Intelligent management system for cable digitization
By introducing long-term short-term memory models and self-attention mechanisms into the cable digital management system, the time dependence of cable sensor data is captured, and the time window is adaptively selected through the multi-head attention mechanism, the problem that traditional systems cannot flexibly deal with different fault types is solved, achieving higher fault diagnosis accuracy and system reliability.
Patent Information
- Application Number
- CN202510311665.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-17
- Publication Date
- 2025-06-27
AI Technical Summary
Traditional cable digital management systems cannot flexibly respond to different types of faults or abnormal situations, and lack effective capture of the time dependence of cable sensor data, resulting in reduced analysis accuracy and reliability.
An intelligent management system for digital cables is designed, including a data analysis module, an adaptive time window selection module, a feature map analysis module and a feature map fusion module. The system captures the time dependence of data through long-term short-term memory models and self-attention mechanisms, and adaptively selects the time window size through the multi-head attention mechanism to generate local and global feature importance graphs, and finally fuses features through graph convolution operations.
The system can flexibly process data at different time scales, capture the long-term and short-term dependencies of data, improve the accuracy and efficiency of fault diagnosis, and significantly improve the overall performance and reliability of the system.
Smart Images
Figure CN120218733A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of digital technologies, and particularly to an intelligent management system for cable digitization. Background Art
[0002] During operation, traditional cable digital management systems may experience various types of faults, and the time scales at which different types of faults occur vary. For example, electrical faults may occur within seconds, while mechanical wear may take weeks or months to manifest; these systems typically use fixed or preset time window sizes to analyze sensor data, which results in an inability to flexibly handle different types of faults or abnormal conditions.
[0003] In addition, cable sensor data has obvious time series characteristics, and the dependency relationship between data at different time points is crucial for fault diagnosis. For example, sudden changes in voltage and current often indicate impending faults. Traditional systems lack effective capture of the data dependency within the time window and often use simple statistical methods or static models to process time series data, ignoring the dynamic characteristics and complex changes of the data over time, which greatly reduces the accuracy and reliability of the analysis. Summary of the Invention
[0004] The purpose of the present invention is to provide an intelligent management system for cable digitization to solve the problems raised in the above background art.
[0005] To achieve the above purpose, the present invention provides the following technical solution: An intelligent management system for cable digitization, including a data analysis module, an adaptive time window selection module, a feature map analysis module, and a feature map fusion module, wherein:
[0006] The data analysis module initializes a time window, captures the dependency of cable sensor data within different time windows, generates a time dependency representation; and calculates the importance score of cable sensor data within different time windows;
[0007] The adaptive time window selection module adaptively selects the corresponding time window size as the priority time window through a multi-head attention mechanism according to the time dependency representation and importance score within different time windows;
[0008] The feature map analysis module provides a local feature importance map within each time window in the priority time window through a gradient method; and performs overall statistical analysis on all time windows in the priority time window using a statistical method to generate a global feature importance map;
[0009] The feature map fusion module constructs feature maps within each time window in the priority time window according to the local feature importance map. Each node in the map represents a feature, and the weight of the node reflects the local importance of the feature within that time window; and uses the global feature importance map to correct the weight of each node; performs graph convolution operations on each corrected feature map to generate graph convolution features, and fuses the graph convolution features of all time windows to generate the final fused features.
[0010] As a further improvement of this technical solution, the process of the data analysis module initializing the time window specifically includes:
[0011] Select a time window size ; select a sliding step , which is used to control the overlapping degree of the time windows;
[0012] From the cable sensor data, slide according to the sliding step , and each time it slides time units, generate a time window with a size of , and use the cable sensor data as , then the number of time windows is: ;
[0013] Format the cable sensor data within each time window into a matrix, where is the number of features in the cable sensor data, and the data matrix within each time window is represented as , where is the index of the time window.
[0014] As a further improvement of this technical solution, the process of the data analysis module generating the time-dependent representation specifically includes:
[0015] Use the cable sensor data matrix within each time window as the input of the long short-term memory model; train the long short-term memory model using historical cable sensor data, and adjust the model parameters through the backpropagation algorithm so that the model can effectively learn the long-term and short-term dependencies of time series data;
[0016] During the training process, each time step of the long short-term memory model will generate a hidden state vector, which contains the dynamic features and dependencies of the data within the time window;
[0017] Use the hidden state vector of each time window as the time-dependent representation of this time window 。
[0018] As a further improvement of this technical solution, the importance score of the cable sensor data within different time windows of the data analysis module is calculated using the self-attention mechanism. The cable sensor data matrix within each time window and its corresponding time dependence representation are used as the input of the self-attention mechanism. A learnable weight matrix and bias term are used to calculate the attention score of each feature at each time step, and the attention weight is used as the importance score of the corresponding time window.
[0019] As a further improvement of this technical solution, the specific process of determining the priority time window by the adaptive time window selection module includes:
[0020] The multi-head attention mechanism captures different types of dependencies through multiple independent attention heads. Each attention head is responsible for calculating the correlation between different time windows and assigns a weight to each time window;
[0021] Each attention head calculates the correlation score between each time window and other time windows according to the time dependence representation and feature importance score. This score reflects the similarity and relevance between time windows;
[0022] The calculated attention scores are normalized to generate the attention weight of each time window. The normalized weight value represents the importance of each time window in the overall time series;
[0023] The attention weight generated by each attention head is multiplied by the corresponding value vector to generate the weighted value vector of each time window; the weighted value vectors of all attention heads are integrated together to generate the final time window weight vector;
[0024] The final time window weight vector is normalized to obtain the normalized weight of each time window. The normalized weight value represents the relative importance of each time window in the diagnostic result; according to the normalized weight, the time window with the highest weight is selected as the priority time window.
[0025] As a further improvement of this technical solution, the specific process of the feature map analysis module for determining the local feature importance map includes:
[0026] The cable sensor data matrix of each time window within the priority time window is input into the trained deep learning model. Through the backpropagation algorithm, the gradient of the model output with respect to each feature at each time step is calculated. This gradient reflects the influence degree of each feature on the model output;
[0027] For each time window, extract the convolutional feature maps of the last layer of the model, calculate the gradients of each feature map at each time step, and perform weighted averaging on the gradients to generate a two-dimensional feature importance map; perform this calculation for each time window in the priority time window to generate a local feature importance map.
[0028] As a further improvement of this technical solution, the process of the feature map analysis module generating the global feature importance map specifically includes:
[0029] Integrate the cable sensor data matrices in all time windows in the priority time window together, summarize the importance scores of each feature within all priority time windows, calculate the average value of the gradient values of each feature within all time windows as the global importance score;
[0030] According to the summarized feature importance scores, sort all features, determine the cable sensor data features that contribute the most to the diagnostic results within the entire priority time window, and visualize the sorted feature importance scores as a global feature importance map. The importance score of each feature in the map can be represented by the depth of color, and the darker the color, the higher the importance of the feature.
[0031] As a further improvement of this technical solution, the feature map fusion module includes a local feature importance map analysis unit. The process of the local feature importance map analysis unit constructing the feature maps within each time window in the priority time window specifically includes:
[0032] Select the data matrix of each time window in the priority time window. Within each time window, use the local feature importance map to construct the feature maps. By defining nodes, each node represents a feature, and according to the gradients of each feature in the local feature importance map, assign weights to each node. The weight of the node reflects the local importance of the feature within this time window.
[0033] As a further improvement of this technical solution, the feature map fusion module includes a correction unit. The process of the correction unit correcting the weight of each node specifically includes:
[0034] Combine the global importance score with the node weights in the local importance map, correct the weights of the nodes, use the global importance score as an adjustment factor to multiply the local importance weights, and use the calculation result as the corrected node weights, and update the feature maps within each time window.
[0035] As a further improvement of this technical solution, the process of the correction unit performing graph convolution operations and fusion specifically includes:
[0036] Perform graph convolution operations on each corrected feature map. The graph convolution operations can capture the spatial relationships and dependencies between features, generating graph convolution features for each time window. By defining a graph convolution kernel, this convolution kernel is used to perform sliding operations on the feature map; perform convolution calculations on the weights of each node and its adjacent nodes in the feature map to generate new node feature representations; summarize the results of the convolution operations to generate graph convolution feature vectors for each time window;
[0037] Fuse the graph convolution feature vectors of all time windows to generate the final fused features; by concatenating the graph convolution feature vectors of all time windows into a single feature vector, and according to the time window weights generated in the adaptive time window selection module, perform weighted averaging on the concatenated feature vector to generate the final fused feature vector.
[0038] Compared with the prior art, the beneficial effects of the present invention are:
[0039] 1. The cable digital intelligent management system initializes time windows and uses a long short-term memory model to capture the dependencies of cable sensor data within different time windows, generating time-dependent representations; meanwhile, this module uses a self-attention mechanism to calculate the importance scores of cable sensor data within different time windows, which can not only flexibly process data at different time scales but also capture the long-term and short-term dependencies of the data;
[0040] 2. The cable digital intelligent management system adaptively selects the optimal time window size as the priority time window according to the time-dependent representations and importance scores within different time windows through a multi-head attention mechanism; the multi-head attention mechanism can capture different types of time dependencies, assign reasonable weights to each time window, and thus select the priority time window that has the most influence on the diagnostic results, improving the accuracy and efficiency of the system;
[0041] 3. The cable digital intelligent management system constructs a feature map for each time window according to the local feature importance map, where each node in the map represents a feature and the weight of the node reflects the local importance of the feature within that time window; by introducing a global feature importance map, the weights of each node are corrected to further enhance the influence of key features; then, perform graph convolution operations on each corrected feature map to capture the spatial relationships and dependencies between features, generating graph convolution feature vectors for each time window; fuse the graph convolution feature vectors of all time windows to generate the final fused feature vector; this process not only considers the local and global feature importance but also can effectively integrate the information of different time windows, providing a more comprehensive and accurate fault diagnosis result, significantly improving the overall performance and reliability of the system. BRIEF DESCRIPTION OF THE DRAWINGS
[0042] Figure 1 It is a schematic diagram of the overall module of the present invention;
[0043] Figure 2 It is a schematic diagram of the feature map fusion module unit of the present invention.
[0044] In the figure: 100, data analysis module; 200, adaptive time window selection module; 300, feature map analysis module; 400, feature map fusion module; 401, local feature importance map analysis unit; 402, correction unit. Detailed implementation manners
[0045] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.
[0046] Next, please refer to Figure 1 - Figure 2 , the present invention provides a technical solution: an intelligent management system for cable digitization, including a data analysis module 100, an adaptive time window selection module 200, a feature map analysis module 300, and a feature map fusion module 400.
[0047] The process of the data analysis module 100 initializing the time window specifically includes:
[0048] Select a time window size , for example, 1 hour, 2 hours, etc.; select a sliding step , used to control the overlapping degree of the time window. For example, select 10 minutes as the sliding step;
[0049] Slide from the cable sensor data according to the sliding step , and each time it slides a time unit, generate a time window with a size of , and use the cable sensor data as , then the number of time windows is: ;
[0050] Format the cable sensor data in each time window into a matrix, where is the number of features in the cable sensor data, and the data matrix in each time window is expressed as , where is the index of the time window.
[0051] The data analysis module 100 captures the dependencies of cable sensor data within different time windows and generates a time-dependent representation, specifically including:
[0052] Taking the cable sensor data matrix within each time window as the input of the long short-term memory model; using historical cable sensor data to train the long short-term memory model, and adjusting the model parameters through the backpropagation algorithm to enable the model to effectively learn the long-term and short-term dependencies of time series data;
[0053] During the training process, each time step of the long short-term memory model generates a hidden state vector, which contains the dynamic features and dependencies of the data within the time window. Specifically, the long short-term memory model controls the flow and storage of information through its internal gating mechanisms (input gate, forget gate, and output gate) to capture the complex changes in sensor data over time;
[0054] Taking each time window 's hidden state vector as the time-dependent representation of this time window , which not only reflects the dynamic features of the data within the time window but also captures the dependencies between different time steps, providing a basis for further feature analysis and adaptive time window selection.
[0055] The data analysis module 100 calculates the importance scores of cable sensor data within different time windows, uses the self-attention mechanism to evaluate the influence of each cable sensor data on the diagnostic results, takes the cable sensor data matrix within each time window and its corresponding time-dependent representation as the input of the self-attention mechanism, uses a learnable weight matrix and bias term to calculate the attention scores of each feature at each time step, and takes the attention weights as the importance scores of the corresponding time window.
[0056] The adaptive time window selection module 200 adaptively selects the corresponding time window size as the preferred time window according to the time-dependent representations and importance scores within different time windows through the multi-head attention mechanism, specifically including:
[0057] The multi-head attention mechanism captures different types of dependencies through multiple independent attention heads. Each attention head is responsible for calculating the correlations between different time windows and assigns a weight to each time window;
[0058] Each attention head calculates the correlation score between each time window and other time windows according to the time-dependent representation and feature importance score, and this score reflects the similarity and relevance between time windows;
[0059] The calculated attention scores are normalized to generate the attention weights for each time window. The normalized weight values represent the importance of each time window in the overall time series;
[0060] Multiply the attention weights generated by each attention head with the corresponding value vectors to generate the weighted value vectors for each time window. The value vectors contain important information within the time windows; Integrate the weighted value vectors of all attention heads together to generate the final time window weight vector, which synthesizes the results of multiple attention heads and more comprehensively reflects the importance of each time window;
[0061] Normalize the final time window weight vector to obtain the normalized weights for each time window. The normalized weight values represent the relative importance of each time window in the diagnostic result; According to the normalized weights, select the time window with the highest weight as the priority time window. The time window with the highest weight has the greatest influence in the diagnostic result and can provide the most valuable information.
[0062] The feature map analysis module 300 provides local feature importance maps within each time window in the priority time window through the gradient method, specifically including:
[0063] For each time window within the priority time window, the cable sensor data matrix is input into the trained deep learning model. Through the backpropagation algorithm, calculate the gradient of the model output with respect to each feature at each time step. This gradient reflects the degree of influence of each feature on the model output;
[0064] For each time window, extract the convolutional feature map of the last layer of the model, calculate the gradient of each feature map at each time step, and perform weighted averaging on the gradients to generate a two-dimensional feature importance map; Each pixel value in the feature importance map represents the relative importance of the corresponding feature within that time window, and the darker the color, the higher the importance of the feature;
[0065] For each time window in the priority time window, generate a local feature importance map, which can help us understand and analyze the decision-making process of the model within each time window and determine which features have the greatest impact on the diagnostic result.
[0066] The feature map analysis module 300 uses statistical methods to perform overall statistical analysis on all time windows in the priority time window to generate a global feature importance map, specifically including:
[0067] Integrate the cable sensor data matrices in all time windows within the priority time window, summarize the importance scores of each feature within all priority time windows, and calculate the average value of the gradient values of each feature within all time windows as the global importance score; Integrate the cable sensor data matrices in all time windows within the priority time window, summarize the importance scores of each feature within all priority time windows, and calculate the average value of the gradient values of each feature within all time windows as the global importance score;
[0068] According to the summarized feature importance scores, sort all features to determine the cable sensor data features that contribute the most to the diagnostic results within the entire priority time window. Visualize the sorted feature importance scores as a global feature importance map, where the importance score of each feature in the map can be represented by the shade of color, and the darker the color, the higher the importance of the feature; The global feature importance map provides an overall view of the feature importance within the entire priority time window, helping us understand which features have the greatest impact on the diagnostic results as a whole.
[0069] Based on the local feature importance map, the local feature importance map analysis unit 401 in the feature map fusion module 400 constructs a feature map for each time window within the priority time window. Each node in the map represents a feature, and the weight of the node reflects the local importance of the feature within that time window, specifically including:
[0070] According to the results of the adaptive time window selection module 200, select the data matrix of each time window within the priority time window. Within each time window, use the local feature importance map to construct a feature map. By defining nodes, each node represents a feature. According to the gradient of each feature in the local feature importance map, assign weights to each node, and the weight of the node reflects the local importance of the feature within that time window; The edges in the feature map represent the correlation or dependency between features, and the weights of the edges are determined by calculating the similarity between features (such as Pearson correlation coefficient, mutual information, etc.).
[0071] The correction unit 402 in the feature map fusion module 400 uses the global feature importance map to correct the weight of each node; Perform graph convolution operations on each corrected feature map to generate graph convolution features, and fuse the graph convolution features of all time windows to generate the final fused features, specifically including:
[0072] Combine the global importance score with the node weights in the local importance map to correct the weights of the nodes. Use the global importance score as an adjustment factor to multiply the local importance weights, and use the calculation result as the corrected node weight to update the feature map within each time window, ensuring that the weight of each node reflects both local importance and global importance;
[0073] Perform graph convolution operations on each corrected feature map. The graph convolution operations can capture the spatial relationships and dependencies between features, generating graph convolution features for each time window. By defining a graph convolution kernel (or graph filter), this convolution kernel is used to perform sliding operations on the feature map; perform convolution calculations on the weights of each node and its adjacent nodes in the feature map to generate new node feature representations; summarize the results of the convolution operations to generate graph convolution feature vectors for each time window;
[0074] Fuse the graph convolution feature vectors of all time windows to generate the final fused features; by concatenating the graph convolution feature vectors of all time windows into a larger feature vector, and according to the importance of each time window in the overall time series (i.e., the time window weights generated in the adaptive time window selection module 200), perform weighted averaging on the concatenated feature vector. The generated final fused feature vector synthesizes the information within all time windows, and takes into account the importance of each time window and the dependencies between features, providing a more comprehensive and accurate feature representation for subsequent fault diagnosis and data analysis.
[0075] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. The above embodiments and descriptions in the specification are only preferred examples of the present invention and are not used to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.
Claims
1. An intelligent management system for cable digitization, characterized in that: The method comprises a data analysis module (100), an adaptive time window selection module (200), a feature map analysis module (300) and a feature map fusion module (400), wherein: The data analysis module (100) initializes the time window, captures the dependency of the cable sensor data in different time windows, generates a time dependency representation, and calculates the importance score of the cable sensor data in different time windows; The adaptive time window selection module (200) adaptively selects the corresponding time window size as the priority time window through a multi-head attention mechanism according to the time dependency representation and importance score in different time windows; The feature map analysis module (300) provides a local feature importance map in each time window in the priority time window by using a gradient method; and uses a statistical method to perform overall statistical analysis on all time windows in the priority time window to generate a global feature importance map; The feature graph fusion module (400) constructs a feature graph in each time window in the priority time window according to the local feature importance graph, wherein each node in the graph represents a feature, and the weight of the node reflects the local importance of the feature in the time window; and uses the global feature importance graph to correct the weight of each node; performs a graph convolution operation on each corrected feature graph to generate a graph convolution feature, and fuses the graph convolution features of all time windows to generate a final fused feature; the fused feature is used to improve the reliability of intelligent management of cables.
2. The cable digital intelligent management system according to claim 1 is characterized in that: The process of initializing the time window of the data analysis module (100) specifically includes: Choose a time window size ; Select a sliding step size , used to control the degree of overlap of time windows; From the cable sensor data, according to the sliding step Slide, every time you slide time units, generating a The cable sensor data is used as the time window of , then the number of time windows for: ; The cable sensor data within each time window is formatted as The matrix of is the number of features in the cable sensor data, and the data matrix in each time window is represented as ,in is the index of the time window.
3. The cable digital intelligent management system according to claim 2 is characterized in that: The process of the data analysis module (100) generating the time-dependent representation specifically includes: The cable sensor data matrix in each time window As the input of the long short-term memory model; the long short-term memory model is trained using the historical cable sensor data, and the model parameters are adjusted through the back propagation algorithm so that the model can effectively learn the long-term and short-term dependencies of the time series data; During the training process, each time step of the LSTM model generates a hidden state vector, which contains the dynamic characteristics and dependencies of the data in the time window; Each time window The hidden state vector As the time dependency of this time window .
4. The cable digital intelligent management system according to claim 3 is characterized in that: The data analysis module (100) calculates the importance scores of the cable sensor data in different time windows using a self-attention mechanism, and converts the cable sensor data matrix in each time window into And its corresponding time-dependent representation , as the input of the self-attention mechanism, using the learnable weight matrix and bias term, calculate the attention score of each feature at each time step, and use the attention weight as the importance score of the corresponding time window.
5. The cable digital intelligent management system according to claim 4 is characterized in that: The process of determining the priority time window of the adaptive time window selection module (200) specifically includes: The multi-head attention mechanism captures different types of dependencies through multiple independent attention heads. Each attention head is responsible for calculating the correlation between different time windows and assigning a weight to each time window. Each attention head calculates the correlation score between each time window and other time windows based on the time dependency representation and feature importance score. The score reflects the similarity and correlation between time windows. The calculated attention scores are normalized to generate the attention weight of each time window. The normalized weight value represents the importance of each time window in the overall time series. Multiply the attention weight generated by each attention head with the corresponding value vector to generate the weighted value vector of each time window; integrate the weighted value vectors of all attention heads together to generate the final time window weight vector; The final time window weight vector is normalized to obtain the normalized weight of each time window. The normalized weight value represents the relative importance of each time window in the diagnosis result. According to the normalized weight, the time window with the highest weight is selected as the priority time window.
6. The cable digital intelligent management system according to claim 5, characterized in that: The process of determining the local feature importance map by the feature map analysis module (300) specifically includes: The cable sensor data matrix for each time window within the priority time window Input into the trained deep learning model, and calculate the gradient of the model output for each feature at each time step through the back propagation algorithm. The gradient reflects the influence of each feature on the model output. For each time window, extract the convolutional feature map of the last layer of the model, calculate the gradient of each feature map at each time step, and perform weighted average on the gradient to generate a two-dimensional feature importance map; perform this calculation on each time window in the priority time window to generate a local feature importance map.
7. The cable digital intelligent management system according to claim 6, characterized in that: The process of generating a global feature importance map by the feature map analysis module (300) specifically includes: Matrix the cable sensor data for all time windows in the priority time window Put it together, summarize the importance scores of each feature in all priority time windows, and calculate the average of the gradient values of each feature in all time windows as the global importance score; According to the summarized feature importance scores, all features are sorted to determine the cable sensor data features that contribute most to the diagnosis results within the entire priority time window. The sorted feature importance scores are visualized as a global feature importance graph. The importance score of each feature in the graph can be represented by color depth. The darker the color, the higher the importance of the feature.
8. The cable digital intelligent management system according to claim 6, characterized in that: The feature map fusion module (400) comprises a local feature importance map analysis unit (401), and the process of the local feature importance map analysis unit (302) constructing a feature map in each time window in the priority time window specifically comprises: Each time window data matrix in the priority time window is selected. In each time window, a feature map is constructed using the local feature importance map. By defining nodes, each node represents a feature. According to the gradient of each feature in the local feature importance map, a weight is assigned to each node. The weight of the node reflects the local importance of the feature in the time window.
9. The cable digital intelligent management system according to claim 7, characterized in that: The feature graph fusion module (400) comprises a correction unit (402), and the process of the correction unit (402) correcting the weight of each node specifically comprises: Combine the global importance score with the node weight in the local importance graph, correct the node weight, use the global importance score as an adjustment factor to multiply the local importance weight, use the calculated result as the corrected node weight, and update the feature graph in each time window.
10. The cable digital intelligent management system according to claim 9, characterized in that: The process of the correction unit (402) performing graph convolution operation and fusion specifically includes: A graph convolution operation is performed on each corrected feature map. The graph convolution operation can capture the spatial relationship and dependency between features and generate graph convolution features for each time window. A graph convolution kernel is defined, which is used to perform sliding operations on the feature map. The weights of each node in the feature map and its adjacent nodes are convolved to generate a new node feature representation. The results of the convolution operation are summarized to generate a graph convolution feature vector for each time window. The graph convolution feature vectors of all time windows are fused to generate a final fused feature; the graph convolution feature vectors of all time windows are spliced into a feature vector, and the spliced feature vectors are weighted averaged according to the time window weights generated in the adaptive time window selection module (200) to generate a final fused feature vector.