Method and system for detecting anomalies in an operational asset and method for repairing an operational asset

Abstract

A method for anomaly detection in an operational asset comprises: collecting a source domain dataset corresponding to a first operating condition of the operational asset, wherein samples from the source domain dataset belong to a healthy class and a defective class; collecting a target domain dataset corresponding to a second operating condition of the operational asset, wherein samples from the target domain dataset belong to the healthy class; inputting the source domain dataset and the target domain dataset as input data to a neural network; extracting features from the input data by the neural network, wherein a first subset of features distinguishes the healthy class and a second subset of features is domain-invariant; reducing a dimensionality of the features into reduced features;and classifying the reduced features into a normal class and an anomaly class using a one-class classifier.;

Description

Cross-reference to related registration

[0001] This application claims priority over the provisional US application with serial number 63 / 431,384, which was filed on December 9, 2022, the entire contents of which are incorporated herein by reference. Technical field

[0002] The present invention relates to tools for monitoring the operational or industrial health or performance of assets. background

[0003] Operational or industrial assets or machines operate in different "domains" or "operating regimes," which include, for example, different rotational speeds, ambient temperatures, and loads. Even with the same machine condition (e.g., healthy, Type I fault, Type II fault, etc.), the data signature can differ due to the operating regime. For example, the vibration of a Type I fault machine at 50 rpm might resemble the vibration of a healthy machine at 100 rpm. This difference is a "domain discrepancy," and domain adaptation must be applied to eliminate it. Transfer learning is about using the knowledge gained from solving a problem in one domain ("source") to extend and improve problem-solving techniques in another domain ("target"). Summary

[0004] Further features and advantages of the present disclosure are set forth in the following detailed description and are partly apparent to those skilled in the art from this description or are recognized by the application of the embodiments described herein, including the detailed description following the claims, and the accompanying drawings.

[0005] In one embodiment, a method for anomaly detection in an operational asset comprises: collecting a source domain dataset corresponding to a first operating condition of the operational asset, wherein samples from the source domain dataset belong to a healthy class and a defective class; collecting a target domain dataset corresponding to a second operating condition of the operational asset, wherein samples from the target domain dataset belong to the healthy class; 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, wherein a first subset of features distinguishes the healthy class and a second subset of features is domain-invariant; and reducing one dimensionality of the features to reduced features.and classifying the reduced features into a normal class and an anomaly class using a single-class classifier.

[0006] In a further embodiment, an anomaly detection system in an operational asset comprises one or more processors; and a non-transitory, computer-readable medium that stores instructions which, when executed by the one or more processors, cause the one or more processors to: receive a source domain data set corresponding to a first operating condition of the operational asset, wherein samples from the source domain data set belong to a healthy class and a faulty class; receive a target domain data set corresponding to a second operating condition of the operational asset, wherein samples from the target domain data set belong to the healthy class; and input the source domain data set and the target domain data set as input data into a neural network.Extracting features from the input data using the neural network, wherein a first subset of features distinguishes the healthy class and a second subset of features is domain-invariant; reducing one dimensionality of the features into reduced features; and classifying the reduced features into a normal class and an anomaly class using a one-class classifier.

[0007] In yet another embodiment, a method for repairing a business asset comprises collecting a source domain dataset corresponding to a first operating condition of the business asset, wherein samples from the source domain dataset belong to a healthy class and a defective class; collecting a target domain dataset corresponding to a second operating condition of the business asset, wherein samples from the target domain dataset belong to the healthy class; 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, wherein a first subset of features distinguishes the healthy class and a second subset of features is domain-invariant; and reducing one dimensionality of the features to reduced features.Classify the reduced features into a normal class and an anomaly class using a single-class classifier; identify at least one operational asset within the anomaly class; and repair the at least one operational asset.

[0008] Both the preceding general description and the following detailed description describe various embodiments and are intended to provide an overview or framework for understanding the nature and character of the claimed subject matter. The accompanying drawings serve to provide a deeper understanding of the various embodiments and are integrated into and form part of this patent application. The drawings illustrate the various embodiments described herein and, together with the description, explain the basic principles and modes of operation of the claimed subject matter. Brief description of the drawings

[0009] The embodiments shown in the drawings are illustrative and exemplary and are not intended to limit the subject matter defined by the claims. The following detailed description of the illustrated embodiments will be understandable when read in conjunction with the following drawings, in which identical structures are identified by the same reference numerals and in which: Fig. Figure 1A schematically shows 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 here; Fig. Figure 1B schematically shows a method for anomaly detection using a single-class support vector machine according to one or more embodiments shown and described herein; Fig. Figure 2A schematically shows an overall architecture of a neural network according to one or more embodiments shown and described here; Fig. Figure 2B schematically shows an overall architecture of an anomaly detection frame according to one or more embodiments shown and described here; Fig. Figure 3 schematically shows a system for anomaly detection in an operational asset, according to one or more embodiments shown and described here; Fig. Figure 4 schematically shows a test setup for a manufacturing system that uses a gearbox, according to one or more embodiments shown and described here; Fig. Figure 5 shows an investigated gearbox and representative wear defect conditions of the gearbox according to one or more embodiments shown and described here; Fig.Figure 6 schematically shows a comparative study of the performance of the exemplary anomaly detection method according to one or more embodiments shown and described herein; Fig. Figure 7A schematically shows the anomaly detection performance of the exemplary anomaly detection method for a task where the value of the term ν = 0.005, where ν influences the number of samples considered as outliers, according to one or more embodiments shown and described here; Fig. Figure 7B schematically shows the anomaly detection performance of the exemplary anomaly detection method for a task where the value of the term ν = 0.5, where ν influences the number of samples considered as outliers, according to one or more embodiments shown and described here; Fig.Figure 7C schematically shows the anomaly detection performance of the exemplary anomaly detection method for a task where the value of the term ν = 0.01, where ν influences the number of samples considered as outliers, according to one or more embodiments shown and described here; Fig. Figure 8 schematically shows a visualization of high-level features of the raw data used in the exemplary anomaly detection method, according to one or more embodiments shown and described herein; and Fig. Figure 9 schematically shows the anomaly detection results for a single-class support vector machine according to one or more embodiments shown and described herein. Detailed description

[0010] The following section details various embodiments of methods and systems for anomaly detection in operational assets (e.g., mechanical or electrical equipment), examples of which are shown in the accompanying drawings. It should be noted that the term "operational asset" is used synonymously with "industrial asset" here, although the embodiments are not limited to industrial plants. Wherever possible, the same reference numerals are used in the drawings to indicate identical or similar parts.

[0011] In industrial environments, assets or systems such as gearboxes often play a crucial role in various applications, including wind turbines, conveyor systems, and industrial robots. These and other applications rely on properly functioning industrial assets to maintain energy production, move parts in a manufacturing hall, or ensure the necessary precision and throughput for material handling and production. Due to the complexity of many industrial assets, which makes them susceptible to various types of deterioration, including wear and corrosion, close monitoring of these assets is essential to prevent unexpected downtime, production losses, material damage, and potentially unsafe conditions that could result from a faulty component.Although industrial assets are designed for maximum lifespan in demanding environments, their constant movement inevitably leads to deterioration over time. Therefore, early and accurate assessment of the health and performance of industrial assets is crucial.

[0012] Traditionally, signal processing techniques are applied to vibration data from industrial assets, such as gearboxes, for prognostics and health management (PHM). However, neural networks can also be used for monitoring industrial assets. For example, convolutional neural networks (CNNs) can be used to analyze the vibration signals of an operational asset. Furthermore, deep learning strategies can be employed, offering advantages over conventional machine learning techniques and neural network approaches. A deep learning architecture, for instance, can perform automatic feature selection in its multiple hidden layers, and bypassing the normally required feature extraction and selection steps, the time required to implement a fault detection procedure is reduced.Furthermore, the features obtained by deep learning architectures are more sensitive and information-rich than those extracted using classical methods.

[0013] Although the data-driven aspects of deep learning are also a crucial advantage, it is not always easy or even possible to obtain data for every possible operating regime and machine state. Therefore, it is important to develop and extend techniques that enable fault identification and health assessment using only a subset of the entire range of possible states. Thus, applications of transfer learning are described here to allow knowledge from one machine state to be used to improve the assessment of other, unseen or less documented states.

[0014] While transfer learning solves the problem of limited data, this disclosure addresses further obstacles in the field of transfer learning. A key challenge in transfer learning is domain discrepancy. When transferring knowledge from the source domain to the target domain, a shift in the underlying domain distribution can affect the accuracy of the results. To avoid misclassifications, this disclosure employs domain matching techniques to overcome domain discrepancy.

[0015] In industrial practice, it is often most practical to perform domain matching using the healthy or basic class from different domains. Once the domain discrepancy is resolved and the healthy machine states are aligned, the deep learning methods disclosed here can be used to identify outlier samples that fall outside the expected range of the healthy state. Such anomaly detection application makes it possible to track the progression of deterioration of industrial assets throughout the component lifecycle, and samples that appear anomalous can trigger necessary maintenance actions to prevent further damage to critical and costly components, such as robot arms.

[0016] As described in various embodiments of this disclosure, the methods and systems disclosed herein utilize deep learning approaches with domain adaptation across different machine states 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 it is not feasible to collect a comprehensive dataset, thereby saving significant time for training data collection and computational resources for processing such training data. That is, the deep learning-based domain adaptation approaches of this disclosure enable the transfer of knowledge obtained from one operating state to another.In this respect, the methods and systems employ a two-stage approach to extract domain-invariant and healthy-class distinguishing features from raw data and subsequently use these features in a single-class classifier for anomaly detection. Furthermore, the deep learning approaches disclosed here are capable of accounting for machine-to-machine variations or the effects of environmental conditions.

[0017] Those skilled in the art will recognize that the various embodiments of the present disclosure can also be carried out without one or more of the specific details described herein or with other substitute and / or additional methods, materials, or components. In other cases, known structures, materials, or processes are not shown or described here in detail to avoid obscuring aspects of the various embodiments of the invention. Similarly, certain numbers, materials, and configurations are given here for explanatory purposes to provide a comprehensive understanding of the various embodiments of the present disclosure. Furthermore, it is self-evident that the various embodiments shown in the figures are illustrative representations and are not necessarily drawn to scale.

[0018] Unless expressly stated otherwise, none of the methods presented herein shall be construed as requiring their steps to be performed in a particular order, nor as requiring any particular orientation with any apparatus. If a method claim does not specify an order of steps, or if a claim for a device or arrangement does not specify an order or orientation of the individual components, or if the claims or description do not expressly state that the steps are restricted to a particular order, or if no particular order or orientation of the components of a device or arrangement is specified, no order or orientation shall be inferred in any way.This applies to any possible non-explicit basis for interpretation, including: questions of logic relating to the arrangement of steps, the workflow, the sequence of components or an alignment of components; a simple meaning that arises from the grammatical structure or punctuation; and the number or nature of the embodiments described in the specification.

[0019] When the present disclosure refers to “a single embodiment” or “an embodiment,” this means that a particular feature, structure, material, or property described in connection with the embodiment is included in at least one embodiment of the invention, but does not mean that it is present in every embodiment. Therefore, the expressions “in one embodiment” or “in another embodiment” at different points in this description do not necessarily refer to the same embodiment of the invention. Furthermore, the singular forms “a,” “an,” and “the” used herein include the plural unless the context clearly indicates otherwise. For example, a reference to “a” component includes aspects with two or more such components unless the context clearly indicates otherwise.Furthermore, the term "a component" can stand for one or more components and can thus be used here in the sense of "at least one".

[0020] In the present disclosure, a transfer learning approach is used to address anomaly detection problems. In this case, the data corresponding to the source domain contain all three health states (e.g., healthy, intermediate, and faulty), while for the target domain dataset, only the healthy state of the system is available. It is true that Ds={(xsi,ysi)}i=1ns and Dt={(xti,yti)}i=1nt The source and target domains should be specified. It should be noted that... xsi,ysi,xti, and yti for the i-th sample and the health label, and n s and n tThe number of samples corresponds to both the source and target domains. The goal is to generate domain-invariant and healthy class distinguishing features that can be used for anomaly detection in both the source and target domain datasets. Because only the healthy class corresponding to the target domain is used for domain fitting, the methodology is called conditional domain fitting (CDA).

[0021] Referring to Fig. 1A and Fig. Figure 1B shows a flowchart illustrating a two-stage anomaly detection methodology or architecture S100 in accordance with the embodiments of the present disclosure. The first stage is generally referred to as feature extraction using a neural network, such as a CNN, and is described in Fig.Figure 1A illustrates the process. The procedure begins in S102 with data preprocessing prior to the first stage. Preprocessing in S102 generally involves preparing input data samples and labels for both a source and a target domain. Once the source and target domain input data samples are prepared and labeled, the samples are fed into a CNN in S104. In S104, the CNN is trained to extract healthy class distinguishing features that differentiate healthy samples from all other classes (e.g., the "in-between" and "defective" classes) and to extract domain-invariant features that are aligned in the feature space for all domains. Specifically, the CNN is trained to extract features that satisfy the healthy class distinguishing and domain-invariant requirements by using a two-part loss function in S106 and S108.The first of the two-part loss functions, cross-entropy loss optimization, is applied in S106 to update the CNN and separate the healthy class from all other classes. That is, the cross-entropy loss optimization in S106 considers all classes for the source domain and only the healthy class for the target domain. The second of the two-part loss functions, domain-match loss optimization, is applied in S108 using the maximum mean discrepancy (MMD) to update the CNN so that the domain discrepancy between the healthy class samples of the source domain and the target domain is minimized. That is, the domain-match loss optimization using an MMD in S108 considers only healthy class samples from the source domain and the target domain.After applying the two-part loss function at S106 and S108, the CNN extracts high-level feature information for both the source and target domains at S110, and the first stage is complete.

[0022] The second stage is generally referred to as anomaly detection using a single-class support vector machine (SVM) and is described in Fig.Figure 1B illustrates this. In the second stage, the high-level features extracted from the trained CNN in the first stage are generally used to perform the actual task of anomaly detection. The second stage begins at S112, where the dimensionality of the problem is reduced by applying t-distributed stochastic neighborhood embedding (t-SNE). At S114, each sample is then assigned health labels by applying a one-class classifier (e.g., one-class SVM). Specifically, at 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. At S116, the one-class SVM then tests the remaining data from the target domain (including healthy and faulty data).In general, S114 and S116 employ the single-class SVM procedure to detect samples that are outliers of the healthy class (i.e., anomalous or abnormal samples corresponding to a likely machine error). The second-stage anomaly detection procedure is capable of detecting abnormal samples in the target domain even if no abnormal samples were seen during training. Steps S102-S116 of the example procedure S100 are now explained in more detail.

[0023] Process step S102 is focused on data preprocessing, which generally includes preparing input data samples and labels for both a source and a target domain. Specifically, the data for the Fixed Cycle Feature Test (FCFT) are obtained by repeatedly executing preset operating patterns at various points in the life cycle of an industrial asset or machine. The preset patterns cover a wide range of operating conditions (e.g., minimum and maximum rotational speeds). Since the machine operates with the same patterns at each point in its life cycle, changes in the collected data are attributable to machine deterioration. According to some embodiments of this disclosure, a method for generating suitable frequency domain samples from the feedback torque current data obtained via FCFT can be defined.The FCFT data are collected for various health states. For each health state, there are multiple repetitions of the FCFT command patterns (characterized by different direction and position commands) executed for each of the main operating regimes (speeds), with multiple patterns occurring for each operating regime (e.g., collecting 100 repetitions of 48 patterns (12 each for 50, 500, 1000, and 3000 rpm)). The collected setpoint speed signal is used to reduce each pattern to contain only the steady-state components of the feedback torque current signal. These are further reduced by removing a predetermined percentage of the initial points to ensure that the steady-state condition is met.All cleaned, steady-state patterns for a health condition are chained together across all repetitions to create a continuous, feedback-driven torque current signal. This torque signal is used to generate the frequency domain samples. A predetermined number of samples of a specific length are obtained by splitting the torque signal into the time domain and then storing 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.

[0024] Process step S104 is directed towards feature extraction using a CNN architecture. An example of a CNN architecture according to the embodiments of the present disclosure is given in Fig.Figure 2A illustrates this. A CNN architecture is used as the feature extraction module to extract useful information from the raw input data. This architecture combines convolutional, pooling, and fully connected (FC) layers. The convolutional layers are primarily responsible for extracting meaningful information from the input data. In this process, the higher-dimensional input data is typically transformed into lower-dimensional output features. In other words, the features are extracted automatically without prior knowledge of the input data. It is true that fjl the j-th feature map on the l-th layer, while the term wijl represents the kernel, which combines the i-th feature map with the j-th feature map on the i-th layer with the bias of bjl connects. By convolution of the feature map with kernels, the features can be extracted as shown in equation (1) below: fjl=Σifil−1∗wijl+bjl where * is the convolution operation. Additionally, a nonlinear activation function is applied to the feature maps to capture the nonlinear correlation between the features and the data. To further reduce the dimensionality of the feature maps, a pooling layer can follow the convolutional layers. The max pooling technique is used to retain the meaningful spatial information of the feature map while simultaneously reducing its dimensionality.

[0025] Regarding MMD, Ehs∼Ps[∅(hs)] the mathematical expectation of Ø(h s ) denotes when the high-level features of the source domain, h s , the probability distribution P s are subject to, and Ehs~Ps[∅(ht)] The mathematical expectation of Ø(h) should be t ) denote when the high-level features of the target domain, h t , the probability distribution P t The interpretation of MMD as the squared distance between the kernel embeddings of marginal distributions in the reproducing kernel Hilbert space (RKHS) can be defined as shown in equation (2) below: MMD2(Ps,Pt)≜‖∅‖H≤1sup‖Ehs∼Ps[∅(hs)]−Eht∼Pt[∅(ht)]‖ where Ø(.) and sup(.) denote the mapping function and supremum of the input aggregate, respectively. Note that ||Ø||ℌ ≤ 1 defines a set of functions in the unit sphere of an RKHS ℌ. In a case where P s = P t Based on the statistical tests, MMD(P) holds true. s , P t) = 0 and vice versa. The MMD calculation method is highly influenced by the choice of kernel. Therefore, the performance of multiple cores (MK) is exploited by a combination of five RBF (Radial Base Function) cores.

[0026] A single-class SVM of level 2 refers to an unsupervised machine learning model that generates a decision function to identify outliers in the dataset. In this approach, the model is trained using only one class of data (typically the normal or healthy class). The knowledge gained from the trained model allows it to identify whether the test data is similar to or different from the training dataset. In most real-world applications, collecting data from anomaly classes is not a simple task. Therefore, when using a single-class SVM for anomaly detection, the model can be trained using only the healthy or normal class dataset, without needing to use other classes.

[0027] The one-class SVM model creates a nonlinear decision boundary by mapping the original feature space to the higher-dimensional spaces. The mapping transformation is performed properly using kernel functions. Assuming that X and Y denote the original and the high-dimensional feature spaces (∅ : X → Y), the one-class SVM model can be trained using the following equations (3): minω,ξ,b12‖ω‖22+1vm∑i=1mξi−b st ∀i=1,2,…,N ν∈(0,1],ξi≥0, ω⋅∅(xi)≥b−ξi. where ξ i The slack variable is represented. The term ν is responsible for excluding a portion of the training dataset as outliers. The following equation (4) is a decision function used to distinguish the anomalies in the dataset: f(x)=sing(ω⋅∅(x)−b)={+1 Normal−1 Outlier

[0028] In the present disclosure, RBF (Radial Base Function) kernels are used to transform the original feature space into the higher-dimensional space.

[0029] As in Fig. As shown in Figure 2A, the CNN architecture of Stage 1 comprises Fig.1A consists of four main modules in two stages. The four main modules in the first stage include a deep learning-based feature extraction module and a domain fitting module, while in the second stage they include a dimensionality reduction module and a single-class classifier module. In the first stage, which involves training and optimizing the network, 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. The HL features are then passed to the domain fitting module. The main task of the domain fitting module is to generate domain-invariant features. In other words, the domain fitting module reduces the discrepancy between the feature representations corresponding to the source and target domains.In the first stage, the network is trained with the source data, which includes all classes (H - healthy, B - intermediate, and F - defective), and with the target domain, which contains only the healthy class. Therefore, the main goal of the first stage is to extract the features that are both domain-invariant and distinctive for the healthy class. It's important to note that in the first stage, all healthy classes of the system are available for the source domain dataset, but only the healthy class is available for the target domain dataset. Missing labels in the target domain can impair the transferability of the CNN, and the features may not be fully distinctive for the healthy class. To address this issue, a second stage is added to the CNN ( . . Fig. 2B) added to improve the anomaly detection capability of the features. In the second stage of the in Fig.In the architecture shown in Figure 2B, the features are used as input for a one-class classifier to fulfill the anomaly detection objectives; however, a dimension reduction module is provided to first reduce the feature dimensions in order to decrease the complexity of the input data while retaining the useful information. In this context, t-distributed stochastic neighborhood embedding (t-SNE) is exploited.

[0030] To achieve cross-domain transfer learning, procedure step S106 is directed towards cross-entropy loss optimization to update the CNN and separate the healthy class from all other classes, while procedure step S108 is directed towards domain-match loss optimization using the maximum mean discrepancy (MMD) to update the CNN so that the domain discrepancy between the healthy class samples of the source domain and the target domain is minimized. Thus, the goal of the first step is to create features that are both healthy class distinguishable and domain-invariant. To satisfy the healthy class distinction, a typical classification loss function, the cross-entropy loss, is used, as shown in equation (5) below: LC=1n∑i=1n∑j=1Nc1{yi=j}log(yi,j') where n, N c , y i , y' ijThe numbers 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 j-th class.

[0031] To transfer the knowledge obtained from the source domain to the target domain, a domain adaptation methodology is used. The data corresponding to the source and target domains are collected under different working regimes, which can affect the data distribution (and the corresponding feature representation) in the domains. To address this issue, the MMD metric is used to measure the discrepancy between the data distribution of source and target domains. The MMD term is added as a loss function to the overall optimization objectives. The MMD loss function is defined in equation (6) below: LDA=MMD(Fsource,Ftarget) where the arguments F source and F targetThe feature representations of the data are those that correspond to the respective source and target domains.

[0032] In summary, the combination of the loss functions from equations (5) and (6) above constitutes the general optimization goal of the exemplary CNN architecture. Therefore, by integrating equations (5) and (6), the following general loss function can be formulated, as shown in equation (7) below: Ltot=αLC+βLDA where α > 0 and β > 0 are the penalty coefficients for L. C and L DA are. During the training stage, the network parameters of the feature extraction module, ϑ f , and of the single-class classifier module, ϑ c , optimized and updated according to the loss function. The update process can be formulated as shown in equation (8) below: θ←θ−δ(α∂LC∂θ+β∂LDA∂θ) where δ indicates the learning rate.

[0033] The high-level feature representations collected in the first stage will be used for anomaly detection purposes with the help of the single-class classifier module. Before the features are fed into the module, the high-dimensional features must be mapped to the 2D feature space. The t-SNE algorithm is used to map the features from the higher-dimensional space to the lower-dimensional space.

[0034] During the training process in the second stage, the parameters of equation (3) are optimized. It should be noted that two hyperparameters also need to be tuned in the second stage. The first hyperparameter is ν, and the second hyperparameter is the RBF kernel hyperparameter, γ, where, according to equation (9) below: k(Φi,Φj)=e−γ‖Φi−Φj‖2 where Φ i , Φ j represents the i-th and j-th feature samples and ||·|| is the L2 norm operator.

[0035] Fig. Figure 3 shows an example of a system that incorporates the anomaly detection system in an operational asset as disclosed herein. Fig. The system shown in Figure 1 can comprise a computer system 1, a user-client 20, a control system 30, and a third-party system 40. The computer system 1 can provide the system for data analysis according to the present disclosure. The computer system 1 can, for example, be implemented with one or more general-purpose computers. As shown in Figure 1, the computer system 1 can be implemented with one or more general-purpose computers. Fig. As shown in Figure 1, the computer system 1 can comprise an application 10 and a data storage device 12. The application 10 can be implemented by a software application containing instructions that cause a computer to perform exemplary processes of the computer system. As shown in Figure 1, the computer system 1 can comprise an application 10 and a data storage device 12. The application 10 can be implemented by a software application containing instructions that cause a computer to perform exemplary processes of the computer system. Fig.As shown in Figure 1, the application 10 can include a CNN 100 with a feature extraction module 102 and a domain adaptation module 104, a dimension reduction module 106, a single-class classifier module 108 and an interface 110.

[0036] The CNN 100 can be the same or similar to the one in Fig. The CNN shown in 2A can be an artificial neural network with an input layer (e.g., input data from Fig. 2A), an output layer (e.g., fully connected layer made of Fig. 2A) and a multitude of hidden layers (e.g., Convolutional Layer 1, Convolutional Layer 2, Pooling Layer 1 and Flatten Layer). Fig.2A) between the input and output layers. The CNN 100 can be trained to process any type of data, such as sensor data from an operational or industrial asset. For example, the CNN 100 can be trained to process data received from the respective sensors, using a training dataset with possible input data for the CNN 100. The training dataset can be stored in the data storage device 12, which the application 10 can access.

[0037] The feature extraction module 102 and the domain fitting module 104 can be connected to and configured with at least one of the multiple hidden layers of the CNN 100 to extract features from the input data, wherein a first subset of features is distinguishing for the healthy class and a second subset of features is domain-invariant. The details of the process performed by the feature extraction module are described above with respect to procedure steps S104-S110. The dimension reduction module 106 can be configured to perform procedure step S112 described above. The single-class classifier module 108 can be configured to perform procedure steps S114-S116 described above.

[0038] Interface 110 can be an interface for application 10 to communicate with various devices that may be provided outside of computer system 1. For example, interface 110 can be configured to communicate information generated by application 10 to these devices. Furthermore, interface 110 can be configured to receive information from these devices addressed to application 10.

[0039] The data storage facility 12 can be configured to store data used by the application 100. Although Fig. Figure 3 shows that the data storage device 12 is part of the computer system 1; however, in some examples, the data storage device 12 may also be located outside the computer system, as long as the data stored in the data storage device 12 is accessible to the application 10.

[0040] User-client 20 can be a client facility connected to computer system 1. User-client 20 can contain a user application 22 that can utilize the predictions and anomaly detection results performed on computer system 1. A specific example of user-client 20 is a workstation remotely connected to a computer server, for example, using SSH (Secure Shell) or HTTP (Hypertext Transfer Protocol) requests. CNN 100 can then be applied to user-provided input on the computer server, and the resulting predictions and anomaly detections can be returned to user-client 20. User-client 20 can be part of the same physical facility as computer system 1 on which application 10 is running, for example, on a workstation configured to perform CNN predictions.

[0041] The control and / or analysis system 30 can control a facility and / or perform further data analysis using the predictions and results of the anomaly detection performed in computer system 1. The control and / or analysis system 30 can be an anomaly detection and / or predictive maintenance system or part of one. An example of the control and / or analysis system 30 could be a control and / or analysis system (e.g., for anomaly detection or predictive maintenance) for an operational asset, such as a machine part or machine, an industrial process or plant, a vehicle (e.g., an autonomous vehicle), a computer network, a financial transaction unit, etc.

[0042] The control and / or analysis system 30 can include a programmatic client 32 that runs in the control and / or analysis system 30, receives an input, performs data analysis, and makes decisions about further application-specific actions, which relate, for example, to the maintenance of the operational asset (e.g., a part of a plant or system) and / or to the control of the operational asset (e.g., a part of a plant or system).

[0043] Furthermore, the subject matter of this disclosure can be implemented as a computer system 1 comprising a processor and a working memory connected to the processor. The working memory can encode one or more programs that cause the processor to perform one or more of the procedures described herein. In some examples, the system 1 can be a general-purpose computer system. In other examples, the system 1 can be a specialized computer system, including an embedded system. Example.

[0044] Gear data from an industrial system was analyzed. As in Fig.As shown in Figure 4, the gearbox is part of a planetary gear system for transmitting power from the motors to the loads. The dataset includes various combinations of gearbox health state and gain settings. The gearbox failure mode is wear, and the states are healthy, intermediate, and faulty. The gearbox under investigation and the representative, wear-induced faulty state of the gearbox are shown in Figure 4. Fig.Figure 5 is shown. The area in between represents an intermediate stage of gear wear between the healthy and defective states. The possible gain settings are low, medium, and high. Gear wear defects are least pronounced in the data with low gain settings, making it most difficult to distinguish between healthy states. Therefore, the procedures described in this example were initially developed using the extreme case of low gain settings.

[0045] Each dataset contains data for the Fixed-Cycle Feature Test (FCFT). For the purposes of this example, the focus of the dataset was on the steady-state components of the FCFT data, rather than the transient parts. A total of 100 iterations (loops) of 63 command patterns were collected for each wear / strength condition. The first 48 patterns are from the four main operating regimes: 50 rpm, 500 rpm, 1000 rpm, and 3000 rpm (12 patterns for each order). These patterns were generated by applying different direction and position commands at the specified speed (rpm). The remaining 15 patterns are similarly characterized by a specific rpm, direction, and position command, but these remaining 15 patterns were not used in this example.Furthermore, in this example, based on input from domain experts, only forward rotations (positive patterns) were considered, i.e., four patterns from each of the operating regimes. That is, backward rotations (negative patterns) were removed from the dataset during preprocessing. Ultimately, 100 loops with 32 patterns (eight from each rotation) were analyzed.

[0046] Each loop contains the command and feedback position, command and feedback rate, and feedback torque current. Samples were collected every 444 µs. Auxiliary information (sender temperature and load rate) was also collected for each loop at a rate of 1 Hz. The time and date stamp of each collected data point and the cumulative experiment time were also available for each sampling frequency. A summary of the data collection information is provided in Table I. Table I. Detailed Data Collection Information signal Unit Sampling rate Main signals collected for each loop of the FCFT Command position encoder pulse count 444 µs Feedback position encoder pulse count 444 µs Command speed rpm 444 µs Feedback rate rpm 444 µs Feedback torque current % of torque rate 444 µs Total travel time s 444 µs Auxiliary signals collected for each loop of the FCFT Sensor temperature °C 1s Load rate % of torque rate 1s Total travel time s 1s

[0047] For preprocessing, the various loops and patterns were separated, resulting in a 100-loop x 32-pattern matrix of samples. The samples initially contain both transient and steady-state information. The first goal of preprocessing is to reduce the samples to contain only steady-state components. To this end, all points of the feedback torque current signal corresponding to the points where the command velocity of each sample is within one revolution per minute of the operating regime velocity were retained. The truncated torque samples were then further reduced to eliminate overshoot. After analyzing several loops and patterns, it was determined that the first 40% of each sample should be removed to retain only the steady-state information.Finally, for each operating regime, the eight cleaned patterns from each loop were chained together, and then these 100 loops were chained together. The result is a cleaned torque signal for each wear and gain condition for each operating regime (a total of 36 signals).

[0048] The final preprocessing step involves organizing the torque signals into samples for the example procedure described here. Each torque signal was down-sampled by taking every 10th data point. The torque signal was then divided into 500 windows, each containing 2000 time domain points (2000 time domain points provide 1000 unique frequency domain points). The step size and percentage overlap for windowing were dynamically determined based on the length of the entire torque signal. The frequency domain information for each window was stored as a sample for the given combination of wear, gain, and operating regime. In total, 500 1000-point samples of frequency data were generated for each of the 36 conditions (three wear cases times three gain settings times four operating regimes) for the cleaned, steady-state torque feedback current signal.

[0049] To evaluate the performance of the methodology (e.g., the S100 procedure described above), various tasks were designed based on different scenarios. Therefore, as shown in Table II below, six different trials were conducted under varying operating conditions. There are two main criteria for task design. The first criterion is the evaluation of domain discrepancy by increasing the distance between the target operating condition and a fixed source condition (e.g., T1 and T2). The second criterion is the evaluation of the selection of the optimal source domain condition by swapping the source and target domains for a given task (e.g., T4 and T5). In all trials, the number of samples in both the source domain and target domain datasets was 1500 (500 samples from each of the three health conditions). Table II. Detailed description of the planned experiments Task assigned Source domain Target domain T1 50 rpm 1000 rpm T2 50 rpm 3000 rpm T3 500 rpm 1000 rpm T4 500 rpm 3000 rpm T5 3000 rpm 500 rpm T6 1000 rpm 50 rpm

[0050] Regarding the in Fig. In the CNN architecture shown in Figure 2A, the feature extraction module is formed from two consecutive convolutional layers with 30 and 20 filters (filter size = 5), respectively. The Leaky Rectified Linear Unit (ReLU) is used as the activation function in both convolutional layers. The convolutional layers are followed by a Max Pooling layer with a pool size of 2. Detailed information about the network implementation and parameters is listed in Table III below. Table III. Specification of network parameters parameter Value parameter Value δ 1e-4 Epochs 500 Neurons of the FC layer 128 Dropout rate 0,5 Batch size for cross-entropy 32 Batch size for MMD 100 Number of samples 1500

[0051] To better assess the efficiency of the methodology presented in this disclosure, a benchmarking study was conducted. Therefore, various anomaly detection methods were applied to the designed transfer tasks and the dataset. Subsequently, the performance and accuracy of the method presented here were compared with the other methods for each task. The benchmarking methods are described below:

[0052] Conventional single-class SVM: The main difference between the method disclosed here and conventional single-class SVM is the classifier input. In conventional single-class SVM, the raw frequency spectrum is used as the input for the classifier. The differences in performance between the method disclosed here and conventional single-class SVM demonstrate the efficiency of the first stage of the method disclosed here.

[0053] Isolation Forest: This method is a non-parametric and 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 the pattern of healthy data. Compared to other similar methods, Isolation Forest offers high-quality anomaly detection capabilities.

[0054] Local Outlier Factor (LOF): This method provides a value that explains the probability of a sample being an outlier. This approach analyzes the local density deviation of the samples under investigation compared to their neighbors. Samples with relatively low density are considered outliers.

[0055] A comprehensive study was conducted to validate the efficiency and effectiveness of the anomaly detection methodology described here compared to other approaches. Fig. Figure 6 shows the comparative evaluation of the investigated methods across various experiments and tasks. As the results demonstrate, the two-stage methodology of the present disclosure clearly outperforms the other methods. It should be noted that the effectiveness of the first stage of the method described here can be validated by comparison with the results of conventional single-class SVM. The only difference between these two approaches lies in the first stage of the method presented here, and the comparative results demonstrate a noticeable performance increase of the network when the single-class classifier is fed with domain-invariant and healthy distinguishing features.

[0056] Based on the first criterion of the transfer task design, experiments were conducted to evaluate the effects of domain discrepancy. A comparison of the results between tasks T1 and T2, as well as between T3 and T4, shows that network performance decreases with increasing divergence between the operating conditions of the source and target domains. Conversely, the choice of source domain dataset also affects network performance. The comparison of results between tasks T4 and T5, and between tasks T1 and T6, reveals that... Fig. Section 6 shows that while the discrepancy between the source and target domains is constant, a change in the source and target domains can affect the performance of the networks. In both cases, the network of the present disclosure is virtually insensitive to this type of phenomenon.

[0057] The value of the term ν influences the number of samples considered outliers. Tuning this hyperparameter is a crucial step in developing the one-class classifier. The effect of varying the value of ν on the anomaly detection capability of the proposed network for task T2 is described in the Fig. Figures 7A-7C are shown. If the value of ν is low, a smaller proportion of the samples are considered outliers. In other words, the model fits too well ( Fig. 7A). In contrast, a higher value of ν leads to an underfit of the model ( Fig. 7B). Based on the hyperparameter tuning study, the optimal value for ν is 0.01, at which the network can properly distinguish the samples corresponding to the healthy state from the other two health states of the system ( Fig. 7C).

[0058] The features extracted in the first stage of the procedure presented here are intended to be domain-invariant and health class-distinguishing. To better understand these two properties of the high-level features, their visualization is shown in [reference to be added]. Fig. 8 shown. As shown. Fig. As can be seen in Figure 8, the features corresponding to the healthy class of both the source and target domains are grouped together. It follows that the network presented here can extract the domain-invariant and healthy class-distinguishing features in the first stage. Since the features are domain-invariant, they can be used for transfer-learning applications, transferring knowledge gained from the source domain to the target domain. However, because the features are healthy-class distinguishing, they are suitable as input for any single-class classifier for the anomaly detection application.

[0059] The extracted high-level features are used as input for the one-class SVM for the anomaly detection application. The goal is to develop a pattern that separates the healthy class samples from the other two classes in the target domain. The visualization of the one-class classifier using the proposed procedure for task T5 is shown in Fig. 9 shown. As in Fig. As shown in Figure 9, the one-class classifier can correctly classify samples from the healthy class of both domains into the same group, while samples corresponding to the other two classes are considered outliers. It should be noted that as the severity of the errors increases and the system health state transitions into the error class, the features move further away from the healthy class boundary.

[0060] In light of the foregoing, it should now be clear that at least some embodiments of the present disclosure aim at a two-stage, deep-learning-based transfer-learning methodology used for anomaly detection. A differentiating feature of the methodology disclosed here lies within the first stage of the architecture, which, as described above, provides the domain-invariant and healthy class-distinguishing features. With the methodology disclosed here, the knowledge obtained from the source domain dataset under one operating condition can be used to detect outliers in the target domain dataset under a different operating condition. In this scenario, only a small amount of data about the health of the system is available for the target domain.

[0061] It should be noted that the terms "essentially" and "approximately" may be used here to represent the inherent degree of uncertainty that can be attributed to any quantitative comparison, value, measurement, or other representation. These terms are also used here to represent the extent to which a quantitative representation may deviate from a given reference without altering the fundamental function of the object in question.

[0062] Although certain embodiments have been presented and described here, it should be clear that various other changes and modifications can be made without affecting the spirit and scope of the claimed subject matter. Furthermore, even though various aspects of the claimed subject matter have been described here, these aspects need not be used in combination. It is therefore intended that the attached claims cover all such changes and modifications that fall within the scope of the claimed subject matter. QUOTES INCLUDED IN THE DESCRIPTION

[0000] This list of documents cited by the applicant was automatically generated and is included solely for the reader's convenience. The list is not part of the German patent or utility model application. The DPMA accepts no liability for any errors or omissions. Cited patent literature

[0000] US 63 / 431,384

[0001]

Claims

[1] Procedure for anomaly detection in an operational asset, comprising: Collecting a source domain dataset that corresponds to a first operating condition of the operational asset, with samples from the source domain dataset belonging to a healthy class and a defective class; Collecting a target domain dataset that corresponds to a second operating condition of the operational asset, with samples from the target domain dataset belonging to the healthy class; 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, where a first subset of features distinguishes the healthy class and a second subset of features is domain-invariant; Reducing the dimensionality of features into reduced features; and Classify the reduced features into a normal class and an anomaly class using a single-class classifier. [2] The method of claim 1, further comprising preprocessing the source domain data set and the target domain data set by data truncation. [3] Method according to claim 1, wherein the first subset of features is extracted by applying cross-entropy loss optimization to the source domain dataset. [4] Method according to claim 1, wherein the second subset of features is extracted by applying domain matching loss optimization to the source domain dataset and the target domain dataset. [5] Method according to claim 4, wherein the domain matching loss optimization uses the maximum mean discrepancy loss. [6] Method according to claim 1, wherein the dimensionality of the features is reduced by t-distributed stochastic neighborhood embedding. [7] Method according to claim 1, wherein the single-class classifier is a support vector machine classifier. [8] System for anomaly detection in an operational asset, comprising: one or more processors; and a non-transitory, computer-readable medium that stores instructions which, when executed by one or more processors, cause one or more processors to: Receiving a source domain dataset that corresponds to a first operating condition of the operational asset, with samples from the source domain dataset belonging to a healthy class and a defective class; Receiving a target domain dataset that meets a second operating condition of the operational asset, with samples from the target domain dataset belonging to the healthy class; 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, where a first subset of features distinguishes the healthy class and a second subset of features is domain-invariant; Reducing the dimensionality of features into reduced features; and Classify the reduced features into a normal class and an anomaly class using a single-class classifier. [9] System according to claim 8, further comprising preprocessing the source domain data set and the target domain data set by data truncation. [10] System according to claim 8, wherein the first subset of features is extracted by applying cross-entropy loss optimization to the source domain dataset. [11] System according to claim 8, wherein the second subset of features is extracted by applying domain matching loss optimization to the source domain dataset and the target domain dataset. [12] System according to claim 11, wherein the domain matching loss optimization uses the maximum mean discrepancy loss. [13] System according to claim 8, wherein the dimensionality of the features is reduced by t-distributed stochastic neighborhood embedding. [14] System according to claim 8, wherein the single-class classifier is a support vector machine classifier. [15] Methods for repairing an operational asset, comprising: Collecting a source domain dataset that corresponds to a first operating condition of the operational asset, with samples from the source domain dataset belonging to a healthy class and a defective class; Collecting a target domain dataset that corresponds to a second operating condition of the operational asset, with samples from the target domain dataset belonging to the healthy class; 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, where a first subset of features distinguishes the healthy class and a second subset of features is domain-invariant; Reducing the dimensionality of features into reduced features; Classifying the reduced features into a normal class and an anomaly class using a single-class classifier; Identify at least one operational asset within the anomaly class; and Repairing at least one operational asset. [16] Method according to claim 15, wherein the first subset of features is extracted by applying cross-entropy loss optimization to the source domain dataset. [17] Method according to claim 15, wherein the second subset of features is extracted by applying domain matching loss optimization to the source domain dataset and the target domain dataset. [18] Method according to claim 17, wherein the domain matching loss optimization uses the maximum mean discrepancy loss. [19] Method according to claim 15, wherein the dimensionality of the features is reduced by t-distributed stochastic neighborhood embedding. [20] Method according to claim 15, wherein the single-class classifier is a support vector machine classifier.

Images