Industrial edge-side open set anomaly identification and physically-inspired active learning method

Abstract

The application provides an industrial edge side open set anomaly recognition and physical heuristic active learning method, statistical offset degree and physical equation residual mutation rate are calculated in parallel, when the statistical offset degree is lower than the threshold value and the physical residual mutation rate exceeds the dynamic stability boundary, the core physical item and the dimension reduction vector are extracted to generate a signature feature vector, so that any hidden anomaly conforming to the data statistical distribution but violating the objective physical law can be captured by the physical residual mutation, thereby effectively solving the high missing report rate problem of the pure data driven method to the hidden anomaly; meanwhile, historical anchor point samples are played back from the scenario memory bank and the elastic weight is calculated to consolidate the penalty term, the penalty term is used for reverse propagation to update the deep representation network parameters, so that the edge side applies damping constraint to the important parameters for maintaining the historical recognition ability when absorbing new class labels, thereby solving the catastrophic forgetting problem in the incremental update of the edge side.

Description

Technical Field

[0001] This invention relates to the field of industrial artificial intelligence and edge computing technology, and in particular to an active learning method for open set anomaly recognition and physical inspiration at the industrial edge. Background Technology

[0002] In industrial IoT edge computing scenarios, device status monitoring typically employs deep learning models to detect anomalies in multi-dimensional time-series data collected by sensors, thereby reducing cloud transmission bandwidth and response latency.

[0003] The closest existing technology to this invention is a purely data-driven open set identification method. This method deploys a pre-trained deep representation network (such as a one-dimensional convolutional neural network or autoencoder) at the edge. Its workflow is as follows: First, multi-dimensional time-series data from industrial equipment is collected in real-time at the edge, and normalized and preprocessed by sequence segmentation to obtain fixed-length time-series data segments. Then, these time-series data segments are input into the deep representation network, mapped to the latent space, and feature vectors are extracted. Next, the Mahalanobis distance or reconstruction error between this feature vector and the cluster centers of known normal operating conditions and known fault categories is calculated. When this distance or error exceeds a preset static threshold, the model determines the current data as an "unknown anomaly"; otherwise, it is determined to be a known category. This method relies solely on the statistical distribution characteristics of the data for open set identification and does not involve any constraints from physical laws.

[0004] However, the aforementioned existing technologies have the following drawbacks: First, the method relies entirely on the statistical manifold characteristics of the data. When a hidden anomaly (such as an early, weak thermal bending fault) occurs in the industrial field that conforms to the statistical distribution law but violates the objective physical laws, the statistical deviation is lower than the alarm threshold. The model will misjudge it as normal, resulting in serious underreporting. Second, when the cloud returns a new category label, if the model is fine-tuned directly on the edge using a small number of new samples, it will lead to a catastrophic forgetting of the ability to identify the original normal working conditions and known fault categories. On the other hand, full retraining requires high computing power and cannot achieve continuous autonomous evolution on resource-constrained edge devices. Summary of the Invention

[0005] This invention provides an active learning method for identifying open set anomalies at the industrial edge and based on physical principles, in order to address the shortcomings of existing purely data-driven methods, such as high false negative rates for hidden anomalies that conform to statistical distributions but violate physical laws, and catastrophic forgetting during incremental updates at the edge.

[0006] On one hand, this invention provides an active learning method for open set anomaly identification and physical inspiration at the industrial edge, comprising: Collect multi-dimensional time-series data of industrial equipment operation, and perform normalization and sequence segmentation preprocessing on the multi-dimensional time-series data to obtain preprocessed time-series data segments; The time series data segments are input into the deep representation network and the physical partial differential equation, respectively, and the statistical offset and the physical equation residual mutation rate are output. The statistical offset is obtained by calculating the Mahalanobis distance after mapping the time series data to the latent space, and the physical equation residual mutation rate is obtained by calculating the residual gradient between the theoretical derivation value and the actual observation value. If the statistical offset is lower than the alarm threshold and the residual mutation rate of the physical equation exceeds the dynamic stability boundary, the core physical terms are extracted from the calculation process of the physical partial differential equation, and the dimension reduction vector is extracted from the latent space of the deep representation network. The core physical term is concatenated with the dimensionality-reduced vector to generate a physical-statistical joint signature feature vector. The new category label is obtained by performing few-sample annotation on the physical-statistical joint signature feature vector; Historical anchor samples are replayed from the locally maintained episode memory bank. A penalty term for elastic weight consolidation is calculated based on the replayed historical anchor samples and the few samples corresponding to the new category label. The parameters of the deep representation network are updated through backpropagation of the penalty term.

[0007] Optionally, the multidimensional time-series data is normalized and preprocessed by sequence segmentation to obtain preprocessed time-series data segments, including: Z-score normalization is performed on each sensor channel data of the multidimensional time series data to make the mean of the sensor channel data 0 and the standard deviation 1, thus obtaining normalized data. The normalized data is divided into non-overlapping sequences according to the preset time window length and sliding step size to obtain multiple time series data segments of fixed length.

[0008] Optionally, the time-series data segment is input into a deep representation network, and the statistical offset is output, including: The time-series data segment is input into a one-dimensional convolutional neural network and mapped to the latent space to obtain the latent space feature vector; The statistical offset is obtained by calculating the Mahalanobis distance between the latent space feature vector and the cluster center of the known category in the latent space.

[0009] Optionally, the time-series data segment is input into the physical partial differential equation, and the residual mutation rate of the physical equation is output, including: Obtain the physical partial differential equations that are pre-constructed and hard-coded based on the physical laws of the target industrial equipment; Substitute the sensor observations in the time-series data segment into the physical partial differential equation to calculate the theoretically derived values. Calculate the residual between the theoretically derived value and the sensor observation value; The gradient of the residual is calculated in the time dimension to obtain the physical equation residual mutation rate output.

[0010] Optionally, core physical terms are extracted from the calculation process of the physical partial differential equations, and dimension reduction vectors are extracted from the latent space of the deep representation network, including: Calculate the contribution of each sub-term in the physical partial differential equation to the residual; One or more sub-items with the largest contribution are identified as core physical items; The latent space feature vectors are subjected to dimensionality reduction processing to obtain dimensionality-reduced vectors.

[0011] Optionally, the core physical term is concatenated with the reduced-dimensional vector to generate a physical-statistical joint signature feature vector, including: The core physical terms are encoded as a first feature vector, and the dimensionality-reduced vector is used as a second feature vector. Align the first feature vector and the second feature vector along the channel dimension and then concatenate them end to end in order to obtain the initial vector; The initial vector is subjected to L2 normalization to generate a physical-statistical joint signature feature vector.

[0012] Optionally, the new category label obtained by performing few-sample annotation on the physical-statistical joint signature feature vector includes: The core physical terms and latent space dimensionality reduction vector in the physical-statistical joint signature feature vector are analyzed through a few-sample manual annotation interface or a large-model automatic annotation interface. By combining the known anomaly categories in the initial knowledge base with the physical model, the anomaly corresponding to the physical statistical joint signature feature vector is determined to be a new anomaly type that has not been included in the database. Based on the new anomaly type, a new category label is generated, which includes the anomaly category name and core physical anomaly representation information.

[0013] Optionally, historical anchor samples are replayed from a locally maintained episodic memory bank. A penalty term for elastic weight consolidation is calculated based on the replayed historical anchor samples and the few samples corresponding to the new category label. The parameters of the deep representation network are then updated through backpropagation using this penalty term, including: Historical anchor point samples are randomly selected from the plot memory database or selected according to priority. The historical anchor point samples include feature anchor points with known normal operating conditions and feature anchor points with known fault categories. The historical anchor samples are merged with the few samples corresponding to the new category label into a training batch; Calculate the prediction loss of the deep representation network for historical anchor samples in the training batch under the current parameters, and the prediction loss for a small number of samples in the training batch; Calculate the diagonal approximation of the Fisher information matrix of the importance of each parameter in the deep representation network to the historical task; Construct a total loss function that includes the prediction loss and the elastic weight consolidation penalty term, wherein the elastic weight consolidation penalty term is expressed as a weighted sum of the importance of each parameter multiplied by the square of the difference between the current parameter and the initial parameter; The parameters of the deep representation network are updated by backpropagation using the total loss function.

[0014] Optionally, it also includes: The control edge intercepts the upload of the original time-series data segment and sends the physical statistics joint signature feature vector as a query packet to the cloud; After receiving the query packet, the cloud parses the core physical terms and dimensionality reduction vector in the physical-statistical joint signature feature vector, performs the few-sample annotation, obtains the new category label, and sends the new category label to the edge. After the control edge receives the new category label, it then executes the replay of historical anchor point samples from the locally maintained episode memory library.

[0015] Optionally, it also includes: Based on the statistical offset and the mutation rate of the physical equation residuals, a joint novelty score is obtained through adaptive weighted calculation. The formula for calculating the joint novelty score is as follows: ; in, For the joint novelty score, To calculate the offset, The mutation rate of the physical equation residuals. and To adaptively adjust the weights, the sum of the two is 1; Based on historical data of normal operating conditions within a sliding time window, the alarm threshold and the dynamic stability boundary are periodically updated. The alarm threshold is determined based on the upper boundary of the normal fluctuation of the statistical offset calculated by extreme value theory, and the dynamic stability boundary is determined based on a preset confidence interval of the historical normal fluctuation range of the residual mutation rate of the physical equation.

[0016] On the other hand, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the active learning method for industrial edge open set anomaly identification and physical heuristics as described above.

[0017] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the active learning method for industrial edge side open set anomaly identification and physical inspiration as described above.

[0018] On the other hand, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the active learning method for industrial edge open set anomaly identification and physical inspiration as described above.

[0019] The present invention provides an active learning method for open set anomaly identification and physical inspiration at the industrial edge. By parallelly calculating the statistical offset and the mutation rate of the physical equation residuals, and extracting core physical terms and dimensionality-reduced vectors to generate signature feature vectors when the statistical offset is below a threshold and the mutation rate of the physical residuals exceeds the dynamic stability boundary, this method ensures that any hidden anomalies that conform to the statistical distribution of data but violate objective physical laws can be captured by physical residual mutations. This effectively solves the problem of high false negative rates for hidden anomalies in pure data-driven methods. At the same time, by replaying historical anchor point samples from the plot memory bank and calculating elastic weights to consolidate the penalty term, and using this penalty term to backpropagate and update the parameters of the deep representation network, this method applies damping constraints to important parameters that maintain historical recognition capabilities when the edge side absorbs new category labels, thereby solving the catastrophic forgetting problem in incremental updates at the edge side. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in this 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 some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating the active learning method for open set anomaly identification and physical inspiration at the industrial edge provided in this embodiment of the invention. Figure 2 This is a schematic diagram of the structure of the electronic device provided in an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0023] Figure 1 This is a flowchart illustrating the active learning method for open set anomaly identification and physical inspiration at the industrial edge provided in this embodiment of the invention.

[0024] like Figure 1 As shown in the embodiment of the present invention, the active learning method for open set anomaly identification and physical inspiration at the industrial edge mainly includes the following steps: 101. Collect multi-dimensional time-series data of industrial equipment operation, normalize and sequence segmentation preprocess the multi-dimensional time-series data to obtain preprocessed time-series data segments.

[0025] Specifically, the multidimensional time-series data undergoes normalization and sequence segmentation preprocessing to obtain preprocessed time-series data fragments, including: Z-score normalization was performed on the data of each sensor channel in the multidimensional time series data to make the mean of the sensor channel data 0 and the standard deviation 1, thus obtaining normalized data. The normalized data is divided into non-overlapping sequences according to the preset time window length and sliding step size to obtain multiple time series data segments of fixed length.

[0026] The multidimensional time-series data consists of time-series sequences from N independent sensor channels. Each channel corresponds to an observable physical quantity, such as vibration acceleration, stator temperature, three-phase current, voltage, etc. Let the nth channel be the... The raw time-series data of each sensor channel is ;in, This represents the total number of sampling points for the time series data in this channel. , This represents the total number of sensor channels.

[0027] Z-score normalization is performed on the raw time-series data of each sensor channel to eliminate numerical magnitude bias caused by differences in physical dimensions and measurement ranges between different sensor channels. This avoids gradient anomalies and large numerical features dominating the model output during subsequent deep representation network training and inference, ensuring a balanced contribution of each channel's data to the model. The calculation formula for this process is as follows: ; in, For the first The average time-series data of each sensor channel within a preset historical normal operating condition benchmark dataset. This benchmark dataset consists of historical time-series data collected under steady-state operation of the equipment, without faults or anomalies. The calculation formula is: ; For the first The standard deviation of time-series data for each sensor channel within the same historical normal operating condition baseline dataset is calculated using the following formula:

[0028] in, This represents the total number of sampling points in the historical normal operating condition benchmark dataset. For the first In the benchmark dataset of the first channel The value of each sampling point For the first The first channel The normalized values ​​of each sampling point are processed to map the raw data of each sensor channel to normalized data with a mean of 0 and a standard deviation of 1. After completing the Z-score normalization of the entire channel, the resulting data has a dimension of [missing value]. The normalized time series data matrix.

[0029] Subsequently, a non-overlapping sequence segmentation operation is performed on the normalized time-series data matrix to decompose the long continuous time-series data into standardized segments that meet the input dimension requirements of subsequent deep representation networks and physical partial differential equations. This segmentation operation first pre-sets a fixed time window length L (unit: number of sampling points) and a sliding step size S based on the operating characteristics of the industrial equipment, the sensor sampling frequency, and the shortest duration of abnormal events. The sliding step size S is equal to the time window length L to achieve non-overlapping segmentation. The window length L must satisfy the following condition: if the sensor sampling frequency is f and the shortest duration of the abnormal event monitored by the target is t seconds, then L ≥ f × t, ensuring that a single time-series segment can cover complete abnormal feature information. Specifically, the segmentation process is as follows: starting from the first sampling point of the normalized time-series data matrix, continuous time-series segments of length L are sequentially truncated according to the step size S. The sampling point interval corresponding to the k-th time-series data segment is [(k-1) × S + 1, [k×S], where k is a positive integer. When the data is truncated to the end, if the number of remaining sampling points is less than the window length L, the tail data that is not long enough is directly discarded. Finally, a fixed-length time series data segment with multiple dimensions unified as N×L is obtained, which serves as the input for subsequent deep representation network and physical partial differential equation calculation.

[0030] 102. Input time series data segments into the deep representation network and the physical partial differential equation respectively, and output the statistical offset and the mutation rate of the physical equation residuals.

[0031] Among them, the statistical offset is obtained by mapping the time series data to the latent space and then calculating the Mahalanobis distance, and the physical equation residual mutation rate is obtained by calculating the residual gradient between the theoretical derivation value and the actual observation value.

[0032] Specifically, the input time-series data segment is fed into the deep representation network, and the output is a statistical offset, including: Time-series data segments are input into a one-dimensional convolutional neural network and mapped to the latent space to obtain latent space feature vectors; The statistical offset is obtained by calculating the Mahalanobis distance between the latent space feature vector and the cluster center of the known class in the latent space.

[0033] Dimension is The preprocessed time series data segments are input into a pre-trained one-dimensional convolutional neural network. The one-dimensional convolutional neural network consists of cascaded one-dimensional convolutional layers, global average pooling layers, and fully connected layers. Through multiple nonlinear transformations, the time series data is mapped to a preset d-dimensional latent space, and the output is a latent space feature vector of fixed dimension. .

[0034] Based on the device history labeled dataset, pre-calculate the cluster centers in the latent space for each known category. The inverse matrix of the latent space feature covariance matrix During the real-time inference phase, the distance between the current feature vector and the centers of all categories is calculated using the Mahalanobis distance formula, and the minimum value is taken as the statistical offset. The calculation formula is: ; ; By rapidly extracting deep temporal features through a one-dimensional convolutional neural network, it is suitable for low-computing scenarios on the edge. Mahalanobis distance can eliminate the correlation of feature dimensions and the influence of units, and accurately quantify the statistical distribution offset of samples from known categories.

[0035] Furthermore, inputting time-series data segments into physical partial differential equations outputs the residual mutation rate of the physical equations, including: Obtain the physical partial differential equations that are pre-constructed and hard-coded based on the physical laws of the target industrial equipment; Substituting sensor observations from time-series data segments into physical partial differential equations, theoretical values ​​are calculated. Calculate the residual between the theoretically derived values ​​and the sensor observations; The gradient of the residuals is calculated in the time dimension to obtain the residual mutation rate output of the physical equation.

[0036] Specifically, based on the physical laws of rotor dynamics, thermodynamics, or energy conservation of the target industrial equipment, physical partial differential equations for the adapted equipment are constructed and hard-coded into edge computing units as a benchmark for physical state determination. Preprocessed time-series data segments are substituted into these equations step-by-step to obtain the theoretically derived values ​​at each time point. The absolute residuals between the theoretically derived values ​​and the actual sensor observations are calculated point-by-point to generate a time-series residual sequence. The first-order difference gradient of this residual sequence is calculated in the time dimension, and the gradient magnitude is output as the residual mutation rate of the physical equation. The core calculation formula is: ; ; in, for The physical theory derivation of the time, for Time sensor observations For single-time residuals, This represents the residual mutation rate.

[0037] Constrained by objective physical laws, it can accurately identify hidden anomalies that have normal statistical characteristics but violate physical laws. The calculation process is lightweight and has no redundant overhead, making it suitable for real-time inference needs at the industrial edge.

[0038] 103. If the statistical offset is lower than the alarm threshold and the mutation rate of the physical equation residual exceeds the dynamic stability boundary, extract the core physical terms from the calculation process of the physical partial differential equation, and extract the dimension reduction vector from the latent space of the deep representation network.

[0039] Specifically, core physical terms are extracted from the calculation process of physical partial differential equations, and dimension reduction vectors are extracted from the latent space of deep representation networks, including: Calculate the contribution of each term in the physical partial differential equation to the residual; One or more sub-items with the largest contribution are identified as core physical items; The dimensionality of the latent space feature vectors is reduced to obtain the dimensionality-reduced vectors.

[0040] In the real-time solution and residual calculation of the physical partial differential equations, the contribution of each physical sub-term in the equation, such as the kinetic energy term, potential energy term, heat dissipation term, and physical coupling term, to the total residual is calculated step by step. The contribution is quantified by the proportion of the residual of each sub-term to the total residual.

[0041] One or more physical sub-items with the largest contribution are identified as the core physical items that trigger residual mutations, accurately locating the physical root cause of the anomaly. At the same time, principal component analysis is used to perform linear dimensionality reduction on the high-dimensional latent space feature vector output by the deep representation network, compressing the high-dimensional features to a preset low-dimensional dimension, resulting in a dimensionality-reduced vector that retains the core statistical features. This extraction method can accurately lock the key factors of physical anomalies, while achieving lightweight compression of statistical features. It ensures the effectiveness of features and significantly reduces the computing power and bandwidth overhead of subsequent feature splicing and transmission, making it suitable for resource-constrained operation scenarios at the industrial edge.

[0042] 104. Concatenate the core physical terms with the dimensionality-reduced vector to generate a physical-statistical joint signature feature vector.

[0043] Specifically, the core physics terms are concatenated with the reduced-dimensional vector to generate a physics-statistical joint signature feature vector, including: Encode the core physical terms as the first feature vector and use the dimensionality-reduced vector as the second feature vector; Align the first feature vector and the second feature vector along the channel dimension and then concatenate them end to end in order to obtain the initial vector; The initial vector is subjected to L2 normalization to generate a physical-statistical joint signature feature vector.

[0044] The extracted core physical terms are converted into a first feature vector through numerical quantization and fixed-dimensional encoding. The dimensionality-reduced latent space vector output by the deep representation network is directly used as the second feature vector. After aligning the first and second feature vectors by channel dimension, they are concatenated end-to-end in the order of physical features first, followed by statistical features, to obtain an initial concatenated vector. L2 normalization is then performed on the initial concatenated vector, using the following normalization formula: ; in, This is the initial concatenation vector. The L2 norm of the initial concatenated vector. This is the final generated physical-statistical joint signature feature vector; The above processing method can integrate physical anomaly root features and statistical representation features, eliminate amplitude differences between different feature dimensions, generate compact and non-redundant joint features, significantly reduce the computing power and bandwidth overhead of subsequent feature transmission and labeling, and adapt to the resource-constrained operation scenarios of industrial edge.

[0045] 105. New category labels obtained by labeling the physical and statistical joint signature feature vector with a small number of samples.

[0046] Specifically, the new category labels obtained after performing few-sample annotation on the physical-statistical joint signature feature vector include: The core physical terms and latent space dimensionality reduction vector in the physical-statistical joint signature feature vector are analyzed through a few-sample manual annotation interface or a large-model automatic annotation interface. By combining the known anomaly categories in the initial knowledge base with the physical model, the anomaly corresponding to the physical-statistical joint signature feature vector is determined to be a new anomaly type that has not been included in the database. Based on the new anomaly type, a new category label is generated, which includes the anomaly category name and core physical anomaly characterization information.

[0047] By using a few-sample manual annotation interface or an industrial large-scale model automatic annotation interface, the physical-statistical joint signature feature vector is parsed, separating and extracting the core physical terms representing the root cause of physical anomalies and the latent space dimensionality-reduced vector representing the statistical distribution of data. The extracted core physical terms and latent space dimensionality-reduced vector are then matched and verified against the known anomaly category features, standard physical models, and fault physical representation information pre-stored in the initial knowledge base. This determines that the anomaly corresponding to the joint signature feature vector is a new anomaly type not included in the knowledge base. Based on this new anomaly type, a unique anomaly category name is created, and the physical anomaly representation information corresponding to the core physical terms is integrated into it to generate a new category label containing the anomaly category name and core physical anomaly representation information. This annotation process relies solely on compact joint signature features, requires no original time-series data, has high annotation efficiency and low resource consumption, and can quickly provide accurate annotation basis for incremental updates of edge-side models.

[0048] 106. Replay historical anchor samples from the locally maintained episode memory bank, calculate the penalty term for elastic weight consolidation based on the replayed historical anchor samples and the few samples corresponding to the new category label, and update the parameters of the deep representation network through backpropagation of the penalty term.

[0049] Specifically, historical anchor samples are replayed from the locally maintained episodic memory bank. A penalty term for elastic weight consolidation is calculated based on the replayed historical anchor samples and the few samples corresponding to the new category labels. The parameters of the deep representation network are then updated through backpropagation of this penalty term, including: Historical anchor samples are randomly selected from the plot memory bank or selected according to priority. The historical anchor samples include feature anchors with known normal operating conditions and feature anchors with known fault categories. The historical anchor samples are merged with a small number of samples corresponding to the new category labels into a training batch; Calculate the prediction loss of the deep representation network for historical anchor samples in the training batch and the prediction loss for a small number of samples in the training batch under the current parameters. Calculate the diagonal approximation of the Fisher information matrix for the importance of each parameter in the deep representation network to the historical task; Construct a total loss function that includes prediction loss and elastic weight consolidation penalty term. The elastic weight consolidation penalty term is expressed as a weighted sum of the importance of each parameter multiplied by the square of the difference between the current parameter and the initial parameter. Backpropagation is performed using the total loss function to update the parameters of the deep representation network.

[0050] In this process, historical anchor samples are randomly selected from the locally maintained episode memory database or selected according to the feature representativeness priority. These anchor samples include known normal operating condition feature anchors and known fault category feature anchors. The selected historical anchor samples are then merged with the small sample data corresponding to the new category labels to construct a unified training batch.

[0051] Based on the current deep representation network parameters, the classification prediction loss for historical anchor samples within the training batch and the classification prediction loss for a small number of new class samples are calculated respectively. The diagonal approximation of the Fisher information matrix of the importance of each parameter of the deep representation network to the historical classification task is calculated based on historical task data, and these approximate values ​​are used as the importance weights of each parameter. A total loss function is constructed that integrates the classification prediction loss and the elastic weight consolidation penalty term. The total loss function is: ; in, To predict loss for classification, The penalty coefficient is... For the first The diagonal approximation of Fisher information for each parameter. For the current parameter, The parameters are the historical best values. Based on the total loss function, the backpropagation algorithm is used to iteratively update the parameters of the deep representation network. While learning new anomaly categories, the historical key parameters are constrained to prevent large shifts, avoiding catastrophic forgetting of the model, and realizing lossless incremental updates and anomaly category expansion of the edge-side model.

[0052] In some embodiments, the active learning method for industrial edge-side open set anomaly identification and physical heuristics provided in this invention further includes: The control edge intercepts the upload of raw time-series data segments and sends the physical statistical joint signature feature vector as a query packet to the cloud; After receiving the query packet in the cloud, the core physical terms and dimensionality reduction vector in the physical-statistical joint signature feature vector are parsed, few-sample labeling is performed to obtain new category labels, and the new category labels are sent to the edge. After receiving the new category label, the control edge then replays historical anchor samples from the locally maintained episode memory.

[0053] Specifically, when the edge determines that a time-series data segment is an unknown open set anomaly, the control local data transmission module intercepts the uplink upload operation of the original multi-dimensional time-series data corresponding to that segment. Only the pre-processed physical statistical joint signature feature vector is encapsulated into a lightweight interrogation packet and transmitted via industrial wireless LAN or 5G. The communication link sends the data to the cloud-based expert assistance system. After receiving the question packet, the cloud uses a high-value question parsing module to extract the core physical terms and latent space dimensionality reduction vectors. It then uses a small-sample manual annotation interface or an industrial large-scale model automatic annotation interface to determine the new anomaly type and generate a new category label containing the anomaly category name and core physical anomaly representation information. This new category label is then sent to the edge device via the communication link. After receiving the new category label from the cloud and caching it locally, the edge device starts the continuous learning module to perform subsequent operations such as replaying historical anchor point samples from the local plot memory. This process only transmits a compact joint signature feature packet, eliminating the need to upload massive amounts of original time-series data. This significantly reduces the communication bandwidth consumption and transmission latency on the edge side. The cloud annotation is highly efficient and consumes little computing power, providing accurate annotation basis for incremental updates of the edge device model and ensuring the efficient and stable operation of edge-cloud collaborative anomaly identification and active learning.

[0054] In some embodiments, the active learning method for industrial edge-side open set anomaly identification and physical heuristics provided in this invention further includes: Based on the statistical bias and the mutation rate of the physical equation residuals, a joint novelty score is obtained through adaptive weighted calculation. The formula for calculating the joint novelty score is: ; in, For the joint novelty score, To calculate the offset, The mutation rate of the physical equation residuals. and To adaptively adjust the weights, the sum of the two is 1; Based on historical data of normal operating conditions within a sliding time window, the alarm threshold and dynamic stability boundary are periodically updated.

[0055] Specifically, the statistical offset corresponding to the current time series data segment is obtained in real time. Mutation rate of residuals from physical equations The weights are adaptively allocated and adjusted based on the real-time operating conditions of the target industrial equipment. and ,make sure and The sum of their values ​​is always 1, according to the calculation formula. The weighted fusion was completed to obtain the joint novelty score. Among them, the increase is due to the equipment operating in a steady state. The weights are increased to reinforce the constraints of physical laws during dynamic processes of operating condition switching. The weighting method reduces false alarms caused by fluctuations in operating conditions. Simultaneously, it collects continuous historical data of normal operating conditions within a fixed-length sliding time window, and updates the alarm threshold and dynamic stability boundary according to a preset period. The alarm threshold is fitted to the normal fluctuation range of statistical deviation using extreme value theory, calculating and determining its upper boundary of normal fluctuation. The dynamic stability boundary is determined based on historical normal fluctuation data of the residual mutation rate of the physical equation, selecting the upper limit of a preset confidence interval. This adaptive weighting method balances the detection weights of statistical characteristics and physical laws, adapting to the complex and ever-changing operating conditions in industrial settings. Periodic dynamic threshold updates eliminate the impact of data drift and gradual changes in operating conditions caused by long-term equipment operation, effectively improving the accuracy and stability of open set anomaly judgment and avoiding false alarms and missed alarms caused by static thresholds.

[0056] Among them, the alarm threshold is determined based on the upper boundary of the normal fluctuation of the statistical offset calculated by the extreme value theory, and the dynamic stability boundary is determined based on the preset confidence interval of the historical normal fluctuation range of the residual mutation rate of the physical equation.

[0057] Figure 2 This is a schematic diagram of the structure of the electronic device provided in an embodiment of the present invention.

[0058] like Figure 2 As shown, the electronic device may include a processor 210, a communications interface 220, a memory 230, and a communication bus 240. The processor 210, communications interface 220, and memory 230 communicate with each other via the communication bus 240. The processor 210 can call logic instructions from the memory 230 to execute an active learning method for open-set anomaly recognition and physical heuristics at the industrial edge.

[0059] Furthermore, the logical instructions in the aforementioned memory 230 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0060] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the active learning method for industrial edge open set anomaly recognition and physical inspiration provided by the above methods.

[0061] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the active learning method for open set anomaly identification and physical inspiration at the industrial edge provided by the methods described above.

[0062] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0063] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0064] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An active learning method for anomaly identification and physical inspiration in open sets at the industrial edge, characterized in that, include: Collect multi-dimensional time-series data of industrial equipment operation, and perform normalization and sequence segmentation preprocessing on the multi-dimensional time-series data to obtain preprocessed time-series data segments; The time series data segments are input into the deep representation network and the physical partial differential equation, respectively, and the statistical offset and the physical equation residual mutation rate are output. The statistical offset is obtained by calculating the Mahalanobis distance after mapping the time series data to the latent space, and the physical equation residual mutation rate is obtained by calculating the residual gradient between the theoretical derivation value and the actual observation value. If the statistical offset is lower than the alarm threshold and the residual mutation rate of the physical equation exceeds the dynamic stability boundary, the core physical terms are extracted from the calculation process of the physical partial differential equation, and the dimension reduction vector is extracted from the latent space of the deep representation network. The core physical term is concatenated with the dimensionality-reduced vector to generate a physical-statistical joint signature feature vector. The new category label is obtained by performing few-sample annotation on the physical-statistical joint signature feature vector; Historical anchor samples are replayed from the locally maintained episode memory bank. A penalty term for elastic weight consolidation is calculated based on the replayed historical anchor samples and the few samples corresponding to the new category label. The parameters of the deep representation network are updated through backpropagation of the penalty term.

2. The active learning method for open set anomaly identification and physical inspiration at the industrial edge as described in claim 1, characterized in that, The multidimensional time-series data is normalized and preprocessed by sequence segmentation to obtain preprocessed time-series data segments, including: Z-score normalization is performed on each sensor channel data of the multidimensional time series data to make the mean of the sensor channel data 0 and the standard deviation 1, thus obtaining normalized data. The normalized data is divided into non-overlapping sequences according to the preset time window length and sliding step size to obtain multiple time series data segments of fixed length.

3. The active learning method for open set anomaly identification and physical inspiration at the industrial edge as described in claim 1, characterized in that, Input the time-series data segment into the deep representation network, and output statistical offsets, including: The time-series data segment is input into a one-dimensional convolutional neural network and mapped to the latent space to obtain the latent space feature vector; The statistical offset is obtained by calculating the Mahalanobis distance between the latent space feature vector and the cluster center of the known category in the latent space.

4. The active learning method for open set anomaly identification and physical inspiration at the industrial edge as described in claim 1, characterized in that, Input the time-series data segment into the physical partial differential equation, and output the residual mutation rate of the physical equation, including: Obtain the physical partial differential equations that are pre-constructed and hard-coded based on the physical laws of the target industrial equipment; Substitute the sensor observations in the time-series data segment into the physical partial differential equation to calculate the theoretically derived values. Calculate the residual between the theoretically derived value and the sensor observation value; The gradient of the residual is calculated in the time dimension to obtain the physical equation residual mutation rate output.

5. The active learning method for open set anomaly identification and physical inspiration at the industrial edge as described in claim 3, characterized in that, The core physical terms are extracted from the calculation process of the physical partial differential equations, and the dimension reduction vectors are extracted from the latent space of the deep representation network, including: Calculate the contribution of each sub-term in the physical partial differential equation to the residual; One or more sub-items with the largest contribution are identified as core physical items; The latent space feature vectors are subjected to dimensionality reduction processing to obtain dimensionality-reduced vectors.

6. The active learning method for open set anomaly identification and physical inspiration at the industrial edge as described in claim 1, characterized in that, The core physical term is concatenated with the reduced-dimensional vector to generate a physical-statistical joint signature feature vector, including: The core physical terms are encoded as a first feature vector, and the dimensionality-reduced vector is used as a second feature vector. Align the first feature vector and the second feature vector along the channel dimension and then concatenate them end to end in order to obtain the initial vector; The initial vector is subjected to L2 normalization to generate a physical-statistical joint signature feature vector.

7. The active learning method for open set anomaly identification and physical inspiration at the industrial edge as described in claim 1, characterized in that, The new category labels obtained by performing few-sample annotation on the physical-statistical joint signature feature vector include: The core physical terms and latent space dimensionality reduction vector in the physical-statistical joint signature feature vector are analyzed through a few-sample manual annotation interface or a large-model automatic annotation interface. By combining the known anomaly categories in the initial knowledge base with the physical model, the anomaly corresponding to the physical statistical joint signature feature vector is determined to be a new anomaly type that has not been included in the database. Based on the new anomaly type, a new category label is generated, which includes the anomaly category name and core physical anomaly representation information.

8. The active learning method for open set anomaly identification and physical inspiration at the industrial edge as described in claim 1, characterized in that, Replaying historical anchor samples from a locally maintained episodic memory, calculating a penalty term for elastic weight consolidation based on the replayed historical anchor samples and the few samples corresponding to the new category label, and updating the parameters of the deep representation network through backpropagation of the penalty term, including: Historical anchor point samples are randomly selected from the plot memory database or selected according to priority. The historical anchor point samples include feature anchor points with known normal operating conditions and feature anchor points with known fault categories. The historical anchor samples are merged with the few samples corresponding to the new category label into a training batch; Calculate the prediction loss of the deep representation network for historical anchor samples in the training batch under the current parameters, and the prediction loss for a small number of samples in the training batch; Calculate the diagonal approximation of the Fisher information matrix of the importance of each parameter in the deep representation network to the historical task; Construct a total loss function that includes the prediction loss and the elastic weight consolidation penalty term, wherein the elastic weight consolidation penalty term is expressed as a weighted sum of the importance of each parameter multiplied by the square of the difference between the current parameter and the initial parameter; The parameters of the deep representation network are updated by backpropagation using the total loss function.

9. The active learning method for open set anomaly identification and physical inspiration at the industrial edge as described in claim 1, characterized in that, Also includes: The control edge intercepts the upload of the time-series data segment and sends the physical statistics joint signature feature vector as a query packet to the cloud; After receiving the query packet, the cloud parses the core physical terms and dimensionality reduction vector in the physical-statistical joint signature feature vector, performs the few-sample annotation, obtains the new category label, and sends the new category label to the edge. After the control edge receives the new category label, it then executes the replay of historical anchor point samples from the locally maintained episode memory library.

10. The active learning method for open set anomaly identification and physical inspiration at the industrial edge as described in claim 1, characterized in that, Also includes: Based on the statistical offset and the mutation rate of the physical equation residuals, a joint novelty score is obtained through adaptive weighted calculation. The formula for calculating the joint novelty score is as follows: ； in, For the joint novelty score, To calculate the offset, The mutation rate of the physical equation residuals. and To adaptively adjust the weights, the sum of the two is 1; Based on historical data of normal operating conditions within a sliding time window, the alarm threshold and the dynamic stability boundary are periodically updated. The alarm threshold is determined based on the upper boundary of the normal fluctuation of the statistical offset calculated by extreme value theory, and the dynamic stability boundary is determined based on a preset confidence interval of the historical normal fluctuation range of the residual mutation rate of the physical equation.

Images