Method and system for anomaly detection in business assets and method for repairing 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 TO RELATED APPLICATIONS This application claims priority to U.S. Provisional Application No. 63 / 431,384, filed December 9, 2022, the entire contents of which are incorporated herein by reference.

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

[0003] Business or industrial assets or machinery operate within different "domains" or "operating regimes," including, for example, different rotational speeds, environmental temperatures, and loads. For the same machine condition (e.g., healthy, fault type I, fault type II, etc.), the data signature may differ depending on the operating regime. For example, the vibration of a fault 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 remove the domain mismatch. Transfer learning uses knowledge gained from solving a problem in one domain (the "source") to enhance and improve problem-solving skills in another domain (the "target"). Summary of the Invention

[0004] Additional features and advantages of the present disclosure will be set forth in the detailed description that follows, and in part will become apparent to those skilled in the art from that description, or will be learned by practicing the embodiments described herein, including the detailed description that follows the claims and the accompanying drawings.

[0005] In one embodiment, 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.

[0006] In another embodiment, a system for anomaly detection in a business asset includes one or more processors and a non-transitory computer-readable medium having instructions stored thereon. The one or more processors execute the instructions to receive a source domain dataset corresponding to a first operating condition of the business asset and 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 to a neural network, extract features from the input data using the neural network, reduce the dimensionality of the features to form reduced features, and classify 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 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 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, 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 a healthy class and a 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 are domain-invariant.

[0008] It is to be understood that both the foregoing general description and the following detailed description 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 into and constitute a part of this specification. The drawings illustrate various embodiments described herein and, together with the description, explain the principles and operation of the claimed subject matter.

[0009] The embodiments set forth in the drawings are illustrative and exemplary in nature and are not intended to limit the subject matter defined by the claims. The following detailed description of exemplary embodiments can be understood when read in conjunction with the following drawings, in which like structure is designated with like reference numerals and in which: [Brief explanation of the drawings]

[0010] [Figure 1A] FIG. 1 illustrates a schematic diagram of a method for feature extraction using a neural network, such as a convolutional neural network (CNN), according to one or more embodiments shown and described herein. [Figure 1B] FIG. 1 illustrates a schematic diagram of a method for anomaly detection using a one-class support vector machine, according to one or more embodiments shown and described herein. [Figure 2A] FIG. 1 is a diagram that schematically illustrates the overall architecture of a neural network, according to one or more embodiments shown and described herein. [Figure 2B] FIG. 1 illustrates a schematic diagram of the overall architecture of an anomaly detection framework, according to one or more embodiments shown and described herein. [Figure 3] 1 is a diagram that schematically illustrates a system for anomaly detection in business assets, according to one or more embodiments shown and described herein. [Figure 4] FIG. 1 is a schematic diagram of an experimental setup for a manufacturing system utilizing a gearbox, according to one or more embodiments shown and described herein. [Figure 5] FIG. 1 illustrates a gearbox under consideration and representative wear-out failure conditions for the gearbox, according to one or more embodiments shown and described herein. [Figure 6] FIG. 10 is a diagram that schematically illustrates a comparative performance study of exemplary anomaly detection methods, according to one or more embodiments shown and described herein. [Figure 7A] FIG. 10 illustrates a schematic diagram of anomaly detection performance of an exemplary anomaly detection method for a task where the value of the term v = 0.005, where v influences the number of samples that are considered outliers, in accordance with one or more embodiments shown and described herein. [Figure 7B]FIG. 10 schematically illustrates the anomaly detection performance of an exemplary anomaly detection method for a task where the value of the term v = 0.5, where v influences the number of samples that are considered outliers, in accordance with one or more embodiments shown and described herein. [Figure 7C] FIG. 10 illustrates a schematic diagram of anomaly detection performance of an exemplary anomaly detection method for a task where the value of the term v = 0.01, where v influences the number of samples that are considered outliers, in accordance with one or more embodiments shown and described herein. [Figure 8] FIG. 2 is a diagram that schematically illustrates a high-level feature visualization of raw data used in an exemplary anomaly detection method, according to one or more embodiments shown and described herein. [Figure 9] FIG. 10 is a schematic illustration of an anomaly detection result of a one-class support vector machine, according to one or more embodiments shown and described herein. DETAILED DESCRIPTION OF THE INVENTION

[0011] Various embodiments of methods and systems for anomaly detection in business assets (e.g., any mechanical or electrical equipment) are detailed below, examples of which are illustrated in the accompanying drawings. Note that the term "business asset" herein is synonymous with the term "industrial asset," but embodiments are not limited to industrial equipment. Wherever possible, the same reference numbers will be used throughout the drawings to refer to the same or like parts.

[0012] In industrial environments, assets or equipment such as gearboxes often play a critical role in various applications, such as wind turbines, conveyors, and industrial robots. These and other applications rely on industrial assets to 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 their complexity, many industrial assets are susceptible to a variety of degradation modes, including wear and corrosion. Therefore, close monitoring of assets is essential to avoid unexpected downtime, production losses, material damage, and potentially hazardous conditions resulting from failed components. While industrial assets are designed to maximize life expectancy in demanding environments, the continuous operation of these assets inevitably leads to degradation over time. Therefore, early and accurate identification of the health status of industrial assets is critical.

[0013] Traditionally, signal processing techniques are applied to vibration data from industrial assets, such as gearboxes, for fault prognostics and health management (PHM). However, neural networks can also be used for industrial asset monitoring. For example, convolutional neural networks (CNNs) can be applied to analyze vibration signals from industrial assets. Deep learning strategies can also be used, offering advantages over traditional machine learning techniques and neural network approaches. For example, deep learning architectures can perform automatic feature selection within their multiple hidden layers, reducing the time it takes to implement fault identification methods by avoiding the typically required feature extraction and selection steps. Furthermore, features extracted by deep learning architectures are more sensitive and have higher information density than features extracted by classical means.

[0014] While the data-driven aspect of deep learning is also an important advantage, obtaining data for all possible operating regimes and machine conditions is not always easy or even possible. Therefore, it is important to develop and extend techniques that allow fault identification and health assessment to be performed for only a subset of the full range of conditions available. Therefore, this specification describes an application of transfer learning that allows knowledge of one machine condition to be used to enhance the assessment of other, less visible or less documented conditions.

[0015] Although transfer learning solves the problem of limited data, the present disclosure addresses a further obstacle in the field of transfer learning. One key challenge 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, the present disclosure applies domain adaptation methods to overcome domain mismatch.

[0016] In real industrial environments, it is often most feasible to perform domain adaptation on healthy or baseline classes from various domains. Once domain inconsistencies are overcome and healthy machine states are aligned, the deep learning methods disclosed herein can be used to identify outlier samples that do not fall within the expected range of healthy states. The application of such anomaly detection allows tracking the progression of industrial asset degradation over the component's lifecycle, and samples that appear anomalous can trigger necessary maintenance actions to avoid further damage to critical and expensive components, such as robotic arms.

[0017] As described in various embodiments of the present disclosure, the methods and systems disclosed herein utilize deep learning techniques with domain adaptation across different machine conditions to improve anomaly detection for industrial assets. The methods and systems disclosed herein eliminate the need to collect data for all operating conditions of an industrial asset and can be extended to other industrial applications where exhaustive data sets are not possible, thereby saving significant time required for training data collection and computational resources for processing such training data. That is, the deep learning-based domain adaptation techniques disclosed herein enable knowledge gained from one operating condition to be transferred to another. In this regard, the methods and systems utilize a two-stage approach to extract domain-invariant, class-robust discriminative features from raw data, followed by using these features in a one-class classifier for anomaly detection. Furthermore, the deep learning techniques disclosed herein can account for the effects of machine-to-machine variations or environmental conditions.

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

[0019] Unless otherwise expressly stated, it is not intended that the methods described herein be construed as requiring that their steps be performed in a particular order or as requiring any specific orientation of any apparatus. Thus, if a method claim does not actually recite the order in which its steps are to be followed, or a device or assembly claim does not actually recite an order or orientation for individual components, or if the claim or description does not otherwise specifically state that the steps are to be limited to a particular order or a particular order or orientation for the components of the device or assembly, then no order or orientation is intended to be inferred in any respect. This applies to any possible implicit basis for interpretation, including logical matters regarding the arrangement of steps, flow of operations, order of components, or orientation of components, the apparent meaning derived from grammatical construction or punctuation, and the number or type of embodiments described herein.

[0020] Throughout this disclosure, the term "one embodiment" or "one embodiment" means that a particular feature, structure, material, or characteristic described in connection with that embodiment is included in at least one embodiment of the invention, but does not necessarily indicate that it is present in all embodiments. Thus, the appearance of the phrase "in an embodiment" or "in another embodiment" in various places throughout this specification does not necessarily refer to the same embodiment of the invention. Furthermore, as used herein, the singular forms "a," "an," and "the" include plural referents unless the context clearly dictates otherwise. Thus, for example, reference to "a" element includes aspects having two or more such elements, unless the context clearly dictates otherwise. Additionally, "a element" may refer to one or more elements and, thus, may be used herein to mean "at least one."

[0021]

number

[0022] 1A and 1B, a flowchart illustrating a two-stage anomaly detection methodology or architecture S100 according to an embodiment of the present disclosure is shown. The first stage, commonly referred to as feature extraction using a neural network such as a CNN, is illustrated in FIG. 1A. The illustrated method begins with data preprocessing in S102 before the first stage. Data preprocessing in S102 generally involves preparing input data samples and labels for both the source and target domains. Once the source and target domain input data samples are prepared and labeled, the data samples are input to the CNN in S104. In S104, the CNN is trained to extract healthy-class discriminative features that separate healthy samples from all other classes (e.g., the “neutral” and “fault” classes) as well as domain-invariant features aligned in the feature space for all domains. In particular, the CNN is trained to extract features that satisfy the requirements of discrimination and domain invariance for the healthy class by using a two-part loss function in S106 and S108. To update the CNN to separate the healthy class from all other classes, a first part of a two-part loss function, cross-entropy error optimization, is applied in S106. That is, the cross-entropy error optimization in S106 considers all classes for the source domain and only the healthy class for the target domain. To update the CNN to minimize the domain discrepancy between healthy class samples in the source and target domains, a second part of the two-part loss function, domain adaptive loss optimization, is applied in S108 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 the 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, commonly referred to as anomaly detection using a one-class support vector machine (SVM), is illustrated in Figure 1B. The second stage typically utilizes high-level features extracted from the CNN trained in the first stage to perform the final task of anomaly detection. 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, a one-class classifier (e.g., a one-class SVM) is applied to assign a health label to each sample. 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 is tested on the remaining data from the target domain (including healthy and faulty data). Generally, in S114 and S116, a one-class SVM method is applied to detect samples that are outliers of the healthy class (i.e., abnormal or non-normal samples that correspond to possible faults in the machine). The second-stage anomaly detection method can detect non-normal samples in the target domain without seeing any non-normal samples during training. Steps S102-S116 of exemplary method S100 will now be discussed in more detail.

[0024] Method step S102 generally relates to data preprocessing, including preparing input data samples and labels for both the source and target domains. In particular, fixed cycle feature test (FCFT) data is acquired by repeatedly executing preset operating patterns at various points in the lifecycle of an industrial asset or machine. The preset patterns cover a wide range of operating conditions (e.g., minimum and maximum rotational speeds). Because the machine executes the same patterns at all points in its lifecycle, variations in the collected data are due to machine degradation. According to some embodiments of the present disclosure, a method for generating appropriate frequency-domain samples from feedback torque current data obtained by FCFT can be defined. FCFT data is collected for various health states. For each health state, an FCFT command pattern (characterized by different rotational direction and position commands) is repeated multiple times for each of the main operating regimes (rotational speeds), resulting in multiple patterns for each operating regime (e.g., collecting 48 patterns repeated 100 times (12 patterns each for 50, 500, 1000, and 3000 RPM)). The collected command speed signal is used to reduce each pattern to include only the steady-state portion of the feedback torque current signal. These patterns are further reduced by removing a preset percentage from the initial point to ensure a steady-state condition is achieved. Once all the steady-state patterns for a single health state are cleaned, they are concatenated across all iterations to generate a single continuous feedback torque current signal. This torque signal is then used to generate frequency-domain samples. A predetermined number of samples of a specified length are obtained by time-domain segmenting the torque signal and then saving the frequency-domain information for each segment as one of the final samples. These frequency-domain samples are then ready for feature extraction using CNN.

[0025]

number

[0026]

number

[0027] where * is the convolution operation. Furthermore, a nonlinear activation function is applied to the feature map to capture the nonlinear correlation between the features and the data. To further reduce the dimensionality of the feature map, a pooling layer can follow the convolution layer. Max-pooling technique is used to preserve meaningful spatial information of the feature map while simultaneously reducing the dimensionality.

[0028]

number

[0029]

number

[0030]

number

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

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

[0033]

number

[0034] However, the following prerequisites are ∀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 variables. The term v serves to exclude parts of the training dataset as outliers. Equation (4) below is the decision function used to distinguish anomalies in the dataset.

[0038]

number

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

[0040] As shown in FIG. 2A, the CNN architecture in stage 1 of FIG. 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 the first stage, and a dimensionality reduction module and a one-class classifier module in the second stage. In the first stage, which involves network training and optimization, the deep learning-based feature extraction module extracts meaningful information from 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 discrepancy between the feature representations corresponding to the source domain and the target domain. In the first stage, the network is trained using source data that encompasses all classes (H healthy, B intermediate, and F faulty) and a target domain that contains only the healthy class. Therefore, the main purpose of the first stage is to extract features that are domain-invariant and discriminate between the healthy class. Note that in the first stage, all health classes of the system are available in the source domain dataset, while only the healthy class is available in the target domain dataset. The lack of labels in the target domain can affect the transferability of the CNN, and the features may not perfectly identify the healthy class. To address this issue, a second stage (Figure 2B) is added to the CNN to improve the features' anomaly detection capability. In the second stage of the architecture shown in Figure 2B, the features are used as input for a one-class classifier to fulfill the anomaly detection objective. However, to reduce the complexity of the input data while retaining useful information, a dimensionality reduction module is first implemented to reduce the feature dimensionality. In this regard, t-SNE (t-SNE) is utilized.

[0041] To achieve transfer learning between domains, method step S106 involves cross-entropy error optimization to update the CNN to separate the healthy class from all other classes. Method step S108 involves domain adaptive loss optimization using maximum mean discrepancy (MMD) to update the CNN to minimize the domain discrepancy between the healthy class samples in the source domain and the healthy class samples in the target domain. Thus, in the first stage, the goal is to create features that distinguish the healthy class and are domain invariant. To achieve healthy class discrimination, a typical classification loss function, cross-entropy error, is used as shown in Equation (5) below.

[0042]

number

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

[0044] A domain adaptation methodology is deployed to transfer knowledge gained from a source domain to a target domain. Data corresponding to the source and target domains are collected under different working regimes, which may affect the data distribution (and corresponding feature representation) across domains. To address this issue, the MMD metric is utilized to measure the discrepancy between the data distributions in the source and target domains. The MMD term is added to the overall optimization objective as a loss function. The MMD loss function is defined as follows:

[0045]

number

[0046] where the 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, the loss functions in Equation (5) and Equation (6) above are combined to form a general optimization objective for the exemplary CNN architecture. Therefore, by integrating Equation (5) and Equation (6), the following general loss function can be formulated as shown in Equation (7) below:

[0048]

number

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

[0050]

number

[0051] In the formula, δ represents the learning rate.

[0052] The high-level feature representation collected from the first stage is expected to be used for anomaly detection purposes using a one-class classifier module. Before feeding the features into the module, the high-dimensional features must be mapped to a 2D feature space. The t-SNE algorithm is utilized to map the features from the high-dimensional space to the low-dimensional space.

[0053] During the second stage training process, the parameters in Equation (3) are optimized. Note that in the second stage, two hyperparameters also need to be adjusted. 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] FIG. 3 illustrates an example of a system for performing anomaly detection on business assets according to the present disclosure. The system illustrated in FIG. 1 may include a computing system 1, a user client 20, a control system 30, and a third-party system 40. The computing system 1 may provide a system for performing data analysis according to the present disclosure. The computing system 1 may be implemented, for example, using one or more general-purpose computers. As illustrated in FIG. 1, the computing system 1 may include an application 10 and a data storage device 12. The application 10 may be implemented by a software application including instructions that cause a computer to perform exemplary processes of the computing system. As illustrated in FIG. 1, the 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] CNN 100 may be the same as or similar to the CNN illustrated in FIG. 2A and may be an artificial neural network having an input layer (e.g., input data in FIG. 2A ), an output layer (e.g., a fully connected layer in FIG. 2A ), and multiple hidden layers between the input layer and the output layer (e.g., convolutional layer 1, convolutional layer 2, pooling layer 1, and flattening layer in FIG. 2A ). CNN 100 may be trained to process any type of data, such as sensor data from business or industrial assets. In one example, CNN 100 may be trained to process data acquired by each sensor using a training dataset that includes data that may be considered as input data to CNN 100. 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 the features identifies a healthy class and a second subset of the 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 be an interface through which application 10 communicates with various devices that may be external to computing system 1. For example, interface 110 may be configured to communicate information generated by application 10 to those devices. Further, for example, interface 110 may be configured to receive information from those devices that is intended for application 10.

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

[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 the anomaly detection made on the computing system 1; an example of a user client 20 may be a workstation remotely connected to a computation server using, for example, SSH (Secure Shell) or HTTP (Hypertext Transfer Protocol) requests. The CNN 100 can then be applied to user-provided inputs on the computation 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, for example, running the application 10 on a workstation configured to make the CNN predictions.

[0062] The control and / or analysis system 30 may use the predictions and results of the anomaly detection made in the computing system 1 to control devices 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 a control and / or analysis system 30 may be a control and / or analysis system (e.g., for anomaly detection or predictive maintenance) of a business asset such as a mechanical component or machine, an industrial process or plant, a vehicle (e.g., an autonomous vehicle), a computer network, a financial trading unit, etc.

[0063] The control and / or analysis system 30 may include a programmatic client 32 operating with the control and / or analysis system 30 to receive input, perform data analysis, and make decisions regarding further application-specific activities, for example, related to the maintenance of and / or control of business assets (e.g., equipment and systems).

[0064] The subject matter of this disclosure may also be implemented as a computing system 1 including a processor and a memory coupled to the processor. The memory may encode one or more programs for causing 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 special-purpose computer system, including an embedded system. Example 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 data set includes various combinations of gear health and gain settings. The gear failure mode is wear, and the states are healthy, intermediate, and faulty. 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 healthy and faulty states. Possible gain settings are low, medium, and high. Gear wear failures are most difficult to distinguish from health states because they are highly opaque in low gain setting data. Therefore, the method described in this example was initially developed using the extreme case of a low gain setting.

[0065] Each dataset contains fixed-cycle feature test (FCFT) data. For the purposes of this example, the dataset focused on the steady-state portion of the FCFT data, rather than the transient portion. In total, 100 repetitions (loops) of 63 command patterns were collected for each wear / gain state. The first 48 patterns were from four main operating 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 a specified rpm. The remaining 15 patterns also feature specific rpm, rotational direction, and position commands, but these were not utilized in this example. Also, based on input from domain experts, this example considered only forward rotation (positive patterns), e.g., four patterns from each operating regime. Reverse rotation (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 included command and feedback position, command and feedback velocity, and feedback torque current. Samples were collected every 444 μs. Ancillary information (encoder temperature and load speed) was also collected for each loop at a rate of 1 Hz. Date and time stamps for each collected data point and cumulative experiment time were also available for each sampling frequency. A summary of the data collection information is shown in Table I below.

[0067] [Table 1]

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

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

[0070] To evaluate the performance of our methodology (e.g., method S100 described above), we designed various tasks based on different scenarios. Therefore, 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 the 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 and target domains for a specific task (e.g., T4 and T5). In all experiments, the source and target domain datasets contained 1,500 samples (500 samples from each of the three health conditions).

[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 leaky rectified linear unit (ReLU) is utilized as the activation function in both convolutional layers. The convolutional layers are followed by a max pooling layer with a pooling size of 2. Detailed information on the network implementation and parameters is provided in Table III below.

[0073] [Table 3]

[0074] In order to better evaluate the efficiency of the methodology of the present disclosure, a benchmark study was conducted. Therefore, various types of anomaly detection methods were applied under a range of designed transfer tasks and datasets. Then, for each task, the performance and accuracy of the method disclosed herein were compared with other methods. The following is a description of the benchmark method.

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

[0076] Isolation 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, it is not easy to discover healthy state patterns. Compared to other similar methods, Isolation Forest provides high-quality anomaly detection capabilities.

[0077] Local Outlier Factor (LOF): This method provides a score that describes how likely a sample is to be an outlier. This technique analyzes the local density deviation of the examined sample relative to neighboring 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 various experiments and tasks. As the results show, the disclosed two-stage methodology clearly outperforms 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 a traditional one-class SVM. The only difference between these two methods is the first stage of the method disclosed herein, and the comparison results confirm a significant increase in the network's performance when domain-invariant and sound class discriminative features are fed into the one-class classifier.

[0079] Based on the first criterion for designing the transfer task, we conducted experiments to evaluate the impact of domain mismatch. Comparing the results between Tasks T1 and T2, and Tasks T3 and T4, we found that the network's performance decreased as the distance between the source and target domains increased. Meanwhile, the choice of source domain dataset also affected the network's performance. As shown in Figure 6, comparing the results of Tasks T4 and T5, and Tasks T1 and T6, we concluded that while the domain mismatch between the source and target domains was constant, changing the source and target domains could affect the network's performance. In both scenarios, the network disclosed herein was largely robust to these types of phenomena.

[0080] The value of the term v affects the number of samples that are considered outliers. Tuning this hyperparameter is a critical step in designing a one-class classifier. The impact of varying the value of v on the anomaly detection ability of the proposed network for Task T2 is demonstrated in Figures 7A-7C. For low values ​​of v, a relatively small percentage of samples are considered outliers. In other words, the model is overfitted (Figure 7A). Conversely, relatively high values ​​of v lead to the model undertraining problem (Figure 7B). Based on hyperparameter tuning studies, the optimal value of v is 0.01, at which point the network can adequately 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 disclosed method are intended to be domain-invariant and discriminate the healthy class. 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, features corresponding to the healthy class in the source and target domains are clustered into the same group. Therefore, it can be concluded that the disclosed network in the first stage is able to extract features that are domain-invariant and discriminate the healthy class. Because the features are domain-invariant, they can be used in transfer learning applications, transferring knowledge gained from the source domain to the target domain. Meanwhile, because the features discriminate the healthy class, 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 for a one-class SVM for anomaly detection applications. The goal is to develop a pattern that separates healthy class samples in the target domain from the other two classes. A visualization of the one-class classifier using the proposed method for task T5 is shown in Figure 9. As shown in Figure 9, the one-class classifier is able to properly cluster healthy class samples from both domains in the same group, while samples corresponding to the other two classes are considered outliers. Note that as the failure becomes more severe and the system health state moves toward the faulty class, the features are located farther away from the healthy class boundary.

[0083] In light of the above, it should be appreciated that at least some embodiments of the present disclosure relate to a two-stage deep learning-based transfer learning methodology to be used for anomaly detection. One feature of the methodology disclosed herein is that within the first stage of the architecture, as described above, it provides features that are domain-invariant and discriminate between healthy classes, making it different 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 a different operating condition. In this scenario, only a small amount of data corresponding to the system's health state is available in the target domain.

[0084] It should be noted that the terms "substantially" and "about" may be utilized herein to express the inherent degree of uncertainty that may result from any quantitative comparison, value, measurement, or other representation. These terms are also utilized herein to express the degree to which a quantitative representation may vary from the stated reference without resulting in a change in the basic functionality of the subject matter at issue.

[0085] While particular embodiments have been illustrated and described herein, it should be understood that various other changes and modifications can be made without departing from the spirit and scope of the claimed subject matter. Moreover, although various aspects of the claimed subject matter have been described herein, such aspects need not be utilized in combination. Accordingly, the appended claims are intended to cover all such changes and modifications that are within the scope of the claimed subject matter.

Claims

1. 1. A method for anomaly detection in a business asset, comprising: 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 by the neural network; reducing the dimensionality of the features to reduce the features; classifying the reduced features into normal and abnormal classes using a one-class classifier; Including, the 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 method wherein a first subset of the features extracted from the input data discriminates the healthy class, and a second subset of the features extracted from the input data are domain invariant.

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

3. The method of claim 1 , wherein the first subset of features is extracted by applying a cross-entropy error optimization to the source domain dataset.

4. The method of claim 1 , wherein the second subset of features is extracted by applying a domain adaptive loss optimization to the source domain dataset and the target domain dataset.

5. The method of claim 4 , wherein the domain adaptive loss optimization utilizes a maximum average discrepancy loss.

6. The method of claim 1 , wherein the dimensionality of the features is reduced by t-distributed stochastic neighbor embedding.

7. The method of claim 1 , wherein the one-class classifier is a support vector machine classifier.

8. 1. A system for anomaly detection in a business asset, comprising: one or more processors; a non-transitory computer-readable medium having instructions stored thereon; Equipped with The one or more processors execute the instructions to: receiving a source domain dataset corresponding to a first operating condition of the business asset; receiving 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 the features to reduced features; classifying the reduced features into normal and abnormal classes using a one-class classifier; the 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 the features extracted from the input data discriminates the healthy class, and a second subset of the features extracted from the input data is domain invariant. system.

9. The system of claim 8 , further comprising preprocessing the source domain dataset and the target domain dataset by data truncation.

10. The system of claim 8 , wherein the first subset of features is extracted by applying a cross-entropy error optimization to the source domain dataset.

11. 10. The system of claim 8, wherein the second subset of features is extracted by applying a domain adaptive loss optimization to the source domain dataset and the target domain dataset.

12. The system of claim 11 , wherein the domain adaptive loss optimization utilizes a maximum average discrepancy loss.

13. The system of claim 8 , wherein the dimensionality of the features is reduced by t-distributed stochastic neighbor embedding.

14. The system of claim 8 , wherein the one-class classifier is a support vector machine classifier.

15. 1. A method for repairing business assets, comprising: 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 by the neural network; reducing the dimensionality of the features to reduce the features; classifying the reduced features into normal and abnormal classes using a one-class classifier; identifying at least one business asset in the anomalous class; Repairing said at least one business asset; Including, the 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 the features extracted from the input data discriminates the healthy class, and a second subset of the features extracted from the input data is domain invariant. method.

16. The method of claim 15 , wherein the first subset of features is extracted by applying a cross-entropy error optimization to the source domain dataset.

17. The method of claim 15 , wherein the second subset of features is extracted by applying a domain adaptive loss optimization to the source domain dataset and the target domain dataset.

18. The method of claim 17 , wherein the domain adaptive loss optimization utilizes maximum average discrepancy loss.

19. The method of claim 15 , wherein the dimensionality of the features is reduced by t-distributed stochastic neighbor embedding.

20. The method of claim 15 , wherein the one-class classifier is a support vector machine classifier.

Images