Multi-mode industrial process anomaly detection method
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-23
- Publication Date
- 2026-04-10
AI Technical Summary
Traditional anomaly detection algorithms cannot effectively cope with the time-varying and nonlinear characteristics of multimodal industrial processes, resulting in high false alarm and false negative rates.
The KFCM-VMD-SVDD algorithm process is adopted. KFCM extracts high-dimensional features for modality segmentation, SVDD is used to detect anomalies in the corresponding modal data, including offline training and online testing phases, and VMD is used to decompose and reconstruct the data and construct a hypersphere to determine the control limits.
It improves the accuracy of anomaly detection in multimodal industrial processes, effectively handles nonlinear problems, reduces false alarm and false negative rates, and is suitable for anomaly diagnosis in ship cooling water systems.
Smart Images
Figure CN121834599A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of state control technology, specifically to a multimodal industrial process anomaly detection method. Background Technology
[0002] As industrial systems continue to develop towards informatization, attack methods are becoming increasingly covert and complex. Anomalies cannot be detected solely at the network layer of an industrial system. However, attacks originating from the network layer can always penetrate to the physical layer, threatening the security of control systems and the personal safety of operators. Therefore, in order to detect this new form of attack, it is necessary to monitor the system status at the physical layer.
[0003] However, physical-layer industrial processes, such as ship cooling water systems, are shifting from single-modal to multi-modal due to changes in operating modes, equipment aging, and the increasing complexity of industrial processes. This poses a challenge to traditional anomaly detection methods for single-modal industrial processes. Furthermore, multi-modal industrial processes exhibit time-varying and nonlinear characteristics, making modeling more complex. Traditional anomaly detection algorithms cannot handle these complex data features, resulting in high false alarm and false negative rates. Therefore, it is necessary to propose an anomaly detection algorithm capable of diagnosing multi-modal industrial processes, based on research into traditional anomaly detection algorithms. Summary of the Invention
[0004] To address the shortcomings of existing technologies, this invention provides a multimodal industrial process anomaly detection method, which solves the problem that traditional anomaly detection algorithms cannot cope with complex data features such as time-varying and nonlinear data, resulting in high false alarm and false negative rates.
[0005] To achieve the above objectives, the present invention is implemented through the following technical solution: a multimodal industrial process anomaly detection method includes a KFCM-VMD-SVDD algorithm flow, which first uses KFCM to extract high-dimensional features for modality division, and then uses SVDD to detect anomalies in the data of the corresponding modality. The KFCM-VMD-SVDD algorithm flow includes an offline training stage and an online testing stage.
[0006] The offline training phase includes the following steps:
[0007] LS1 and KFCM first project the raw data into a high-dimensional space and cluster it into different modalities, while retaining the cluster centers;
[0008] LS2. Use VMD to decompose and reconstruct the data for each modality;
[0009] LS3. Using the SVDD algorithm, a minimum hypersphere is constructed to enclose most of the normal data.
[0010] LS4. Calculate and determine the control limit R;
[0011] The online testing phase includes the following steps:
[0012] ZS1, Calculate x new The membership degree to each cluster center is used to determine which cluster each point belongs to;
[0013] ZS2: Use VMD to decompose and reconstruct the data for each modality;
[0014] ZS3, Calculate the distance D from each data point in the feature space to the center of the hypersphere;
[0015] ZS4. Compare D and R to determine if it is abnormal.
[0016] Preferably, in the offline training phase, step LS1 generates normal data of different modalities through the TE simulation platform, producing a mixed dataset X∈R. m*n Mixed dataset X∈R m*n The kernel function φ maps to a high-dimensional feature space, which is then used for FCM clustering, and the cluster center positions C of the final result are preserved. i .
[0017] Preferably, the newly generated data x in the ZS1 step of the online testing phase new ∈R 1*m The kernel function φ maps the data to a high-dimensional space, and the cluster center positions C are determined during the training phase. i Determine the new data x new Which mode does it belong to?
[0018] Preferably, in the online testing phase ZS4 step, the control limit R determined in the offline training phase is used to determine whether the new data x is abnormal.
[0019] Preferably, the input to the KFCM algorithm is a mixed dataset X = [x1, x2, ..., xn] ∈ R of various modalities. m*n Where m represents the number of variables to be detected, and n represents the number of samples, assuming U∈R c*n Where c represents the number of cluster centers, and the constraint is expressed as: Because of the non-linear characteristics of the dataset, FCM, which incorporates the kernel method, is adopted. The kernel method can project the original data x into a high-dimensional space through a non-linear kernel function φ. Since the original data x is non-linear and does not have an ellipsoidal shape in the original space, and FCM can generally only cluster ellipsoidal data, the data becomes ellipsoidal in the high-dimensional space, which allows for good FCM clustering.
[0020] Preferably, the TE model is an industrial simulation model based on industrial field data. It can obtain various fault data and normal data through a large number of simulations. This model is widely used in the field of industrial process anomaly detection, and the efficiency of many models has also been verified. The KFCM-VMD-SVDD algorithm proposed in this invention is applied to the data generated by the TE model, and this method has been verified to have strong adaptability to this model.
[0021] Preferably, the algorithm formula for VMD is as follows:
[0022]
[0023] Constraints:
[0024] in:
[0025] uk(t) is the kth model component;
[0026] wk is the center frequency of the kth modal component;
[0027] δ(t) is the Dirac function;
[0028] * indicates convolution;
[0029] f(t) is the original signal.
[0030] Preferably, the VMD algorithm can decompose complex multimodal data into multiple modal components with different frequency characteristics to reveal the intrinsic features of the data.
[0031] Preferably, the algorithm formula for SVDD is as follows:
[0032]
[0033] in:
[0034] R is the radius of the hypersphere;
[0035] α is the center of the hypersphere;
[0036] ξ is a relaxation factor used to allow some samples to be outside the hypersphere;
[0037] C is a penalty parameter that balances the hypersphere volume and the error rate.
[0038] Preferably, the SVDD algorithm aims to find a hypersphere that contains all or almost all target samples and has the smallest volume. This hypersphere is determined by the support vectors, i.e., the data points located on the boundary of the hypersphere. Through nonlinear mapping, the original training samples are mapped to a high-dimensional feature space to facilitate spherical description.
[0039] This invention discloses a multimodal industrial process anomaly detection method, which has the following beneficial effects:
[0040] 1. This multimodal industrial process anomaly detection method proposes a multimodal anomaly detection algorithm based on KFCM-VMD-SVDD. KFCM extracts high-dimensional features for modality segmentation, and SVDD is used to detect anomalies in the corresponding modalities. In the offline training phase, KFCM first projects the original data into a high-dimensional space and clusters it into different modalities. Then, VMD decomposes and reconstructs the data for each modality. Finally, the SVDD algorithm constructs a minimal hypersphere to enclose most of the normal data, thus determining the control limits. In the online testing phase, the KFCM algorithm divides each test data into corresponding modalities and uses the corresponding control limits to determine whether it is an anomaly. This method can handle the nonlinearity of multimodal industrial data and has a higher anomaly detection accuracy compared to traditional anomaly detection algorithms. This method is applied to the cooling water system of ships to diagnose relevant anomaly patterns.
[0041] 2. This multimodal industrial process anomaly detection method, through the VMD algorithm, can decompose complex multimodal data into multiple modal components with different frequency characteristics, thereby better revealing the intrinsic characteristics of the data and providing a more accurate data basis for subsequent anomaly detection.
[0042] 3. This multimodal industrial process anomaly detection method, considering the nonlinear characteristics of the dataset, adopts FCM that integrates kernel methods. Kernel methods can map and project the original data x to a high-dimensional space through a nonlinear kernel function φ. Since the original data x is nonlinear and does not have an ellipsoidal shape in the original space, and FCM can generally only cluster ellipsoidal data, the data becomes ellipsoidal in the high-dimensional space, which can be well clustered by FCM. Attached Figure Description
[0043] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0044] Figure 1 This is a flowchart of the offline training and online testing phases of the algorithm of this invention;
[0045] Figure 2 This is a process flow diagram of the TE process of the present invention. Detailed Implementation
[0046] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are described clearly and completely. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0047] This invention discloses a multimodal industrial process anomaly detection method.
[0048] Example 1: According to the appendix Figure 1-2 As shown, a multimodal industrial process anomaly detection method includes the KFCM-VMD-SVDD algorithm flow. First, KFCM is used to extract high-dimensional features for modality division, and then SVDD is used to detect anomalies in the corresponding modal data. The KFCM-VMD-SVDD algorithm flow includes an offline training stage and an online testing stage.
[0049] The offline training phase includes the following steps:
[0050] LS1 and KFCM first project the raw data into a high-dimensional space and cluster it into different modalities, while retaining the cluster centers;
[0051] LS2. Use VMD to decompose and reconstruct the data for each modality. Through the VMD algorithm, complex multimodal data can be decomposed into multiple modal components with different frequency characteristics, thereby better revealing the intrinsic characteristics of the data and providing a more accurate data foundation for subsequent anomaly detection.
[0052] LS3 utilizes the SVDD algorithm to reconstruct a minimal hypersphere that encloses most of the normal data. The SVDD algorithm aims to find a hypersphere with the smallest volume that contains all or almost all target samples. This hypersphere is determined by the support vectors, i.e., the data points located on the hypersphere's boundary. Through nonlinear mapping, the original training samples are mapped to a high-dimensional feature space. In a high-dimensional space, data is more easily described spherically, thus more accurately defining the range of normal data.
[0053] LS4. Calculate and determine the control limit R; The control limit R is an important threshold for judging whether the data is abnormal. It integrates the analysis and processing results of the data characteristics in the previous steps and provides a key judgment standard for subsequent online testing.
[0054] The online testing phase includes the following steps:
[0055] ZS1, Calculate x new The membership degree to each cluster center determines which cluster each point belongs to; when new data x is generated new ∈R1*m Upon arrival, it is first mapped to a high-dimensional space using a kernel function φ, and then the cluster center positions C determined during the training phase are used. i Calculate the new data x new The membership degree to each cluster center is used to accurately determine the membership degree of the new data x. new Which modality does it belong to? This step ensures that new data can be correctly classified into the corresponding modality, providing accurate modal information for subsequent anomaly detection.
[0056] ZS2: Use VMD to decompose and reconstruct the data for each modality;
[0057] ZS3. Calculate the distance D from each data point in the feature space to the center of the hypersphere. This distance D reflects the degree of deviation of the data point from the normal data range and is an important indicator for judging whether the data is abnormal. By accurately calculating the distance D, abnormal changes in the data can be detected in a timely manner.
[0058] ZS4. Compare D and R to determine anomalies. New data x is evaluated by comparing the distance D with the control limit R determined during offline training. new Is it abnormal? If D>R, it means that the data point deviates from the normal range and is judged as abnormal data; otherwise, it is judged as normal data. This judgment method is accurate and reliable, and can effectively detect abnormal situations in multimodal industrial processes.
[0059] Furthermore, in the offline training phase, step LS1 generates normal data of different modalities through the TE simulation platform, producing a mixed dataset X∈R. m*n Mixed dataset X∈R m*n The kernel function φ maps to a high-dimensional feature space, which is then used for FCM clustering, and the cluster center positions C of the final result are preserved. i .
[0060] Furthermore, the newly generated data x in the ZS1 step of the online testing phase new ∈R 1*m The kernel function φ maps the data to a high-dimensional space, and the cluster center positions C are determined during the training phase. i Determine the new data x new Which mode does it belong to?
[0061] In the online testing phase (ZS4 step), the new data x is determined using the control limits R determined in the offline training phase. new Is it abnormal?
[0062] Furthermore, the input to the KFCM algorithm is a mixed dataset X = [x1, x2, ..., xn] ∈ R of various modalities. m*nWhere m represents the number of variables to be detected, and n represents the number of samples, assuming U∈R c*n Where c represents the number of cluster centers, and the constraint is expressed as: Because of the non-linear characteristics of the dataset, FCM, which incorporates the kernel method, is adopted. The kernel method can project the original data x into a high-dimensional space through a non-linear kernel function φ. Since the original data x is non-linear and does not have an ellipsoidal shape in the original space, and FCM can generally only cluster ellipsoidal data, the data becomes ellipsoidal in the high-dimensional space, which allows for good FCM clustering.
[0063] Furthermore, the algorithm formula for VMD is as follows:
[0064]
[0065] Constraints:
[0066] in:
[0067] uk(t) is the kth model component;
[0068] wk is the center frequency of the kth modal component;
[0069] δ(t) is the Dirac function;
[0070] * indicates convolution;
[0071] f(t) is the original signal. The VMD algorithm can decompose complex multimodal data into multiple modal components with different frequency characteristics, thereby better revealing the intrinsic characteristics of the data and providing a more accurate data basis for subsequent anomaly detection.
[0072] Furthermore, the algorithm formula for SVDD is as follows:
[0073]
[0074] in:
[0075] R is the radius of the hypersphere;
[0076] α is the center of the hypersphere;
[0077] ξ is a relaxation factor used to allow some samples to be outside the hypersphere;
[0078] C is a penalty parameter that balances the hypersphere volume and the misclassification rate. The SVDD algorithm aims to find a hypersphere that contains all or almost all target samples and has the smallest volume. This hypersphere is determined by the support vectors, i.e., the data points located on the hypersphere's boundary. A nonlinear mapping is used to map the original training samples to a high-dimensional feature space, making it easier to describe the spherical shape.
[0079] This paper proposes a multimodal anomaly detection algorithm based on KFCM-VMD-SVDD. This method can handle the nonlinearity of multimodal industrial data and has a higher anomaly detection accuracy compared with traditional anomaly detection algorithms. It solves the problem that traditional anomaly detection algorithms cannot cope with complex data features such as time-varying and nonlinearity, resulting in high false alarm and false negative rates. This method can be applied to the cooling water system of ships to diagnose relevant anomaly patterns.
[0080] Example 2: According to the appendix Figure 1-2 As shown, a multimodal industrial process anomaly detection method includes the KFCM-VMD-SVDD algorithm flow. First, KFCM is used to extract high-dimensional features for modality division, and then SVDD is used to detect anomalies in the corresponding modal data. The KFCM-VMD-SVDD algorithm flow includes an offline training stage and an online testing stage.
[0081] The offline training phase includes the following steps:
[0082] LS1 and KFCM first project the raw data into a high-dimensional space and cluster it into different modalities, while retaining the cluster centers;
[0083] LS2. Use VMD to decompose and reconstruct the data for each modality;
[0084] LS3. Using the SVDD algorithm, a minimum hypersphere is constructed to enclose most of the normal data.
[0085] LS4. Calculate and determine the control limit R;
[0086] The online testing phase includes the following steps:
[0087] ZS1, Calculate x new The membership degree to each cluster center is used to determine which cluster each point belongs to;
[0088] ZS2: Use VMD to decompose and reconstruct the data for each modality;
[0089] ZS3, Calculate the distance D from each data point in the feature space to the center of the hypersphere;
[0090] ZS4. Compare D and R to determine if it is abnormal.
[0091] Furthermore, the TE model is an industrial simulation model based on industrial field data, capable of obtaining various fault and normal data through extensive simulations. This model is widely used in the field of industrial process anomaly detection, and its efficiency has been verified in many cases. The KFCM-VMD-SVDD algorithm proposed in this invention is applied to the data generated by the TE model, and this method has been verified to have strong adaptability to this model.
[0092] Working principle: First, KFCM is used to extract high-dimensional features for modality segmentation. Then, SVDD is used to detect anomalies in the data of the corresponding modalities. In the offline training phase, KFCM first projects the original data into a high-dimensional space and performs clustering to divide it into different modalities. Then, VMD is used to decompose and reconstruct the data of each modality. Finally, the SVDD algorithm is used to construct a minimum hypersphere to enclose most of the normal data to determine the control limits. In the online testing phase, the KFCM algorithm divides each test data into the corresponding modality and uses the corresponding control limits to judge whether it is abnormal.
[0093] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.
Claims
1. A multimodal industrial process anomaly detection method, characterized in that, The algorithm process includes KFCM-VMD-SVDD, which first uses KFCM to extract high-dimensional features for modality segmentation, and then uses SVDD to perform anomaly detection on the data of the corresponding modality. The KFCM-VMD-SVDD algorithm process includes an offline training phase and an online testing phase. The offline training phase includes the following steps: LS1 and KFCM first project the raw data into a high-dimensional space and cluster it into different modalities, while retaining the cluster centers; LS2. Use VMD to decompose and reconstruct the data for each modality; LS3. Using the SVDD algorithm, a hypersphere is reconstructed to enclose the normal data. LS4. Calculate and determine the control limit R; The online testing phase includes the following steps: ZS1, Calculate x new The membership degree to each cluster center is used to determine which specific cluster each point belongs to; ZS2: Use VMD to decompose and reconstruct the data for each modality; ZS3, Calculate the distance D from each data point in the feature space to the center of the hypersphere; ZS4. Compare D and R to determine if it is abnormal.
2. The multimodal industrial process anomaly detection method according to claim 1, characterized in that, In the offline training phase, step LS1 generates normal data of different modalities through the TE simulation platform, producing a mixed dataset X∈R. m*n Mixed dataset X∈R m*n The kernel function φ maps to a high-dimensional feature space, which is then... Perform FCM clustering and retain the cluster center positions C of the final result. i .
3. The multimodal industrial process anomaly detection method according to claim 1, characterized in that, New data x generated during the ZS1 step of the online testing phase new ∈R 1*m The kernel function φ maps the data to a high-dimensional space, and the cluster center positions C are determined during the training phase. i Determine the new data x new Which mode does it belong to? 4. The multimodal industrial process anomaly detection method according to claim 1, characterized in that, In the online testing phase ZS4 step, the control limit R determined in the offline training phase is used to determine whether the new data x is an anomaly.
5. The multimodal industrial process anomaly detection method according to claim 1, characterized in that, The input to the KFCM algorithm is a mixed dataset of various modalities, X = [x1, x2, ..., xn] ∈ R. m*n Where m represents the number of variables to be detected, and n represents the number of samples, assuming U∈R c*n Where c represents the number of cluster centers, and the constraint is expressed as:
6. The multimodal industrial process anomaly detection method according to claim 2, characterized in that, The TE model is an industrial simulation model based on industrial field data, which can obtain various fault data and normal data through multiple simulations.
7. The multimodal industrial process anomaly detection method according to claim 1, characterized in that, The algorithm formula for VMD is as follows: Constraints: in: uk(t) is the kth model component; wk is the center frequency of the kth modal component; δ(t) is the Dirac function; * indicates convolution; f(t) is the original signal.
8. The multimodal industrial process anomaly detection method according to claim 7, characterized in that, The VMD algorithm can decompose complex multimodal data into multiple modal components with different frequency characteristics to reveal the intrinsic features of the data.
9. The multimodal industrial process anomaly detection method according to claim 1, characterized in that, The algorithm formula for SVDD is as follows: in: R is the radius of the hypersphere; α is the center of the hypersphere; ξ is a relaxation factor used to allow some samples to be outside the hypersphere; C is a penalty parameter that balances the hypersphere volume and the error rate.
10. The multimodal industrial process anomaly detection method according to claim 9, characterized in that, The SVDD algorithm maps the original training samples to a high-dimensional feature space through nonlinear mapping.