An anomaly detection method using a system for detecting anomalies in business assets, a system for detecting anomalies in business assets, and a method for using the system for detecting anomalies in business assets to repair business assets.

Abstract

A method for anomaly detection in a business asset includes collecting a source domain dataset corresponding to a first operating condition of the business asset and collecting a target domain dataset corresponding to a second operating condition of the business asset, inputting the source domain dataset and the target domain dataset as input data to a neural network, extracting features from the input data using the neural network, reducing the dimensionality of the features to form reduced features, and classifying the reduced features into a normal class and an abnormal class using a one-class classifier. Samples from the source domain dataset belong to a healthy class and a faulty class. Samples from the target domain dataset belong to the healthy class. A first subset of features extracted from the input data identifies the healthy class, and a second subset of features extracted from the input data is domain-invariant.

Description

[Technical Field]

[0001] Cross-reference of related applications This application claims priority to U.S. Provisional Application No. 63 / 431,384, filed on 9 December 2022, the entire contents of which are incorporated herein by reference.

[0002] This invention relates to a tool for monitoring the health of business assets or industrial assets. [Background technology]

[0003] Business or industrial assets or machinery operate within various “domains” or “operating systems,” including, for example, different rotational speeds, ambient temperatures, and loads. For the same machine condition (e.g., healthy, failure type I, failure type II, etc.), data signatures may differ depending on the operating system. For example, the vibration of a failure type I machine at 50 RPM may be similar to the vibration of a healthy machine at 100 RPM. The difference between the two is a “domain mismatch,” and domain adaptation must be applied to eliminate the domain mismatch. Transfer learning uses knowledge gained from problem-solving in one domain ("source") to enhance and improve problem-solving skills in another domain ("target"). [Overview of the Initiative]

[0004] Further features and advantages of this disclosure are described below in detail and will be partially apparent to those skilled in the art from that description, or will be recognized by carrying out the embodiments described herein, including the detailed description that follows the claims and the accompanying drawings.

[0005] In one embodiment, a method for detecting anomalies in business assets includes: collecting a source domain dataset corresponding to a first operating condition of the business asset; collecting a target domain dataset corresponding to a second operating condition of the business asset; inputting the source domain dataset and the target domain dataset as input data into a neural network; extracting features from the input data using the neural network; reducing the dimensionality of the features to reduce them into reduced features; and classifying the reduced features into normal and abnormal classes using a one-class classifier. Samples from the source domain dataset belong to the healthy and faulty classes. Samples from the target domain dataset belong to the healthy class. A first subset of the features extracted from the input data identifies the healthy class, and a second subset of the features extracted from the input data is domain-invariant.

[0006] In another embodiment, a system for anomaly detection in business assets comprises one or more processors and a non-temporary computer-readable medium storing instructions. The one or more processors, by executing the instructions, receive a source domain dataset corresponding to a first operating condition of the business asset, receive a target domain dataset corresponding to a second operating condition of the business asset, input the source domain dataset and the target domain dataset as input data into a neural network, the neural network extracts features from the input data, reduces the dimensionality of the features to make them reduced features, and uses a one-class classifier to classify the reduced features into normal and abnormal classes. Samples from the source domain dataset belong to the healthy and faulty classes. Samples from the target domain dataset belong to the healthy class. A first subset of the features extracted from the input data identifies the healthy class, and a second subset of the features extracted from the input data is domain-invariant.

[0007] In yet another embodiment, a method for repairing a business asset includes: collecting a source domain dataset corresponding to a first operating condition of the business asset; collecting a target domain dataset corresponding to a second operating condition of the business asset; inputting the source domain dataset and the target domain dataset as input data into a neural network; having the neural network extract features from the input data; reducing the dimensionality of the features to make them reduced features; classifying the reduced features into a normal class and an abnormal class using a one-class classifier; identifying at least one business asset in the abnormal class; and repairing the at least one business asset. Samples from the source domain dataset belong to the healthy class and the faulty class. Samples from the target domain dataset belong to the healthy class. A first subset of the features extracted from the input data identifies the healthy class, and a second subset of the features extracted from the input data is domain-invariant.

[0008] It should be understood that both the general description above and the detailed description below describe various embodiments and are intended to provide an overview or framework for understanding the nature and features of the claimed subject matter. The accompanying drawings are included to provide a further understanding of the various embodiments and are incorporated herein and constitute part of this specification. The drawings illustrate the various embodiments described herein and, in conjunction with the description, illustrate the principles and operation of the claimed subject matter.

[0009] The embodiments shown in the drawings are essentially illustrative for illustrative purposes and are not intended to limit the subject matter defined by the claims. The following detailed description of the exemplary embodiments can be understood in conjunction with the following drawings, in which similar structures are indicated by the same reference numerals. [Brief explanation of the drawing]

[0010] [Figure 1A] This figure schematically illustrates a method for feature extraction using a neural network, such as a convolutional neural network (CNN), according to one or more embodiments illustrated and described herein. [Figure 1B] This figure schematically illustrates a method for anomaly detection using a Class 1 support vector machine, according to one or more embodiments illustrated and described herein. [Figure 2A] This figure schematically illustrates the overall architecture of a neural network according to one or more embodiments illustrated and described herein. [Figure 2B] This figure schematically illustrates the overall architecture of an anomaly detection framework according to one or more embodiments illustrated and described herein. [Figure 3] This figure schematically illustrates a system for anomaly detection in business assets, according to one or more embodiments illustrated and described herein. [Figure 4] This figure schematically illustrates an experimental setup of a gearbox-based manufacturing system according to one or more embodiments illustrated and described herein. [Figure 5] This figure shows a gearbox under consideration and typical wear failure conditions of the gearbox, according to one or more embodiments illustrated and described herein. [Figure 6] This figure schematically illustrates a comparative study of the performance of exemplary anomaly detection methods according to one or more embodiments illustrated and described herein. [Figure 7A] This figure schematically illustrates the anomaly detection performance of an exemplary anomaly detection method for a task, according to one or more embodiments illustrated and described herein, where the value of term v is = 0.005, and v affects the number of samples considered to be outliers. [Figure 7B]This figure schematically illustrates the anomaly detection performance of an exemplary anomaly detection method for a task, according to one or more embodiments illustrated and described herein, where the value of term v is = 0.5, and v affects the number of samples considered to be outliers. [Figure 7C] This figure schematically illustrates the anomaly detection performance of an exemplary anomaly detection method for a task, according to one or more embodiments illustrated and described herein, where the value of term v is = 0.01, and v affects the number of samples considered to be outliers. [Figure 8] This figure schematically illustrates a high-level feature visualization of raw data used in an exemplary anomaly detection method, according to one or more embodiments illustrated and described herein. [Figure 9] This figure schematically illustrates the anomaly detection results of a Class 1 support vector machine according to one or more embodiments illustrated and described herein. [Modes for carrying out the invention]

[0011] This document details various embodiments of methods and systems for detecting anomalies in business assets (e.g., any machinery or electrical equipment), with an example shown in the accompanying drawings. Note that the term "business assets" as used herein is synonymous with "industrial assets," but the embodiments are not limited to industrial equipment. Wherever possible, the same reference numerals are used throughout the drawings to refer to the same or similar parts.

[0012] In an industrial environment, assets or equipment such as gearboxes often play an important role in various applications, such as wind turbines, conveyors, and industrial robots. These applications and others rely on industrial assets that function properly, for example, to sustain energy output, move parts on a manufacturing floor, or maintain the precise motion and throughput required for material handling and production. Due to the complexity of many industrial assets, they are susceptible to a wide range of degradation modes, including wear and corrosion. Therefore, meticulous monitoring of assets is essential to avoid unexpected downtime, production losses, material damage, and potentially dangerous conditions resulting from failed components. Industrial assets are designed to maximize their average lifespan in demanding environments, but the continuous operation of these assets inevitably leads to degradation over time. Thus, early and accurate identification of the health state of industrial assets is important.

[0013] Conventionally, signal processing techniques have been applied to the vibration data of industrial assets such as gearboxes, and fault prediction and health management (prognostics and health management (PHM)) have been performed. However, neural networks can also be used for industrial asset monitoring. For example, a convolutional neural network (CNN) can be applied to the analysis of vibration signals obtained from business assets. In addition, deep learning strategies can be utilized that offer benefits superior to conventional machine learning techniques and neural network approaches. For example, deep learning architectures can perform automatic feature selection within their multiple hidden layers, shortening the time required to implement a fault identification method by avoiding the feature extraction and selection steps that are usually required. Furthermore, the features extracted by deep learning architectures are more confidential and have a higher information density than features extracted by classical means.

[0014] The data-driven approach of deep learning is also an important advantage, but it is not always easy or even possible to obtain data for all possible operating regimes and machine states. Therefore, it is important to develop and expand techniques that enable fault identification and health assessment to be performed only on a subset of the full range of available conditions. Accordingly, this specification describes the application of transfer learning that enables the use of knowledge of one machine state to enhance the assessment of other unseen or less-documented conditions.

[0015] Transfer learning solves the problem of limited data, but this disclosure addresses a further obstacle in the field of transfer learning. One important issue in transfer learning is domain mismatch. When transferring knowledge from a source domain to a target domain, a shift in the underlying domain distribution can affect the accuracy of the results. To avoid misclassification, this disclosure applies domain adaptation methods to overcome domain mismatch.

[0016] In a real industrial environment, it is often most practical to perform domain adaptation on healthy classes or baseline classes from various domains. Once domain mismatch is overcome and healthy machine states are aligned, outlier samples that do not fall within the expected range of healthy states can be identified using the deep learning methods disclosed herein. Such applications of anomaly detection enable tracking of the progression of industrial asset degradation over the component life cycle, and samples that appear as anomalies can trigger maintenance activities necessary to avoid further damage to critical and expensive components such as robotic arms.

[0017] As described in various embodiments of this disclosure, the methods and systems disclosed herein utilize deep learning techniques with domain adaptation across different machine conditions for the purpose of improving anomaly detection in industrial assets. The methods and systems disclosed herein eliminate the need to collect data for all operating conditions of industrial assets and can be extended to other industrial applications where it is not possible to collect a complete dataset, thereby saving considerable time required for training data collection and computational resources for processing such training data. In other words, the deep learning-based domain adaptation techniques of this disclosure enable the transfer of knowledge gained from one operating condition to another. In this regard, the methods and systems utilize a two-stage technique to extract domain-invariant and healthy class-identifier features from raw data, and then use these features in a one-class classifier to perform anomaly detection. Furthermore, the deep learning techniques disclosed herein can take into account the effects of machine-to-machine variation or environmental conditions.

[0018] Those skilled in the art will recognize that various embodiments of this disclosure may be carried out without one or more of the specific details described herein, or using other substitute and / or additional methods, materials, or components. In other instances, well-known structures, materials, or operations are not illustrated or described in detail herein to avoid obscuring the aspects of the various embodiments of the invention. Similarly, certain numbers, materials, and configurations are described herein for illustrative purposes and to provide a complete understanding of the various embodiments of this disclosure. Furthermore, it should be understood that the various embodiments shown in the figures are illustrative and not necessarily drawn to scale.

[0019] Unless otherwise specified, the methods described herein are not intended to be construed as requiring the steps to be performed in a specific order or requiring any device-specific orientation. Therefore, if a method claim does not actually enumerate the order in which the steps should be followed, or if a device or assembly claim does not actually enumerate the order or orientation of the individual components, or otherwise, if the claim or description does not specifically state that the steps should be limited to a particular order, or if no specific order or orientation of the components of the device or assembly is described, no order or orientation is intended to be inferred in any way. This applies to any possible implicit grounds for interpretation, including logical matters relating to the arrangement of steps, the flow of operations, the order of components, or the orientation of components, the obvious meaning derived from grammatical construction or punctuation, and the number or type of embodiments described herein.

[0020] Throughout this disclosure, “one embodiment” or “one embodiment” means that certain features, structures, materials, or properties described in relation to that embodiment are included in at least one embodiment of the present invention, but not that they are present in all embodiments. Thus, the occurrence of the phrase “in an embodiment” or “in another embodiment” in various parts of this specification does not necessarily refer to the same embodiment of the present invention. Furthermore, as used herein, the singular forms “a,” “an,” and “the” include multiple referents unless the context clearly indicates otherwise. Thus, for example, a reference to “a” component includes embodiments having two or more such components unless the context clearly indicates otherwise. Also, “a” component may represent one or more components and may therefore be used herein to mean “at least one.”

[0021]

number

[0022] Referring to Figures 1A and 1B, flowcharts illustrating a two-stage anomaly detection methodology or architecture S100 according to embodiments of the present disclosure are shown. The first stage is generally called feature extraction using a neural network such as a CNN and is shown in Figure 1A. The illustrated method begins in S102 with data preprocessing before performing the first stage. Data preprocessing in S102 generally involves preparing input data samples and labels for both the source domain and the target domain. Once the input data samples for the source and target domains are prepared and labeled, in S104 the data samples are input to the CNN. In S104 the CNN is trained to extract healthy class discrimination features that separate healthy samples from all other classes (e.g., "intermediate" class and "failed" class) and to extract domain-invariant features that are aligned in the feature space for all domains. In particular, the CNN is trained to extract features that satisfy the requirements of healthy class discrimination and domain invariance by using a two-part loss function in S106 and S108. To update the CNN so that healthy classes are isolated from all other classes, in S106, the first part of the two-part loss function, cross-entropy error optimization, is applied. That is, the cross-entropy error optimization in S106 considers all classes for the source domain and only healthy classes for the target domain. To update the CNN so that the domain mismatch between healthy class samples in the source and target domains is minimized, in S108, the second part of the two-part loss function, domain adaptive loss optimization, is applied using maximum mean discrepancy (MMD). That is, the domain adaptive loss optimization using MMD in S108 considers only healthy class samples from the source and target domains.After applying a two-part loss function in S106 and S108, the CNN extracts high-level feature information for both the source and target domains in S110, completing the first stage.

[0023] The second stage is generally called anomaly detection using a one-class support vector machine (SVM), as shown in Figure 1B. The second stage generally performs the final task of anomaly detection using high-level features extracted from the CNN trained in the first stage. The second stage begins in S112, where the dimensionality of the problem is reduced by applying t-distributed stochastic neighbor embedding (t-SNE). Then, in S114, each sample is assigned a health label by applying a one-class classifier (e.g., a one-class SVM). Specifically, in S114, the one-class SVM is trained on healthy data from the source domain and healthy data from the target domain, and the one-class SVM is validated on healthy data from the target domain for fine-tuning. Then, in S116, the one-class SVM tests the remaining data from the target domain (including healthy and faulty data). Generally, in S114 and S116, the one-class SVM method is applied to detect samples that are outliers of the healthy class (i.e., abnormal or non-normal samples corresponding to possible failures in the machine). The second-stage anomaly detection method can detect non-normal samples within the target domain without seeing any non-normal samples during training. Here, steps S102 to S116 of the exemplary method S100 are discussed in more detail.

[0024] Method step S102 generally relates to data preprocessing, which includes preparing input data samples and labels for both the source domain and the target domain. In particular, fixed cycle feature test (FCFT) data is acquired by repeatedly executing a pre-defined operating pattern at various points in the lifecycle of an industrial asset or machine. The pre-defined pattern covers a wide range of operating conditions (e.g., minimum and maximum rotational speeds). Since the machine will execute the same pattern at all points in its lifecycle, any changes in the collected data are due to machine degradation. According to some embodiments of this disclosure, a method can be defined for generating appropriate frequency domain samples from feedback torque current data obtained by FCFT. FCFT data is collected for various health conditions. For each health condition, an FCFT command pattern (characterized by different rotational direction and position commands) is repeated multiple times for each of the primary operating states (rotational speeds), resulting in multiple patterns for each operating state (e.g., 48 patterns are collected after 100 repetitions (12 patterns each for 50, 500, 1000, and 3000 RPM)). The collected command velocity signals are used to reduce each pattern so that it contains only the steady-state portion of the feedback torque current signal. These patterns are further reduced by removing a predetermined percentage from the initial point to ensure that the steady-state condition is achieved. The steady-state patterns for one health state, all cleaned, are concatenated over all iterations to generate a single continuous feedback torque current signal. This torque signal is used to generate frequency-domain samples. A predetermined number of samples of a specified length are obtained by dividing the torque signal in the time domain and then saving the frequency-domain information of each segment as one of the final samples. These frequency-domain samples are then ready for feature extraction using a CNN.

[0025]

number

[0026]

number

[0027] In the formula, * represents a convolution operation. Furthermore, a nonlinear activation function is applied to the feature map to capture the nonlinear correlation between features and data. To further reduce the dimensionality of the feature map, a pooling layer can follow the convolution layer. Maximum pooling techniques are used to preserve meaningful spatial information in the feature map while simultaneously reducing dimensionality.

[0028]

number

[0029]

number

[0030]

number

[0031] For a one-class SVM model in Stage 2, an unsupervised machine learning model is shown that generates a decision function for estimating outliers from a dataset. In this method, the model is trained using one class of data (usually the normal class or healthy class). The knowledge gained from the trained model can be used to determine whether the test data is similar to or different from the training dataset. In most real-world applications, collecting data from the anomalous class is not a straightforward task. Therefore, when using a one-class SVM for anomaly detection, the model can be trained using only the healthy class or normal class dataset without requiring the use of other classes.

[0032] A one-class SVM model generates a nonlinear decision boundary by mapping the original feature space to a higher-dimensional space. The mapping transformation is appropriately performed using a kernel function. A one-class SVM model can be trained using equation (3) below, assuming that X and Y represent the original feature space and the higher-dimensional feature space (φ: X → Y).

[0033]

number

[0034] However, the following conditions apply. ∀i=1,2,...,N

[0035] v∈(0,1],ξ i ≥0,

[0036] ω·φ(X i ) ≧b-ξ i·

[0037] In the formula, ξ i represents the slack variable. Term v plays a role in excluding a portion of the training dataset as outliers. Equation (4) below is the decision function used to distinguish anomalies within the dataset.

[0038]

number

[0039] In this disclosure, a radial basis function (RBF) kernel is used to transform the original feature space into a higher-dimensional space.

[0040] As shown in Figure 2A, the Stage 1 CNN architecture in Figure 1A includes four main modules in two stages. The four main modules include a deep learning-based feature extraction module and a domain adaptation module in Stage 1, and a dimensionality reduction module and a one-class classifier module in Stage 2. In Stage 1, which includes network training and optimization, the deep learning-based feature extraction module extracts meaningful information from the raw input data in the form of low-dimensional high-level (HL) features. The frequency spectrum of the collected data is used as input data. The HL features are then fed to the domain adaptation module. The main role of the domain adaptation module is to generate domain-invariant features. In other words, the domain adaptation module reduces the discrepancies between feature representations corresponding to the source domain and the target domain. In Stage 1, the network is trained using source data that includes all classes (H healthy, B intermediate, and F fault) and a target domain that has only the healthy class. Therefore, the main objective of Stage 1 is to extract features that are domain-invariant and identify the healthy class. In the first stage, all health status classes of the system are available in the source domain dataset, but only the healthy class is available in the target domain dataset. The absence of labels in the target domain can affect the transferability of the CNN, and features may not perfectly identify the healthy class. To address this problem, a second stage (Figure 2B) is added to the CNN to improve the anomaly detection capability of the features. In the second stage of the architecture shown in Figure 2B, features are used as input to a one-class classifier to satisfy the intent of anomaly detection, but a dimensionality reduction module is provided to first reduce the dimensionality of the features in order to reduce the complexity of the input data while maintaining useful information. In this regard, t-distribution stochastic nearest neighbor embedding (t-SNE) is used.

[0041] To achieve transfer learning between domains, method step S106 relates to cross-entropy error optimization to update the CNN so that healthy classes are separated from all other classes. Method step S108 relates to domain adaptive loss optimization using maximum mean mismatch (MMD) to update the CNN so that the domain mismatch between healthy class samples from the source domain and healthy class samples from the target domain is minimized. Thus, the objective of the first stage is to create features that identify healthy classes and are domain invariant. To achieve identification of healthy classes, the cross-entropy error, a typical classification loss function, is used as shown in equation (5) below.

[0042]

number

[0043] In the formula, n, N c , y i , y' ij These represent the number of samples, the number of classes in each domain, the label corresponding to the i-th sample, and the predicted labels corresponding to the i-th sample and the j-th class, respectively.

[0044] A domain adaptation methodology is developed to transfer knowledge gained from the source domain to the target domain. Data corresponding to the source and target domains are collected under different working arrangements, and these arrangements can affect the data distribution (and corresponding feature representations) across the domains. To address this problem, the MMD metric is used to measure the discrepancy between the data distributions of the source and target domains. The MMD term is added to the overall optimization goal as a loss function. The MMD loss function is defined by equation (6) below.

[0045]

number

[0046] In the expression, argument Fsource and argument F target are the feature representations of the data corresponding to the source domain and the target domain, respectively.

[0047] In summary, by combining the loss functions of the above equations (5) and (6), a general optimization goal for an exemplary CNN architecture is formed. Therefore, by integrating equations (5) and (6), the following general loss function can be formulated as shown in the following equation (7).

[0048]

Equation

[0049] where α > 0 and β > 0 are the penalty coefficients of L C and L DA respectively. During the training phase, the network parameters θ f of the feature extraction module and the network parameters θ c of the one-class classifier module are optimized and updated according to the loss function. The update process can be formulated as shown in the following equation (8).

[0050]

Equation

[0051] where δ represents the learning rate.

[0052] The high-level feature representations collected from the first stage are assumed to be used for the purpose of anomaly detection using the one-class classifier module. Before feeding the features to the module, the high-dimensional features must be mapped to a 2D feature space. The t-SNE algorithm is used for mapping the features from a high-dimensional space to a low-dimensional space.

[0053] During the second stage of the training process, the parameters in equation (3) are optimized. In addition, two hyperparameters also need to be adjusted in the second stage. The first hyperparameter is v, and the second hyperparameter is the RBF kernel hyperparameter γ, which follows equation (9) below.

[0054]

number

[0055] In the formula, Φ i , Φ j represents the i-th and j-th feature samples, and ||·|| is the L2 norm operator.

[0056] Figure 3 shows an example of a system including a system for detecting anomalies in business assets as disclosed herein. The system shown in Figure 1 may include a computing system 1, a user client 20, a control system 30, and a third-party system 40. Computing system 1 may provide a system for performing data analysis as disclosed herein. Computing system 1 may be implemented using, for example, one or more general-purpose computers. As shown in Figure 1, computing system 1 may include an application 10 and a data storage device 12. Application 10 may be implemented by a software application that includes instructions for causing the computer to perform exemplary processes of the computing system. As shown in Figure 1, application 10 may include a CNN 100 having a feature extraction module 102 and a domain adaptation module 104, a dimensionality reduction module 106, a one-class classifier module 108, and an interface 110.

[0057] CNN100 may be the same as or similar to the CNN illustrated in Figure 2A, and may be an artificial neural network having an input layer (e.g., the input data in Figure 2A), an output layer (e.g., the fully connected layer in Figure 2A), and a number of hidden layers between the input and output layers (e.g., convolutional layer 1, convolutional layer 2, pooling layer 1, and flattening layer in Figure 2A). CNN100 can be trained to process any type of data, such as sensor data from business or industrial assets. In one example, CNN100 may be trained to process data acquired by each sensor using a training dataset containing data that could be considered as input data to CNN100. The training dataset may be stored in a data storage device 12 accessible by application 10.

[0058] The feature extraction module 102 and the domain adaptation module 104 may be connected to at least one of the multiple hidden layers of the CNN 100 and configured to extract features from the input data, where a first subset of these features identifies healthy classes and a second subset of these features is domain-invariant. Details of the process performed by the feature extraction module are described above with respect to method steps S104-S110. The dimensionality reduction module 106 may be configured to perform method step S112 described above. The one-class classifier module 108 may be configured to perform method steps S114-S116 described above.

[0059] Interface 110 may also be an interface for communication between application 10 and various devices that may be located outside the computing system 1. For example, interface 110 may be configured to communicate information generated by application 10 to those devices. Furthermore, interface 110 may be configured to receive information directed from those devices to application 10.

[0060] The data storage device 12 may be configured to store data used by application 100. Figure 3 shows the data storage device 12 as part of computing system 1, but in some examples, the data storage device 12 may be located outside the computing system, provided that the data stored in the data storage device 12 is accessible by application 10.

[0061] The user client 20 may be a client device connected to the computing system 1. The user client 20 may include a user application 22 that can use the predictions and results of anomaly detection performed on the computing system 1. A concrete example of the user client 20 may be a workstation remotely connected to the computing server using SSH (Secure Shell) or HTTP (Hypertext Transfer Protocol) requests. In this case, the CNN 100 can be applied to user-provided inputs on the computing server and the resulting predictions and anomaly detections can be returned to the user client 20. The user client 20 may be part of the same physical device as the computing system 1 running application 10 on a workstation configured to perform CNN predictions, for example.

[0062] The control and / or analysis system 30 may use the predictions and results of anomaly detection performed by the computing system 1 to control the device and / or perform further data analysis. The control and / or analysis system 30 may constitute or be part of a system for anomaly detection and / or predictive maintenance. An example of the control and / or analysis system 30 may be a control and / or analysis system (for anomaly detection or predictive maintenance, etc.) for business assets such as mechanical components or machines, industrial processes or plants, vehicles (such as autonomous vehicles), computer networks, or financial trading units.

[0063] The control and / or analysis system 30 may include a programmatic client 32 operating on the control and / or analysis system 30 that receives inputs, performs data analysis, and makes decisions regarding further application-specific activities relating to the maintenance of business assets (e.g., equipment and systems) and / or the control of business assets (e.g., equipment and systems).

[0064] Furthermore, the subject matter of this disclosure can also be realized as a computing system 1 including a processor and memory coupled to the processor. The memory may encode one or more programs to cause the processor to perform one or more of the methods described herein. In some examples, system 1 may be a general-purpose computer system. In other examples, system 1 may be a dedicated computer system including an embedded system. Examples Gearbox data from an industrial system was analyzed. As shown in Figure 4, the gearbox is part of a planetary gearbox system for transmitting power from a motor to a load. The dataset includes various combinations of gear health and gain settings. The gear failure mode is wear, and the states are healthy, intermediate, and failed. Representative wear failure states for the gearbox and gears under consideration are shown in Figure 5. Intermediate corresponds to an intermediate stage of gear wear between the healthy and failed states. Possible gain settings are low, medium, and high. Gear wear failures are extremely opaque in low-gain setting data, making it most difficult to distinguish between healthy and worn states. Therefore, the method described in this embodiment was initially developed using extreme examples of low-gain settings.

[0065] Each dataset contains Fixed-Cycle Feature Test (FCFT) data. For the purposes of this embodiment, the datasets were focused on the steady-state portion of the FCFT data rather than the transient portion. In total, 100 iterations (loops) of 63 command patterns were collected for each wear / gain state. The first 48 patterns are from four main working regimes: 50 rpm, 500 rpm, 1000 rpm, and 3000 rpm (12 patterns per regime). These patterns were generated by applying various rotational direction and position commands at the specified rpm. The remaining 15 patterns similarly feature specific rpm, rotational direction, and position commands, but these 15 patterns were not used in this embodiment. Also, based on input from domain experts, this embodiment considered only forward rotations (positive patterns), such as the four patterns from each working regime. That is, backward rotations (negative patterns) were removed from the dataset during preprocessing. Finally, 100 loops of 32 patterns (8 patterns from each rpm) were analyzed.

[0066] All loops include command and feedback position, command and feedback speed, and feedback torque current. Samples were collected every 444 μs. Auxiliary information (encoder temperature and load speed) was also collected for each loop at a speed of 1 Hz. The date and time stamp and cumulative experimental time for each collected data point were also available for each sampling frequency. A summary of the collected information is shown in Table I below.

[0067] [Table 1]

[0068] For preprocessing, different loops and patterns were separated, resulting in a matrix of 100 loops × 32 patterns of samples. Initially, the samples contained both transient and steady-state information. The primary objective of preprocessing was to reduce the samples to contain only the steady-state component. To achieve this, all points of the feedback torque current signal corresponding to the point where the command speed of each sample was within 1 rpm of the working system speed were retained. The shortened torque samples were then further reduced to eliminate overshoot. After analyzing multiple loops and patterns, it was determined that the first 40% of each sample should be removed to retain only steady-state information. Finally, for each working system, eight cleaned patterns from each loop were concatenated, and then these 100 loops were concatenated. The result is one cleaned torque signal for each wear and gain condition of each working system (a total of 36 signals).

[0069] The final step of preprocessing is to organize the torque signals into samples for the methods of the embodiments described herein. Each torque signal was downsampled by taking every 10th data point. The torque signals were then divided into 500 windows of 2000 time-domain points each (the 2000 time-domain points provide 1000 unique frequency-domain points). The step size and overlap rate of the windowing process were dynamically determined based on the length of the entire torque signal. Frequency-domain information for each window was obtained and stored as a sample for a given combination of wear, gain, and working state. In total, 1000-point samples of frequency data for 500 of the cleaned steady-state torque feedback current signals were generated for each of the 36 conditions (3 wear cases × 3 gain settings × 4 working states).

[0070] To evaluate the performance of the methodology (e.g., method S100 described above), various tasks were designed based on various scenarios. Accordingly, six different experiments were conducted under different operating conditions, as shown in Table II below. There were two main criteria for designing the tasks. The first criterion was to evaluate domain mismatch by increasing the distance of the target operating condition from a fixed source condition (e.g., T1 and T2). The second criterion was to evaluate the selection of the optimal source domain condition by swapping the source domain and target domain for a particular task (e.g., T4 and T5). In all experiments, the sample size of the source domain dataset and target domain dataset was 1500 (500 samples from each of the three health states).

[0071] [Table 2]

[0072] Referring to the CNN architecture illustrated in Figure 2A, the feature extraction module consists of two consecutive convolutional layers, each with 30 and 20 filters (filter size = 5), respectively. A rectified linear unit (ReLU) is used as the activation function in both convolutional layers. Following the convolutional layers is a maximum pooling layer with a pooling size of 2. Detailed information on the network implementation and parameters is shown in Table III below.

[0073] [Table 3]

[0074] To better evaluate the efficiency of the methodology disclosed herein, a benchmark study was conducted. Therefore, various types of anomaly detection methods were applied within the scope of the designed transfer tasks and datasets. Subsequently, for each task, the performance and accuracy of the methods disclosed herein were compared with other methods. The benchmark methods are described below.

[0075] Conventional One-Class SVM: The main difference between the method disclosed herein and a conventional one-class SVM is the classifier input. In the scope of a conventional one-class SVM, the raw frequency spectrum is used as the classifier input. The difference between the performance of the method disclosed herein and a conventional one-class SVM underscores the efficiency of the first stage of the method disclosed herein.

[0076] Isolated Forest: This method is a non-parametric, unsupervised algorithm that can be used to detect outliers in a dataset. When the input data space is high-dimensional, finding healthy state patterns is not straightforward. Compared to other similar methods, isolated forests offer high-quality anomaly detection capabilities.

[0077] Local Outlier Factor (LOF): This method provides a score that explains how likely a sample is to be an outlier. This technique analyzes the local density deviation of the surveyed sample relative to adjacent samples. Samples with relatively low density are considered outliers.

[0078] A comprehensive study was conducted to verify the efficiency and effectiveness of the anomaly detection methodology described herein compared to other methods. Figure 6 shows a comparative evaluation of the methods under consideration under various experiments and tasks. As the results show, the two-stage methodology of this disclosure is clearly superior to other methods. Note that the effectiveness of the first stage of the method described herein can be verified by comparing it with the results of conventional one-class SVMs. The only difference between these two methods is the first stage of the method disclosed herein, and the comparative results support a significant increase in network performance when domain-invariant and healthy class identification features are supplied to the one-class classifier.

[0079] Based on the first criterion for designing the transfer task, experiments were conducted to evaluate the impact of domain mismatch. In this regard, comparing the results between Task T1 and Task T2, and between Task T3 and Task T4, the network performance degrades as the distance between the operating conditions of the source domain and the target domain increases. On the other hand, the selection of source, domain, and dataset also affects the network performance. As shown in Figure 6, comparing the results of Task T4 and Task T5, as well as the results of Task T1 and Task T6, it can be concluded that although the domain mismatch between the source and target domains is constant, changing the source and target domains can affect the network performance. In both scenarios, the network of this disclosure is substantially robust to these types of phenomena.

[0080] The value of term v affects the number of samples considered outliers. Tuning this hyperparameter is a crucial step in designing a one-class classifier. The effect of variations in v value on the proposed network's anomaly detection capability for task T2 is demonstrated in Figures 7A-7C. When v is low, a relatively small proportion of samples are considered outliers. In other words, the model is overfitting (Figure 7A). Conversely, a relatively high value of v leads to the model underfitting problem (Figure 7B). Based on hyperparameter tuning studies, the optimal value of v is 0.01, in which case the network can properly distinguish samples corresponding to the healthy state from the other two healthy states of the system (Figure 7C).

[0081] The features extracted from the first stage of the method of this disclosure are intended to be domain-invariant and to identify healthy classes. To better understand these two properties of high-level features, a visualization of these features is shown in Figure 8. As can be seen from Figure 8, the features corresponding to healthy classes in the source and target domains are clustered into the same group. Therefore, it can be concluded that the network of this disclosure in the first stage can extract features that are domain-invariant and identify healthy classes. Because the features are domain-invariant, they can be used for transfer learning applications, allowing knowledge gained from the source domain to be transferred to the target domain. On the other hand, because the features identify healthy classes, they can be suitable candidates for input to any one-class classifier for anomaly detection applications.

[0082] The extracted high-level features are used as input to a one-class SVM for anomaly detection. The objective is to develop a pattern that separates healthy class samples within a target domain from the other two classes. Figure 9 shows a visualization of the one-class classifier using the proposed method for task T5. As shown in Figure 9, the one-class classifier can properly cluster healthy class samples from both domains within the same group, with samples corresponding to the other two classes being considered outliers. Note that as the failure becomes more severe and the system health status moves into the failure class, the features will be located further away from the healthy class boundary.

[0083] In view of the foregoing, it should be understood that at least some embodiments of this disclosure relate to a two-stage deep learning-based transfer learning methodology to be used for anomaly detection. One characteristic of the methodology disclosed herein is that, within the first stage of the architecture, it provides a domain-invariant feature that identifies healthy classes, as described above, and thus distinguishes it from other methods. Using the methodology disclosed herein, knowledge gained from a source domain dataset under one operating condition can be used to distinguish outliers in a target domain dataset under another operating condition. In this scenario, only a small amount of data corresponding to the health of the system is available to the target domain.

[0084] It should be noted that the terms “substantially” and “about” may be used herein to express the degree of inherent uncertainty that may arise from any quantitative comparison, value, measurement, or other expression. These terms are also used herein to express the degree to which a quantitative expression may deviate from the reference given without altering the fundamental function of the subject matter in question.

[0085] While specific embodiments have been illustrated and described herein, it should be understood that various other changes and modifications may be made without departing from the spirit and scope of the claimed subject matter. Furthermore, while various aspects of the claimed subject matter have been described herein, it is not necessary to use such aspects in combination. Accordingly, the attached claims are intended to encompass all such changes and modifications that fall within the scope of the claimed subject matter.

Claims

1. An anomaly detection method using a system for detecting anomalies in business assets, The system's neural network is input with a source domain dataset corresponding to the first operating condition of the business asset and a target domain dataset corresponding to the second operating condition of the business asset as input data. The feature extraction module and domain adaptation module of the neural network extract features from the input data, The dimensionality reduction module of the system reduces the dimension of the feature, thereby reducing the feature to a reduced feature. Using the one-class classifier of the aforementioned system, the reduced features are classified into normal and abnormal classes. Includes, The samples from the aforementioned source domain dataset belong to the healthy class and the faulty class. The samples from the aforementioned target domain dataset belong to the healthy class, The features extracted from the input data are domain-invariant and identify the healthy class. The extraction of features from the aforementioned input data is performed using cross-entropy error optimization and domain adaptive loss optimization. Using the cross-entropy error optimization involves extracting a first subset of the features by applying the cross-entropy error optimization to the source domain dataset, the first subset of which identifies the healthy class, Using the domain adaptive loss optimization method involves extracting a second subset of the features by applying domain adaptive loss optimization to the source domain dataset and the target domain dataset, wherein the second subset is domain invariant, and the domain adaptive loss optimization utilizes the maximum mean mismatch loss such that the domain mismatch between the healthy class samples of the source domain dataset and the healthy class samples of the target domain dataset is minimized.

2. The method according to claim 1, further comprising preprocessing the source domain dataset and the target domain dataset by data truncation.

3. The method according to claim 1, wherein the dimension of the feature is reduced by a t-distribution type stochastic neighbor embedding method.

4. The method according to claim 1, wherein the one-class classifier is a support vector machine classifier.

5. An anomaly detection method using a system for detecting anomalies in business assets, The system's neural network is input with a source domain dataset corresponding to the first operating condition of the business asset and a target domain dataset corresponding to the second operating condition of the business asset as input data. The feature extraction module and domain adaptation module of the neural network extract features from the input data, The dimensionality reduction module of the system reduces the dimension of the feature, thereby reducing the feature to a reduced feature. Using the one-class classifier of the aforementioned system, the reduced features are classified into normal and abnormal classes. Includes, The samples from the aforementioned source domain dataset belong to the healthy class and the faulty class. The samples from the aforementioned target domain dataset belong to the healthy class, The features extracted from the input data are domain-invariant and identify the healthy class. Extracting features from the aforementioned input data involves using domain adaptive loss optimization. The domain adaptive loss optimization is applied in such a way that the domain mismatch between the healthy class samples of the source domain dataset and the healthy class samples of the target domain dataset is minimized.

6. A system for detecting anomalies in business assets, One or more processors, A non-temporary computer-readable medium that stores instructions, Equipped with, The one or more processors execute the instruction, The source domain dataset corresponding to the first operating conditions of the aforementioned business asset is received. The target domain dataset corresponding to the second operating condition of the aforementioned business asset is received. The source domain dataset and the target domain dataset are input to the neural network as input data. The neural network extracts features from the input data, By reducing the dimension of the aforementioned feature, the aforementioned feature is reduced to a reduced feature. A single-class classifier is used to classify the reduced features into normal and abnormal classes. The samples from the aforementioned source domain dataset belong to the healthy class and the faulty class. The samples from the aforementioned target domain dataset belong to the healthy class, The features extracted from the input data are domain-invariant and identify the healthy class. The extraction of features from the aforementioned input data is performed using cross-entropy error optimization and domain adaptive loss optimization. Using the cross-entropy error optimization involves extracting a first subset of the features by applying the cross-entropy error optimization to the source domain dataset, the first subset of which identifies the healthy class, Using the domain adaptive loss optimization involves extracting a second subset of the features by applying the domain adaptive loss optimization to the source domain dataset and the target domain dataset, wherein the second subset is domain-invariant, and the domain adaptive loss optimization utilizes the maximum mean mismatch loss such that the domain mismatch between the healthy class samples of the source domain dataset and the healthy class samples of the target domain dataset is minimized. system.

7. The system according to claim 6, further comprising preprocessing the source domain dataset and the target domain dataset by data truncation.

8. The system according to claim 6, wherein the dimension of the feature is reduced by a t-distribution type stochastic neighbor embedding method.

9. The system according to claim 6, wherein the one-class classifier is a support vector machine classifier.

10. A system for detecting anomalies in business assets, One or more processors, A non-temporary computer-readable medium that stores instructions, Equipped with, The one or more processors execute the instruction, The source domain dataset corresponding to the first operating conditions of the aforementioned business asset is received. The target domain dataset corresponding to the second operating condition of the aforementioned business asset is received. The source domain dataset and the target domain dataset are input to the neural network as input data. The neural network extracts features from the input data, By reducing the dimension of the aforementioned feature, the aforementioned feature is reduced to a reduced feature. A single-class classifier is used to classify the reduced features into normal and abnormal classes. The samples from the aforementioned source domain dataset belong to the healthy class and the faulty class. The samples from the aforementioned target domain dataset belong to the healthy class, The features extracted from the input data are domain-invariant and identify the healthy class. Extracting features from the aforementioned input data involves using domain adaptive loss optimization. The domain adaptive loss optimization is applied in such a way that the domain mismatch between the healthy class samples of the source domain dataset and the healthy class samples of the target domain dataset is minimized. system.

11. A method for using a system for detecting anomalies in business assets in order to repair business assets, The system's neural network is input with a source domain dataset corresponding to the first operating condition of the business asset and a target domain dataset corresponding to the second operating condition of the business asset as input data. The feature extraction module and domain adaptation module of the neural network extract features from the input data, The dimensionality reduction module of the system reduces the dimension of the feature, thereby reducing the feature to a reduced feature. By using the one-class classifier of the system, the reduced features are classified into normal and abnormal classes, thereby identifying at least one business asset in the abnormal class. Includes, The samples from the aforementioned source domain dataset belong to the healthy class and the faulty class. The samples from the aforementioned target domain dataset belong to the healthy class, The features extracted from the input data are domain-invariant and identify the healthy class. The extraction of features from the aforementioned input data is performed using cross-entropy error optimization and domain adaptive loss optimization. Using the cross-entropy error optimization involves extracting a first subset of the features by applying the cross-entropy error optimization to the source domain dataset, the first subset of which identifies the healthy class, Using the domain adaptive loss optimization involves extracting a second subset of the features by applying the domain adaptive loss optimization to the source domain dataset and the target domain dataset, wherein the second subset is domain-invariant, and the domain adaptive loss optimization utilizes the maximum mean mismatch loss such that the domain mismatch between the healthy class samples of the source domain dataset and the healthy class samples of the target domain dataset is minimized. method.

12. The method according to claim 11, wherein the dimension of the feature is reduced by a t-distribution type stochastic neighbor embedding method.

13. The method according to claim 11, wherein the one-class classifier is a support vector machine classifier.

14. A method for using a system for detecting anomalies in business assets in order to repair business assets, The system's neural network is input with a source domain dataset corresponding to the first operating condition of the business asset and a target domain dataset corresponding to the second operating condition of the business asset as input data. The feature extraction module and domain adaptation module of the neural network extract features from the input data, The dimensionality reduction module of the system reduces the dimension of the feature, thereby reducing the feature to a reduced feature. By using the one-class classifier of the system, the reduced features are classified into normal and abnormal classes, thereby identifying at least one business asset in the abnormal class. Includes, The samples from the aforementioned source domain dataset belong to the healthy class and the faulty class. The samples from the aforementioned target domain dataset belong to the healthy class, The features extracted from the input data are domain-invariant and identify the healthy class. Extracting features from the aforementioned input data involves using domain adaptive loss optimization. The domain adaptive loss optimization is applied in such a way that the domain mismatch between the healthy class samples of the source domain dataset and the healthy class samples of the target domain dataset is minimized. method.

Images