Servo drive abnormality detection system based on multi-sensor fusion

Abstract

The application discloses a servo drive abnormality detection system based on multi-sensor fusion, comprising: a data processing module for collecting and preprocessing intelligent sensor data; a network construction module for constructing a graph neural network and generating node features and edge weights; a message passing module for multi-round message passing to update node embedding to reflect local and global states; an attention fusion module for multi-head self-attention fusion of cross-node and time sequence features to form a global state; an abnormality detection module for adaptively identifying abnormal states and outputting types, probabilities and development trends; an abnormality marking module for classifying and marking abnormalities and generating early warning and evolution reports; and a data distribution module for distributing abnormal data to a control unit and a monitoring terminal to realize distributed monitoring. The application realizes servo drive abnormality identification and early warning through multi-sensor fusion and a graph neural network, and improves precision and monitoring efficiency.

Description

Technical Field

[0001] This invention relates to the field of servo drive control technology, and in particular to a servo drive anomaly detection system based on multi-sensor fusion. Background Technology

[0002] In modern industrial automation, servo drive systems are widely used in metal cutting, welding, electronic component manufacturing, and other high-precision processing equipment. Their operating status directly affects production efficiency and product quality. Traditional servo drive anomaly detection methods often rely on single sensor signals or simple threshold judgments, such as monitoring only current, voltage, or position feedback, and determining whether the equipment is abnormal by setting a fixed threshold. Although these methods are simple to implement, they are easily affected by load changes, environmental interference, and multi-node coupling under complex operating conditions, resulting in low anomaly detection sensitivity, high false alarm rate, and serious false negative rate, making it impossible to identify potential servo drive faults at an early stage.

[0003] To improve the reliability of anomaly detection, some studies in recent years have introduced multi-sensor data fusion technology, which comprehensively analyzes the operating status of equipment by simultaneously collecting multi-dimensional signals such as current, rotational speed, vibration, and temperature. However, most existing methods use simple weighted averaging or statistical feature aggregation, lacking sufficient consideration of the spatial topological relationships and dynamic evolution of time series between sensors. This makes it difficult to characterize the complex coupling between nodes and global operational behavior, resulting in limitations in anomaly judgment. Especially in distributed control and IoT environments, it is impossible to achieve synchronous identification and dynamic tracking of node-level and global-level states.

[0004] Furthermore, traditional anomaly detection systems suffer from low computational efficiency in data processing and feature modeling, relying mostly on CPU serial processing or single-threaded computation on edge devices, making them ill-suited for high-density sensor deployments and high-speed sampling scenarios. In terms of graph structured modeling and deep learning applications, existing technologies have not fully utilized graph neural networks and multi-head self-attention mechanisms to fuse node embedding vectors across nodes and across time series, thus limiting the applicability and real-time performance of anomaly detection algorithms in high-dimensional, multi-node systems.

[0005] Therefore, how to provide a servo drive anomaly detection system based on multi-sensor fusion is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a servo drive anomaly detection system based on multi-sensor fusion. This invention fully utilizes graph neural networks, multi-head self-attention mechanisms, and adaptive anomaly detection algorithms. It describes in detail the implementation method of collecting multi-dimensional operating data of servo drives through intelligent sensors, constructing a graph structure model, performing multi-round message passing and time series encoding, fusing cross-node and cross-time series features, and generating node-level and global anomaly states. It has the advantages of high anomaly recognition accuracy, strong adaptability to complex working conditions, and high real-time monitoring efficiency.

[0007] A servo drive anomaly detection system based on multi-sensor fusion according to an embodiment of the present invention includes: The data processing module is used to perform data acquisition through intelligent sensors deployed in the servo drive loop, and to perform preprocessing to generate structured operational data; The network construction module is used to build a graph neural network, using smart sensors as graph nodes. The node features are preprocessed running data, and the edge weights between nodes are generated by signal correlation and spatial topology. The message passing module is used to perform multi-round message passing operations in the graph neural network. It generates the embedding vector of each node through time series encoding and node feature update, and establishes a node representation that reflects the local and global state of the servo drive. The attention fusion module is used to fuse cross-node and cross-time series features based on the embedded vector using a multi-head self-attention mechanism to form a global state feature representation, which is used to characterize the dynamic behavior of servo-driven operation and the coupling relationship between nodes. The anomaly detection module is used to build an adaptive anomaly detection algorithm to identify abnormal states in distributed control and IoT environments, and output anomaly type sequence, anomaly probability and anomaly development trend. The anomaly labeling module is used to classify and label anomaly type sequences and anomaly probabilities, generate anomaly early warning data and anomaly evolution reports, and organize the data according to nodes, timestamps and anomaly development trends; The data distribution module is used to distribute abnormal early warning data and abnormal evolution reports to the production site control unit and the upper monitoring terminal, so as to realize distributed monitoring and data acquisition visualization management of servo drive abnormalities.

[0008] Optionally, modules can be integrated using the following methods: S1. In the production sites of metal cutting and welding equipment manufacturing and electronic component manufacturing, data acquisition is performed by intelligent sensors deployed in the servo drive circuit, and preprocessing is carried out to generate structured operation data. S2. Construct a graph neural network on a GPU or edge computing device, using smart sensors as graph nodes. The node features are pre-processed running data, and the edge weights between nodes are generated by signal correlation and spatial topology. S3. Perform multi-round message passing operations in the graph neural network, generate the embedding vector of each node through time series encoding and node feature update, and establish a node representation that reflects the local and global state of the servo drive. S4. Based on the embedded vector, a multi-head self-attention mechanism is used to fuse cross-node and cross-time series features to form a global state feature representation, which is used to characterize the dynamic behavior of servo drive operation and the coupling relationship between nodes. S5. Based on global state feature representation, construct an adaptive anomaly detection algorithm to identify abnormal states in distributed control and IoT environments, and output anomaly type sequence, anomaly probability and anomaly development trend. S6. Classify and label the abnormal type sequence and abnormal probability, generate abnormal early warning data and abnormal evolution report, and organize the data according to nodes, timestamps and abnormal development trends; S7. Distribute abnormal early warning data and abnormal evolution reports to the production site control unit and the upper monitoring terminal to realize distributed monitoring and data acquisition visualization management of servo drive abnormalities.

[0009] Optionally, the data acquisition includes current signal, voltage signal, rotation speed signal, position feedback signal, vibration signal, and temperature signal. The preprocessing sequentially performs time synchronization, outlier removal, noise suppression, and amplitude normalization on the acquired signals.

[0010] Optionally, the process of generating the edge weights between nodes specifically includes: Establish a unified time-aligned buffer for the running data of each smart sensor on a GPU or edge computing device, and load the multi-dimensional signal sequence within the same time window into shared video memory. The correlation calculation process is performed in parallel on the data sequences of any two sensor nodes. By comparing the signal amplitude variation trend, frequency domain energy distribution, and phase synchronization characteristics through a sliding window, a correlation matrix characterizing the signal coupling strength is generated, specifically including: Each pair of sensor nodes is allocated an independent parallel computing thread block, and the multidimensional signal sequence of the corresponding node in the unified time alignment buffer is expanded by channel and mapped to a continuous video memory area. Within each thread block, a segmented sliding window slicing operation is performed on the signal sequence. Trend feature encoding is performed on the signal amplitude sequence within each window. At the same time, frequency domain transformation is performed on the signals within the same window to extract energy distribution features, and synchronization matching is performed on the phase change sequence of cross-node signals. After completing window-level feature extraction, the trend similarity results, frequency domain energy matching results, and phase synchronization matching results obtained from each window are weighted and aggregated in shared memory to generate a description of the coupling strength of the node pair within the current time window. Parallel accumulation and smoothing are performed on the coupling strength descriptions for all time windows to form a stable representation of inter-node coupling strength, forming a correlation matrix, thereby completing the signal correlation modeling process; Based on this, the physical structure topology information of the servo drive system is retrieved, and the installation positions of the sensors in the drive loop, the control link connections, and the electrical adjacency relationships are encoded into a spatial topology matrix, specifically including: Load the structural configuration data of the servo drive system onto the GPU or edge computing device, parse the physical installation position, functional unit and control loop hierarchy information of each smart sensor in the drive loop into discrete topology node descriptions, and organize the electrical connection relationship, control signal flow relationship and mechanical coupling relationship between sensors into directed or undirected connection records. For each pair of sensor nodes, a topology relationship determination operation is performed in parallel. Based on their relative level in the control loop, physical distance range, and whether there is a direct or indirect electrical connection, a topology association identifier is generated for the corresponding node pair. Based on this, the topological association identifier is mapped to a spatial adjacency weight description, and a node-pair level topological association matrix is ​​constructed in shared memory; Parallel normalization is performed on the topological association matrix to express different types of topological relationships at a uniform scale, and pruning operations are performed on node pairs that do not meet the preset topological constraints, thereby forming a spatial topological matrix that retains only the effective physical and control association relationships. An element-wise fusion operation is performed on the correlation matrix and the spatial topology matrix, and the fusion result is normalized and thresholded to form a set of edge weights for edge connections in the graph neural network.

[0011] Optionally, the execution process of the multi-round message passing operation specifically includes: The initial feature vectors and edge weight matrices of the graph nodes are loaded into the video memory, and an independent computation thread block is allocated to each node; In each round of message passing, each node receives messages from its neighboring nodes in parallel. The message content includes the embedding vector of the neighboring node and the weighted value of the corresponding edge weight. During the receiving process, vector concatenation and weighted summation operations are performed to integrate neighborhood information. Perform time-series encoding on the node's own feature vector to map the state features of historical time steps into a unified dimension representation, and then fuse it with the received neighborhood messages; The fused feature vectors are then updated with node embeddings through node-level parallel nonlinear activation functions and normalization operations, so that the embedding vector of each node absorbs local neighborhood information in round by round and reflects the dynamic changes across time steps. After completing message passing for all rounds, the final embedding vector of each node is cached in video memory, forming a node representation that reflects the local operating state and global coupling relationship of the servo drive system.

[0012] Optionally, the formation process of the global state feature representation specifically includes: The embedding vector of each node is copied in parallel to the memory area of ​​multiple attention heads on a GPU or edge computing device, and each attention head independently processes the interaction relationship between nodes. A linear mapping is performed on the node embedding vector within each attention head to generate query vector, key vector, and value vector. Attention scores for each pair of nodes are computed in parallel in memory at the node pair level. Attention weights are obtained by the dot product or similarity measure of the embedding vectors between nodes. The obtained attention weights are normalized within each attention head so that the weights of each node pair form a probability distribution in the local neighborhood or global node set. Then, the normalized attention weights are weighted and summed with the corresponding value vectors to generate the fusion representation of each node under the current attention head. All attention heads' node representations are concatenated or superimposed in video memory, and after nonlinear activation and normalization operations, they form the final global state feature representation, which is used to characterize the dynamic behavior of servo drive operation and the coupling relationship between nodes.

[0013] Optionally, the adaptive anomaly detection algorithm specifically includes: The global state features of each node are represented in the video memory and divided into continuous tensor blocks according to time steps. An independent computation thread block is allocated to each node to achieve parallel processing. A weighted sliding window operation is performed on the state features of each node within a continuous time window. The feature vector within the window is compared with the historical normal operation feature set element by element to generate a node-level anomaly score sequence. The edge weight matrix between nodes is used for neighborhood feature aggregation, and the anomaly score sequence of each node is subjected to neighborhood weighted averaging to dynamically adjust the anomaly sensitivity and form an adaptive anomaly threshold. The GPU is used to normalize and dynamically threshold the abnormal score sequence in parallel. Time steps that are higher than the adaptive threshold are marked as abnormal. The abnormal labels are then clustered in parallel, and the abnormal states of adjacent time steps and neighboring nodes are merged to generate an abnormal type sequence. When generating anomaly probabilities, the local anomaly score of a node is combined with the neighborhood weighted result, and the probability distribution is calculated through Softmax mapping. Furthermore, smoothing filtering is applied to the probability changes over consecutive time steps to capture anomaly development trends. Specifically, this includes: In video memory, a neighborhood index table cache is allocated for each node to record its directly connected nodes and the corresponding edge weights. Each node's computation thread reads its own abnormal score sequence and the abnormal score sequences of its neighboring nodes in parallel. By multiplying the scores of neighboring nodes by the corresponding edge weights and accumulating them element by element within the sliding time window, a weighted fusion of neighborhood features is achieved. After the accumulation is completed, the fusion result is normalized so that the weighted anomaly score of each node reflects its own and its neighbors' comprehensive anomaly status on a unified scale. To dynamically adjust anomaly sensitivity, the local standard deviation of the weighted anomaly score for each node is calculated and compared element by element with the fluctuation range of the historical normal operation window to generate an adaptive anomaly threshold sequence. The adaptive threshold sequence is written back to the video memory for subsequent anomaly detection at each time step, enabling node-level anomaly detection to adapt to neighborhood coupling relationships and dynamic fluctuations. The sequence of anomaly types, anomaly probabilities, and anomaly development trends of each node are organized into a global anomaly matrix in the video memory according to time and node order. This matrix is ​​then distributed synchronously to distributed control and IoT terminals, enabling node-level and global-level anomaly identification and tracking.

[0014] The beneficial effects of this invention are: First, by deploying multiple intelligent sensors in the servo drive circuit and fusing the data, this invention achieves comprehensive acquisition and preprocessing of multi-dimensional operating data such as current, speed, position feedback, vibration, and temperature. This generates structured node features, providing an accurate and reliable data foundation for subsequent graph neural network modeling and anomaly detection, effectively improving the sensitivity and accuracy of anomaly identification.

[0015] Secondly, this invention utilizes graph neural networks to construct a topological relationship model between nodes, and combines multi-round message passing and time series encoding methods to enable the embedding vector of each node to reflect the local and global states. At the same time, it integrates cross-node and cross-time series features through a multi-head self-attention mechanism to form a global state feature representation, thereby achieving an accurate characterization of the dynamic behavior of the servo drive system and the coupling relationship between nodes, thus significantly enhancing the robustness of anomaly detection and the ability to adapt to complex working conditions.

[0016] Finally, this invention designs an adaptive anomaly detection algorithm that combines the local anomaly score of a node with neighborhood weighted information and generates anomaly type sequences, anomaly probabilities, and anomaly development trends within a continuous time window. This enables dynamic anomaly identification and distributed early warning at both the node and global levels, allowing for real-time monitoring of servo drive systems in distributed control and IoT environments. This improves system operational safety, monitoring efficiency, and anomaly response speed. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a block diagram of the servo drive anomaly detection system based on multi-sensor fusion proposed in this invention. Figure 2 This is a flowchart of the servo drive anomaly detection system based on multi-sensor fusion proposed in this invention. Figure 3 This is a flowchart of the multi-round message passing and time series encoding process of the servo drive anomaly detection system based on multi-sensor fusion proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figure 1 A servo drive anomaly detection system based on multi-sensor fusion includes: The data processing module is used to perform data acquisition through intelligent sensors deployed in the servo drive loop, and to perform preprocessing to generate structured operating data; The network construction module is used to build a graph neural network, using smart sensors as graph nodes. The node features are preprocessed running data, and the edge weights between nodes are generated by signal correlation and spatial topology. The message passing module is used to perform multi-round message passing operations in the graph neural network. It generates the embedding vector of each node through time series encoding and node feature update, and establishes a node representation that reflects the local and global state of the servo drive. The attention fusion module is used to fuse cross-node and cross-time series features based on the embedded vector using a multi-head self-attention mechanism to form a global state feature representation, which is used to characterize the dynamic behavior of servo-driven operation and the coupling relationship between nodes. The anomaly detection module is used to build an adaptive anomaly detection algorithm to identify abnormal states in distributed control and IoT environments, and output anomaly type sequence, anomaly probability and anomaly development trend. The anomaly labeling module is used to classify and label anomaly type sequences and anomaly probabilities, generate anomaly early warning data and anomaly evolution reports, and organize the data according to nodes, timestamps and anomaly development trends; The data distribution module is used to distribute abnormal early warning data and abnormal evolution reports to the production site control unit and the upper monitoring terminal, so as to realize distributed monitoring and data acquisition visualization management of servo drive abnormalities.

[0020] refer to Figure 2-3 In this embodiment, the modules are interconnected using the following method: S1. In the production sites of metal cutting and welding equipment manufacturing and electronic component manufacturing, data acquisition is performed by intelligent sensors deployed in the servo drive circuit, and preprocessing is carried out to generate structured operation data. S2. Construct a graph neural network on a GPU or edge computing device, using smart sensors as graph nodes. The node features are pre-processed running data, and the edge weights between nodes are generated by signal correlation and spatial topology. S3. Perform multi-round message passing operations in the graph neural network, generate the embedding vector of each node through time series encoding and node feature update, and establish a node representation that reflects the local and global state of the servo drive. S4. Based on the embedded vector, a multi-head self-attention mechanism is used to fuse cross-node and cross-time series features to form a global state feature representation, which is used to characterize the dynamic behavior of servo drive operation and the coupling relationship between nodes. S5. Based on global state feature representation, construct an adaptive anomaly detection algorithm to identify abnormal states in distributed control and IoT environments, and output anomaly type sequence, anomaly probability and anomaly development trend. S6. Classify and label the abnormal type sequence and abnormal probability, generate abnormal early warning data and abnormal evolution report, and organize the data according to nodes, timestamps and abnormal development trends; S7. Distribute abnormal early warning data and abnormal evolution reports to the production site control unit and the upper monitoring terminal to realize distributed monitoring and data acquisition visualization management of servo drive abnormalities.

[0021] In this embodiment, the data acquisition includes current signal, voltage signal, rotation speed signal, position feedback signal, vibration signal, and temperature signal. The preprocessing sequentially performs time synchronization, outlier removal, noise suppression, and amplitude normalization on the acquired signals.

[0022] In this embodiment, the process of generating the edge weights between nodes specifically includes: Establish a unified time-aligned buffer for the running data of each smart sensor on a GPU or edge computing device, and load the multi-dimensional signal sequence within the same time window into shared video memory. The correlation calculation process is performed in parallel on the data sequences of any two sensor nodes. By comparing the signal amplitude variation trend, frequency domain energy distribution, and phase synchronization characteristics through a sliding window, a correlation matrix characterizing the signal coupling strength is generated, specifically including: Each pair of sensor nodes is allocated an independent parallel computing thread block, and the multidimensional signal sequence of the corresponding node in the unified time alignment buffer is expanded by channel and mapped to a continuous video memory area. Within each thread block, a segmented sliding window slicing operation is performed on the signal sequence. Trend feature encoding is performed on the signal amplitude sequence within each window. At the same time, frequency domain transformation is performed on the signals within the same window to extract energy distribution features, and synchronization matching is performed on the phase change sequence of cross-node signals. After completing window-level feature extraction, the trend similarity results, frequency domain energy matching results, and phase synchronization matching results obtained from each window are weighted and aggregated in shared memory to generate a description of the coupling strength of the node pair within the current time window. Parallel accumulation and smoothing are performed on the coupling strength descriptions for all time windows to form a stable representation of inter-node coupling strength, forming a correlation matrix, thereby completing the signal correlation modeling process; Based on this, the physical structure topology information of the servo drive system is retrieved, and the installation positions of the sensors in the drive loop, the control link connections, and the electrical adjacency relationships are encoded into a spatial topology matrix, specifically including: Load the structural configuration data of the servo drive system onto the GPU or edge computing device, parse the physical installation position, functional unit and control loop hierarchy information of each smart sensor in the drive loop into discrete topology node descriptions, and organize the electrical connection relationship, control signal flow relationship and mechanical coupling relationship between sensors into directed or undirected connection records. For each pair of sensor nodes, a topology relationship determination operation is performed in parallel. Based on their relative level in the control loop, physical distance range, and whether there is a direct or indirect electrical connection, a topology association identifier is generated for the corresponding node pair. Based on this, the topological association identifier is mapped to a spatial adjacency weight description, and a node-pair level topological association matrix is ​​constructed in shared memory; Parallel normalization is performed on the topological association matrix to express different types of topological relationships at a uniform scale, and pruning operations are performed on node pairs that do not meet the preset topological constraints, thereby forming a spatial topological matrix that retains only the effective physical and control association relationships. An element-wise fusion operation is performed on the correlation matrix and the spatial topology matrix, and the fusion result is normalized and thresholded to form a set of edge weights for edge connections in the graph neural network.

[0023] In this embodiment, the execution process of the multi-round message passing operation specifically includes: The initial feature vectors and edge weight matrices of the graph nodes are loaded into the video memory, and an independent computation thread block is allocated to each node; In each round of message passing, each node receives messages from its neighboring nodes in parallel. The message content includes the embedding vector of the neighboring node and the weighted value of the corresponding edge weight. During the receiving process, vector concatenation and weighted summation operations are performed to integrate neighborhood information. Perform time-series encoding on the node's own feature vector to map the state features of historical time steps into a unified dimension representation, and then fuse it with the received neighborhood messages; The fused feature vectors are then updated with node embeddings through node-level parallel nonlinear activation functions and normalization operations. This allows each node's embedding vector to absorb local neighborhood information round by round and reflect dynamic changes across time steps, specifically including: Allocate an adjacency node index cache and a message buffer for each node, and preload the embedding vectors of adjacency nodes and their corresponding edge weights into shared memory; Each node's computation thread block reads the embedding vector and edge weights in parallel according to the adjacent node index, and multiplies the adjacent node embedding vector and edge weights element by element to generate a weighted message; The weighted messages are summed in parallel by channel in shared memory. At the same time, the summation results are concatenated element by element with the time series encoding features of the node itself to form a temporary fusion vector. After fusion is completed, the thread block performs a node-level non-linear activation function on the fused vector and performs a batch normalization operation in the video memory to update the node embedding vector. This process is repeated in parallel across all nodes, ensuring that the embedding vector of each node absorbs information from its neighboring nodes in each round of message passing, while maintaining a dynamic representation of time-series features and completing the progressive encoding of local and global states. After completing message passing for all rounds, the final embedding vector of each node is cached in video memory, forming a node representation that reflects the local operating state and global coupling relationship of the servo drive system.

[0024] In this embodiment, the formation process of the global state feature representation specifically includes: The embedding vector of each node is copied in parallel to the memory area of ​​multiple attention heads on a GPU or edge computing device, and each attention head independently processes the interaction relationship between nodes. A linear mapping is performed on the node embedding vector within each attention head to generate a query vector, key vector, and value vector. Attention scores for each node pair are computed in parallel at the node-pair level in GPU memory. Attention weights are obtained through the dot product or similarity metric of the embedding vectors between nodes, specifically including: Allocate contiguous storage space for the embedding vector of each node, and load the embedding vector and the weight matrix of the attention head into shared memory in parallel; Each computation thread block is responsible for performing matrix multiplication operations on the node embedding vector and the corresponding weight matrix to generate the node's query vector, key vector, and value vector, and writing the results to an independent cache in the video memory to maintain data isolation between different attention heads; During the node-pair granular parallel computation stage, each thread block reads the query vector and the key vector of the corresponding node according to the node pair, performs dot product or similarity measurement operations in the video memory to obtain the attention score, and temporarily stores the score in shared memory. Within each attention head, the attention score is normalized. The attention weights of the node pairs are converted into local probability distributions through parallel computation. At the same time, the normalized weights and the corresponding value vectors are summed element-wise in the video memory to generate the fusion representation of each node under the current attention head. The fusion result is then stored back in the video memory. The obtained attention weights are normalized within each attention head so that the weights of each node pair form a probability distribution in the local neighborhood or global node set. Then, the normalized attention weights are weighted and summed with the corresponding value vectors to generate the fusion representation of each node under the current attention head. All attention heads' node representations are concatenated or superimposed in video memory, and after nonlinear activation and normalization operations, they form the final global state feature representation, which is used to characterize the dynamic behavior of servo drive operation and the coupling relationship between nodes.

[0025] In this embodiment, the adaptive anomaly detection algorithm specifically includes: The global state features of each node are represented in the video memory and divided into continuous tensor blocks according to time steps. An independent computation thread block is allocated to each node to achieve parallel processing. A weighted sliding window operation is performed on the state features of each node within a continuous time window. The feature vector within the window is compared with the historical normal operation feature set element by element to generate a node-level anomaly score sequence. The edge weight matrix between nodes is used for neighborhood feature aggregation, and the anomaly score sequence of each node is processed by neighborhood weighted averaging to dynamically adjust the anomaly sensitivity and form an adaptive anomaly threshold. The GPU is used to normalize and dynamically threshold the abnormal score sequence in parallel. Time steps that are higher than the adaptive threshold are marked as abnormal. The abnormal labels are then clustered in parallel, and the abnormal states of adjacent time steps and neighboring nodes are merged to generate an abnormal type sequence. When generating anomaly probabilities, the local anomaly score of a node is combined with the neighborhood weighted result, and the probability distribution is calculated through Softmax mapping. Furthermore, smoothing filtering is applied to the probability changes over consecutive time steps to capture anomaly development trends. Specifically, this includes: In video memory, a neighborhood index table cache is allocated for each node to record its directly connected nodes and the corresponding edge weights. Each node's computation thread reads its own abnormal score sequence and the abnormal score sequences of its neighboring nodes in parallel. By multiplying the scores of neighboring nodes by the corresponding edge weights and accumulating them element by element within the sliding time window, a weighted fusion of neighborhood features is achieved. After the accumulation is completed, the fusion result is normalized so that the weighted anomaly score of each node reflects its own and its neighbors' comprehensive anomaly status on a unified scale. To dynamically adjust anomaly sensitivity, the local standard deviation of the weighted anomaly score for each node is calculated and compared element by element with the fluctuation range of the historical normal operation window to generate an adaptive anomaly threshold sequence. The adaptive threshold sequence is written back to the video memory for subsequent anomaly detection at each time step, enabling node-level anomaly detection to adapt to neighborhood coupling relationships and dynamic fluctuations. The sequence of anomaly types, anomaly probabilities, and anomaly development trends of each node are organized into a global anomaly matrix in the video memory according to time and node order. This matrix is ​​then distributed synchronously to distributed control and IoT terminals, enabling node-level and global-level anomaly identification and tracking.

[0026] Example 1: To verify the feasibility of this invention in practice, it was applied to a servo drive monitoring scenario for high-precision metal processing equipment. In this scenario, the servo drive system needs to operate for extended periods under high loads and complex conditions while maintaining processing accuracy and equipment stability. Existing technologies typically rely on a single sensor or fixed threshold for anomaly detection, which cannot accurately capture potential equipment anomalies, resulting in missed and false alarms. This leads to delays in equipment fault location and increased maintenance costs.

[0027] In this embodiment, multiple intelligent sensors, including those for current, rotational speed, position feedback, vibration, and temperature, are first deployed in the servo drive loop. Structured node features are generated through unified time alignment and signal preprocessing. During data acquisition, the standard deviation of current fluctuations ranges from 0.05A to 0.12A, rotational speed deviations from 0.3 to 0.8 RPM, peak vibration accelerations from 0.02 to 0.05g, and temperature variations from 0.6 to 1.2℃. These multidimensional data comprehensively reflect the operating status of the servo drive. Subsequently, these node features are input into a graph neural network constructed using a GPU or edge computing device. The edge weights between nodes, combined with signal correlation and spatial topological relationships, generate a graph structure. Multiple rounds of message passing and time-series encoding are performed within the graph, enabling the embedding vector of each node to reflect its local state and its coupling relationship with neighboring nodes. A multi-head self-attention mechanism fuses cross-node and cross-time-series features to form a global state feature representation, further enhancing the ability to characterize dynamic anomaly patterns.

[0028] Based on global state features, an adaptive anomaly detection algorithm is constructed. This algorithm calculates a neighborhood-weighted anomaly score for each node and dynamically adjusts the anomaly threshold using a sliding window, enabling the prediction and labeling of anomaly types, probabilities, and development trends. In this embodiment, the system detected 52 servo-driven anomaly events, including 42 early warnings, 5 missed detections, and 5 false alarms. Compared to traditional single-sensor methods, the early warning rate is increased by 30%, the missed detection rate is reduced by 40%, and the false alarm rate is reduced by 25%. The average delay time for anomaly identification is 0.8 seconds, a reduction of approximately 60% compared to traditional methods.

[0029] In practical applications, the system can distribute abnormal information to field control units and upper-level monitoring terminals in real time, enabling distributed monitoring and visual management of servo drive equipment. Monitoring data from continuous operation shows that during a 500-hour high-load continuous operation test cycle, the system achieved an anomaly identification accuracy of over 94% for key nodes and 100% global status monitoring coverage, effectively preventing the escalation of potential equipment failures and improving production efficiency and safety.

[0030] To clearly demonstrate the effectiveness of this embodiment, a table showing some of the acquired and anomaly detection data is provided below, illustrating the multi-dimensional sensor signal characteristics, anomaly detection results, and performance indicators: Table 1. Example data for high-precision servo drive anomaly monitoring

[0031] Table 1 fully demonstrates the specific performance of each node's features under multi-sensor fusion, as well as the anomaly probability and anomaly type generated by the adaptive anomaly detection algorithm. It also reflects the early warning capability and intuitively verifies the effectiveness and reliability of the present invention in a practical servo drive system.

[0032] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A servo drive anomaly detection system based on multi-sensor fusion, characterized in that, include: The data processing module is used to perform data acquisition through intelligent sensors deployed in the servo drive loop, and to perform preprocessing to generate structured operational data; The network construction module is used to build a graph neural network, using smart sensors as graph nodes. The node features are preprocessed running data, and the edge weights between nodes are generated by signal correlation and spatial topology. The message passing module is used to perform multi-round message passing operations in the graph neural network. It generates the embedding vector of each node through time series encoding and node feature update, and establishes a node representation that reflects the local and global state of the servo drive. The attention fusion module is used to fuse cross-node and cross-time series features based on the embedded vector using a multi-head self-attention mechanism to form a global state feature representation, which is used to characterize the dynamic behavior of servo-driven operation and the coupling relationship between nodes; The anomaly detection module is used to build an adaptive anomaly detection algorithm to identify abnormal states in distributed control and IoT environments, and output anomaly type sequence, anomaly probability and anomaly development trend. The anomaly labeling module is used to classify and label anomaly type sequences and anomaly probabilities, generate anomaly early warning data and anomaly evolution reports, and organize the data according to nodes, timestamps and anomaly development trends; The data distribution module is used to distribute abnormal early warning data and abnormal evolution reports to the production site control unit and the upper monitoring terminal, so as to realize distributed monitoring and data acquisition visualization management of servo drive abnormalities.

2. The servo drive anomaly detection system based on multi-sensor fusion according to claim 1, characterized in that, The modules are connected in the following way: S1. In the production sites of metal cutting and welding equipment manufacturing and electronic component manufacturing, data is collected and preprocessed by intelligent sensors deployed in the servo drive circuit to generate structured operating data. S2. Construct a graph neural network on a GPU or edge computing device, using smart sensors as graph nodes. The node features are pre-processed running data, and the edge weights between nodes are generated by signal correlation and spatial topology. S3. Perform multi-round message passing operations in the graph neural network, generate the embedding vector of each node through time series encoding and node feature update, and establish a node representation that reflects the local and global state of the servo drive. S4. Based on the embedded vector, a multi-head self-attention mechanism is used to fuse cross-node and cross-time series features to form a global state feature representation, which is used to characterize the dynamic behavior of servo drive operation and the coupling relationship between nodes. S5. Based on global state feature representation, construct an adaptive anomaly detection algorithm to identify abnormal states in distributed control and IoT environments, and output anomaly type sequence, anomaly probability and anomaly development trend; S6. Classify and label the abnormal type sequence and abnormal probability, generate abnormal early warning data and abnormal evolution report, and organize the data according to nodes, timestamps and abnormal development trends; S7. Distribute abnormal early warning data and abnormal evolution reports to the production site control unit and the upper monitoring terminal to realize distributed monitoring and data acquisition visualization management of servo drive abnormalities.

3. The servo drive anomaly detection system based on multi-sensor fusion according to claim 2, characterized in that, The data acquisition includes current signals, voltage signals, rotational speed signals, position feedback signals, vibration signals, and temperature signals. The preprocessing sequentially performs time synchronization, outlier removal, noise suppression, and amplitude normalization on the acquired signals.

4. The servo drive anomaly detection system based on multi-sensor fusion according to claim 2, characterized in that, The process of generating the edge weights between nodes specifically includes: Establish a unified time-aligned buffer for the running data of each smart sensor on a GPU or edge computing device, and load the multi-dimensional signal sequence within the same time window into shared video memory. The correlation calculation process is performed in parallel on the data sequences of any two sensor nodes. By comparing the signal amplitude variation trend, frequency domain energy distribution, and phase synchronization characteristics through a sliding window, a correlation matrix characterizing the signal coupling strength is generated, specifically including: Each pair of sensor nodes is allocated an independent parallel computing thread block, and the multidimensional signal sequence of the corresponding node in the unified time alignment buffer is expanded by channel and mapped to a continuous video memory area. Within each thread block, a segmented sliding window slicing operation is performed on the signal sequence. Trend feature encoding is performed on the signal amplitude sequence within each window. At the same time, frequency domain transformation is performed on the signals within the same window to extract energy distribution features, and synchronization matching is performed on the phase change sequence of cross-node signals. After completing window-level feature extraction, the trend similarity results, frequency domain energy matching results, and phase synchronization matching results obtained from each window are weighted and aggregated in shared memory to generate a description of the coupling strength of the node pair within the current time window. Parallel accumulation and smoothing are performed on the coupling strength descriptions for all time windows to form a stable representation of inter-node coupling strength, forming a correlation matrix, thereby completing the signal correlation modeling process; Based on this, the physical structure topology information of the servo drive system is retrieved, and the installation positions of the sensors in the drive loop, the control link connections, and the electrical adjacency relationships are encoded into a spatial topology matrix, specifically including: Load the structural configuration data of the servo drive system onto the GPU or edge computing device, parse the physical installation position, functional unit and control loop hierarchy information of each smart sensor in the drive loop into discrete topology node descriptions, and organize the electrical connection relationship, control signal flow relationship and mechanical coupling relationship between sensors into directed or undirected connection records. For each pair of sensor nodes, a topology relationship determination operation is performed in parallel. Based on their relative level in the control loop, physical distance range, and whether there is a direct or indirect electrical connection, a topology association identifier is generated for the corresponding node pair. Based on this, the topological association identifier is mapped to a spatial adjacency weight description, and a node-pair level topological association matrix is ​​constructed in shared memory; Parallel normalization is performed on the topological association matrix to express different types of topological relationships at a uniform scale, and pruning operations are performed on node pairs that do not meet the preset topological constraints, thereby forming a spatial topological matrix that retains only the effective physical and control association relationships. An element-wise fusion operation is performed on the correlation matrix and the spatial topology matrix, and the fusion result is normalized and thresholded to form a set of edge weights for edge connections in the graph neural network.

5. The servo drive anomaly detection system based on multi-sensor fusion according to claim 2, characterized in that, The execution process of the multi-round message passing operation specifically includes: The initial feature vectors and edge weight matrices of the graph nodes are loaded into the video memory, and an independent computation thread block is allocated to each node; In each round of message passing, each node receives messages from its neighboring nodes in parallel. The message content includes the embedding vector of the neighboring node and the weighted value of the corresponding edge weight. During the receiving process, vector concatenation and weighted summation operations are performed to integrate neighborhood information. Perform time-series encoding on the node's own feature vector to map the state features of historical time steps into a unified dimension representation, and then fuse it with the received neighborhood messages; The fused feature vectors are then updated with node embeddings through node-level parallel nonlinear activation functions and normalization operations, so that the embedding vector of each node absorbs local neighborhood information in round by round and reflects the dynamic changes across time steps. After completing message passing for all rounds, the final embedding vector of each node is cached in video memory, forming a node representation that reflects the local operating state and global coupling relationship of the servo drive system.

6. The servo drive anomaly detection system based on multi-sensor fusion according to claim 2, characterized in that, The formation process of the global state feature representation specifically includes: The embedding vector of each node is copied in parallel to the memory area of ​​multiple attention heads on a GPU or edge computing device, and each attention head independently processes the interaction relationship between nodes. A linear mapping is performed on the node embedding vector within each attention head to generate query vector, key vector, and value vector. Attention scores for each pair of nodes are computed in parallel at the node pair level in memory. Attention weights are obtained by the dot product or similarity measure of the embedding vectors between nodes. The obtained attention weights are normalized within each attention head so that the weights of each node pair form a probability distribution in the local neighborhood or global node set. Then, the normalized attention weights are weighted and summed with the corresponding value vectors to generate the fusion representation of each node under the current attention head. All attention heads' node representations are concatenated or superimposed in parallel in video memory. After nonlinear activation and normalization operations, they form the final global state feature representation, which is used to characterize the dynamic behavior of servo drive operation and the coupling relationship between nodes.

7. The servo drive anomaly detection system based on multi-sensor fusion according to claim 2, characterized in that, The adaptive anomaly detection algorithm specifically includes: The global state features of each node are represented in the video memory and divided into continuous tensor blocks according to time steps. An independent computation thread block is allocated to each node to achieve parallel processing. A weighted sliding window operation is performed on the state features of each node within a continuous time window. The feature vector within the window is compared with the historical normal operation feature set element by element to generate a node-level anomaly score sequence. The edge weight matrix between nodes is used for neighborhood feature aggregation, and the anomaly score sequence of each node is processed by neighborhood weighted averaging to dynamically adjust the anomaly sensitivity and form an adaptive anomaly threshold. The GPU is used to normalize and dynamically threshold the abnormal score sequence in parallel. Time steps that are higher than the adaptive threshold are marked as abnormal. The abnormal labels are then clustered in parallel, and the abnormal states of adjacent time steps and neighboring nodes are merged to generate an abnormal type sequence. When generating anomaly probabilities, the local anomaly score of a node is combined with the neighborhood weighted result, and the probability distribution is calculated through Softmax mapping. Furthermore, smoothing filtering is applied to the probability changes over consecutive time steps to capture anomaly development trends. Specifically, this includes: In video memory, a neighborhood index table cache is allocated for each node to record its directly connected nodes and the corresponding edge weights. Each node's computation thread reads its own abnormal score sequence and the abnormal score sequences of its neighboring nodes in parallel. By multiplying the scores of neighboring nodes by the corresponding edge weights and accumulating them element by element within the sliding time window, a weighted fusion of neighborhood features is achieved. After the accumulation is completed, the fusion result is normalized so that the weighted anomaly score of each node reflects its own and its neighbors' comprehensive anomaly status on a unified scale. To dynamically adjust anomaly sensitivity, the local standard deviation of the weighted anomaly score for each node is calculated and compared element by element with the fluctuation range of the historical normal operation window to generate an adaptive anomaly threshold sequence. The adaptive threshold sequence is written back to the video memory for subsequent anomaly detection at each time step, realizing the adaptive response of node-level anomaly detection to neighborhood coupling relationships and dynamic fluctuations. The sequence of anomaly types, probabilities, and trends of each node are organized into a global anomaly matrix in memory according to time and node order. This matrix is ​​then distributed synchronously to distributed control and IoT terminals, enabling node-level and global-level anomaly identification and tracking.

Images