in cases where anomalies were previously unknown. Moreover, the runtime efficiency of the present invention enables fault isolation and real-time anomaly analysis. Thus, the present invention can not only perform accurate anomaly detection but also perform rapid analysis to identify faults. Rapid fault diagnosis reduces downtime and increases the overall efficiency of the industrial processes being performed.

[0012] Before describing the proposed invention in more detail, it should be understood that throughout this patent document, various definitions are provided for certain words and phrases, and one of ordinary skill in the art will understand that these definitions apply to the prior and future use of these defined words and phrases in many, if not most, instances. Although some terms may encompass multiple embodiments, the appended claims may specifically limit these terms to a particular embodiment. It should also be understood that features explained in the context of the proposed method can also be included by the proposed system by appropriately configuring and adjusting the system, and vice versa.

[0013] As used herein, "industrial process" refers to one or more processes performed in an industrial facility. "Industrial facility" refers to a facility for manufacturing, production that can be semi-automated or fully automated. For example, an industrial facility can include a laboratory facility, a construction facility, a manufacturing facility, etc. An industrial facility can also refer to a combination of the above facilities. Examples of industrial processes can include painting the body of an automobile, assembling an automobile or a part of an automobile, manufacturing a compound, transporting materials in a warehouse, etc.

[0014] Industrial facilities include machines, such as machine tools, mobile robots, automated transportation systems, etc. The machines are computer-controlled machines (also referred to as machines herein) and can thus be controlled using software executed on the machine itself or on computing resources separate from the machine. The present invention is described in detail from the perspective of a machine tool such as a drill. One of ordinary skill in the art will understand that a drill is an example. The method disclosed by the present invention can be applied to any computer-controlled machine, such as a robotic arm or a 3D printer.

[0015] The present invention proposes a GAN that learns the target state of a machine based on the default signal expected from the machine and the previous signals generated by the machine. At runtime, the trained discriminator of the GAN classifies the signals and operations performed by the machine. Through this classification, previously unknown anomalies can be detected and classified to locate faults. The method of the present invention is described in detail below.

[0016] The method includes: receiving at least one signal associated with a computer-controlled machine, the at least one signal indicating a runtime operation performed by the computer-controlled machine. The signal includes at least one of a spectrum and data points generated by a sensor associated with the computer-controlled machine. In an embodiment, the signal includes a tuple of data points generated by a sensor associated with the machine and a spectrum. An example of a spectrum is the Fourier transform of the data points. The spectrum is the output of a frequency analysis of the data points received from the sensor. In another embodiment, the signal includes data points from the sensor, a spectrum, and a spectrogram. A spectrogram is a visual representation of the signal intensity over time at different frequencies. Thus, the method can include: generating a spectrum and a spectrogram using Fourier analysis.

[0017] The method further includes: classifying the signal as valid or invalid by a trained discriminator. The trained discriminator is generated using a generative adversarial network (GAN) architecture. In an embodiment, the GAN architecture uses an auxiliary classifier and is an ACGAN architecture. The ACGAN architecture is a type of GAN, specifically a conditional GAN that trains the discriminator to predict classes. In the present invention, the ACGAN architecture enables the trained discriminator to classify the signal into valid and invalid classes. Thus, the method can include: classifying the signal as valid or invalid by a trained discriminator generated using the ACGAN architecture. The trained discriminator is generated based on standard signals and simulated signals generated by a generator of the GAN architecture. The standard signals include at least one of a default signal expected from the computer-controlled machine and a previous signal generated from the computer-controlled machine.

[0018] The method can further include: generating the trained discriminator by training the discriminator of the GAN architecture based on standard signals and simulated signals generated by the generator of the GAN architecture. In an embodiment using the ACGAN architecture, the method can include: generating the trained discriminator by training the discriminator of the GAN architecture based on standard signals and simulated signals generated by the generator of the GAN architecture.

[0019] The method can include: updating discriminator weights of the discriminator based on at least one of a discriminator loss function and a generator loss function. The discriminator loss function quantifies misclassifying a standard signal as invalid. The generator loss function quantifies classifying a simulated signal as invalid. In fact, the generator loss function quantifies the accuracy of the discriminator. The trained discriminator is trained to account for characteristics of the machine by using the trained standard signals and by updating the discriminator weights based on the discriminator loss function and the generator loss function.

[0020] In some embodiments, the method can further include: training the generator based on the output of the discriminator of the GAN. Additionally, the method can include: updating the generator weights of the generator based on a generator loss function. The generator weights are updated such that the generator is penalized when the discriminator classifies the analog signal as invalid. The generator loss function is used to train the generator to induce the discriminator to classify the analog signal as valid. The interaction between the generator loss function and the discriminator loss function improves the quality of the trained discriminator.

[0021] The method can further include: training the discriminator of the GAN architecture using the updated training dataset. The updated training dataset includes updated standard operations and updated standard signals generated during the operation of the computer-controlled machine. With the updated training dataset, the discriminator stays up-to-date with the operating conditions of the machine. In some embodiments, the update of the training dataset is performed at fixed intervals. In other embodiments, the upgrade of the training dataset is performed only when an anomaly is detected.

[0022] The method can further include: comparing the standard operations and standard signals with the updated standard operations and updated standard signals to identify creep wear in the machine; and retraining the discriminator based on the creep wear. Thus, the present invention advantageously detects creep wear in the machine and addresses the creep wear while detecting anomalies.

[0023] The method includes: predicting, by the trained discriminator, an operation performed by the computer-controlled machine in response to the signal; as used herein, an "operation" performed by the machine refers to an action or function performed by the machine. For example, operations for a machine tool include a non-exhaustive set of actions such as hold, idle, up cut, down cut, up feed, down feed, etc. The type and nature of the operation vary based on the machine. In another example where the machine is a robotic arm, the operations can include up, down, forward, backward, rotate, insert, etc. The nature of the operation affects the generation of the trained discriminator. Thus, the method can include: generating, by the generator, simulated operations, and the discriminator of the GAN architecture is trained to predict the actions performed / performable by the machine. Thus, considering the example of a machine tool, the simulated operations include actions such as hold, idle, up cut, down cut, up feed, down feed performable by the machine tool.

[0024] By predicting the operations of the trained discriminator, the reliance on the signal validity is balanced. Thus, even if the predicted action is not the action performed by the machine and the signal is classified as valid, an anomaly can be detected.

[0025] The method includes detecting an anomaly when a signal is classified as invalid or when a predicted operation is different from a runtime operation performed by the machine. The present invention advantageously improves the accuracy of anomaly detection and detects an anomaly when either condition (invalid signal or incorrect operation prediction) occurs.

[0026] The method further includes: determining a fault in a computer-controlled machine based on the detected anomaly.

[0027] The fault is determined by isolating the anomaly in the spectrum and / or data points. To isolate the anomaly, the method can include: identifying at least a portion of the spectrum and / or anomalous data points in the signal classified as invalid. Thus, the present invention identifies the data points that lead to the detected anomaly. These data points are referred to as anomalous data points. Additionally, the method isolates the spectral portion that causes the signal to be classified as spurious or causes an incorrect operation prediction.

[0028] The identification of the anomalous data points and / or portion of the spectrum can be used to identify the runtime operations performed that led to the detected anomaly. These operations are referred to as anomalous runtime operations. Thus, the method can include: mapping the portion of the spectrum and / or anomalous data points to at least one anomalous runtime operation. The method can further include: determining the control commands provided to the computer-controlled machine associated with the anomalous runtime operation.

[0029] In an embodiment, an explanation generated for the signal is used to determine the anomalous runtime operation. Techniques such as saliency maps, layer-wise relevance propagation, gradient-based convex optimization, etc. can be used to generate the explanation. Thus, the method can include: generating an explanation that illustrates the impact of the signal classified as invalid on the trained discriminator, and mapping the portion of the spectrum and / or anomalous data points to the anomalous runtime operation based on the explanation. The method can further include: analyzing the explanation to determine the root cause of at least one of the anomaly and the anomalous runtime operation.

[0030] Since it is known which action was taken at runtime and the signal is genuine, a saliency map (or LRP, GradCAM, or other method) can be used to determine which input now contributes to the anomalous runtime operation. Specifically, it can be determined what needs to be changed in the signal in order to classify the signal as "genuine" or predict the correct action.

[0031] By using interpretive execution for root cause analysis, the functionality of the trained discriminator is analyzable and thus anomalies can also be analyzed and isolated to identify faults. The analysis of anomalies can further be used to reduce future anomalies. In one embodiment, the method includes: determining a modification to a signal classified as invalid, wherein the trained discriminator classifies the modified signal as valid; and validating the modified signal when the predicted modified action for the modified signal is one of the standard actions associated with a standard signal, wherein the standard action is performed to generate the standard signal. Thus, the present invention is advantageously capable of performing prognostics and reduction operations to avoid anomalies.

[0032] The above method is implemented using a system that includes a processing unit that executes the steps disclosed herein in a distributed computing environment. The system may include: a data hub configured to store operations, signals, standard operations, and standard signals associated with the industrial process. The processing unit is configured to train the generator and discriminator of the GAN using the standard operations and standard signals. The training may be performed by computing resources that can handle high computational requirements, such as a cloud computing platform. Additionally, the detection of runtime anomalies may be done by computing resources with low latency. For example, on the machine itself or on the industrial controller that controls the machine.

[0033] The technical features of the present disclosure have been outlined quite broadly above so that those skilled in the art may better understand the detailed description below. Additional features and advantages of the present disclosure that form the subject matter of the claims will be described hereinafter. Those skilled in the art will understand that they can readily use the disclosed concepts and specific embodiments as a basis for modifying or designing other structures for the same purpose of implementing the present disclosure. Those skilled in the art will also recognize that such equivalent constructions do not depart from the scope of the present disclosure in its broadest form. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Hereinafter, the present invention will be described using the embodiments shown in the drawings.

[0035] Figure 1 An industrial environment including a system for detecting anomalies in an industrial process according to an embodiment of the present invention is shown;

[0036] Figure 2A -C shows training a generative adversarial network (GAN) to detect anomalies in an industrial process according to an embodiment of the present invention;

[0037] Figure 3A -C shows an example of detecting anomalies in an industrial process according to an embodiment of the present invention;

[0038] Figure 4A method for detecting at least one anomaly in an industrial process according to an embodiment of the present invention is shown. Detailed implementation

[0039] Hereinafter, embodiments for implementing the present invention are described in detail. Various embodiments are described with reference to the accompanying drawings, in which the same reference numerals are used throughout to refer to the same elements. In the following description, for the purpose of explanation, numerous specific details are set forth to provide a thorough understanding of one or more embodiments. Obviously, these embodiments can be implemented without these specific details.

[0040] Figure 1 An industrial environment 130 according to an embodiment of the present invention is shown. The industrial environment includes a system 120 for detecting anomalies in an industrial process 110. The industrial process 110 is described by computer-controlled machines 112 and 114 and sensors 116. The system 120 is a control system including one or more processors (processing units), which are configured to directly run applications for data acquisition, preprocessing, and analysis at the computer-controlled machines 112 and 114. The control system 120 has two separate Ethernet interfaces. One is for a cloud computing platform 160 via the Internet 150, and the other is for connecting to the industrial environment 130. In addition to data processing, the system 120 is configured to generate control signals in response to detected anomalies, such as stopping the computer-controlled machines 112 and 114, in order to minimize damage.

[0041] Figure 1 A user device 170 capable of displaying a graphical user interface (GUI) is also shown, and the detected anomalies are displayed on the GUI. Figure 1 A data hub 180 is also shown, which is configured to store operations, signals, standard operations, and standard signals associated with the industrial process 130. In an embodiment, the functions of the system 120 can be partially implemented on the cloud platform 160 and partially in the industrial environment 130. In such an embodiment, one or more processors are configured to train the generator and discriminator of a generative adversarial network (GAN) with the standard operations and standard signals stored in the data hub 180.

[0042] Anomaly detection by system 120 can be divided into three phases. When the computer-controlled machines 112 and 114 are installed, necessary data is collected as part of a calibration run to train the GAN architecture. In an embodiment, calibration can be performed at regular intervals after the installation of machines 112 and 114. The old calibration data is archived / stored in the data concentrator 180 for continuous comparison and detection of whether creep wear has occurred. After calibration, the GAN architecture is used to monitor the computer-controlled machines 112 and 114 during operation and detect anomalies. If an anomaly has been detected, the system 120 analyzes the anomaly / enables analysis via the user device 170. For example, the system 120 can transmit data to the GUI to display the interference frequencies derived from the detected anomalies. A technician using the user device 170 can further interpret the displayed frequencies.

[0043] The operation of system 120 for detecting anomalies in the industrial process 110 is further shown in Figures 2A-2C FIG. Figures 2A-2C FIG. shows the training of a GAN according to an embodiment of the present invention to monitor anomalies in the industrial process 110.

[0044] Figure 2A FIG. shows that the industrial process performed by machine 220 can be classified as action "a", which is performed to generate signal "x". The action is an operation that is performed or executable by machine 220. Machine 220 receives an instruction to perform the action from controller 210. In some embodiments, controller 210 is part of machine 220. In an embodiment, action a includes holding, idling, up cutting, down cutting, up feeding, and down feeding that can be performed by machine 220. In an embodiment, signal x includes a tuple of data points generated from sensors associated with machine 220 and a spectrum. The spectrum is the Fourier transform of the sensor data points. Signal x is referred to as a standard signal or a real signal. Similarly, action a is referred to as a standard operation / standard action. Records of standard signals and standard operations for machine 220 can also be created using simulated data by a virtual machine that replicates the operation of machine 220.

[0045] Figure 2B FIG. shows the process of training the generator 230 and discriminator 240 of the GAN architecture. Generator 230 is shown as an artificial intelligence module executed on device 230. Discriminator 240 is shown as an artificial intelligence module executed on device 240.

[0046] In Figure 2BIn it, an Auxiliary Classifier GAN (ACGAN) is shown. The ACGAN architecture is a modified GAN architecture. The generator 230 receives an action a' and a noise vector z. The generator 230 generates a signal x', also referred to as an analog signal x'. The discriminator 240 receives the analog signal x', a standard signal x, and a standard operation a. In response, the discriminator 240 predicts the authenticity of the standard signal x and / or the analog signal x'. In addition, the discriminator predicts the corresponding action that causes x or x'.

[0047] After training the generator 230 and the discriminator 240 are pooled, the generator 230 is capable of generating a random signal x' for an associated action a'. The training of the discriminator 240 is performed based on the standard signal x and the analog signal x' generated by the generator 230. The discriminator 240 classifies the standard signal x as valid or invalid during training. The weights of the discriminator 240 are updated based on the discriminator loss function and / or the generator loss function. The discriminator loss function quantifies misclassifying the standard signal x as invalid. The generator loss function quantifies classifying the analog signal x' as invalid. The training of the generator 230 is performed based on the output of the discriminator 240. The generator weights of the generator 230 are updated based on the generator loss function, wherein the generator weights are updated such that the generator 230 is penalized when the discriminator 240 classifies the analog signal as invalid / false.

[0048] In one embodiment, the discriminator 240 is trained using an updated training dataset (i.e., updated standard signals with associated actions). Thus, the updated training dataset includes updated standard operations and updated standard signals generated during the runtime of the machine 220. The standard operations (i.e., action a and standard signal x) are compared with the updated standard operations and updated standard signals to identify creep wear in the machine. Then the discriminator 240 is trained based on the creep wear.

[0049] Figure 2C An operation phase for detecting an anomaly in an industrial process during runtime is shown. The operation is to be performed by the machine 220, and the operation is shown by an input of action A as a control signal to the machine 220. In response to action A, the machine 220 generates a signal X. Then the trained discriminator D classifies the signal X as real or false. Additionally, the trained discriminator D predicts the action When the signal X is classified as false or if the action is not action A, an anomaly is detected.

[0050] Figure 3A -C shows an example of detecting an anomaly in an industrial process according to an embodiment of the present invention. In Figure 3AIn [case 0], the action is "hold" and signal X1 is generated from machine 220. The trained discriminator D predicts that X1 is fake and the action is "up cut". Since X1 is classified as fake and the additionally predicted action is not "hold", an anomaly is detected. In Figure 3B In [case 1], the action is "up cut" and signal X2 is generated from machine 220. Even though signal X2 is classified as real, the predicted action is not "up cut". Thus, an anomaly is detected. In Figure 3C In [case 2], the action is "hold" and signal X3 is generated from machine 220. Even though the predicted action is "hold", signal X3 is classified as fake. Thus, an anomaly is detected.

[0051] When an anomaly is detected, signals X1, X2, and X3 are analyzed. Signals X1, X2, and X3 include tuples of time series data generated by sensors associated with machine 220 and associated spectra. In some embodiments, spectrograms, which are also part of signals X1, X2, and X3, are formed in addition to the time series data and spectra. This analysis includes determining exactly where the anomaly in signals X1, X2, and X3 has manifested. Given which action a is made at runtime and signals X1, X2, and X3 are real, saliency maps (or LRP, GradCAM, or other methods) can be used to determine which inputs now contribute to the prediction. Specifically, the analysis includes determining what needs to be changed in signals X1, X2, and X3 in order to classify the signals as "real" or predict the correct action. Input gradients precisely provide this information. For example, the input gradient can precisely determine the location of a break in signal X1 because the input gradient will specifically indicate a peak in signal X1. Since this peak is at a very high frequency, the operator will correspondingly find a difference in this spectrum. In another example, an unwanted vibration (oscillation) can be detected in the spectrum, and the gradient can also be used to determine which frequency (or frequency band) has an oscillation that interferes with machine 220.

[0052] Figure 4 A method for detecting at least one anomaly in an industrial process is shown, the industrial process being at least partially performed by a computer-controlled machine.

[0053] The method begins at step 410 by receiving at least one signal associated with the machine, the at least one signal indicating a runtime operation performed by the machine. The signal includes at least one of one or more spectra and one or more data points generated by sensors associated with the machine.

[0054] Step 420 includes classifying the signal as valid or invalid by a trained discriminator. The trained discriminator is generated using a generative adversarial network (GAN) architecture and is generated based on a standard signal and a simulated signal generated by a generator of the GAN architecture. The standard signal includes at least one of a default signal expected from the machine and a previous signal generated from the machine.

[0055] Step 430 includes predicting an operation performed by the machine by the trained discriminator in response to the signal.

[0056] Step 440 includes detecting an anomaly when the signal is classified as invalid or when the predicted operation is different from the operating-time operation performed by the machine.

[0057] Step 450 includes determining a fault in the computer-controlled machine based on the detected anomaly, wherein the fault is determined by isolating the anomaly in the spectrum and / or data points. Step 450 can also include identifying at least a part of the spectrum and / or anomalous data points in the signal classified as invalid. Mapping the part of the spectrum and / or anomalous data points to at least one anomalous operating-time operation. Determining an operation control command provided to the machine from the anomalous operating-time. Additionally, step 450 can include generating an explanation illustrating the impact of the signal classified as invalid on the trained discriminator. Based on the explanation, mapping the part of the spectrum and / or anomalous data points to the anomalous operating-time operation. Step 450 can also include analyzing the explanation to determine the root cause of the anomaly and the anomalous operating-time operation.

[0058] The present invention can also include, at step 450, determining a modification to the signal classified as invalid, wherein the trained discriminator classifies the modified signal as valid. Validating the modified signal when the modified operation predicted for the modified signal is one of the standard operations associated with the standard signal, wherein the standard operation is performed to generate the standard signal.

[0059] As shown in the above figures, the present invention shows a method in which the system 120 can independently learn the standard operations of machines (e.g., 112, 114, and 120) based on processing or sensor signals and then detect anomalies. The detected anomalies can be interpreted and classified in a resource-efficient manner at runtime. The present invention also allows for a more precise analysis of anomalies such that interference frequencies or anomalies in the signals can be identified. Additionally, the implementation of the present invention is by machine-readable instructions and can thus be retrofitted in all machines via software updates.

[0060] The present invention can take the form of a computer program product that includes program modules accessible from a computer-usable or computer-readable medium that stores program code for use by or in conjunction with one or more computers, processors, or instruction execution systems. For the purposes of this specification, a computer-usable or computer-readable medium can be any device that can contain, store, communicate, propagate, or transport a program for use by or in conjunction with an instruction execution system, apparatus, or device. The medium can be an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system (or apparatus or device) or a propagation medium, which, as such, is not included in the definition of a physical computer-readable medium, which includes semiconductor or solid state memories, magnetic tape, removable computer disks, random access memory (RAM), read-only memory (ROM), hard disk, and optical disks (such as compact disc read-only memory (CD-ROM), optical disk read / write, and DVD). As is known to those skilled in the art, the processors and program code for implementing each aspect of the present technology can be centralized or distributed (or a combination thereof).

[0061] Although the present invention has been described in detail with reference to certain embodiments, it should be understood that the present invention is not limited to these embodiments. Based on the present disclosure, many modifications and variations will be apparent to those skilled in the art, as described above, without departing from the scope of the different embodiments of the present invention. Accordingly, the scope of the present invention is indicated by the following claims, rather than by the foregoing description. All advantageous embodiments claimed in the method claims can also be applied to the system / apparatus claims.

Claims

1. A method for detecting at least one anomaly in an industrial process (110), wherein, The industrial process (110) is at least partially performed by computer-controlled machines (112, 114, and 220), and the method includes: Receiving at least one signal associated with the computer-controlled machines (112, 114, and 220), the at least one signal indicating a runtime operation performed by the computer-controlled machines (112, 114, and 220), where the signal (X1, X2, X3) includes at least one of one or more spectra and one or more data points, the one or more data points being generated by sensors associated with the computer-controlled machines (112, 114, and 220), and the spectrum being an output of a frequency analysis of the data points received from the sensors; Classifying the signal (X1, X2, X3) as valid or invalid by a trained discriminator, where the trained discriminator is generated using a generative adversarial network (GAN) architecture, and the trained discriminator is generated based on a standard signal (x) and a simulated signal (x') generated by a generator of the GAN architecture, and the standard signal (x) includes at least one of a default signal expected from the computer-controlled machines (112, 114, and 220) and a previous signal generated from the computer-controlled machines (112, 114, and 220); Predicting, in response to the signal (X1, X2, X3), an operation performed by the computer-controlled machines (112, 114, and 220) by the trained discriminator; Detecting the anomaly when the signal (X1, X2, X3) is classified as invalid or when the predicted operation is different from the runtime operation performed by the computer-controlled machines (112, 114, and 220); and Determining a fault in the computer-controlled machines (112, 114, and 220) based on the detected anomaly, where the fault is determined by isolating the anomaly in the spectrum and / or the data points.

2. The method according to claim 1, wherein, Isolating the anomaly includes: Identifying at least a part of the spectrum and / or anomalous data points in the signal (X1, X2, X3) classified as invalid; Mapping the part of the spectrum and / or the anomalous data points to at least one anomalous runtime operation; and Determining a control command provided to the computer-controlled machines (112, 114, and 220) associated with the anomalous runtime operation.

3. The method according to claim 2, further comprising: Generating an explanation illustrating the impact of the signal (X1, X2, X3) classified as invalid on the trained discriminator; Based on the explanation, mapping the part of the spectrum and / or the anomalous data points to the anomalous runtime operation; And Analyzing the explanation to determine a root cause of at least one of the anomaly and the anomalous runtime operation.

4. The method according to claim 1, further comprising: Determine a modification to the signals (X1, X2, X3) classified as invalid, where the trained discriminator classifies the modified signal as valid; and Verify the modified signal when the predicted modified operation for the modified signal is one of the standard operations associated with the standard signal (x), where the standard operation is performed to generate the standard signal (x).

5. The method according to any one of claims 1 and 2, further comprising: Generate the trained discriminator by: Training the discriminator of the GAN architecture based on the standard signal (x) and the simulated signal (x’) generated by the generator of the GAN architecture, where the trained discriminator classifies the standard signal (x) as valid or invalid; and Updating the discriminator weights of the discriminator based on at least one of a discriminator loss function and a generator loss function, where the discriminator loss function quantifies misclassifying the standard signal (x) as invalid, and where the generator loss function quantifies classifying the simulated signal (x’) as invalid.

6. The method according to claim 5, wherein, Generating the trained discriminator further comprises: Training the generator based on the output of the discriminator; Updating the generator weights of the generator based on the generator loss function, where the generator weights are updated such that the generator is penalized when the discriminator classifies the simulated signal as invalid.

7. The method according to claim 5, further comprising: Training the discriminator of the GAN architecture with an updated training dataset, where the updated training dataset contains updated standard operations and updated standard signals (x) generated during the operation of the computer-controlled machines (112, 114, and 220); Comparing the standard operations and the standard signal (x) with the updated standard operations and the updated standard signal (x) to identify creep wear in the machine; and Retraining the discriminator based on the creep wear.

8. The method according to at least one of the preceding claims, wherein, The signal further comprises a spectrogram of the data points generated from the sensors associated with the computer-controlled machines (112, 114, and 220).

9. The method according to at least one of the preceding claims, wherein, The signal further comprises a tuple of the data points and the frequency spectrum generated from the sensors associated with the computer-controlled machines (112, 114, and 220), where the frequency spectrum is the Fourier transform of the data points.

10. The method according to at least one of the preceding claims, wherein, The computer-controlled machines (112, 114, and 220) are machine tools, and where the operations performed by the machines and the simulated operations include actions that can be performed by the machine tools including holding, idling, up cutting, down cutting, up feeding, and down feeding.

11. A computer-readable medium having machine-readable instructions stored therein, which when executed by a processing unit cause the processing unit to perform the steps according to claims 1 to 10.

12. A system (120) for detecting anomalies in an industrial process (110), the system comprising a server including at least one processing unit configured to coordinate the execution of one or more method steps according to claims 1 to 10 that can be performed in a distributed computing environment.

13. The system (120) according to claim 12, further comprising a data hub configured to store operations, signals, standard operations, and standard signals (x) associated with the industrial process (110), and wherein, The processing unit is configured to train the generator and the discriminator of the generative adversarial neural network (GAN) with the standard operations and the standard signal (x).

Images