Thermal power equipment real-time monitoring method and system based on edge calculation

Abstract

The invention provides a thermal power equipment real-time monitoring method and system based on edge computing, and relates to the technical field of thermal power equipment real-time monitoring, and the method comprises the steps: deploying an edge computing node array to collect multi-source heterogeneous data of thermal power equipment, and carrying out the preprocessing through data screening, multi-scale adaptive filtering and wavelet packet decomposition, a conditional variation auto-encoder is used to extract features, a hierarchical attention mechanism and a deep feature fusion network are combined to generate mixed feature representation, refined distribution estimation and abnormal mode recognition are performed on equipment states, cooperative monitoring modeling is performed based on a multi-scale spatial-temporal feature fusion network and a hierarchical depth deterministic policy gradient network, and a multi-scale spatial-temporal feature fusion network is established. The real-time monitoring accuracy and efficiency of the thermal power equipment can be effectively improved, the equipment failure rate is reduced, and safe and stable operation of a thermal power plant is guaranteed.

Description

Technical Field

[0001] The present invention relates to the technical field of real-time monitoring of thermal power equipment, and particularly to a real-time monitoring method and system for thermal power equipment based on edge computing. Background Art

[0002] The real-time monitoring of thermal power equipment is crucial for ensuring the safe and stable operation of the power system. With the continuous increase in the scale and complexity of thermal power equipment, traditional monitoring methods face many challenges. Traditional monitoring systems usually adopt a centralized architecture, transmitting all data to a central server for processing, which has problems such as large data transmission delays, high bandwidth occupancy, and the risk of single-point failures;

[0003] In addition, traditional monitoring methods are difficult to effectively process the massive heterogeneous data generated by thermal power equipment, and it is difficult to extract valuable information from it for fault diagnosis and prediction. Existing real-time monitoring technologies for thermal power equipment have problems such as difficulty in effectively processing multi-source heterogeneous data, low accuracy of anomaly detection, and lack of collaborative monitoring capabilities;

[0004] Therefore, there is an urgent need for a solution to solve the problems existing in the prior art. Summary of the Invention

[0005] Embodiments of the present invention provide a real-time monitoring method and system for thermal power equipment based on edge computing, which can at least solve some of the problems existing in the prior art.

[0006] In the first aspect of the embodiments of the present invention, a real-time monitoring method for thermal power equipment based on edge computing is provided, including:

[0007] Deploy an edge computing node array and collect multi-source heterogeneous data of thermal power equipment through a joint sensing mechanism, calculate the credibility score of the multi-source heterogeneous data based on preset quality indicators and screen to obtain valid data, transmit the valid data as original monitoring data to the control center, perform multi-scale adaptive filtering on the original monitoring data to obtain filtered data and calculate the corresponding information entropy through adaptive wavelet packet decomposition, select the optimal decomposition layer based on the information entropy criterion and decompose the filtered data into sub-signals of different frequency bands, calculate the time-frequency energy distribution and statistical features for each sub-signal, generate a threshold surface through an anomaly detection model based on the Isolation Forest and Support Vector Data Description algorithms and perform anomaly correction to obtain preprocessed data and add it to an adaptive migration network for feature alignment, construct an adjacency matrix in combination with the physical connection relationship of the equipment, extract spatio-temporal correlation features and construct a fusion feature matrix;

[0008] Add the fused feature matrix to the conditional variational autoencoder, process the equipment state features and working condition features through the dual-stream encoder, generate conditional latent variable distribution in combination with the reparameterization module, reconstruct features through the recursive neural network, calculate the reconstruction error and KL divergence and optimize the parameters of the conditional variational autoencoder to obtain conditional probability distribution features, extract features from the conditional probability distribution features based on the hierarchical attention mechanism to obtain local fine-grained features and global macro features, obtain temporal context features in combination with the adaptive weight fuser, evaluate the importance and dynamically fuse the conditional probability distribution features and the temporal context features through the deep feature fusion network, generate a hybrid feature representation and perform multi-scale prediction through recursive dilated convolution and hierarchical discriminant structure to obtain refined distribution estimation results and abnormal pattern recognition results corresponding to the equipment state;

[0009] The refined distribution estimation result and the abnormal pattern recognition result are added to the multi-scale spatiotemporal feature fusion network, the importance weights of the device state features at different time scales are calculated and the spatial dependencies are determined in combination with the graph convolutional neural network to obtain the overall state representation of the device group, the overall state representation is added to the hierarchical deep deterministic policy gradient network, the task allocation plan between edge computing nodes is generated through the main network, the resource scheduling strategy within the node is optimized through the auxiliary network, the update step size of the optimized resource scheduling strategy is adjusted through the adaptive learning rate to obtain the initial monitoring data, the initial monitoring data is probabilistically modeled by the Gaussian process regression algorithm, the modeled data is quickly retrieved and matched with abnormal patterns in combination with the Bloom filter and the locally sensitive hashing algorithm, the abnormal propagation probability is generated and combined with the time-varying graph neural network for collaborative monitoring modeling to obtain the optimal monitoring plan.

[0010] In an optional embodiment,

[0011] Deploy edge computing node arrays and collect multi-source heterogeneous data of thermal power equipment through a joint perception mechanism. Calculate the credibility score of the multi-source heterogeneous data based on pre-set quality indicators and screen to obtain valid data. Transmit the valid data as original monitoring data to the control center. Perform multi-scale adaptive filtering on the original monitoring data to obtain filtered data and calculate the corresponding information entropy through adaptive wavelet packet decomposition. Select the optimal decomposition layer based on the information entropy criterion and decompose the filtered data into sub-signals of different frequency bands. Calculate the time-frequency energy distribution and statistical characteristics for each sub-signal. Generate a threshold surface and perform anomaly correction through an anomaly detection model based on isolation forest and support vector data description algorithm. Obtain preprocessed data and add it to the adaptive migration network for feature alignment. Construct an adjacency matrix based on the physical connection relationship of the equipment. Extract spatiotemporal correlation features and construct a fusion feature matrix, including:

[0012] Deploy an edge computing node array, achieve time alignment among nodes by configuring the clock synchronization module of the edge computing node array, schedule the vibration sensors, temperature sensors, pressure sensors, and displacement sensors in the edge computing node array to work together to form a joint sensing mechanism, collect multi-source heterogeneous data of thermal power equipment, and normalize each type of data in the multi-source heterogeneous data;

[0013] Traverse the sampling sequence of the multi-source heterogeneous data according to the pre-set quality indicators, count the number of missing consecutive sampling points to mark invalid segments, calculate the change rate by dividing the numerical difference between adjacent sampling points by the sampling interval and compare it with the standard deviation of historical data to mark suspicious points, record the time difference from the acquisition end to the processing end of the data to obtain the transmission delay and mark delay points, multiply each marking result by the corresponding weight and superimpose and calculate to obtain the credibility score;

[0014] Transmit the valid data with a credibility score exceeding the preset threshold to the control center, divide the processing window of the original monitoring data by a fixed number of points, calculate the mean square value of the signal within each processing window to obtain the signal variance, calculate the mean square value of the residual to obtain the noise variance, divide the signal variance by the noise variance to obtain the signal-to-noise ratio, and adaptively adjust the cut-off frequency and order of the filter according to the signal-to-noise ratio to perform multi-scale adaptive filtering;

[0015] Perform wavelet packet decomposition on the filtered data to obtain wavelet coefficients of each layer, calculate the probability distribution of the wavelet coefficients of each layer to obtain the information entropy, subtract the information entropy of the corresponding sub-band of the (k - 1)th layer from the information entropy of the wavelet coefficients of the kth layer to obtain the entropy value difference, compare the entropy value difference with the preset threshold based on the information entropy criterion, when the entropy value difference is less than the preset threshold, determine the optimal decomposition layer of the sub-band and stop decomposition, and decompose the filtered data into sub-signals of different frequency bands;

[0016] Calculate the mean, standard deviation, skewness, kurtosis, waveform factor, and margin factor of each sub-signal to obtain time-domain statistical features, calculate the power spectrum by fast Fourier transform and integrate to obtain time-frequency energy distribution, and extract the frequency band energy ratio, main frequency component, and spectral kurtosis to obtain frequency-domain statistical features;

[0017] Input the time-frequency energy distribution and statistical features into an isolation forest detector, construct a decision tree by randomly selecting features and split points to calculate the isolation depth of the sample to obtain the anomaly score, input the anomaly score as the sample weight into the support vector data description algorithm, map the sample to a high-dimensional feature space through a kernel function and solve a quadratic programming problem to generate a minimum enclosing hypersphere as the threshold surface, and perform anomaly correction on the abnormal points exceeding the threshold surface by local mean replacement to obtain preprocessed data;

[0018] Input the preprocessed data into a feature extractor composed of a multi-layer perceptron network, extract deep feature representations through a non-linear activation function and fully connected layers, input the deep feature representations into a domain classifier to calculate the feature distribution differences between the source domain and the target domain, and alternately optimize the feature extraction loss function and the domain discrimination loss function to achieve feature alignment;

[0019] Construct an initial adjacency matrix based on the physical pipeline connections and control logic relationships of the devices, calculate the conditional probability of the feature sequences between device nodes to obtain the mutual information, calculate the transition probability of the feature sequences to obtain the transfer entropy, update the connection relationships in the adjacency matrix according to the values of the mutual information and the transfer entropy, extract spatio-temporal correlation features, and combine the feature vectors of each device node with the weighted feature vectors of its adjacent nodes to construct a fusion feature matrix.

[0020] In an alternative embodiment,

[0021] Construct an initial adjacency matrix based on the physical pipeline connections and control logic relationships of the devices, calculate the conditional probability of the feature sequences between device nodes to obtain the mutual information, calculate the transition probability of the feature sequences to obtain the transfer entropy, update the connection relationships in the adjacency matrix according to the values of the mutual information and the transfer entropy, and the spatio-temporal correlation features extracted include:

[0022] Obtain the physical pipeline connection diagram and the control system logic diagram of the thermal power plant equipment, extract the water pipe connection relationship, the steam pipe connection relationship and the oil pipe connection relationship from the physical pipeline connection diagram to form a physical connection relationship, extract the cascade control relationship, the feedforward control relationship and the decoupling control relationship from the control system logic diagram to form a control association relationship, merge the physical connection relationship and the control association relationship to construct an initial adjacency matrix, and assign the value of 1 to the corresponding adjacency matrix element when there is the physical connection relationship or the control association relationship between adjacent device nodes, and assign the value of 0 to the corresponding adjacency matrix element when there is no physical connection relationship and no control association relationship;

[0023] Select a time window with a preset time length to perform probability statistics on the feature sequences of adjacent device nodes, count the joint occurrence frequency of the feature sequences in a preset value range and normalize it to obtain a joint probability distribution, count the individual occurrence frequency of the feature sequences in a preset value range and normalize it to obtain a marginal probability distribution, and calculate the conditional probability based on the joint probability distribution and the marginal probability distribution to obtain a mutual information value;

[0024] Perform delay processing on the feature sequences at a preset time interval to obtain delayed feature sequences, count the occurrence frequency of the feature values at the current moment of the feature sequences and the feature values of the delayed feature sequences and normalize it to obtain a transition probability distribution, and calculate the transfer entropy value based on the transition probability distribution;

[0025] Comparing the mutual information value with a first preset threshold to obtain a first comparison result, comparing the transfer entropy value with a second preset threshold to obtain a second comparison result, establishing a connection relationship in the initial adjacency matrix when the first comparison result is greater than the first preset threshold and the second comparison result is greater than the second preset threshold, and writing the sum of the mutual information value multiplied by a first weight coefficient and the transfer entropy value multiplied by a second weight coefficient as a connection weight value into the initial adjacency matrix to obtain an updated adjacency matrix;

[0026] Extract the feature vector of each device node as its own feature, obtain the feature vectors of all device nodes adjacent to each device node as adjacent features, weight the adjacent features according to the connection weight values ​​in the updated adjacency matrix to obtain weighted features, and combine the own features with the weighted features to obtain multiple spatiotemporal correlation features.

[0027] In an optional embodiment,

[0028] The fused feature matrix is ​​added to the conditional variational autoencoder, the equipment state features and working condition features are processed by the dual-stream encoder, the conditional latent variable distribution is generated by combining the reparameterization module, the feature reconstruction is performed by the recursive neural network, the reconstruction error and KL divergence are calculated and the parameters of the conditional variational autoencoder are optimized to obtain the conditional probability distribution features, the conditional probability distribution features are extracted based on the hierarchical attention mechanism to obtain local fine-grained features and global macro features, and the temporal context features are obtained by combining the adaptive weight fuser, the importance of the conditional probability distribution features and the temporal context features are evaluated and dynamically fused by the deep feature fusion network, a hybrid feature representation is generated and multi-scale prediction is performed by recursive dilated convolution and hierarchical discriminant structure, and the refined distribution estimation results and abnormal pattern recognition results corresponding to the equipment state are obtained, including:

[0029] Add the fused feature matrix to the conditional variational autoencoder, process the device state feature and the working condition feature through the dual-stream encoder in the conditional variational autoencoder, the first stream of the dual-stream encoder maps the device state feature to the first latent space to generate the first latent space feature, the second stream of the dual-stream encoder maps the working condition feature to the second latent space to generate the second latent space feature, the first latent space feature and the second latent space feature are spliced ​​and input into the reparameterization module;

[0030] The reparameterization module generates a mean vector and a variance vector, samples a noise vector, multiplies the noise vector by the variance vector and then superimposes the mean vector to generate a conditional latent variable distribution, and inputs the conditional latent variable distribution into a recursive neural network for feature reconstruction;

[0031] Calculate the reconstruction error between the fused feature matrix and the reconstructed feature, calculate the KL divergence between the conditional latent variable distribution and the standard normal distribution, combine the reconstruction error and the KL divergence to form a loss function, and optimize the parameters of the conditional variational autoencoder to obtain conditional probability distribution features;

[0032] The conditional probability distribution features are extracted based on the hierarchical attention mechanism, the feature sequence is segmented by a sliding window through the local attention layer, the attention score and feature weight of the time point in the window are calculated to obtain the local fine-grained features, the feature sequence is divided into time segments through the global attention layer, and the attention score and feature weight of the time segment are calculated to obtain the global macro features;

[0033] Combined with an adaptive weight fusion device, statistical characteristics of the local fine-grained features and the global macro features are extracted to generate fusion weights, and the fusion weights are weighted with the local fine-grained features and the global macro features to obtain temporal context features. The conditional probability distribution features and the temporal context features are evaluated for importance through a deep feature fusion network to generate dynamic fusion weights, and the features are multiplied by the dynamic fusion weights and then connected in series to generate a hybrid feature representation;

[0034] The hybrid feature representation is input into the recursive dilated convolution and hierarchical discriminant structure for multi-scale prediction. The receptive field is expanded by multiple layers of dilated convolution to obtain multi-scale features. The output of each layer of dilated convolution is connected to the discriminator for state recognition. The output of each layer of discriminator is weighted to obtain the refined distribution estimation result and abnormal pattern recognition result corresponding to the device state.

[0035] In an optional embodiment,

[0036] The hybrid feature representation is input into the recursive dilated convolution and hierarchical discriminant structure for multi-scale prediction. The receptive field range is expanded by multi-layer dilated convolution to obtain multi-scale features. The output of each layer of dilated convolution is connected to the discriminator for state recognition. The output of each layer of discriminator is weighted to obtain the refined distribution estimation result and abnormal pattern recognition result corresponding to the device state, including:

[0037] Normalizing the mixed feature representation, dividing the mixed feature representation into equipment state parameter features and operating condition parameter features, generating a normalized feature vector, constructing a recursive dilated convolution and hierarchical discriminant structure, and setting the dilation rates of the multi-layer dilated convolution layers to increase in sequence to expand the receptive field range;

[0038] Input the normalized eigenvector into the first dilated convolutional layer, select the eigenvectors at each time point and adjacent time points in the input sequence for convolutional operation to generate the first multi-scale features, input the first multi-scale features into the first discriminator, and perform feature dimensionality reduction and state mapping through the multi-layer fully connected network in the first discriminator to generate the first state probability distribution;

[0039] Input the normalized eigenvector into the second dilated convolutional layer, select the eigenvectors at each time point and their dilated interval positions based on the second dilation rate for convolutional operation to generate the second multi-scale features, input the second multi-scale features into the second discriminator for state recognition to generate the second state probability distribution;

[0040] Input the normalized eigenvector into different dilated convolutional layers in sequence, select the eigenvectors at time points and their corresponding dilated interval positions based on the increasing dilation rate for convolutional operation, generate the corresponding multi-scale features for each layer, input the multi-scale features into the corresponding discriminators for state recognition to generate the state probability distributions for each layer;

[0041] Calculate the entropy value of the state probability distribution output by each discriminator to obtain the prediction reliability index, determine the fusion weights of different state probability distributions based on the prediction reliability index, multiply the fusion weights by the corresponding state probability distributions and sum them to generate the refined distribution estimation result corresponding to the device state;

[0042] Set multiple state probability thresholds, compare the refined distribution estimation result with the multiple state probability thresholds to divide the state levels, input the divided state level information into the abnormal pattern recognizer for abnormal type recognition, and output the abnormal pattern recognition result.

[0043] In an alternative embodiment,

[0044] Add the refined distribution estimation result and the abnormal pattern recognition result to the multi-scale spatio-temporal feature fusion network, calculate the importance weights of the device state features at different time scales and determine the spatial dependence relationship in combination with the graph convolutional neural network to obtain the overall state representation of the device group, add the overall state representation to the hierarchical deep deterministic policy gradient network, generate the task allocation scheme among edge computing nodes through the main network, optimize the resource scheduling strategy within the nodes through the auxiliary network, adjust the update step size of the optimized resource scheduling strategy through the adaptive learning rate to obtain the initial monitoring data, perform probability modeling on the initial monitoring data through the Gaussian process regression algorithm, and perform fast retrieval and abnormal pattern matching on the modeled data in combination with the Bloom filter and the locality-sensitive hashing algorithm to generate the abnormal propagation probability and perform collaborative monitoring modeling in combination with the time-varying graph neural network to obtain the optimal monitoring scheme including:

[0045] Add the refined distribution estimation results and anomaly pattern recognition results to the multi-scale spatio-temporal feature fusion network, construct multiple temporal convolutional branches to extract state features at different temporal scales, select the state data of the most recent twenty-four time points for one-dimensional convolution to obtain hourly-level features, select the state data of the corresponding moments in the most recent seven days for one-dimensional convolution to obtain daily-level features, select the state data of the corresponding moments in the most recent four weeks for one-dimensional convolution to obtain weekly-level features, map the hourly-level features to generate hourly-level query vectors, hourly-level key vectors, and hourly-level value vectors, map the daily-level features to generate daily-level query vectors, daily-level key vectors, and daily-level value vectors, map the weekly-level features to generate weekly-level query vectors, weekly-level key vectors, and weekly-level value vectors, calculate the similarity between the hourly-level query vector and the hourly-level key vector to obtain the hourly-level attention score, calculate the similarity between the daily-level query vector and the daily-level key vector to obtain the daily-level attention score, calculate the similarity between the weekly-level query vector and the weekly-level key vector to obtain the weekly-level attention score, and generate an attention weight matrix for the corresponding temporal scale according to the hourly-level attention score, the daily-level attention score, and the weekly-level attention score;

[0046] Encode the connection relationship between devices as an adjacency matrix, set the element values of the adjacency matrix to represent the association strength between devices, perform an aggregation operation on the node features and neighbor node features through multiple graph convolutional layers, and implement a residual connection on the graph convolutional features of different layers to obtain spatial dimension features representing spatial dependence relationships;

[0047] Fuse the temporal dimension features corresponding to the attention weight matrix and the spatial dimension features, use a multi-layer perceptron to adaptively adjust the fusion ratio of the temporal dimension features and the spatial dimension features to obtain the overall state representation of the device group, add the overall state representation to the hierarchical deep deterministic policy gradient network, map the overall state representation to a low-dimensional feature space through the four-layer fully connected structure of the main network and extract task-related features, generate a task allocation plan based on the task-related features, add the task allocation plan to the auxiliary network, calculate the resource requirements, storage resource requirements, and communication bandwidth requirements according to the task-related features, generate a task priority allocation strategy and optimize the resource scheduling strategy within the node, iteratively optimize the resource scheduling strategy, monitor the historical change trend of the policy gradient, increase the learning rate when the policy gradients in consecutive multiple iterations are in the same direction, decrease the learning rate when the policy gradient direction changes frequently, obtain the initial monitoring data, perform normalization processing on the initial monitoring data, calculate the kernel function similarity matrix between samples and establish a joint probability distribution model of the monitoring data;

[0048] Map the feature vectors of the monitoring data to the bitmap of the Bloom filter for coarse-grained screening, map similar feature vectors to the same or adjacent hash buckets for abnormal pattern matching, establish a conditional probability model based on the physical association relationship and state correlation between devices, and use the probability propagation algorithm to estimate the abnormal diffusion risk and generate the abnormal propagation probability;

[0049] Use a gated recurrent unit to model the time dynamic evolution process of edge weights, combine static weights and dynamic weights to generate edge weight values, use the static weights to represent the physical coupling strength between devices, use the dynamic weights to represent the degree of correlation of device states and the abnormal propagation probability, map node features to multiple attention heads, calculate the attention scores between nodes using the multiple attention heads respectively, determine the transmission strength of information in the graph structure according to the attention scores, use a long short-term memory network structure to predict the future state of nodes, and obtain the optimal monitoring scheme according to the prediction results and the current system state.

[0050] In an alternative embodiment,

[0051] Mapping the feature vectors of the monitoring data to the bitmap of the Bloom filter for coarse-grained screening, mapping similar feature vectors to the same or adjacent hash buckets for abnormal pattern matching, establishing a conditional probability model based on the physical association relationship and state correlation between devices, and using the probability propagation algorithm to estimate the abnormal diffusion risk and generate the abnormal propagation probability includes:

[0052] Set a dynamically expandable bitmap space and multiple independent hash functions based on the Bloom filter, set the expansion threshold of the dynamically expandable bitmap space, normalize and extract features from the monitoring data to obtain feature vectors, map the feature vectors to the dynamically expandable bitmap space through the hash functions for coarse-grained screening, detect the usage rate of the dynamically expandable bitmap space, and when the usage rate exceeds the expansion threshold, expand the dynamically expandable bitmap space according to a preset expansion ratio and remap the existing data;

[0053] Divide the multi-level hash structure into a first-level hash structure and a second-level hash structure, set multiple main hash buckets in the first-level hash structure, configure multiple sub-hash buckets for each main hash bucket in the second-level hash structure, construct an orthogonal basis vector group of the feature space, project the feature vectors to the orthogonal basis vector group to obtain projection coefficients, and map the feature vectors with similar features to the same or adjacent hash buckets according to the similarity of the projection coefficients for abnormal pattern matching;

[0054] Establish a multi - layer device association graph based on the physical association relationship between devices, record device identifiers and device status parameters in the nodes of the multi - layer device association graph, set a time decay factor to weight the historical status data of adjacent device pairs, dynamically adjust the time decay factor according to a preset time interval, discretize the state space into a preset number of levels and construct a conditional probability model;

[0055] Use the probability propagation algorithm to estimate the abnormal diffusion risk, divide the abnormal propagation process, set a first propagation depth for short - range propagation and use real - time updated conditional probability values, set a second propagation depth for long - range propagation and introduce a distance decay coefficient, set different probability decay coefficients according to the number of propagation layers, calculate the abnormal propagation probability of each propagation path, and stop the propagation of the corresponding propagation path when the abnormal propagation probability is lower than a preset cut - off threshold.

[0056] In the second aspect of the embodiments of the present invention, a real - time monitoring system for thermal power equipment based on edge computing is provided, including:

[0057] A first unit for deploying an edge - computing node array and collecting multi - source heterogeneous data of thermal power equipment through a joint sensing mechanism, calculating the credibility score of the multi - source heterogeneous data based on pre - set quality indicators and screening to obtain valid data, transmitting the valid data as original monitoring data to a control center, performing multi - scale adaptive filtering on the original monitoring data, obtaining filtered data and calculating the corresponding information entropy through adaptive wavelet packet decomposition, selecting the optimal decomposition layer based on the information entropy criterion and decomposing the filtered data into sub - signals of different frequency bands, calculating the time - frequency energy distribution and statistical features for each sub - signal, generating a threshold surface through an anomaly detection model based on the isolation forest and support vector data description algorithms and performing anomaly correction, obtaining pre - processed data and adding it to an adaptive migration network for feature alignment, constructing an adjacency matrix in combination with the device physical connection relationship, and extracting spatio - temporal correlation features and constructing a fusion feature matrix;

[0058] The second unit is used to add the fused feature matrix to the conditional variational autoencoder, process the equipment state features and working condition features through the dual-stream encoder, generate conditional latent variable distribution in combination with the reparameterization module, reconstruct features through a recursive neural network, calculate the reconstruction error and KL divergence and optimize the parameters of the conditional variational autoencoder to obtain conditional probability distribution features, extract features of the conditional probability distribution features based on the hierarchical attention mechanism to obtain local fine-grained features and global macro features, obtain temporal context features in combination with an adaptive weight fuser, evaluate the importance and dynamically fuse the conditional probability distribution features and the temporal context features through a deep feature fusion network, generate a hybrid feature representation, and perform multi-scale prediction through recursive dilated convolution and hierarchical discriminant structures to obtain refined distribution estimation results and abnormal pattern recognition results corresponding to the equipment state;

[0059] The third unit is used to add the refined distribution estimation results and abnormal pattern recognition results to the multi-scale spatiotemporal feature fusion network, calculate the importance weights of the device state features at different time scales and determine the spatial dependencies in combination with the graph convolutional neural network to obtain the overall state representation of the device group, add the overall state representation to the hierarchical deep deterministic policy gradient network, generate the task allocation plan between edge computing nodes through the main network, optimize the resource scheduling strategy within the node through the auxiliary network, adjust the update step size of the optimized resource scheduling strategy through the adaptive learning rate to obtain initial monitoring data, perform probability modeling on the initial monitoring data through the Gaussian process regression algorithm, combine the Bloom filter and the local sensitive hashing algorithm to quickly retrieve and match the modeled data, generate the abnormal propagation probability and combine the time-varying graph neural network for collaborative monitoring modeling to obtain the optimal monitoring plan.

[0060] According to a third aspect of the embodiments of the present invention,

[0061] An electronic device is provided, comprising:

[0062] processor;

[0063] a memory for storing processor-executable instructions;

[0064] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0065] A fourth aspect of the embodiments of the present invention is:

[0066] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0067] In the present invention, through multi-scale adaptive filtering, selecting the optimal decomposition layer number by the information entropy criterion, and an anomaly detection model based on the isolation forest and support vector data description algorithm, the influence of noise and abnormal data on the monitoring results is effectively reduced, and the refined distribution estimation of the device state is realized. By using the conditional variational autoencoder, hierarchical attention mechanism, and deep feature fusion network technology, the deep features of the device state and working conditions can be extracted, and multi-scale prediction is carried out by combining recursive dilated convolution and hierarchical discriminant structure, so as to realize the accurate identification of abnormal patterns. By using the multi-scale spatio-temporal feature fusion network, hierarchical deep deterministic policy gradient network, Gaussian process regression algorithm, Bloom filter, and locality-sensitive hashing algorithm, the optimization of task allocation between edge computing nodes and the optimization of resource scheduling strategies within nodes are realized, and the optimal monitoring scheme is obtained through collaborative monitoring modeling, thereby improving the monitoring efficiency and reducing resource consumption. Description of the Drawings

[0068] Figure 1 It is a schematic flow chart of the real-time monitoring method for thermal power equipment based on edge computing according to an embodiment of the present invention;

[0069] Figure 2 It is a schematic structural diagram of the real-time monitoring system for thermal power equipment based on edge computing according to an embodiment of the present invention. Detailed Embodiments

[0070] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0071] The technical solutions of the present invention will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0072] Figure 1 It is a schematic flow chart of the real-time monitoring method for thermal power equipment based on edge computing according to an embodiment of the present invention, as Figure 1 shown, and the method includes:

[0073] S1. Deploy an edge computing node array and collect multi-source heterogeneous data of thermal power equipment through a joint sensing mechanism. Calculate the credibility score of the multi-source heterogeneous data based on pre-set quality indicators, screen out the valid data, and transmit the valid data to the control center as the original monitoring data. Perform multi-scale adaptive filtering on the original monitoring data to obtain filtered data, calculate the corresponding information entropy through adaptive wavelet packet decomposition, select the optimal decomposition level based on the information entropy criterion, and decompose the filtered data into sub-signals of different frequency bands. Calculate the time-frequency energy distribution and statistical features for each sub-signal, generate a threshold surface through an anomaly detection model based on the Isolation Forest and Support Vector Data Description algorithms, and perform anomaly correction to obtain preprocessed data and add it to the adaptive transfer network for feature alignment. Construct an adjacency matrix in combination with the physical connection relationship of the equipment, extract spatio-temporal correlation features, and construct a fusion feature matrix;

[0074] The joint sensing mechanism is a mechanism that fuses and processes multiple sensing information sources, usually used to improve the sensing accuracy and robustness of the system. The multi-scale adaptive filtering is a technology that can adaptively adjust filtering parameters at different scales, used to process multi-resolution data, reduce noise interference, and improve feature extraction ability. The information entropy is a quantity used to measure the uncertainty or randomness of a system or information source, often used to evaluate the richness of information. The information entropy criterion is a decision criterion based on information entropy, which optimizes the system design or decision-making process by minimizing the entropy value of the system or maximizing the information gain. The Isolation Forest is an unsupervised learning method based on tree structure, mainly used for anomaly detection. It constructs isolation trees to evaluate the isolation degree of data points and then discovers abnormal data. The Support Vector Data Description algorithm is an unsupervised learning method based on the idea of support vector machines, used to find the decision boundary of data in high-dimensional space, mainly used for anomaly detection and data classification. The threshold surface is a geometric shape used to represent the decision boundary or response change under different parameter thresholds, usually used in decision analysis and model optimization. The adaptive transfer network is a neural network architecture that can perform adaptive transfer between different tasks or domains, used to solve cross-domain learning problems.

[0075] In an alternative embodiment,

[0076] Deploy an edge computing node array and collect multi-source heterogeneous data of thermal power equipment through a joint sensing mechanism. Calculate the credibility score of the multi-source heterogeneous data based on pre-set quality metrics and filter to obtain valid data. Transmit the valid data as original monitoring data to the control center, perform multi-scale adaptive filtering on the original monitoring data, obtain filtered data and calculate the corresponding information entropy through adaptive wavelet packet decomposition. Select the optimal decomposition level based on the information entropy criterion and decompose the filtered data into sub-signals of different frequency bands. Calculate the time-frequency energy distribution and statistical features for each sub-signal, generate a threshold surface through an anomaly detection model based on the Isolation Forest and Support Vector Data Description algorithms and perform anomaly correction, obtain preprocessed data and add it to the adaptive migration network for feature alignment, construct an adjacency matrix in combination with the physical connection relationship of the equipment, and extract spatio-temporal correlation features and construct a fusion feature matrix including:

[0077] Deploy an edge computing node array, achieve time alignment between nodes by configuring the clock synchronization module of the edge computing node array, schedule the vibration sensors, temperature sensors, pressure sensors, and displacement sensors in the edge computing node array to work collaboratively to form a joint sensing mechanism, collect multi-source heterogeneous data of thermal power equipment, and normalize each type of data in the multi-source heterogeneous data;

[0078] Traverse the sampling sequence of the multi-source heterogeneous data according to the pre-set quality metrics, count the number of missing consecutive sampling points to mark invalid segments, calculate the numerical difference between adjacent sampling points divided by the sampling interval to obtain the change rate and compare it with the standard deviation of historical data to mark suspicious points, record the time difference from the acquisition end to the processing end of the data to obtain the transmission delay and mark delay points, multiply each marking result by the corresponding weight and sum them up to calculate the credibility score;

[0079] Transmit the valid data with a credibility score exceeding the preset threshold as the original monitoring data to the control center, divide the processing window of the original monitoring data with a fixed number of points, calculate the mean square value of the signal in each processing window to obtain the signal variance, calculate the mean square value of the residual to obtain the noise variance, divide the signal variance by the noise variance to obtain the signal-to-noise ratio, and adaptively adjust the cut-off frequency and order of the filter according to the signal-to-noise ratio to perform multi-scale adaptive filtering;

[0080] Perform wavelet packet decomposition on the filtered data to obtain wavelet coefficients of each layer, calculate the probability distribution of the wavelet coefficients of each layer to obtain the information entropy, subtract the information entropy of the corresponding sub-band of the (k - 1)-th layer from the information entropy of the k-th layer wavelet coefficients to obtain the entropy value difference, compare the entropy value difference with the preset threshold based on the information entropy criterion, when the entropy value difference is less than the preset threshold, determine the optimal decomposition level of the sub-band and stop decomposition, and decompose the filtered data into sub-signals of different frequency bands;

[0081] Calculate the mean, standard deviation, skewness, kurtosis, waveform factor, and margin factor of each sub-signal to obtain time-domain statistical features. Calculate the power spectrum through the fast Fourier transform and integrate it to obtain the time-frequency energy distribution. Extract the proportion of band energy, main frequency component, and spectral kurtosis to obtain frequency-domain statistical features;

[0082] Input the time-frequency energy distribution and statistical features into the isolation forest detector. Construct decision trees by randomly selecting features and split points to calculate the isolation depth of samples to obtain the anomaly score. Input the anomaly score as the sample weight into the support vector data description algorithm. Map the samples to a high-dimensional feature space through the kernel function and solve the quadratic programming problem to generate the minimum enclosing hypersphere as the threshold surface. Perform anomaly correction on the outlier points beyond the threshold surface through local mean replacement to obtain the preprocessed data;

[0083] Input the preprocessed data into the feature extractor composed of a multi-layer perceptron network. Extract the deep feature representation through the non-linear activation function and the fully connected layer. Input the deep feature representation into the domain classifier to calculate the feature distribution difference between the source domain and the target domain, and alternately optimize the feature extraction loss function and the domain discrimination loss function to achieve feature alignment;

[0084] Construct an initial adjacency matrix based on the physical pipeline connections and control logic relationships of the devices. Calculate the conditional probability of the feature sequences between device nodes to obtain the mutual information. Calculate the transition probability of the feature sequences to obtain the transfer entropy. Update the connection relationships in the adjacency matrix according to the values of the mutual information and the transfer entropy. Extract the spatio-temporal correlation features and combine the feature vectors of each device node with the weighted feature vectors of its adjacent nodes to construct a fusion feature matrix.

[0085] The fast Fourier transform is an efficient algorithm for calculating the discrete Fourier transform and is widely used in fields such as signal processing and image processing. The power spectrum is a representation method in the frequency-domain analysis of signals, used to display the power distribution of different frequency components and is often used to analyze the frequency characteristics of signals. The isolation depth is a metric in the isolation forest algorithm, used to measure the degree to which a data point is isolated. The smaller the depth, the easier it is for the data point to be isolated and is usually used for anomaly detection. The quadratic programming problem is a type of optimization problem where the objective function is a quadratic function and the constraint conditions are linear functions, and is widely applied in optimization and machine learning. The minimum enclosing hypersphere is a concept in geometry, representing the smallest hypersphere that can enclose a set of data points and is often used in data clustering and geometric optimization problems. The domain classifier is a machine learning model used to determine which domain the data belongs to and is widely applied in transfer learning and domain adaptation.

[0086] Deploy an edge computing node array. Select a suitable installation location, integrate various sensors such as vibration, temperature, pressure, and displacement sensors into each edge computing node, and ensure the stable and reliable power supply and network connection for each node. To achieve time synchronization between nodes, configure a clock synchronization module, such as using methods like the Network Time Protocol (NTP) or the Precision Time Protocol (PTP), to keep the time of all nodes consistent and eliminate data errors caused by time deviation.

[0087] Collect multi-source heterogeneous data of thermal power equipment through a joint sensing mechanism, and schedule various sensors in the edge computing node array to work together through a preset acquisition plan. For example, when it is necessary to monitor the operating status of a certain device, the vibration, temperature, and pressure sensors near the device can be started simultaneously for data acquisition. To ensure the comparability of different types of data, normalize each type of collected data. For example, scale the data to the range of 0 to 1. For instance, assuming the collected temperature data is 80°C and 100°C, the normalized data is 0.8 and 1 respectively.

[0088] Evaluate the credibility of the collected multi-source heterogeneous data. According to the preset quality indicators, screen the data and remove invalid data. For example, the number of missing consecutive sampling points can be counted. If the missing number exceeds the preset threshold, this segment of data is marked as an invalid segment. The ratio of the numerical difference between adjacent sampling points to the sampling interval, that is, the change rate, can also be calculated and compared with the standard deviation of historical data. If the change rate is too large, this sampling point is marked as a suspicious point. In addition, the time difference from the data collection end to the processing end, that is, the transmission delay, can be recorded. If the transmission delay is too long, this data is marked as a delayed point. Multiply each marking result by the corresponding weight and sum them up to obtain the credibility score of each data. For example, assuming the missing weight is 0.5, the change rate weight is 0.3, and the delay weight is 0.2, and the missing marking of a certain data is 1, the change rate marking is 0, and the delay marking is 1, then the credibility score of this data is 0.5×1 + 0.3×0 + 0.2×1 = 0.7.

[0089] Transmit the valid data with a credibility score exceeding the preset threshold to the control center as the original monitoring data. At the control center, divide the original monitoring data into processing windows with a fixed number of points. For example, every 1000 data points form a processing window. Within each processing window, calculate the mean square value and the residual mean square value of the signal to obtain the signal variance and the noise variance respectively. Divide the signal variance by the noise variance to get the signal-to-noise ratio. Adaptively adjust the cut-off frequency and order of the filter according to the signal-to-noise ratio, and perform multi-scale adaptive filtering on the data to remove noise interference.

[0090] Perform adaptive wavelet packet decomposition on the filtered data to obtain wavelet coefficients at each layer, calculate the probability distribution of the wavelet coefficients at each layer to obtain the information entropy. Subtract the information entropy of the corresponding subband at the (k - 1)-th layer from the information entropy of the wavelet coefficients at the k-th layer to obtain the entropy value difference. Based on the information entropy criterion, compare the entropy value difference with a preset threshold. If the entropy value difference is less than the preset threshold, determine the optimal decomposition level of the subband and stop the decomposition. Finally, decompose the filtered data into sub-signals in different frequency bands.

[0091] For each sub-signal, calculate its time-frequency energy distribution and statistical features. In the time domain, calculate statistical features such as mean, standard deviation, skewness, kurtosis, waveform factor, and margin factor. In the frequency domain, calculate the power spectrum through fast Fourier transform and integrate to obtain the time-frequency energy distribution, and extract statistical features such as the proportion of band energy, main frequency component, and spectral kurtosis.

[0092] Construct an anomaly detection model using the Isolation Forest and Support Vector Data Description algorithms. Input the calculated time-frequency energy distribution and statistical features into the Isolation Forest detector, calculate the anomaly score for each sample, and input the anomaly score as the sample weight into the Support Vector Data Description algorithm to generate the minimum enclosing hypersphere as the threshold surface. For the anomaly points beyond the threshold surface, use local mean replacement for anomaly correction to obtain the preprocessed data.

[0093] Input the preprocessed data into the Adaptive Transfer Network for feature alignment. The Adaptive Transfer Network consists of a multi-layer perceptron network, and extracts deep feature representations through non-linear activation functions and fully connected layers. Then, input the deep feature representations into the domain classifier, calculate the feature distribution difference between the source domain and the target domain, and alternately optimize the feature extraction loss function and the domain discrimination loss function to achieve feature alignment.

[0094] Construct a fused feature matrix. Based on the physical pipeline connections and control logic relationships of the devices, construct an initial adjacency matrix. Calculate the conditional probability of the feature sequences between device nodes to obtain the mutual information, and calculate the transition probability of the feature sequences to obtain the transfer entropy. Update the connection relationships in the adjacency matrix according to the values of the mutual information and the transfer entropy. Extract spatio-temporal correlation features, and combine the feature vectors of each device node with the weighted feature vectors of its adjacent nodes to construct a fused feature matrix for subsequent fault diagnosis or state assessment.

[0095] In this embodiment, through multi-source heterogeneous data fusion and credibility evaluation, noise and abnormal data are effectively removed, the data quality is improved, providing a reliable data basis for subsequent analysis. Through multi-scale adaptive filtering, wavelet packet decomposition, and the Adaptive Transfer Network, more representative features are extracted, enhancing the feature expression ability and improving the accuracy of fault diagnosis and state assessment. Through feature alignment and spatio-temporal correlation feature extraction, the correlation relationships between devices are considered, enhancing the generalization ability of the model, enabling it to adapt to different working conditions and different devices.

[0096] In an alternative embodiment,

[0097] Construct an initial adjacency matrix based on the physical pipeline connections and control logic relationships of the devices, calculate the conditional probability of the characteristic sequences between device nodes to obtain the mutual information, calculate the transition probability of the characteristic sequences to obtain the transfer entropy, and update the connection relationships in the adjacency matrix according to the values of the mutual information and transfer entropy. The spatio-temporal correlation features extracted include:

[0098] Obtain the physical pipeline connection diagram and the control system logic diagram of the thermal power plant equipment. Extract the water pipe connection relationship, steam pipe connection relationship, and oil pipe connection relationship from the physical pipeline connection diagram to form a physical connection relationship. Extract the cascade control relationship, feedforward control relationship, and decoupling control relationship from the control system logic diagram to form a control association relationship. Combine the physical connection relationship and the control association relationship to construct an initial adjacency matrix. When there is the physical connection relationship or the control association relationship between adjacent device nodes, the corresponding adjacency matrix element is assigned a value of 1. When there is no physical connection relationship and no control association relationship, the corresponding adjacency matrix element is assigned a value of 0;

[0099] Select a time window with a preset time length to perform probability statistics on the characteristic sequences of adjacent device nodes. Statistically count the joint occurrence frequency of the characteristic sequences in a preset value range and normalize it to obtain a joint probability distribution. Statistically count the individual occurrence frequency of the characteristic sequences in a preset value range and normalize it to obtain a marginal probability distribution. Calculate the conditional probability based on the joint probability distribution and the marginal probability distribution to obtain the mutual information value;

[0100] Perform delay processing on the characteristic sequences at a preset time interval to obtain delay characteristic sequences. Statistically count the occurrence frequency of the characteristic values at the current moment of the characteristic sequences and the characteristic values of the delay characteristic sequences and normalize it to obtain a transition probability distribution. Calculate the transfer entropy value based on the transition probability distribution;

[0101] Compare the mutual information value with a first preset threshold to obtain a first comparison result, compare the transfer entropy value with a second preset threshold to obtain a second comparison result. When the first comparison result is greater than the first preset threshold and the second comparison result is greater than the second preset threshold, establish a connection relationship in the initial adjacency matrix. Write the sum of the mutual information value multiplied by a first weight coefficient and the transfer entropy value multiplied by a second weight coefficient as the connection weight value into the initial adjacency matrix to obtain an updated adjacency matrix;

[0102] Extract the feature vectors of each device node as its own features, obtain the feature vectors of all device nodes adjacent to each device node as adjacent features, weight the adjacent features according to the connection weight values in the updated adjacency matrix to obtain weighted features, and combine the own features and the weighted features to obtain multiple spatio-temporal correlation features.

[0103] The control system logic diagram is a tool for graphically representing the working principle of a control system. By showing the relationships between various control modules, it helps analyze system behavior and design control strategies. The marginal probability distribution is a probability distribution that describes a random variable or event and is commonly used in multivariate statistical analysis to represent the probability distribution of a variable without considering other variables.

[0104] Obtain the physical pipeline connection diagram and control system logic diagram of the thermal power plant equipment. Extract the connection relationships of water pipes, steam pipes, and oil pipes from the physical pipeline connection diagram to form a set of physical connection relationships. For example, the set of water pipe connection relationships includes: (boiler, feed water pump), (steam turbine, condenser), etc. Extract the cascade control, feedforward control, and decoupling control relationships from the control system logic diagram to form a set of control association relationships. For example, the set of cascade control relationships includes: (feed water pump speed controller, feed water pump), (boiler main steam pressure controller, burner), etc.

[0105] Merge the physical connection relationships and control association relationships to construct an initial adjacency matrix. If there is a physical connection relationship or control association relationship between two devices, the corresponding element of the adjacency matrix is assigned a value of 1; otherwise, it is assigned a value of 0. For example, if there is a water pipe connection relationship between the boiler and the feed water pump, the corresponding element value in the adjacency matrix is 1; if there is no direct physical connection or control relationship between the boiler and the steam turbine, the corresponding element value is 0. Assuming there are 5 devices in total, the initial adjacency matrix can be represented as a 5x5 matrix.

[0106] Select a time window of a preset time length, such as 1 hour, and perform probability statistics on the feature sequences of adjacent device nodes. Assume the feature we are interested in is the temperature of the device, and the preset value range is [100, 200]. Divide the range into 10 sub-intervals. Statistically count the joint occurrence frequencies of the feature sequences of adjacent device nodes in each sub-interval and normalize them to obtain the joint probability distribution. At the same time, statistically count the individual occurrence frequencies of the feature sequences of each device node in each sub-interval and normalize them to obtain the marginal probability distribution. Calculate the conditional probability based on the joint probability distribution and the marginal probability distribution, and then obtain the mutual information value. For example, the mutual information value of the temperature feature sequences of the boiler and the feed water pump is 0.8.

[0107] The characteristic sequence is delayed at a preset time interval (e.g., 1 minute) to obtain a delayed characteristic sequence. The occurrence frequencies of the characteristic values at the current moment of the characteristic sequence and the characteristic values of the delayed characteristic sequence are counted, and the transition probability distribution is obtained by normalization. The transfer entropy value is calculated based on the transition probability distribution. For example, the transfer entropy value of the boiler temperature to the feed pump temperature is 0.6.

[0108] The calculated mutual information value is compared with a first preset threshold (e.g., 0.5), and the transfer entropy value is compared with a second preset threshold (e.g., 0.3). If the mutual information value is greater than the first preset threshold and the transfer entropy value is greater than the second preset threshold, a connection relationship is established in the initial adjacency matrix, and the sum of the mutual information value multiplied by a first weight coefficient (e.g., 0.7) and the transfer entropy value multiplied by a second weight coefficient (e.g., 0.3) is used as the connection weight value and written into the adjacency matrix to obtain an updated adjacency matrix. For example, the mutual information value of the boiler and the feed pump is 0.8 which is greater than 0.5, and the transfer entropy value is 0.6 which is greater than 0.3, then a connection relationship is established in the adjacency matrix, and the connection weight value is 0.8×0.7 + 0.6×0.3 = 0.74.

[0109] The feature vector of each device node is extracted as its own feature. For example, the feature vector of the boiler may include temperature, pressure, water level, etc. The feature vectors of all device nodes adjacent to each device node are obtained as adjacent features. The adjacent features are weighted according to the connection weight values in the updated adjacency matrix to obtain weighted features. The own feature and the weighted features are combined to obtain multiple spatio-temporal correlation features. For example, the own feature of the boiler and the weighted features of the devices such as the feed pump and the burner adjacent to it are combined to obtain the spatio-temporal correlation feature of the boiler.

[0110] In this embodiment, not only the physical connection relationship and control logic relationship between devices are considered, but also the hidden correlation relationship between devices is further explored through data-driven mutual information and transfer entropy analysis, so that the actual correlation between devices can be more accurately described, avoiding the information loss caused by relying only on static connection relationships. By considering the time dependence of the characteristic sequence and the mutual influence between devices, spatio-temporal correlation features containing time and space information are extracted, which can more comprehensively reflect the operating state and interaction relationship of devices. The extracted spatio-temporal correlation features can be used for various downstream tasks, such as fault diagnosis, performance optimization, anomaly detection, etc., improving the accuracy and efficiency of the tasks.

[0111] S2. Add the fused feature matrix to the conditional variational autoencoder, process the equipment state features and working condition features through the dual-stream encoder, generate conditional latent variable distribution in combination with the reparameterization module, reconstruct features through a recursive neural network, calculate the reconstruction error and KL divergence, and optimize the parameters of the conditional variational autoencoder to obtain conditional probability distribution features, extract features from the conditional probability distribution features based on the hierarchical attention mechanism, obtain local fine-grained features and global macro features, obtain temporal context features in combination with an adaptive weight fuser, evaluate the importance and dynamically fuse the conditional probability distribution features and temporal context features through a deep feature fusion network, generate a hybrid feature representation, and perform multi-scale prediction through recursive dilated convolution and hierarchical discriminant structures to obtain refined distribution estimation results and abnormal pattern recognition results corresponding to the equipment state;

[0112] The dual-stream encoder is a neural network architecture that uses two independent encoders to process different input data streams and then fuses the resulting representations to capture different types of information features. It is often used in multimodal learning. The reparameterization module is a method for optimizing model training, which is often used in models such as variational autoencoders. It reparameterizes the sampling process of latent variables into differentiable operations so that gradients can propagate through random sampling paths. The conditional latent variable distribution refers to the probability distribution of latent variables learned by the model under certain given conditions. It is often used in generative models to generate data under specific input conditions. The deep feature fusion network is a network architecture that combines features from multiple different levels. It improves the expressiveness of the model by fusing features at different levels. It is often used in deep learning tasks in the fields of images and speech. The hierarchical discriminant structure is a network architecture that processes problems in a hierarchical manner. Each layer performs different discriminant tasks, and there is a certain dependency between the levels. It is often used in multi-level classification tasks.

[0113] In an optional embodiment,

[0114] The fused feature matrix is ​​added to the conditional variational autoencoder, the equipment state features and working condition features are processed by the dual-stream encoder, the conditional latent variable distribution is generated by combining the reparameterization module, the feature reconstruction is performed by the recursive neural network, the reconstruction error and KL divergence are calculated and the parameters of the conditional variational autoencoder are optimized to obtain the conditional probability distribution features, the conditional probability distribution features are extracted based on the hierarchical attention mechanism to obtain local fine-grained features and global macro features, and the temporal context features are obtained by combining the adaptive weight fuser, the importance of the conditional probability distribution features and the temporal context features are evaluated and dynamically fused by the deep feature fusion network, a hybrid feature representation is generated and multi-scale prediction is performed by recursive dilated convolution and hierarchical discriminant structure, and the refined distribution estimation results and abnormal pattern recognition results corresponding to the equipment state are obtained, including:

[0115] Add the fused feature matrix to the conditional variational autoencoder, process the device state feature and the working condition feature through the dual-stream encoder in the conditional variational autoencoder, the first stream of the dual-stream encoder maps the device state feature to the first latent space to generate the first latent space feature, the second stream of the dual-stream encoder maps the working condition feature to the second latent space to generate the second latent space feature, the first latent space feature and the second latent space feature are spliced ​​and input into the reparameterization module;

[0116] The reparameterization module generates a mean vector and a variance vector, samples a noise vector, multiplies the noise vector by the variance vector and then superimposes the mean vector to generate a conditional latent variable distribution, and inputs the conditional latent variable distribution into a recursive neural network for feature reconstruction;

[0117] Calculate the reconstruction error between the fused feature matrix and the reconstructed feature, calculate the KL divergence between the conditional latent variable distribution and the standard normal distribution, combine the reconstruction error and the KL divergence to form a loss function, and optimize the parameters of the conditional variational autoencoder to obtain conditional probability distribution features;

[0118] The conditional probability distribution features are extracted based on the hierarchical attention mechanism, the feature sequence is segmented by a sliding window through the local attention layer, the attention score and feature weight of the time point in the window are calculated to obtain the local fine-grained features, the feature sequence is divided into time segments through the global attention layer, and the attention score and feature weight of the time segment are calculated to obtain the global macro features;

[0119] Combined with an adaptive weight fusion device, statistical characteristics of the local fine-grained features and the global macro features are extracted to generate fusion weights, and the fusion weights are weighted with the local fine-grained features and the global macro features to obtain temporal context features. The conditional probability distribution features and the temporal context features are evaluated for importance through a deep feature fusion network to generate dynamic fusion weights, and the features are multiplied by the dynamic fusion weights and then connected in series to generate a hybrid feature representation;

[0120] The hybrid feature representation is input into the recursive dilated convolution and hierarchical discriminant structure for multi-scale prediction. The receptive field is expanded by multiple layers of dilated convolution to obtain multi-scale features. The output of each layer of dilated convolution is connected to the discriminator for state recognition. The output of each layer of discriminator is weighted to obtain the refined distribution estimation result and abnormal pattern recognition result corresponding to the device state.

[0121] The recursive dilated convolution is a form of extended convolution. By introducing a recursive structure and a dilation rate in the convolution operation, the network can capture a wider range of contextual information. It is often used to process time series data or image data. The receptive field refers to the area size of the input data that a neuron in a neural network can "perceive". The larger the receptive field, the more contextual information the model can capture. It is often used in convolutional neural networks.

[0122] Construct a fusion feature matrix. This matrix integrates multi-source heterogeneous data during the operation of the equipment, such as sensor data (such as temperature, pressure, current, etc.), operating condition data (such as load, speed, ambient temperature, etc.) and other related information. For example, the five features of temperature, pressure, current, load, and speed can be collected at 100 time points to form a matrix of 100 rows and 5 columns as the fusion feature matrix.

[0123] The fused feature matrix is ​​input into the conditional variational autoencoder. The autoencoder contains a dual-stream encoder structure. The first stream of the dual-stream encoder processes device state features, such as temperature, pressure, current, etc. Assume that there are three state features: temperature, pressure, and current. The second stream of the dual-stream encoder processes operating condition features, such as load and speed. Assume that there are two operating condition features: load and speed. The first stream maps the three-dimensional state features to a first latent space with a dimension of 16, generating a 16-dimensional first latent space feature. The second stream maps the two-dimensional operating condition features to a second latent space with a dimension of 8, generating an 8-dimensional second latent space feature. Then the 16-dimensional first latent space feature and the 8-dimensional second latent space feature are concatenated into a 24-dimensional vector.

[0124] The concatenated 24-dimensional vector is input to the reparameterization module. The reparameterization module generates a 24-dimensional mean vector and a 24-dimensional variance vector. At the same time, a 24-dimensional noise vector is sampled from the standard normal distribution. The noise vector is multiplied by the corresponding elements of the variance vector and then added to the mean vector to generate a 24-dimensional conditional latent variable distribution. This distribution is input into the recurrent neural network for feature reconstruction to obtain the reconstructed feature matrix.

[0125] Calculate the reconstruction error between the fused feature matrix and the reconstructed feature matrix, for example, using the mean square error. At the same time, calculate the KL divergence between the conditional latent variable distribution and the standard normal distribution. Perform a weighted sum of the reconstruction error and the KL divergence to form a loss function. Use an optimization algorithm (such as the Adam optimizer) to optimize the parameters of the conditional variational autoencoder, minimize the loss function, and finally obtain the conditional probability distribution feature. Assume that the obtained conditional probability distribution feature is a matrix with 100 rows and 32 columns.

[0126] The obtained conditional probability distribution features are subjected to feature extraction. A hierarchical attention mechanism is adopted, including a local attention layer and a global attention layer. The local attention layer adopts a sliding window mechanism, for example, the window size is 5. The feature sequence is divided into multiple windows, the attention score of each time point in the window is calculated, and then the attention score is weighted and summed with the features of the corresponding time point to obtain local fine-grained features. The global attention layer divides the feature sequence into multiple time segments, for example, 100 time points are divided into 10 segments, each of which contains 10 time points. The attention score of each time segment is calculated, and then the attention score is weighted and summed with the features of the segment to obtain global macro features.

[0127] The local fine-grained features and global macro features are input into the adaptive weight fuser. The fuser extracts the statistical characteristics of the local fine-grained features and the global macro features, such as the mean and variance, and generates fusion weights. The fusion weights are then weighted and summed with the local fine-grained features and the global macro features to obtain the temporal context features.

[0128] The conditional probability distribution features and temporal context features are input into the deep feature fusion network. The network evaluates the importance of the two features and generates dynamic fusion weights. The features are multiplied by the corresponding dynamic fusion weights and then concatenated to generate a hybrid feature representation.

[0129] The mixed feature representation is input into the recursive dilated convolution and hierarchical discriminant structure for multi-scale prediction. Multi-layer dilated convolution expands the receptive field range and obtains multi-scale features. The output of each layer of dilated convolution is connected to a discriminator for state recognition. The output of each layer of discriminators is weighted and summed to obtain the refined distribution estimation result and abnormal pattern recognition result corresponding to the device state.

[0130] In this embodiment, by fusing multi-source heterogeneous data and adopting a hierarchical attention mechanism, the device state information can be captured more comprehensively, thereby improving the accuracy of state estimation. Combined with the conditional variational autoencoder and recursive dilated convolution, the characteristics of the device abnormal mode can be effectively extracted to achieve more refined abnormal pattern recognition. The adaptive weight fuser and deep feature fusion network can dynamically adjust the model parameters according to different devices and working conditions, thereby improving the robustness and generalization ability of the model.

[0131] In an optional embodiment,

[0132] The hybrid feature representation is input into the recursive dilated convolution and hierarchical discriminant structure for multi-scale prediction. The receptive field range is expanded by multi-layer dilated convolution to obtain multi-scale features. The output of each layer of dilated convolution is connected to the discriminator for state recognition. The output of each layer of discriminator is weighted to obtain the refined distribution estimation result and abnormal pattern recognition result corresponding to the device state, including:

[0133] Normalizing the mixed feature representation, dividing the mixed feature representation into equipment state parameter features and operating condition parameter features, generating a normalized feature vector, constructing a recursive dilated convolution and hierarchical discriminant structure, and setting the dilation rates of the multi-layer dilated convolution layers to increase in sequence to expand the receptive field range;

[0134] Input the normalized feature vector into the first dilated convolution layer, select the feature vectors of each time point and adjacent time points in the input sequence for convolution operation to generate a first layer of multi-scale features, input the first layer of multi-scale features into the first discriminator, perform feature dimension reduction and state mapping through the multi-layer fully connected network in the first discriminator to generate a first layer of state probability distribution;

[0135] Input the normalized feature vector into the second dilated convolution layer, select the feature vector of each time point and its dilated interval position based on the second dilation rate to perform convolution operation to generate a second layer of multi-scale features, and input the second layer of multi-scale features into the second discriminator for state recognition to generate a second layer of state probability distribution;

[0136] The normalized feature vectors are sequentially input into different dilated convolutional layers, and feature vectors of time points and corresponding dilated interval positions are selected based on the increasing dilation rate to perform convolution operations, so as to generate multi-scale features corresponding to each layer, and the multi-scale features are respectively input into corresponding discriminators for state recognition to generate state probability distribution of each layer;

[0137] Calculate the entropy value of the state probability distribution output by each discriminator to obtain the prediction reliability index. Based on the prediction reliability index, determine the fusion weights of different state probability distributions, multiply the fusion weights by the corresponding state probability distributions and sum them to generate a refined distribution estimation result corresponding to the device state;

[0138] Set multiple levels of state probability thresholds, compare the refined distribution estimation result with the multiple levels of state probability thresholds to divide the state levels, and input the divided state level information into the abnormal pattern recognizer for abnormal type recognition to output the abnormal pattern recognition result.

[0139] The dilation interval position refers to the interval size between the convolutional kernel elements in dilated convolution. By using dilated convolution, the receptive field can be increased without increasing the computational amount, which is usually used for the processing of time series data or image data.

[0140] Preprocess the collected device operation data, including data cleaning, missing value filling and other operations to ensure the quality and integrity of the data. For example, if there are some outliers in the data sequence collected by a sensor, median filtering or linear interpolation methods can be used for correction. After preprocessing, fuse different types of feature data, such as device state parameters (such as temperature, pressure, current, etc.) and working condition parameters (such as rotation speed, load, ambient temperature, etc.) to generate a mixed feature representation. Suppose there are 3 device state parameters and 2 working condition parameters, then the values of these 5 parameters can be concatenated into a 5-dimensional feature vector as the mixed feature representation.

[0141] Perform normalization processing on the mixed feature representation. For example, the min-max normalization method can be used to scale the numerical range of each feature to between 0 and 1. Divide the normalized feature vector into two parts: device state parameter features and working condition parameter features. Suppose the first 3 dimensions of the 5-dimensional feature vector are device state parameters and the last 2 dimensions are working condition parameters, then it can be divided into a 3-dimensional vector and a 2-dimensional vector. Re-concatenate the divided feature vectors to generate a normalized feature vector.

[0142] Construct a recursive dilated convolution and hierarchical discriminant structure. Set multiple dilated convolution layers and make their dilation rates increase sequentially. For example, the dilation rate of the first layer is 1, the dilation rate of the second layer is 2, the dilation rate of the third layer is 4, and so on to expand the receptive field range. Connect a discriminator after each dilated convolution layer. The discriminator is composed of multiple fully connected networks and is used for feature dimensionality reduction and state mapping.

[0143] Input the normalized feature vector into the first dilated convolutional layer. The convolution operation will select the feature vectors at each time point and adjacent time points in the input sequence. For example, if the length of the input sequence is 10, the feature vector at each time point will be convolved with the adjacent feature vectors before and after it to generate the first-layer multi-scale features. Input the first-layer multi-scale features into the first discriminator to generate the first-layer state probability distribution.

[0144] Input the normalized feature vector into the second dilated convolutional layer. Based on the dilation rate of the second layer, select the feature vectors at each time point and their dilated interval positions for convolution operation. For example, if the dilation rate is 2, the feature vectors at each time point and the second time point after it will be selected for convolution operation to generate the second-layer multi-scale features. Input the second-layer multi-scale features into the second discriminator to generate the second-layer state probability distribution.

[0145] And so on, input the normalized feature vector into different dilated convolutional layers in turn, generate the corresponding multi-scale features for each layer, and input the multi-scale features into the corresponding discriminators to generate the state probability distributions for each layer. For example, assuming there are three dilated convolutional layers, three layers of multi-scale features and three layers of state probability distributions will be generated.

[0146] Calculate the entropy value of the state probability distribution output by each discriminator as the prediction reliability index. For example, cross-entropy can be used to calculate the entropy value. Determine the fusion weights of different state probability distributions based on the prediction reliability index. For example, the state probability distribution with a smaller entropy value can be assigned a larger weight. Multiply the fusion weights by the corresponding state probability distributions and sum them to generate the refined distribution estimation result corresponding to the device state.

[0147] Set multiple state probability thresholds. For example, three thresholds can be set: 0.8, 0.5, and 0.2, corresponding to the three state levels of "normal", "warning", and "abnormal" respectively. Compare the refined distribution estimation result with the multiple state probability thresholds to divide the state levels. For example, if the refined distribution estimation result is greater than 0.8, it is determined to be in the "normal" state; if it is between 0.5 and 0.8, it is determined to be in the "warning" state; if it is less than 0.2, it is determined to be in the "abnormal" state. Input the divided state level information into an anomaly pattern recognizer, such as a support vector machine or a random forest, for anomaly type recognition and output the anomaly pattern recognition result.

[0148] In this embodiment, through the multi-layer dilated convolution and hierarchical discrimination structure, richer state information can be extracted from multi-scale features, so as to more accurately estimate the probability distribution of the device state, realize refined state assessment. Combining multi-level state probability thresholds and anomaly pattern recognizers can more effectively identify different types of anomaly patterns, improve the accuracy and reliability of anomaly detection. By calculating the entropy value output by each discriminator as a prediction reliability index and fusing the state probability distribution based on this index, the uncertainty of the prediction result can be effectively reduced, and the reliability of the prediction can be improved.

[0149] S3. Add the refined distribution estimation result and anomaly pattern recognition result to the multi-scale spatio-temporal feature fusion network, calculate the importance weights of the device state features at different time scales and determine the spatial dependence relationship in combination with the graph convolutional neural network to obtain the overall state representation of the device group. Add the overall state representation to the hierarchical deep deterministic policy gradient network. Generate a task allocation scheme between edge computing nodes through the main network, optimize the resource scheduling strategy within the nodes through the auxiliary network, adjust the update step size of the optimized resource scheduling strategy through the adaptive learning rate to obtain the initial monitoring data. Perform probability modeling on the initial monitoring data through the Gaussian process regression algorithm, and perform fast retrieval and anomaly pattern matching on the modeled data in combination with the Bloom filter and the locality-sensitive hashing algorithm to generate an anomaly propagation probability and perform collaborative monitoring modeling in combination with the time-varying graph neural network to obtain an optimal monitoring scheme.

[0150] The hierarchical deep deterministic policy gradient network is a deep reinforcement learning architecture, which includes multiple levels of networks, and each level is responsible for processing different tasks or decisions to improve the optimization effect of the policy. The auxiliary network refers to introducing auxiliary tasks or auxiliary networks to assist the learning of the main network during the training process of the main model, usually used to increase the expressive ability of the model or help the main model overcome the local optimum problem. The Gaussian process regression algorithm is a non-parametric regression method that models functions based on Gaussian processes as prior distributions and can provide a measure of uncertainty, commonly used in regression problems in machine learning, especially when the data is scarce. The Bloom filter is a space-efficient probabilistic data structure used to detect whether an element belongs to a set. The locality-sensitive hashing algorithm is an approximate nearest neighbor search algorithm for high-dimensional data. By mapping similar input data to the same bucket, the cost of calculating distances is reduced, and it is widely used in clustering and classification problems of high-dimensional data. The collaborative monitoring modeling is a modeling method for collaboration and monitoring between multiple systems or components. By analyzing the relationships, data, and behaviors between systems, it predicts, optimizes, and manages their collaborative effects, and is usually applied to distributed systems or multi-agent systems.

[0151] In an alternative embodiment,

[0152] Add the refined distribution estimation result and the abnormal pattern recognition result to the multi-scale spatio-temporal feature fusion network, calculate the importance weights of the device state features at different time scales and determine the spatial dependence relationship by combining with the graph convolutional neural network to obtain the overall state representation of the device group. Add the overall state representation to the hierarchical deep deterministic policy gradient network, generate a task allocation scheme between edge computing nodes through the main network, optimize the resource scheduling strategy within the nodes through the auxiliary network, adjust the update step size of the optimized resource scheduling strategy through the adaptive learning rate to obtain the initial monitoring data, perform probability modeling on the initial monitoring data through the Gaussian process regression algorithm, and combine the Bloom filter and the locality-sensitive hashing algorithm to perform fast retrieval and abnormal pattern matching on the modeled data, generate the abnormal propagation probability and perform collaborative monitoring modeling by combining with the time-varying graph neural network to obtain the optimal monitoring scheme, including:

[0153] Add the refined distribution estimation result and the abnormal pattern recognition result to the multi-scale spatio-temporal feature fusion network, construct multiple time convolution branches to extract the state features at different time scales, select the state data of the most recent twenty-four time points for one-dimensional convolution to obtain the hourly-level features, select the state data of the corresponding moments in the most recent seven days for one-dimensional convolution to obtain the daily-level features, select the state data of the corresponding moments in the most recent four weeks for one-dimensional convolution to obtain the weekly-level features, map the hourly-level features to generate the hourly-level query vector, hourly-level key vector and hourly-level value vector, map the daily-level features to generate the daily-level query vector, daily-level key vector and daily-level value vector, map the weekly-level features to generate the weekly-level query vector, weekly-level key vector and weekly-level value vector, calculate the similarity between the hourly-level query vector and the hourly-level key vector to obtain the hourly-level attention score, calculate the similarity between the daily-level query vector and the daily-level key vector to obtain the daily-level attention score, calculate the similarity between the weekly-level query vector and the weekly-level key vector to obtain the weekly-level attention score, and generate the attention weight matrix corresponding to the time scale according to the hourly-level attention score, the daily-level attention score and the weekly-level attention score;

[0154] Encode the connection relationship between devices as an adjacency matrix, set the element values of the adjacency matrix to represent the association strength between devices, perform an aggregation operation on the node features and the neighbor node features through multiple graph convolutional layers, and perform a residual connection on the graph convolutional features of different layers to obtain the spatial dimension features representing the spatial dependence relationship;

[0155] Fuse the time - dimension features and the space - dimension features corresponding to the attention weight matrix, and use a multi - layer perceptron to adaptively adjust the fusion ratio of the time - dimension features and the space - dimension features to obtain the overall state representation of the device group. Add the overall state representation to the hierarchical deep deterministic policy gradient network, map the overall state representation to a low - dimensional feature space through the four - layer fully - connected structure of the main network and extract task - related features. Generate a task assignment plan based on the task - related features, add the task assignment plan to the auxiliary network, calculate the resource requirements, storage resource requirements, and communication bandwidth requirements according to the task - related features, generate a task priority assignment strategy and optimize the resource scheduling strategy within the node. Iteratively optimize the resource scheduling strategy, monitor the historical change trend of the policy gradient, increase the learning rate when the policy gradients in consecutive multiple iterations are in the same direction, and decrease the learning rate when the policy gradient direction changes frequently to obtain the initial monitoring data. Standardize the initial monitoring data, calculate the kernel function similarity matrix between samples and establish a joint probability distribution model of the monitoring data;

[0156] Map the feature vectors of the monitoring data into the bitmap of the Bloom filter for coarse - grained screening, map similar feature vectors into the same or adjacent hash buckets for abnormal pattern matching, establish a conditional probability model according to the physical association relationship and state correlation between devices, and use the probability propagation algorithm to estimate the abnormal diffusion risk and generate the abnormal propagation probability;

[0157] Use a gated recurrent unit to model the time - dynamic evolution process of the edge weights, combine the static weights and the dynamic weights to generate the edge weight values. Use the static weights to represent the physical coupling strength between devices, use the dynamic weights to represent the degree of correlation of device states and the abnormal propagation probability. Map the node features to multiple attention heads, calculate the attention scores between nodes using the multiple attention heads respectively, determine the transmission strength of information in the graph structure according to the attention scores, use the long - short - term memory network structure to predict the future state of the nodes, and obtain the optimal monitoring plan according to the prediction results and the current system state.

[0158] The hash bucket is a data structure in the locality - sensitive hashing algorithm, which is used to store elements mapped to the same bucket by the hash function. The probability propagation algorithm is an algorithm used to infer the uncertainty distribution in the probabilistic graphical model, and is often used in graphical models such as Bayesian networks to achieve global inference through local calculation and message passing.

[0159] Perform refined distribution estimation and anomaly pattern recognition on device status data. For example, collect the temperature, vibration frequency, and current data of 10 devices on a factory assembly line in the past month, use the Gaussian mixture model to model the probability distribution of each indicator for each device, and identify anomaly patterns based on the degree of deviation from the normal distribution. Use these results as input features.

[0160] Construct a multi-scale spatio-temporal feature fusion network. Taking one device as an example, select its temperature data for the most recent 24 hours, and use a one-dimensional convolutional kernel (e.g., convolutional kernel size is 3) to extract hourly-level features. Similarly, select the temperature data at the same moment in the most recent 7 days to extract daily-level features, and select the temperature data at the same moment in the most recent 4 weeks to extract weekly-level features. Map these features to query vectors, key vectors, and value vectors respectively, calculate the similarity between them, and obtain hourly-level, daily-level, and weekly-level attention scores. Generate the attention weight matrix corresponding to the corresponding time scale based on these scores. For example, if the hourly-level attention score is higher, then assign a higher weight to the hourly-level features.

[0161] Determine the spatial dependence relationship between devices. Encode the connection relationship between devices as an adjacency matrix, where the element value represents the association strength between devices. For example, if device A and device B are directly connected, the corresponding element value in the adjacency matrix is 1; if device A and device C are indirectly connected, the corresponding element value is 0.5. Aggregate the node features and neighbor node features through multiple graph convolutional layers. For example, perform a weighted average on the features of device A and its neighbor devices B and C, and the weights are determined by the adjacency matrix. Implement residual connections for the graph convolutional features of different layers to obtain the spatial dimension features representing the spatial dependence relationship.

[0162] Fuse the time dimension features and the spatial dimension features. Concatenate the features of different time scales corresponding to the attention weight matrix with the spatial dimension features, and use a multi-layer perceptron to adaptively adjust their fusion ratio to obtain the overall state representation of the device group. For example, if the spatial dependence is stronger during a certain period, then assign a higher weight to the spatial dimension features.

[0163] Input the overall state representation into a hierarchical deep deterministic policy gradient network. The main network consists of a four-layer fully connected structure, which maps the overall state representation to a low-dimensional feature space and extracts task-related features. For example, extract features such as device load and task priority. Generate a task assignment plan based on these features. Input the task assignment plan into the auxiliary network, calculate the resource requirements (e.g., CPU, memory, bandwidth) based on the task-related features, generate a task priority assignment policy, and optimize the resource scheduling policy within the node. Iteratively optimize the resource scheduling policy through an adaptive learning rate. For example, if the policy gradient directions of consecutive multiple iterations are the same, then increase the learning rate; if the policy gradient directions change frequently, then decrease the learning rate.

[0164] The initial monitoring data is obtained through the above steps. The initial monitoring data is standardized, the kernel function similarity matrix between samples is calculated, for example, using the Gaussian kernel function, and the joint probability distribution model of the monitoring data is established, for example, using Gaussian process regression.

[0165] The Bloom filter and the locality-sensitive hashing algorithm are used to quickly retrieve and match abnormal patterns in the modeled data. The feature vectors of the monitoring data are mapped to the bitmap of the Bloom filter for coarse-grained screening, and the similar feature vectors are mapped to the same or adjacent hash buckets for abnormal pattern matching. For example, the feature vectors are hashed into 100 buckets, and if two feature vectors are hashed into the same bucket, they are considered similar. A conditional probability model is established according to the physical association relationship and state correlation between devices, for example, using a Bayesian network. The probability propagation algorithm is used to estimate the abnormal diffusion risk and generate the abnormal propagation probability.

[0166] The time-varying graph neural network is used for collaborative monitoring modeling. The gated recurrent unit is used to model the time dynamic evolution process of the edge weights, and the static weights and dynamic weights are combined to generate the edge weight values. The static weight represents the physical coupling strength between devices, and the dynamic weight represents the correlation degree of device states and the abnormal propagation probability. The node features are mapped to multiple attention heads, and the attention scores between nodes are calculated respectively by the multiple attention heads. The transmission intensity of information in the graph structure is determined according to the attention scores. The long short-term memory network structure is used to predict the future state of the nodes, and the optimal monitoring scheme is obtained according to the prediction results and the current system state. For example, if it is predicted that a certain device is about to fail, preventive measures are taken in advance.

[0167] In this embodiment, through multi-scale spatio-temporal feature fusion and graph neural network, the device state information and abnormal patterns can be captured more comprehensively, the monitoring accuracy can be improved, the resource scheduling strategy can be optimized through the hierarchical deep deterministic policy gradient network, the resource utilization rate of the edge computing nodes can be improved, and potential risks can be early warned through collaborative monitoring modeling and abnormal propagation probability prediction, enhancing the system reliability.

[0168] In an alternative embodiment,

[0169] Mapping the feature vectors of the monitoring data to the bitmap of the Bloom filter for coarse-grained screening, mapping the similar feature vectors to the same or adjacent hash buckets for abnormal pattern matching, establishing a conditional probability model according to the physical association relationship and state correlation between devices, and using the probability propagation algorithm to estimate the abnormal diffusion risk and generate the abnormal propagation probability includes:

[0170] Set a dynamically expandable bitmap space and multiple independent hash functions based on a Bloom filter, set an expansion threshold for the dynamically expandable bitmap space, normalize and extract features from the monitoring data to obtain feature vectors, map the feature vectors into the dynamically expandable bitmap space through the hash functions for coarse-grained screening, detect the usage rate of the dynamically expandable bitmap space, and when the usage rate exceeds the expansion threshold, expand the dynamically expandable bitmap space according to a preset expansion ratio and remap the existing data;

[0171] Divide the multi-level hash structure into a first-level hash structure and a second-level hash structure, set multiple main hash buckets in the first-level hash structure, configure multiple sub-hash buckets for each main hash bucket in the second-level hash structure, construct an orthogonal basis vector group of the feature space, project the feature vectors onto the orthogonal basis vector group to obtain projection coefficients, and map the feature vectors with similar features into the same or adjacent hash buckets according to the similarity of the projection coefficients for abnormal pattern matching;

[0172] Establish a multi-layer device association graph based on the physical association relationship between devices, record device identifiers and device status parameters in the nodes of the multi-layer device association graph, set a time decay factor to weight the historical status data of adjacent device pairs, dynamically adjust the time decay factor according to a preset time interval, discretize the state space into a preset number of levels and construct a conditional probability model;

[0173] Use the probability propagation algorithm to estimate the abnormal diffusion risk, divide the abnormal propagation process, set a first propagation depth for short-range propagation and use a condition probability value updated in real time, set a second propagation depth for long-range propagation and introduce a distance decay coefficient, set different probability decay coefficients according to the number of propagation layers, calculate the abnormal propagation probability of each propagation path, and stop the propagation of the corresponding propagation path when the abnormal propagation probability is lower than a preset cut-off threshold.

[0174] The dynamically expandable bitmap space is a data structure that expands the bitmap space by dynamically allocating memory, enabling effective processing when the data volume increases. It is commonly used for the storage and retrieval of large-scale data. The orthogonal basis vector group is a concept in linear algebra, referring to a group of mutually orthogonal vectors that can form a basis for a space and can represent any vector in the space. It is commonly used in signal processing and dimensionality reduction methods. The time decay factor refers to a factor in certain algorithms where certain values or weights gradually decrease over time, used to emphasize recent events or data and reduce dependence on the past. The distance decay coefficient refers to a coefficient in certain distance-based algorithms used to adjust the impact of points far from the target on the result. It is commonly used in the similarity measurement of spatial or graph data. The probability decay coefficient is a factor that controls the change of probability distribution over time or distance, usually used in models to handle probabilities that decrease over time or with increasing spatial distance, such as in reinforcement learning and graph networks.

[0175] Collect real-time data from various monitoring devices, such as the CPU usage rate, memory occupancy rate, network traffic, etc. of servers, as well as the packet loss rate, latency, etc. of network devices. Clean these raw data to remove noise and invalid data. Then, perform normalization processing on the cleaned data to convert data with different dimensions into a unified numerical range, such as between 0 and 1. Finally, extract key features from the normalized data to form a feature vector. For example, the mean, variance, maximum value, minimum value, etc. of the CPU usage rate can be extracted as features. Suppose the feature vector obtained after processing the monitoring data of a server is [0.8, 0.2, 0.9, 0.1].

[0176] Use a dynamically expandable Bloom filter for coarse-grained screening. Initialize a bitmap space and set multiple independent hash functions. Map the preprocessed feature vector into the bitmap space through these hash functions. Set the usage rate threshold of the bitmap space, such as 80%. Regularly detect the usage rate of the bitmap space. When the usage rate exceeds the threshold, expand the bitmap space according to a preset expansion ratio, such as 1.5 times, and remap the existing feature vectors into the new bitmap space. Suppose the initial bitmap size is 1000, and currently 750 feature vectors have been mapped, with a usage rate reaching 75%. When continuing to map feature vectors and causing the usage rate to reach 80%, expand the bitmap space to 1500 and remap the previous 750 feature vectors and the newly arrived feature vectors into the new bitmap space.

[0177] Perform anomaly pattern matching. Divide the multi-level hash structure into a first level and a second level. Set multiple main hash buckets in the first-level hash structure, such as 10. In the second-level hash structure, configure multiple sub-hash buckets for each main hash bucket, such as 5. Construct an orthogonal basis vector group for the feature space, for example, obtain the orthogonal basis vector group using the principal component analysis (PCA). Project the feature vectors onto the orthogonal basis vector group to obtain projection coefficients. According to the similarity of the projection coefficients, map the feature vectors with similar features into the same or adjacent hash buckets. For example, if the distance between the projection coefficients of two feature vectors in the Euclidean space is less than a certain threshold, they are considered similar and mapped into the same sub-hash bucket. Suppose the projection coefficients of two feature vectors are [0.1, 0.2, 0.3] and [0.11, 0.21, 0.29] respectively. If the distance threshold is set to 0.1, these two feature vectors are considered similar and will be mapped into the same or adjacent sub-hash buckets.

[0178] Establish a device association model. Based on the physical association relationship between devices, such as the network topology structure, establish a multi-layer device association graph. Record the device identifier and device status parameters in the nodes of the association graph. Set a time decay factor, such as 0.9, to weight the historical state data of adjacent device pairs, and dynamically adjust the time decay factor according to a preset time interval, such as 1 hour. Discretize the state space into a preset number of levels, such as 5 levels. Then, construct a conditional probability model, such as using a Bayesian network, to represent the dependence relationship between device states. Suppose device A and device B are adjacent, and the historical state data of device A is [1, 2, 3], and the time decay factor is 0.9, then the weighted historical state data is [0.9, 1.8, 2.7].

[0179] Perform anomaly diffusion risk assessment. Use a probability propagation algorithm, such as the belief propagation algorithm, to estimate the anomaly diffusion risk. Divide the anomaly propagation process, such as into short-range propagation and long-range propagation. Set a first propagation depth for short-range propagation, such as 2, and adopt real-time updated conditional probability values. Set a second propagation depth for long-range propagation, such as 5, and introduce a distance decay coefficient, such as 0.8. Set different probability decay coefficients according to the number of propagation layers. Calculate the anomaly propagation probability of each propagation path. When the anomaly propagation probability is lower than a preset cut-off threshold, such as 0.1, stop the propagation of the corresponding propagation path. Suppose device A has an anomaly, its anomaly state is 3, and the conditional probability of the adjacent device B is 0.5, then the probability of the anomaly spreading to device B is 3×0.5 = 1.5. If the distance decay coefficient is 0.8 and the propagation depth is 2, then the probability of spreading to device C adjacent to B is 1.5×0.8×conditional probability.

[0180] In this embodiment, coarse-grained screening is performed through Bloom filters, which can quickly exclude most normal data, reduce the amount of data to be processed subsequently, and improve the efficiency of anomaly detection. Through multi-level hash structures and feature vector projections, similar feature vectors can be mapped to the same or adjacent hash buckets, thereby improving the accuracy of anomaly pattern matching. By constructing a device association model and a probability propagation algorithm, the risk of anomaly diffusion can be more accurately assessed, providing a basis for taking timely preventive measures.

[0181] Figure 2 Schematic diagram of the structure of the real-time monitoring system of thermal power equipment based on edge computing according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0182] The first unit is used to deploy an edge computing node array and collect multi-source heterogeneous data of thermal power equipment through a joint perception mechanism, calculate the credibility score of the multi-source heterogeneous data based on a pre-set quality indicator and screen to obtain valid data, transmit the valid data as original monitoring data to the control center, perform multi-scale adaptive filtering on the original monitoring data to obtain filtered data and calculate the corresponding information entropy through adaptive wavelet packet decomposition, select the optimal decomposition layer based on the information entropy criterion and decompose the filtered data into sub-signals of different frequency bands, calculate the time-frequency energy distribution and statistical characteristics for each sub-signal, generate a threshold surface and perform anomaly correction through an anomaly detection model based on an isolation forest and support vector data description algorithm, obtain preprocessed data and add it to an adaptive migration network for feature alignment, construct an adjacency matrix based on the physical connection relationship of the equipment, extract spatiotemporal correlation features and construct a fusion feature matrix;

[0183] The second unit is used to add the fused feature matrix to the conditional variational autoencoder, process the equipment state features and working condition features through the dual-stream encoder, generate conditional latent variable distribution in combination with the reparameterization module, reconstruct features through a recursive neural network, calculate the reconstruction error and KL divergence and optimize the parameters of the conditional variational autoencoder to obtain conditional probability distribution features, extract features of the conditional probability distribution features based on the hierarchical attention mechanism to obtain local fine-grained features and global macro features, obtain temporal context features in combination with an adaptive weight fuser, evaluate the importance and dynamically fuse the conditional probability distribution features and the temporal context features through a deep feature fusion network, generate a hybrid feature representation, and perform multi-scale prediction through recursive dilated convolution and hierarchical discriminant structures to obtain refined distribution estimation results and abnormal pattern recognition results corresponding to the equipment state;

[0184] The third unit is used to add the refined distribution estimation result and the abnormal pattern recognition result to the multi-scale spatio-temporal feature fusion network, calculate the importance weights of the device state features at different time scales, and determine the spatial dependence relationship by combining with the graph convolutional neural network to obtain the overall state representation of the device group. Add the overall state representation to the hierarchical deep deterministic policy gradient network, generate a task allocation scheme between edge computing nodes through the main network, optimize the resource scheduling strategy within the nodes through the auxiliary network, adjust the update step size of the optimized resource scheduling strategy through the adaptive learning rate to obtain the initial monitoring data, perform probability modeling on the initial monitoring data through the Gaussian process regression algorithm, and perform fast retrieval and abnormal pattern matching on the modeled data by combining the Bloom filter and the locality-sensitive hashing algorithm, generate the abnormal propagation probability, and perform collaborative monitoring modeling by combining with the time-varying graph neural network to obtain the optimal monitoring scheme.

[0185] In the third aspect of the embodiments of the present invention,

[0186] A kind of electronic device is provided, including:

[0187] A processor;

[0188] A memory for storing instructions executable by the processor;

[0189] Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0190] In the fourth aspect of the embodiments of the present invention,

[0191] A computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0192] The present invention can be a method, a device, a system, and / or a computer program product. The computer program product may include a computer-readable storage medium, on which computer-readable program instructions for executing various aspects of the present invention are uploaded.

[0193] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: it is still possible to modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A real-time monitoring method for thermal power equipment based on edge computing, characterized in that: include: Deploy edge computing node arrays and collect multi-source heterogeneous data of thermal power equipment through a joint perception mechanism. Calculate the credibility score of the multi-source heterogeneous data based on pre-set quality indicators and screen to obtain valid data. Transmit the valid data as original monitoring data to the control center, perform multi-scale adaptive filtering on the original monitoring data to obtain filtered data and calculate the corresponding information entropy through adaptive wavelet packet decomposition. Select the optimal number of decomposition layers based on the information entropy criterion and decompose the filtered data into sub-signals of different frequency bands. Calculate the time-frequency energy distribution and statistical characteristics for each sub-signal. Generate a threshold surface and perform anomaly correction through an anomaly detection model based on isolation forest and support vector data description algorithm. Obtain preprocessed data and add it to an adaptive migration network for feature alignment. Construct an adjacency matrix based on the physical connection relationship of the equipment, extract spatiotemporal correlation features and construct a fusion feature matrix. Add the fused feature matrix to the conditional variational autoencoder, process the equipment state features and working condition features through the dual-stream encoder, generate conditional latent variable distribution in combination with the reparameterization module, reconstruct features through the recursive neural network, calculate the reconstruction error and KL divergence and optimize the parameters of the conditional variational autoencoder to obtain conditional probability distribution features, extract features from the conditional probability distribution features based on the hierarchical attention mechanism to obtain local fine-grained features and global macro features, obtain temporal context features in combination with the adaptive weight fuser, evaluate the importance and dynamically fuse the conditional probability distribution features and the temporal context features through the deep feature fusion network, generate a hybrid feature representation and perform multi-scale prediction through recursive dilated convolution and hierarchical discriminant structure to obtain refined distribution estimation results and abnormal pattern recognition results corresponding to the equipment state; The refined distribution estimation result and the abnormal pattern recognition result are added to the multi-scale spatiotemporal feature fusion network, the importance weights of the device state features at different time scales are calculated and the spatial dependencies are determined in combination with the graph convolutional neural network to obtain the overall state representation of the device group, the overall state representation is added to the hierarchical deep deterministic policy gradient network, the task allocation plan between edge computing nodes is generated through the main network, the resource scheduling strategy within the node is optimized through the auxiliary network, the update step size of the optimized resource scheduling strategy is adjusted through the adaptive learning rate to obtain the initial monitoring data, the initial monitoring data is probabilistically modeled by the Gaussian process regression algorithm, the modeled data is quickly retrieved and matched with abnormal patterns in combination with the Bloom filter and the locally sensitive hashing algorithm, the abnormal propagation probability is generated and combined with the time-varying graph neural network for collaborative monitoring modeling to obtain the optimal monitoring plan.

2. The method according to claim 1, characterized in that Deploy edge computing node arrays and collect multi-source heterogeneous data of thermal power equipment through a joint perception mechanism. Calculate the credibility score of the multi-source heterogeneous data based on pre-set quality indicators and screen to obtain valid data. Transmit the valid data as original monitoring data to the control center. Perform multi-scale adaptive filtering on the original monitoring data to obtain filtered data and calculate the corresponding information entropy through adaptive wavelet packet decomposition. Select the optimal decomposition layer based on the information entropy criterion and decompose the filtered data into sub-signals of different frequency bands. Calculate the time-frequency energy distribution and statistical characteristics for each sub-signal. Generate a threshold surface and perform anomaly correction through an anomaly detection model based on isolation forest and support vector data description algorithm. Obtain preprocessed data and add it to the adaptive migration network for feature alignment. Construct an adjacency matrix based on the physical connection relationship of the equipment. Extract spatiotemporal correlation features and construct a fusion feature matrix, including: Deploy an edge computing node array, configure a clock synchronization module of the edge computing node array to achieve time alignment between nodes, schedule vibration sensors, temperature sensors, pressure sensors, and displacement sensors in the edge computing node array to work together to form a joint perception mechanism, collect multi-source heterogeneous data of thermal power equipment, and normalize each type of data in the multi-source heterogeneous data; According to the preset quality index, the sampling sequence of the multi-source heterogeneous data is traversed, the number of missing continuous sampling points is counted to mark invalid segments, the numerical difference of adjacent sampling points is calculated and divided by the sampling interval to obtain the change rate and compared with the standard deviation of historical data to mark suspicious points, the time difference of data from the acquisition end to the processing end is recorded to obtain the transmission delay and mark the delay point, and each marking result is multiplied by the corresponding weight and superimposed to calculate to obtain the credibility score; Transmit valid data whose credibility score exceeds a preset threshold as original monitoring data to the control center, divide the processing window of the original monitoring data into fixed points, calculate the signal mean square value in each processing window to obtain the signal variance, calculate the residual mean square value to obtain the noise variance, divide the signal variance by the noise variance to obtain the signal-to-noise ratio, and adaptively adjust the cutoff frequency and order of the filter according to the signal-to-noise ratio to perform multi-scale adaptive filtering; Perform wavelet packet decomposition on the filtered data to obtain wavelet coefficients of each layer, calculate the probability distribution of the wavelet coefficients of each layer to obtain information entropy, subtract the information entropy of the subband corresponding to the k-1 layer from the information entropy of the wavelet coefficients of the kth layer to obtain an entropy value difference, compare the entropy value difference with a preset threshold value based on the information entropy criterion, determine the optimal decomposition layer number of the subband and stop decomposition when the entropy value difference is less than the preset threshold value, and decompose the filtered data into sub-signals of different frequency bands; The mean, standard deviation, skewness, kurtosis, waveform factor and margin factor of each sub-signal are calculated to obtain the time-domain statistical characteristics. The power spectrum is calculated by fast Fourier transform and integrated to obtain the time-frequency energy distribution. The frequency band energy proportion, main frequency component and spectrum kurtosis are extracted to obtain the frequency domain statistical characteristics. Input the time-frequency energy distribution and statistical features into an isolation forest detector, construct a decision tree by randomly selecting features and splitting points to calculate the isolation depth of the sample to obtain an abnormality score, input the abnormality score as a sample weight into a support vector data description algorithm, map the sample to a high-dimensional feature space through a kernel function, and solve a quadratic programming problem to generate a minimum enclosing hypersphere as a threshold surface, and perform abnormal correction on abnormal points exceeding the threshold surface by local mean replacement to obtain preprocessed data; Input the preprocessed data into a feature extractor composed of a multi-layer perceptron network, extract deep feature representation through a nonlinear activation function and a fully connected layer, input the deep feature representation into a domain classifier to calculate the feature distribution difference between the source domain and the target domain, and alternately optimize the feature extraction loss function and the domain discrimination loss function to achieve feature alignment; An initial adjacency matrix is ​​constructed based on the physical pipeline connection and control logic relationship of the equipment, the conditional probability of the feature sequence between the equipment nodes is calculated to obtain the mutual information, and the transfer probability of the feature sequence is calculated to obtain the transfer entropy. The connection relationship in the adjacency matrix is ​​updated according to the numerical values ​​of the mutual information and the transfer entropy, and the spatiotemporal correlation features are extracted. The feature vector of each equipment node is combined with the weighted feature vector of its adjacent nodes to construct a fusion feature matrix.

3. The method according to claim 2, characterized in that An initial adjacency matrix is ​​constructed based on the physical pipeline connection and control logic relationship of the equipment, the conditional probability of the feature sequence between the equipment nodes is calculated to obtain the mutual information, the transition probability of the feature sequence is calculated to obtain the transfer entropy, and the connection relationship in the adjacency matrix is ​​updated according to the values ​​of the mutual information and the transfer entropy. Extracting spatiotemporal correlation features includes: Obtain a physical pipeline connection diagram and a control system logic diagram of thermal power plant equipment, extract a water pipe connection relationship, a steam pipe connection relationship and an oil pipe connection relationship from the physical pipeline connection diagram to form a physical connection relationship, extract a cascade control relationship, a feedforward control relationship and a decoupling control relationship from the control system logic diagram to form a control association relationship, merge the physical connection relationship and the control association relationship to construct an initial adjacency matrix, and assign a value of 1 to a corresponding adjacency matrix element when the physical connection relationship or the control association relationship exists between adjacent device nodes, and assign a value of 0 to a corresponding adjacency matrix element when the physical connection relationship and the control association relationship do not exist; Select a time window of a preset time length to perform probability statistics on the feature sequences of adjacent device nodes, count the joint occurrence frequencies of the feature sequences in a preset value interval and normalize them to obtain a joint probability distribution, count the individual occurrence frequencies of the feature sequences in a preset value interval and normalize them to obtain an edge probability distribution, and calculate the conditional probability based on the joint probability distribution and the edge probability distribution to obtain a mutual information value; Delaying the feature sequence according to a preset time interval to obtain a delayed feature sequence, counting the occurrence frequencies of the feature value of the feature sequence at the current moment and the feature value of the delayed feature sequence and normalizing them to obtain a transfer probability distribution, and calculating a transfer entropy value based on the transfer probability distribution; Comparing the mutual information value with a first preset threshold to obtain a first comparison result, comparing the transfer entropy value with a second preset threshold to obtain a second comparison result, establishing a connection relationship in the initial adjacency matrix when the first comparison result is greater than the first preset threshold and the second comparison result is greater than the second preset threshold, and writing the sum of the mutual information value multiplied by a first weight coefficient and the transfer entropy value multiplied by a second weight coefficient as a connection weight value into the initial adjacency matrix to obtain an updated adjacency matrix; Extract the feature vector of each device node as its own feature, obtain the feature vectors of all device nodes adjacent to each device node as adjacent features, weight the adjacent features according to the connection weight values ​​in the updated adjacency matrix to obtain weighted features, and combine the own features with the weighted features to obtain multiple spatiotemporal correlation features.

4. The method according to claim 1, characterized in that: The fused feature matrix is ​​added to the conditional variational autoencoder, the equipment state features and working condition features are processed by the dual-stream encoder, the conditional latent variable distribution is generated by combining the reparameterization module, the feature reconstruction is performed by the recursive neural network, the reconstruction error and KL divergence are calculated and the parameters of the conditional variational autoencoder are optimized to obtain the conditional probability distribution features, the conditional probability distribution features are extracted based on the hierarchical attention mechanism to obtain local fine-grained features and global macro features, and the temporal context features are obtained by combining the adaptive weight fuser, the importance of the conditional probability distribution features and the temporal context features are evaluated and dynamically fused by the deep feature fusion network, a hybrid feature representation is generated and multi-scale prediction is performed by recursive dilated convolution and hierarchical discriminant structure, and the refined distribution estimation results and abnormal pattern recognition results corresponding to the equipment state are obtained, including: Add the fused feature matrix to the conditional variational autoencoder, process the device state feature and the working condition feature through the dual-stream encoder in the conditional variational autoencoder, the first stream of the dual-stream encoder maps the device state feature to the first latent space to generate the first latent space feature, the second stream of the dual-stream encoder maps the working condition feature to the second latent space to generate the second latent space feature, the first latent space feature and the second latent space feature are spliced ​​and input into the reparameterization module; The reparameterization module generates a mean vector and a variance vector, samples a noise vector, multiplies the noise vector by the variance vector and then superimposes the mean vector to generate a conditional latent variable distribution, and inputs the conditional latent variable distribution into a recursive neural network for feature reconstruction; Calculate the reconstruction error between the fused feature matrix and the reconstructed feature, calculate the KL divergence between the conditional latent variable distribution and the standard normal distribution, combine the reconstruction error and the KL divergence to form a loss function, and optimize the parameters of the conditional variational autoencoder to obtain conditional probability distribution features; The conditional probability distribution features are extracted based on the hierarchical attention mechanism, the feature sequence is segmented by a sliding window through the local attention layer, the attention score and feature weight of the time point in the window are calculated to obtain the local fine-grained features, the feature sequence is divided into time segments through the global attention layer, and the attention score and feature weight of the time segment are calculated to obtain the global macro features; Combined with an adaptive weight fusion device, statistical characteristics of the local fine-grained features and the global macro features are extracted to generate fusion weights, and the fusion weights are weighted with the local fine-grained features and the global macro features to obtain temporal context features. The conditional probability distribution features and the temporal context features are evaluated for importance through a deep feature fusion network to generate dynamic fusion weights, and the features are multiplied by the dynamic fusion weights and then connected in series to generate a hybrid feature representation; The hybrid feature representation is input into the recursive dilated convolution and hierarchical discriminant structure for multi-scale prediction. The receptive field is expanded by multiple layers of dilated convolution to obtain multi-scale features. The output of each layer of dilated convolution is connected to the discriminator for state recognition. The output of each layer of discriminator is weighted to obtain the refined distribution estimation result and abnormal pattern recognition result corresponding to the device state.

5. The method according to claim 4, characterized in that The hybrid feature representation is input into the recursive dilated convolution and hierarchical discriminant structure for multi-scale prediction. The receptive field range is expanded by multi-layer dilated convolution to obtain multi-scale features. The output of each layer of dilated convolution is connected to the discriminator for state recognition. The output of each layer of discriminator is weighted to obtain the refined distribution estimation result and abnormal pattern recognition result corresponding to the device state, including: Normalizing the mixed feature representation, dividing the mixed feature representation into equipment state parameter features and operating condition parameter features, generating a normalized feature vector, constructing a recursive dilated convolution and hierarchical discriminant structure, and setting the dilation rates of the multi-layer dilated convolution layers to increase in sequence to expand the receptive field range; Input the normalized feature vector into the first dilated convolution layer, select the feature vectors of each time point and adjacent time points in the input sequence for convolution operation to generate a first layer of multi-scale features, input the first layer of multi-scale features into the first discriminator, perform feature dimension reduction and state mapping through the multi-layer fully connected network in the first discriminator to generate a first layer of state probability distribution; Input the normalized feature vector into the second dilated convolution layer, select the feature vector of each time point and its dilated interval position based on the second dilation rate to perform convolution operation to generate a second layer of multi-scale features, and input the second layer of multi-scale features into the second discriminator for state recognition to generate a second layer of state probability distribution; The normalized feature vectors are sequentially input into different dilated convolutional layers, and feature vectors of time points and corresponding dilated interval positions are selected based on the increasing dilation rate to perform convolution operations, so as to generate multi-scale features corresponding to each layer, and the multi-scale features are respectively input into corresponding discriminators for state recognition to generate state probability distribution of each layer; Calculate the entropy value of the state probability distribution output by each discriminator to obtain a prediction reliability index, determine the fusion weight of different state probability distributions based on the prediction reliability index, multiply the fusion weight with the corresponding state probability distribution and sum them to generate a refined distribution estimation result corresponding to the device state; A multi-level state probability threshold is set, the refined distribution estimation result is compared with the multi-level state probability threshold to divide the state level, the divided state level information is input into the abnormal pattern recognizer for abnormal type recognition, and the abnormal pattern recognition result is output.

6. The method according to claim 1, characterized in that The refined distribution estimation result and the abnormal pattern recognition result are added to the multi-scale spatiotemporal feature fusion network, the importance weights of the device state features at different time scales are calculated and the spatial dependencies are determined in combination with the graph convolutional neural network to obtain the overall state representation of the device group, the overall state representation is added to the hierarchical deep deterministic policy gradient network, the task allocation scheme between edge computing nodes is generated through the main network, the resource scheduling strategy within the node is optimized through the auxiliary network, the update step size of the optimized resource scheduling strategy is adjusted through the adaptive learning rate to obtain the initial monitoring data, the initial monitoring data is probabilistically modeled by the Gaussian process regression algorithm, the modeled data is quickly retrieved and abnormal pattern matched in combination with the Bloom filter and the local sensitive hashing algorithm, the abnormal propagation probability is generated and the collaborative monitoring modeling is performed in combination with the time-varying graph neural network, and the optimal monitoring scheme includes: Add the refined distribution estimation results and the abnormal pattern recognition results to the multi-scale spatiotemporal feature fusion network, construct multiple time convolution branches to extract state features of different time scales, select the state data of the latest twenty-four time points for one-dimensional convolution to obtain hourly features, select the state data of the corresponding moments of the latest seven days for one-dimensional convolution to obtain day-level features, select the state data of the corresponding moments of the latest four weeks for one-dimensional convolution to obtain week-level features, map the hourly features to generate hourly query vectors, hourly key vectors and hourly value vectors, map the day-level features to generate day-level query vectors, day-level key vectors and day-level value vectors, map the week-level features to generate week-level query vectors, week-level key vectors and week-level value vectors, calculate the similarity between the hourly query vector and the hourly key vector to obtain the hourly attention score, calculate the similarity between the day-level query vector and the day-level key vector to obtain the day-level attention score, calculate the similarity between the week-level query vector and the week-level key vector to obtain the week-level attention score, and generate the attention weight matrix of the corresponding time scale according to the hourly attention score, the day-level attention score and the week-level attention score; Encode the connection relationship between devices into an adjacency matrix, set the element value of the adjacency matrix to represent the association strength between devices, aggregate node features and neighbor node features through multiple layers of graph convolution layers, and perform residual connection on graph convolution features of different layers to obtain spatial dimension features representing spatial dependency relationships; The time dimension features and the space dimension features corresponding to the attention weight matrix are fused, and the fusion ratio of the time dimension features and the space dimension features is adaptively adjusted by using a multi-layer perceptron to obtain the overall state representation of the device group, and the overall state representation is added to the hierarchical deep deterministic policy gradient network. The overall state representation is mapped to a low-dimensional feature space through the four-layer fully connected structure of the main network and task-related features are extracted. A task allocation scheme is generated based on the task-related features, and the task allocation scheme is added to the auxiliary network. The resource requirements, storage resource requirements and communication bandwidth requirements are calculated according to the task-related features, a task priority allocation strategy is generated, and the resource scheduling strategy within the node is optimized. The resource scheduling strategy is iteratively optimized, and the historical change trend of the policy gradient is monitored. When the policy gradient of multiple consecutive iterations remains in the same direction, the learning rate is increased, and when the policy gradient direction changes frequently, the learning rate is reduced to obtain initial monitoring data, and the initial monitoring data is standardized, and the kernel function similarity matrix between samples is calculated and a joint probability distribution model of monitoring data is established; Map the feature vectors of monitoring data to the bitmap of the Bloom filter for coarse-grained screening, map similar feature vectors to the same or adjacent hash buckets for abnormal pattern matching, establish a conditional probability model based on the physical association relationship and state correlation between devices, and use the probability propagation algorithm to estimate the abnormal diffusion risk and generate the abnormal propagation probability; The time dynamic evolution process of edge weights is modeled using gated recurrent units, and static weights are combined with dynamic weights to generate edge weights. The static weights are used to characterize the physical coupling strength between devices, and the dynamic weights are used to characterize the correlation between device states and the probability of abnormal propagation. Node features are mapped to multiple attention heads, and the attention scores between nodes are calculated using the multiple attention heads. The transmission intensity of information in the graph structure is determined based on the attention scores, and the long short-term memory network structure is used to predict the future state of the node. The optimal monitoring plan is obtained based on the prediction results and the current system state.

7. The method according to claim 6, characterized in that Map the feature vectors of the monitoring data to the bitmap of the Bloom filter for coarse-grained screening, map similar feature vectors to the same or adjacent hash buckets for abnormal pattern matching, establish a conditional probability model based on the physical association relationship and state correlation between devices, and use the probability propagation algorithm to estimate the abnormal diffusion risk and generate the abnormal propagation probability, including: A dynamic extension bitmap space and a plurality of mutually independent hash functions are set based on a Bloom filter, an extension threshold of the dynamic extension bitmap space is set, monitoring data is normalized and feature extracted to obtain a feature vector, the feature vector is mapped to the dynamic extension bitmap space through the hash function for coarse-grained screening, the utilization rate of the dynamic extension bitmap space is detected, and when the utilization rate exceeds the extension threshold, the dynamic extension bitmap space is expanded according to a preset expansion ratio and the existing data is remapped; Divide the multi-level hash structure into a first-level hash structure and a second-level hash structure, set a plurality of main hash buckets in the first-level hash structure, configure a plurality of sub-hash buckets for each of the main hash buckets in the second-level hash structure, construct an orthogonal basis vector group of a feature space, project the feature vectors onto the orthogonal basis vector group to obtain projection coefficients, and map the feature vectors with similar features to the same or adjacent hash buckets for abnormal pattern matching according to the similarity of the projection coefficients; A multi-layer device association graph is established based on the physical association relationship between devices, device identifications and device state parameters are recorded in nodes of the multi-layer device association graph, a time decay factor is set to weight the historical state data of adjacent device pairs, the time decay factor is dynamically adjusted according to a preset time interval, the state space is discretized into a preset number of levels, and a conditional probability model is constructed; The probability propagation algorithm is used to estimate the risk of abnormal diffusion, divide the abnormal propagation process, set the first propagation depth for short-range propagation and adopt the conditional probability value updated in real time, set the second propagation depth for long-range propagation and introduce the distance attenuation coefficient, set different probability attenuation coefficients according to the number of propagation layers, calculate the abnormal propagation probability of each propagation path, and stop the propagation of the corresponding propagation path when the abnormal propagation probability is lower than the preset cutoff threshold.

8. A real-time monitoring system for thermal power equipment based on edge computing, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to deploy an edge computing node array and collect multi-source heterogeneous data of thermal power equipment through a joint perception mechanism, calculate the credibility score of the multi-source heterogeneous data based on a pre-set quality indicator and screen to obtain valid data, transmit the valid data as original monitoring data to the control center, perform multi-scale adaptive filtering on the original monitoring data to obtain filtered data and calculate the corresponding information entropy through adaptive wavelet packet decomposition, select the optimal decomposition layer based on the information entropy criterion and decompose the filtered data into sub-signals of different frequency bands, calculate the time-frequency energy distribution and statistical characteristics for each sub-signal, generate a threshold surface and perform anomaly correction through an anomaly detection model based on an isolation forest and support vector data description algorithm, obtain preprocessed data and add it to an adaptive migration network for feature alignment, construct an adjacency matrix based on the physical connection relationship of the equipment, extract spatiotemporal correlation features and construct a fusion feature matrix; The second unit is used to add the fused feature matrix to the conditional variational autoencoder, process the equipment state features and working condition features through the dual-stream encoder, generate conditional latent variable distribution in combination with the reparameterization module, reconstruct features through a recursive neural network, calculate the reconstruction error and KL divergence and optimize the parameters of the conditional variational autoencoder to obtain conditional probability distribution features, extract features of the conditional probability distribution features based on the hierarchical attention mechanism to obtain local fine-grained features and global macro features, obtain temporal context features in combination with an adaptive weight fuser, evaluate the importance and dynamically fuse the conditional probability distribution features and the temporal context features through a deep feature fusion network, generate a hybrid feature representation, and perform multi-scale prediction through recursive dilated convolution and hierarchical discriminant structures to obtain refined distribution estimation results and abnormal pattern recognition results corresponding to the equipment state; The third unit is used to add the refined distribution estimation results and abnormal pattern recognition results to the multi-scale spatiotemporal feature fusion network, calculate the importance weights of the device state features at different time scales and determine the spatial dependencies in combination with the graph convolutional neural network to obtain the overall state representation of the device group, add the overall state representation to the hierarchical deep deterministic policy gradient network, generate the task allocation plan between edge computing nodes through the main network, optimize the resource scheduling strategy within the node through the auxiliary network, adjust the update step size of the optimized resource scheduling strategy through the adaptive learning rate to obtain initial monitoring data, perform probability modeling on the initial monitoring data through the Gaussian process regression algorithm, combine the Bloom filter and the local sensitive hashing algorithm to quickly retrieve and match the modeled data, generate the abnormal propagation probability and combine the time-varying graph neural network for collaborative monitoring modeling to obtain the optimal monitoring plan.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images