Regional fire early warning method and system based on multiple sensors

Abstract

The invention relates to the technical field of fire early warning, and discloses a regional fire early warning method and system based on multiple sensors, and the method comprises the steps: obtaining a multi-source data stream, and carrying out the preprocessing of the multi-source data stream, and obtaining a time sequence feature vector; a graph neural network is adopted to construct and analyze an inter-node dynamic relation graph, and hidden node association strength is determined; if the association strength exceeds a threshold value, extracting a time association sequence of the strong association nodes, and constructing and weighting according to the time association sequence to obtain a comprehensive risk score; dynamically correcting a score threshold based on the environment humidity and the insulation thermal aging risk coefficient, and generating an early warning signal when the risk score exceeds the threshold; carrying out continuity inspection and noise filtering on the early warning signal to obtain a refined early warning sequence; and finally, generating a risk prediction result through a long-short-term memory network model, and distributing the risk prediction result to a response system to complete early warning after packaging verification. According to the invention, early and accurate early warning of complex electrical fire hazards is realized, and the false alarm rate and the missing report rate are effectively reduced.

Description

Technical Field

[0001] This invention relates to the field of fire early warning technology, and in particular to a regional fire early warning method and system based on multiple sensors. Background Technology

[0002] Research in the field of fire early warning has become a top priority in social safety management.

[0003] In one existing technology, fire early warning systems mainly rely on a single sensor or a few sensors to detect fire characteristics and use rule-based or statistical model-based methods to process the sensor data. When the detected value exceeds a threshold, an alarm is triggered.

[0004] However, existing technologies typically focus only on data performance at a single point in time, failing to fully explore the temporal correlations and dynamic coupling relationships between multi-source data in sensor networks. This is especially true in complex fire scenarios such as electrical faults, where it is difficult to capture the gradual change patterns between different signals before a fire occurs. Therefore, existing technologies suffer from high false alarm rates and significant missed alarm risks due to neglecting the temporal correlation characteristics of multi-source information. Summary of the Invention

[0005] This invention provides a regional fire early warning method and system based on multiple sensors to solve the problems of high false alarm rate and high risk of missed alarm caused by ignoring the time correlation characteristics of multi-source information in the prior art.

[0006] In a first aspect, to address the aforementioned technical problems, the present invention provides a multi-sensor-based regional fire early warning method, comprising: A multi-source information data stream is acquired and preprocessed to obtain a time-series feature vector; the multi-source information data stream includes current fluctuation data, temperature signal data, and environmental humidity data. Based on the time series feature vectors and the physical connection relationships between sensors, a graph neural network is used to construct a dynamic connection graph between nodes, and the correlation strength between nodes in the hidden mode is determined through parsing. If the correlation strength between the nodes exceeds a preset correlation strength threshold, then the time correlation characteristic sequence of the corresponding node is extracted from the dynamic connection graph between the nodes. A risk prediction vector is constructed based on the time-related characteristic sequence, and a weight allocation matrix is ​​constructed based on the node topological centrality corresponding to the risk prediction vector. The risk prediction vector and the weight allocation matrix are weighted and summed to generate a comprehensive risk score. The comprehensive risk score is corrected based on the environmental humidity data, and the insulation thermal aging risk coefficient is calculated in combination with the temperature signal data. If the insulation thermal aging risk coefficient exceeds the corresponding preset risk threshold, a preliminary warning signal is generated. The noise interference in the preliminary early warning signal is removed through time sequence processing to obtain a candidate early warning sequence, and continuity inspection and logical comparison are performed, and if the logical comparison result meets a preset logical threshold, a refined early warning sequence is obtained; According to the refined early warning sequence, a final risk prediction output is generated in combination with a preset long short-term memory network model, and is distributed to a response system for execution.

[0007] In a second aspect, the present application provides a regional fire early warning system based on multiple sensors, comprising: A multi-source data acquisition and preprocessing module acquires a multi-source information data stream, and obtains a time sequence feature vector after preprocessing; the multi-source information data stream comprises current fluctuation data, temperature signal data and environmental humidity data; A dynamic association graph construction and analysis module adopts a graph neural network to construct a dynamic contact graph between nodes based on the time sequence feature vector and the physical connection relationship between sensors, and determines the association strength between nodes in a hidden mode through analysis; An association strength threshold judgment and feature extraction module extracts a time association characteristic sequence of the corresponding node from the dynamic contact graph between nodes if the association strength between nodes exceeds a preset association strength threshold; A risk feature fusion and evaluation module constructs a risk prediction vector according to the time association characteristic sequence, and constructs a weight distribution matrix according to the node topology centrality corresponding to the risk prediction vector, and performs weighted summation on the risk prediction vector and the weight distribution matrix to generate a comprehensive risk score; An adaptive threshold adjustment and preliminary early warning module corrects the comprehensive risk score based on the environmental humidity data, and calculates an insulation thermal aging risk coefficient in combination with the temperature signal data, and generates a preliminary early warning signal if the insulation thermal aging risk coefficient exceeds a corresponding preset risk threshold; An early warning signal purification and refining module removes noise interference in the preliminary early warning signal through time sequence processing to obtain a candidate early warning sequence, and performs continuity inspection and logical comparison, and obtains a refined early warning sequence if the logical comparison result meets a preset logical threshold; A risk prediction output and response distribution module generates a final risk prediction output in combination with a preset long short-term memory network model according to the refined early warning sequence, and distributes it to a response system for execution.

[0008] Compared with the prior art, the present application has the following beneficial effects: (1) The application can deeply mine the spatio-temporal correlation and hidden coupling mode among multi-source sensor data such as current, temperature and vibration by constructing and analyzing the dynamic contact graph between nodes by using the graph neural network, breaking through the limitation of traditional methods which only focus on single time point data or simple statistical threshold, and significantly improving the early identification ability of progressive and complex electrical fire hazards.

[0009] (2) The application constructs a multi-level and adaptive dynamic evaluation system by introducing weight distribution based on node topology centrality, dynamically correcting risk score threshold combined with environmental humidity, and adaptively adjusting the correlation strength threshold by using insulation thermal aging risk coefficient and partial discharge data, so that the early warning system can more accurately respond to the real operating state of the equipment and environmental interference, effectively filter false alarms caused by transient interference, and at the same time enhance the sensitivity to potential risks.

[0010] (3) The application forms a complete purification process from raw data to refined early warning sequence by using sliding window segmentation, principal component analysis dimensionality reduction to extract time sequence features, and sequence density analysis, continuity test and feature filtering on early warning signals. Combined with long short-term memory network for time series prediction and generating risk level with confidence interval, and finally through encapsulating and checking the message to distribute reliably, the intelligent early warning from multi-source information perception, intelligent analysis to accurate response closed loop is realized, the system has high overall reliability and good engineering application value. BRIEF DESCRIPTION OF DRAWINGS

[0011] Figure 1 is a flowchart of a regional fire early warning method based on multiple sensors provided by the first embodiment of the application; Figure 2 is a structural diagram of a regional fire early warning system based on multiple sensors provided by the second embodiment of the application. DETAILED DESCRIPTION

[0012] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the application.

[0013] With reference to Figure 1 , the first embodiment of the application provides a regional fire early warning method based on multiple sensors, including the following steps: S11, acquiring multi-source information data stream to obtain time sequence feature vector after preprocessing; the multi-source information data stream includes current fluctuation data, temperature signal data and environmental humidity data; S12, based on the time sequence feature vector and the physical connection relationship between sensors, a graph neural network is used to construct a dynamic contact graph between nodes, and the correlation strength between nodes in the hidden mode is determined by analysis; S13, if the correlation strength between nodes exceeds a preset correlation strength threshold, a time correlation characteristic sequence of the corresponding node is extracted from the dynamic contact graph between nodes; S14, a risk prediction vector is constructed according to the time correlation characteristic sequence, and a weight distribution matrix is constructed according to the node topology center degree corresponding to the risk prediction vector, and the risk prediction vector and the weight distribution matrix are weighted and summed to generate a comprehensive risk score; S15, based on the environmental humidity data, the comprehensive risk score is corrected, and an insulation thermal aging risk coefficient is calculated combined with the temperature signal data, if the insulation thermal aging risk coefficient exceeds the corresponding preset risk threshold, a preliminary warning signal is generated; S16, through time sequence processing, noise interference in the preliminary warning signal is removed to obtain a candidate warning sequence, and continuity test and logical comparison are performed, if the logical comparison result meets the preset logical threshold, a refined warning sequence is obtained; S17, according to the refined warning sequence, a long short-term memory network model is generated to generate a final risk prediction output, and is distributed to a response system for execution.

[0014] It should be noted that before deploying the present application, the initialization and parameter calibration of the system need to be completed. Long time sequence multi-sensor data of target equipment or similar equipment under normal working conditions and typical fault working conditions are collected, and it is recommended that the coverage time be not less than 10 months. The data should include current fluctuation, temperature signal, environmental humidity, vibration signal (if applicable) and other multi-source information, and the sampling frequency should not be less than the system design requirement. The data set should cover the complete evolution process of at least 3 typical fault modes (such as overload, insulation aging, mechanical wear, etc.) and the data segments before and after the corresponding faults occur.

[0015] The specific parameter values given in the embodiment, such as sampling frequency, window length, hidden layer dimension, cumulative contribution rate threshold, etc. are exemplary, and in actual application, they should be calibrated and optimized based on the specific equipment characteristics, data distribution and system requirements through historical data analysis, grid search, cross-validation or simulation experiment.

[0016] The weight coefficients (such as , etc.) in the formula are determined by grid search or regression analysis based on historical fault cases. For example, in the risk feature weighted aggregation, the node topology center degree weight is obtained after Softmax normalization; the temperature and risk score weight in the insulation thermal aging risk coefficient is obtained by shallow neural network training.

[0017] In step S11, a multi-source information data stream is acquired, and a time sequence feature vector is obtained after preprocessing; the multi-source information data stream includes current fluctuation data, temperature signal data and environmental humidity data, and includes: S1101, a multi-source information data stream containing current fluctuation data, temperature signal data and environmental humidity data is acquired through a sensor; S1102, low-pass filtering is used on the multi-source information data stream to remove noise and perform signal alignment to obtain a standardized data stream; S1103, a sliding window is used to segment and intercept the standardized data stream to obtain a time sequence segment to be analyzed; S1104, frequency domain components and time domain statistics of the time sequence segment to be analyzed are extracted to obtain a multi-dimensional feature set; S1105, principal component analysis is performed on the multi-dimensional feature set and dimension compression is completed to obtain a time sequence feature vector.

[0018] In step S1101, a multi-source information data stream containing current fluctuation data, temperature signal data and environmental humidity data is acquired through a sensor.

[0019] It should be noted that the multi-source information data stream containing current fluctuation data, temperature signal data and environmental humidity data is collected in real time by a plurality of sensors deployed on the device. The current sensor, the temperature sensor and the humidity sensor respectively acquire the measurement values of each physical quantity at a set sampling frequency synchronously, and generate corresponding time sequence data.

[0020] In an embodiment, the sensor is configured to collect current and temperature data once per second. The current effective value collected by a certain type of motor under rated load fluctuates in the range of 11A to 11.8A, and the temperature value changes between 42.5 degrees and 48.3 degrees. At the same time, the humidity sensor collects environmental humidity data at the same frequency. All collected data is uploaded to the subsequent processing unit in real time through the communication interface. It should be noted that the current and temperature values are only illustrative, and the actual collected values vary depending on the device model, load state and environmental conditions.

[0021] In step S1102, low-pass filtering is used on the multi-source information data stream to remove noise and perform signal alignment to obtain a standardized data stream.

[0022] It should be noted that a low-pass filter is used to filter high-frequency interference in the current signal; signal alignment unifies different sensor data to the same time reference to obtain a standardized data stream.

[0023] In an embodiment, low-pass filtering is applied to the current signal to smooth the burst spike interference, for example, filtering out the abnormal peak of 15.7 A in one collection and smoothing to 12.5 A; at the same time, the timestamps of the current and temperature signals are aligned to ensure that the data corresponds on the time axis. Considering the system processing pressure, the sampling frequency can be adjusted according to the actual application to balance the data accuracy and system load.

[0024] In step S1103, the standardized data stream is segmented and intercepted by using a sliding window to obtain a time sequence segment to be analyzed.

[0025] It should be noted that the standardized data stream is segmented and intercepted by using a sliding window, a window of a fixed time length is set, and the window is slid forward at a fixed time interval, and each time the data in the period covered by the window is intercepted as an analysis unit. For example, the window length is set to 1 minute and the sliding step is 10 seconds.

[0026] In an embodiment, a certain segment intercepted by this method shows that the current value therein presents regular periodic fluctuations, and the temperature value presents a slow rising trend.

[0027] In step S1104, the frequency domain components and time domain statistics of the time sequence segment to be analyzed are extracted to obtain a multi-dimensional feature set.

[0028] It should be noted that when the frequency domain components and time domain statistics of the time sequence segment to be analyzed are extracted, the current signal and other time sequence data in the segment are subjected to fast Fourier transform to convert them to the frequency domain, and the main frequency components thereof are extracted as frequency domain features; at the same time, the statistics of the segment are calculated in the time domain, including the mean, variance, peak value, etc.

[0029] In an embodiment, a current time sequence segment is analyzed, and the mean is calculated to be 11.2 A and the variance is 0.3 A²; through Fourier transform, it is found that the main frequency component is 0.5 Hz. These extracted features such as 11.2 A, 0.3 A², and 0.5 Hz together constitute a multi-dimensional feature vector describing the running state of the device in this period as the multi-dimensional feature set.

[0030] In step S1105, principal component analysis is performed on the multi-dimensional feature set and dimension compression is completed to obtain a time sequence feature vector.

[0031] It should be noted that, in order to perform principal component analysis on the multi-dimensional feature set to realize dimension reduction processing, first, the covariance matrix of the multi-dimensional feature set is calculated, the eigenvalue decomposition is performed on the matrix, the eigenvalues are sorted from large to small, the first s principal components are selected, so that the cumulative contribution rate of the principal components reaches a predetermined cumulative contribution rate threshold (for example, 95%), and finally the original features are projected onto a low-dimensional subspace spanned by the s principal components. The setting of the cumulative contribution rate threshold is usually based on the inflection point analysis of the eigenvalue distribution. For example, in the motor monitoring scene, by analyzing the eigenvalue distribution of the historical data, it is found that when the cumulative contribution rate of the first s principal components reaches 95%, the corresponding eigenvalue curve appears an obvious inflection point, and the key mode of the original data can be effectively retained, so the threshold is set to 95%. The value can be adjusted in the range of 85%-99% according to the feature dimension of different monitoring objects.

[0032] In an embodiment, the original feature set contains 10-dimensional frequency domain and time domain indicators, which are compressed into 3-dimensional feature vectors through principal component analysis, retaining 95% of the information amount of the original data.

[0033] In step S12, based on the time series feature vector and the physical connection relationship between sensors, a graph neural network is used to construct a dynamic contact graph between nodes, and the correlation strength between nodes in a hidden mode is determined through analysis, including: S1201, combine the time series feature vector and the physical connection relationship between sensors to generate an initial adjacency matrix; S1202, calculate the feature similarity between nodes in the initial adjacency matrix to construct a weighted graph structure; S1203, input the weighted graph structure into a graph neural network layer to update the corresponding hidden layer state vector of each node; S1204, adjust the dynamic weight on the propagation path in the weighted graph structure according to the difference value between the hidden layer state vectors to reconstruct a dynamic contact graph between nodes; S1205, analyze the connection tightness of the dynamic contact graph between nodes to determine the correlation strength between nodes in a hidden mode.

[0034] In step S1201, the time series feature vector and the physical connection relationship between sensors are combined to generate an initial adjacency matrix.

[0035] It should be noted that the data source of each sensor, such as current and temperature, is abstracted as a node in the graph, and then the edges between nodes are defined according to the physical installation position and connection relationship between sensors. If the sensors corresponding to two nodes have a direct physical connection or are installed on the same device unit, the corresponding element in the adjacency matrix is set to 1, otherwise it is set to 0, thereby generating a sparse binary matrix.

[0036] In an embodiment, the monitoring system comprises three sensor nodes of current, temperature and vibration. Among them, the current sensor and the temperature sensor are installed on the same motor and directly connected, so the element corresponding to the current-temperature pair in the matrix is set to 1; the temperature sensor and the vibration sensor are close in physical position, so the temperature-vibration element is also set to 1; and the current sensor and the vibration sensor have no direct physical association, so the current-vibration element is set to 0.

[0037] In step S1202, the feature similarity between nodes in the initial adjacency matrix is calculated, and a weighted graph structure is constructed.

[0038] It should be noted that the node represents a different source signal in the multi-source information data stream. For each edge with a value of 1 in the initial adjacency matrix, i.e. for each pair of nodes with physical connection, the similarity measure between the time series feature vectors corresponding to the two nodes is calculated using the Pearson correlation coefficient; then the correlation coefficient is used as a weight to update the value of the edge in the initial adjacency matrix, thereby converting the binary adjacency matrix into a weighted adjacency matrix and forming a weighted graph structure.

[0039] In an embodiment, the correlation coefficient between the feature vectors of the current node and the temperature node is 0.93, the correlation coefficient between the temperature node and the vibration node is 0.93, and the correlation coefficient between the current node and the vibration node is 0.35; then in the weighted adjacency matrix, the weight of the current-temperature edge is updated to 0.82, the weight of the temperature-vibration edge is updated to 0.91, and the weight of the current-vibration edge is updated to 0.35.

[0040] In step S1203, the weighted graph structure is input into a graph neural network layer, and the hidden layer state vector corresponding to each node is updated.

[0041] It should be noted that the weighted graph structure is input into a graph neural network layer for processing. The graph neural network adopts a two-layer graph convolution network architecture, the input layer dimension is consistent with the dimension of the generated time series feature vector, for example, 3, the hidden layer dimension of the first layer and the second layer is set to 64, and the ReLU function is used as the activation function between layers. The hidden layer dimension such as 64 is determined by grid search combined with cross-validation. The historical monitoring data is collected to build a dataset containing 5000 weighted graph structure-node state label data pairs, of which 4000 are normal state samples and 1000 are typical fault state samples. The training set, validation set and test set are randomly divided in the ratio of 7:2:1; grid search is performed in the hidden layer dimension set {32, 64, 128, 256}, and other hyperparameters such as learning rate and Dropout rate remain unchanged; the node state classification accuracy is taken as the main evaluation index, and the model reasoning delay is considered to meet the real-time requirement; the experimental results show that when the hidden layer dimension is 64, the node state classification accuracy of the model on the validation set reaches 96.2%, and the reasoning delay is less than 50ms, which meets the real-time requirement of the system. Considering the accuracy and computational efficiency, the hidden layer dimension is selected as 64.

[0042] The network is trained in an end-to-end manner, and the training target is node state classification (normal / fault). During the training process, the cross-entropy loss function and the Adam optimizer are used for parameter optimization, the learning rate is set to 0.001, the batch size is 32, and the training period is 100 rounds. The accuracy of the trained model on the test set reaches 95.8%, indicating that it has good generalization ability. The hidden layer state vector corresponding to each node in the graph is updated through the information propagation and aggregation of network iteration. Each layer of graph convolution network aggregates feature information from its adjacent nodes according to the connection relationship and edge weight defined by the weighted adjacency matrix; the aggregated information is combined with the state of the node itself in the last layer, and after the above learnable nonlinear transformation, a new hidden layer state vector of the node in the current layer is generated. After two layers of stacking and feature transformation, the final state vector of each node can integrate the information of related nodes in the multi-hop neighborhood.

[0043] In an embodiment, a trained GCN model is used, which has a node state classification accuracy of 95% on the validation set. The trained model is applied to the current weighted graph structure, in which the state of the current node is aggregated with the information of the temperature node which is strongly related (weight 0.82) to it, and the feature of thermal influence is integrated; at the same time, the state of the temperature node also aggregates the load information from the current node, so that its state vector can better reflect the temperature rise pattern caused by the current change; the state of the vibration node is significantly strengthened with the coupling of the temperature feature due to the high-weight connection (0.91) with the temperature node.

[0044] In step S1204, the dynamic weights on the propagation path in the weighted graph structure are adjusted according to the difference values between the hidden layer state vectors, and a dynamic contact graph between nodes is reconstructed and generated.

[0045] It should be noted that the Euclidean distance between the hidden layer state vectors of each two nodes connected in the calculation graph is calculated to measure the state difference; then, a predefined attenuation function is used to map the difference value to a weight adjustment coefficient, and the original edge weight is updated. The principle is that the smaller the difference is, the larger the coefficient is, and the corresponding edge weight is improved, and vice versa.

[0046] In an embodiment, the hidden layer state vectors of the current nodes and the temperature nodes show high consistency after being aggregated by the graph neural network, and the Euclidean distance is small; while the hidden layer state difference between the current nodes and the vibration nodes is large. By iteratively performing this process, a dynamic contact graph between nodes is finally generated, in which the edge weight changes adaptively with the node state.

[0047] In step S1205, the connection tightness of the dynamic contact graph between nodes is analyzed to determine the correlation strength between nodes in the hidden mode.

[0048] It should be noted that the correlation strength between nodes in the hidden mode is determined by quantifying the connection tightness, calculating the shortest path length between each pair of nodes in the dynamic contact graph between nodes, and calculating the average clustering coefficient of the entire graph; the shorter the average path is, the higher the clustering coefficient is, which indicates that the connection of the node community in the graph is tighter, and the hidden correlation behind the sensor data stream is stronger.

[0049] In an embodiment, it is found that the edge weight between the current nodes and the temperature nodes is continuously and stably above 0.9, the shortest path length between the two nodes is 1, and the clustering coefficient of the subgraph to which they belong is as high as 0.88, which collectively indicates that they have significant and stable strong correlation during equipment operation, especially in potential fault modes.

[0050] In step S13, if the correlation strength between nodes exceeds a preset correlation strength threshold, the time correlation characteristic sequence of the corresponding node is extracted from the dynamic contact graph between nodes, including: S1301, if the correlation strength between nodes does not exceed the preset correlation strength threshold, a supplementary signal is obtained from the multi-source information data stream; S1302, performing feature alignment processing on the supplementary signal, and obtaining an updated topology result of the dynamic contact graph between nodes according to the alignment processing result; S1303, based on the updated topology result, the correlation strength between nodes is recalculated, if the updated correlation strength exceeds the preset correlation strength threshold, the time sequence feature data of the corresponding node is extracted from the dynamic contact graph between nodes, and is arranged and spliced in time sequence to generate a time correlation characteristic sequence.

[0051] In step S1301, if the correlation strength between nodes does not exceed the preset correlation strength threshold, a supplementary signal is obtained from the multi-source information data stream.

[0052] It should be noted that the preset correlation strength threshold is used to determine whether the dynamic contact graph between nodes has sufficiently represented the equipment state, and the threshold is determined based on statistical analysis of the correlation strength values of all node pairs in the historical normal operation period. The distribution of the correlation strength of all node pairs in the dynamic contact graph of the computing device in the fault-free operation state is taken as the threshold, and the specific high quantile such as the 95th percentile is taken as the threshold for distinguishing between accidental fluctuations and potential abnormal strong correlation in normal operation. If the correlation strength value of the whole or the key path does not exceed the correlation strength threshold, the asynchronous acquisition instruction of the multi-source information data stream is triggered, the supplementary signal is synchronously obtained from the standby sensor deployed on the equipment, and the original multi-source information data stream is injected. The types of the supplementary signal can include higher precision vibration spectrum, acoustic emission signal, partial discharge signal, infrared thermal image data, etc., and the priority is preset according to the type of the equipment and the monitoring scene; the supplementary signal includes but is not limited to higher precision vibration spectrum, acoustic emission signal or partial discharge signal, and the priority is preset according to the type of the equipment. When the features are aligned, the supplementary signal needs to be resampled to the same frequency as the main data stream, and the weight between nodes is updated by the Pearson correlation coefficient.

[0053] In one embodiment, the calculated current and vibration nodes are 0.65 (lower than the threshold). The system starts the vibration auxiliary sensor located at the motor bearing to collect higher frequency vibration spectrum data as a supplementary signal.

[0054] In step S1302, the supplementary signal is processed for feature alignment, and the updated topology result of the dynamic contact graph between nodes is obtained according to the alignment result.

[0055] It should be noted that the supplementary signal collected asynchronously is pre-processed, including aligning its timestamp with the main data stream, resampling at the main data stream sampling frequency, normalization, and missing value interpolation. Subsequently, the feature vector is extracted from the processed signal. If the supplementary signal represents a brand new monitoring dimension, it is added as a new node to the graph, and based on the correlation of its features with the existing node features, such as calculating the Pearson correlation coefficient, the connection edge is established and the initial weight is assigned; if the signal is an enhancement of the existing monitoring node, the new feature vector is used to update the state of the corresponding node, and the edge weight between the node and other connected nodes in the graph is recalculated accordingly, thereby obtaining the updated topology result.

[0056] In an embodiment, the 1 kHz vibration spectrum data collected from the bearing as a supplementary signal is first resampled to 100 Hz to align with the original vibration signal, and the main frequency amplitude and other features are extracted after normalization. Since the signal is from the same physical location, the system regards it as an enhanced feature of the original "vibration node", and recalculates the feature correlation between the vibration node and the temperature node and the current node accordingly, updates the weight of the vibration-temperature edge from 0.92 to 0.96, and updates the weight of the vibration-current edge from 0.65 to 0.78, thereby completing the topology update.

[0057] In step S1303, the correlation strength between nodes is recalculated based on the updated topology result, and if the updated correlation strength exceeds the preset correlation strength threshold, the time sequence feature data of the corresponding nodes is extracted from the dynamic contact graph between nodes, and arranged and spliced in time sequence to generate a time correlation characteristic sequence.

[0058] It should be noted that the Pearson correlation coefficient is used to recalculate the correlation coefficient between each pair of nodes as the updated correlation strength value based on the latest feature vector. If the correlation strength value of a particular node pair exceeds the threshold after recalculation, the feature change data of the corresponding nodes on the strongly correlated path is extracted from the dynamic contact graph between nodes, and arranged and spliced in time sequence to generate a time correlation characteristic sequence describing the coordinated evolution mode of multiple signals.

[0059] In an embodiment, after recalculation, the correlation strength between the current node and the temperature node is increased to 0.93, and the correlation strength between the vibration node and the ambient temperature node is 0.90, both of which exceed the threshold value of 0.85. The system then extracts the feature data of the nodes involved in the "current-temperature" and "vibration-ambient temperature" strong connection paths, arranges the current feature values, temperature feature values, etc. in the order of acquisition time to form a multi-dimensional time sequence.

[0060] In step S14, a risk prediction vector is constructed according to the time correlation characteristic sequence, a weight distribution matrix is constructed according to the topological centrality of the node corresponding to the risk prediction vector, the risk prediction vector and the weight distribution matrix are weighted and summed to generate a comprehensive risk score, including: In S1401, a multi-dimensional fault feature matrix is constructed according to the time correlation characteristic sequence. In S1402, the multi-dimensional fault feature matrix is subjected to feature space mapping transformation to generate a risk prediction vector. In S1403, a weight distribution matrix is constructed by normalizing the weight coefficients based on the topological centrality of the node corresponding to the risk prediction vector through a Softmax function and determining the main diagonal elements. In S1404, a risk feature set is obtained by performing weighted operation on the risk prediction vector and the weight distribution matrix. In S1405, a comprehensive risk score is obtained by performing additive aggregation on the risk feature set.

[0061] In step S1401, a multi-dimensional fault feature matrix is constructed according to the time correlation characteristic sequence.

[0062] It should be noted that when constructing the multi-dimensional fault feature matrix according to the time correlation characteristic sequence, the characteristic values in the sequence arranged in time sequence and from different signals are integrated to construct a two-dimensional matrix. Each row of the matrix corresponds to a specific sampling time point, and each column corresponds to a specific feature dimension extracted from the strong correlation path, for example, the main frequency amplitude of current, the change rate of temperature, and the effective value of vibration. Each element in the matrix is the specific value of the feature dimension at the time point.

[0063] In an embodiment, a matrix is constructed from a sequence containing current and temperature correlation. At time point T1, the current main frequency amplitude column is 0.8 (normalized value) and the temperature change rate column is 0.2 in the corresponding matrix row. At time point T2, the current main frequency amplitude becomes 0.7 and the temperature change rate is 0.3. Finally, the data of all time points together form an N-row and M-column feature matrix.

[0064] In step S1402, the multi-dimensional fault feature matrix is subjected to feature space mapping transformation to generate a risk prediction vector.

[0065] It should be noted that the feature space mapping transformation is performed on the multi-dimensional fault feature matrix to generate the risk prediction vector, and principal component analysis is used for feature dimension reduction to obtain a g-dimensional risk prediction vector. First, the covariance matrix of the matrix is calculated and eigenvalue decomposition is performed, and then the eigenvectors corresponding to the first k largest eigenvalues are selected as principal components, wherein the selection of g is based on the cumulative contribution rate of the selected principal components being not less than a set value (for example, 95%) determined based on historical data statistical analysis, so as to construct a projection matrix from the original high-dimensional feature space to the low-dimensional potential risk space. Finally, the original multi-dimensional fault feature matrix is multiplied by the projection matrix to map it into a g-dimensional risk prediction vector, wherein the value of each dimension represents the risk intensity of the equipment in a certain specific potential fault mode (such as current abnormality, overheating, and mechanical vibration).

[0066] In an embodiment, a three-dimensional vector is obtained after mapping, representing the potential risk values of current abnormality, temperature abnormality, and vibration abnormality, such as 0.7, 0.4, and 0.6.

[0067] In step S1403, based on the topological centrality of the node corresponding to the risk prediction vector, a weight coefficient is obtained by normalization through a Softmax function and is determined as a main diagonal element to construct a weight distribution matrix.

[0068] It should be noted that the node topological centrality is obtained by calculating its weighted degree centrality, that is, the sum of the weights of all adjacent edges of the node. Based on the topological structure of the inter-node dynamic contact graph, the weighted degree centrality of each node is calculated, that is, the sum of the weights of the edges directly connected to the node; then, the weighted degree centralities of all nodes are normalized by using a Softmax function, so that the sum of the weight coefficients of each node is 1, to obtain a weight distribution vector; finally, these normalized weight coefficients are converted into a diagonal matrix as a weight distribution matrix.

[0069] In an embodiment, the weighted degree centralities of the temperature node, the vibration node, and the current node in the inter-node dynamic contact graph are calculated as 1.86, 1.11, and 0.93, respectively. After Softmax normalization, the weight coefficients of the three are 0.6, 0.3, and 0.1, respectively. The 3x3 diagonal weight distribution matrix constructed therefrom has elements on the main diagonal line as 0.6, 0.3, and 0.1, respectively.

[0070] In step S1404, a risk feature set is obtained after weighted operation of the risk prediction vector and the weight distribution matrix.

[0071] It should be noted that the risk prediction vector is weighted with the weight distribution matrix, matrix multiplication is performed, and the risk prediction vector is right multiplied by the weight distribution matrix. The mathematical essence is that each dimension value of the risk prediction vector is multiplied by the normalized topology centrality weight of the corresponding node. The operation result obtains a new vector, i.e., a risk feature set, wherein each element represents a subdivided dimension risk quantization value weighted by the network topology importance.

[0072] In an embodiment, the risk prediction vector is (0.7, 0.4, 0.6) corresponding to the risk values of the current, temperature, and vibration nodes, respectively; and the weight distribution matrix is diag(0.1, 0.6, 0.3). The risk feature set obtained after the weighted operation is (0.07, 0.24, 0.18). As can be seen, the originally low absolute value of the temperature node risk value (0.4) is significantly improved in the contribution value (0.24) after weighting due to its central position in the topology (weight 0.6); and the contribution of the vibration node is relatively reduced.

[0073] In step S1405, the risk feature set is added and aggregated to obtain a comprehensive risk score.

[0074] It should be noted that all elements in the risk feature set vector are summed to obtain an original comprehensive risk score, and the original score is mapped to the interval of 0, 1 through a preset nonlinear mapping function to obtain a final comprehensive risk score. The nonlinear mapping function aims to improve the discrimination of high-risk states, amplify abnormal risk signals, and maintain stability in the normal risk range. For example, a variant of the S-shaped function can be used for mapping, wherein, is the original comprehensive risk score, is the mapped final score, and a and b are parameters determined by fitting historical data through logistic regression, for example, using the LogisticRegression model of Scikit-learn to control the steepness and center offset of the curve, respectively. By reasonably setting the parameters, the low risk interval (such as<0.3) can be mapped to change gently, and the high risk interval (such as>0.6) can be mapped to quickly approach 1, thereby significantly improving the warning sensitivity.

[0075] In one embodiment, the risk feature set is (0.07, 0.24, 0.18), representing the weighted risk values of the current, temperature, and vibration dimensions, respectively. After aggregation by addition, the original comprehensive risk score is 0.49. Assuming that the calibrated parameters are a = 5 and b = 0.5, the final comprehensive risk score after nonlinear mapping is 0.49; if the original score increases to 0.7 due to an anomaly, the mapped score is 0.73.

[0076] In step S15, the comprehensive risk score is corrected based on the ambient humidity data, and an insulation thermal aging risk coefficient is calculated in combination with the temperature signal data. If the insulation thermal aging risk coefficient exceeds the corresponding preset risk threshold, a preliminary warning signal is generated, including: S1501, obtaining ambient humidity data from the multi-source information data stream and mapping it in the comprehensive risk score to generate an ambient-corrected comprehensive risk score; S1502, calculating insulation strength by associating the ambient-corrected comprehensive risk score with the temperature signal data, and obtaining an insulation thermal aging risk coefficient by fusing the insulation strength with the harmonic distortion rate; S1503, generating a preliminary warning signal if the insulation thermal aging risk coefficient exceeds the preset insulation thermal aging risk threshold.

[0077] In step S1501, ambient humidity data is obtained from the multi-source information data stream and mapped in the comprehensive risk score to generate an ambient-corrected comprehensive risk score.

[0078] It should be noted that the comprehensive risk score generally reflects multiple potential problems in the operation of the equipment, while the ambient humidity data may affect the load fluctuation of the equipment. The synchronously collected ambient humidity time series is extracted from the obtained multi-source information data stream, and its average value in the current analysis time window is calculated. Subsequently, based on a predefined humidity-risk correction coefficient mapping table, the average humidity value is mapped to a specific correction coefficient; the mapping relationship between the humidity and the correction coefficient, i.e., mapping H to a specific correction coefficient k. This mapping relationship is obtained by fitting historical data, and a piecewise linear model can be used for calibration. For example, for a certain type of motor, based on one year of operation data, the following piecewise correction model is established, where H is the ambient humidity percentage. This model gives a slight suppression factor (0.95) at low humidity (H less than 40%), a linear correction in the normal humidity range (40%-80%), and further amplifies the risk at high humidity (H greater than 80%). In practical applications, the model should be calibrated according to the specific equipment type, installation environment, and historical operation data; the segment thresholds 40% and 80% refer to the IEC 60085 standard, and the coefficients 0.004 and 0.005 are based on linear regression of 1000 sets of historical humidity-failure data.

[0079] In one embodiment, the average ambient humidity in the current time window is calculated to be 75%, falling within the normal humidity interval, and the formula is substituted to obtain the correction factor k = 1.06. If the original comprehensive risk score is 0.72, then the comprehensive risk score after humidity correction is 0.763.

[0080] In step S1502, the ambient corrected comprehensive risk score is associated with the temperature signal data to calculate the insulation strength, and the insulation strength is fused with the harmonic distortion rate to obtain the insulation thermal aging risk coefficient.

[0081] It should be noted that the training samples are constructed based on historical insulation failure case data. The input features of each sample include the temperature average value, temperature variance extracted from the temperature time series data within a set time window before failure, and the corresponding ambient corrected comprehensive risk score. The target value of insulation degradation degree of each sample is calculated by the percentage decrease of the actual insulation resistance value relative to the rated insulation resistance value of the device, and is normalized to the interval [0, 1] (0 represents no loss, and 1 represents complete failure). Using the sample set constructed as described above, a regression model for calculating the insulation degradation degree coefficient A is trained. The model maps the aforementioned three input features to the insulation degradation degree coefficient A (value range 0-1).

[0082] The regression model is a neural network with a single hidden layer, the input layer contains 4 features such as temperature average value, temperature variance, comprehensive risk score, and the interaction term of comprehensive risk score and temperature average value, the hidden layer has 8 neurons, and the output layer is one neuron using Sigmoid activation function to ensure the output in the interval (0, 1). A dataset containing 500 historical samples (including 200 samples before insulation failure and 300 normal samples) is used for training, and the mean square error is used as the loss function. The output of the model is called the insulation thermal aging risk coefficient A (value range 0-1).

[0083] The input features of the neural network model include temperature mean value, variance, environment-corrected risk score, and interaction term; the output label is insulation resistance reduction percentage (normalized to 0-1); the model optimization uses Adam algorithm, the loss function is mean square error, and the training is stable to the validation set error.

[0084] The harmonic distortion rate THD is calculated from the frequency domain features of the current signal in the same monitoring period, and its value is also normalized to the interval [0, 1]. The insulation deterioration degree coefficient A and the harmonic distortion rate THD are fused by weighted summation to calculate the insulation thermal aging risk coefficient, and the specific formula is, Where, α and β are weight coefficients preset according to the contribution of the two types of risks in historical fault data, and satisfy α+β=1. Collect the monitoring data before all historical insulation fault cases occur, calculate the exceeding amplitude of the insulation deterioration degree coefficient A and the harmonic distortion rate THD relative to their normal operation reference values in each case; then, the average exceeding amplitude of A and the average exceeding amplitude of THD in all cases are calculated; finally, the average exceeding amplitude of A and the average exceeding amplitude of THD are divided by the sum of the two, and the ratio obtained is taken as the weight coefficients α and β respectively.

[0085] In one embodiment, when the past 5-minute sliding average temperature value is 65℃ and the corrected comprehensive risk score is 0.792, the insulation thermal aging risk coefficient calculated by inputting the model is 0.75; at the same time, the harmonic distortion rate THD extracted from the current signal is 0.4. The system records the two risk indicators as A=0.75, THD=0.4. Based on the analysis of 10 historical fault cases, the average exceeding amplitude of the insulation deterioration degree coefficient A is 0.5, and the average exceeding amplitude of the harmonic distortion rate THD is 0.2. Set α=0.71, β=0.29, then the insulation thermal aging risk coefficient is 0.648. In step S1503, if the insulation thermal aging risk coefficient exceeds the preset insulation thermal aging risk threshold, a preliminary warning signal is generated.

[0086] It should be noted that a preset insulation thermal aging risk threshold value is 0.5, which is determined according to the equipment insulation safety specification and based on statistical analysis of the historical failure case library. For example, through analysis of historical data, it is found that more than 80% of insulation failure cases have a continuous insulation thermal aging risk coefficient higher than 0.5 before the failure occurs, so the value is set as the early warning threshold. The calculated insulation thermal aging risk coefficient is compared with the threshold value. If it exceeds, it indicates that the insulation deterioration risk has reached a level that needs to be vigilant; at this time, the system automatically generates and outputs a preliminary early warning signal. The early warning signal is a structured data packet, which at least contains the risk triggering time, the risk type such as insulation deterioration risk, the risk level (which can be divided according to the percentage of exceeding the threshold value) and the key sensor node identifier involved such as the current node and the temperature node. In an embodiment, the calculated insulation thermal aging risk value is 0.61, which exceeds the preset threshold value 0.5. The system immediately generates an early warning signal, the content of which is "time: [time stamp]; risk type: composite electrical-thermal risk; risk level: intermediate; associated node: current sensor A, temperature sensor B." In step S16, noise interference in the preliminary early warning signal is removed by time sequence processing, continuity test and logical comparison are performed, and if the logical comparison result meets the preset logical threshold, a refined early warning sequence is obtained, including: S1601, mapping the preliminary early warning signal to a preset time window to form an original early warning sequence; S1602, calculating the sequence density of the original early warning sequence, and identifying the discrete distributed noise interference according to a preset discrimination standard; S1603, obtaining a candidate early warning sequence after removing the noise interference points, and obtaining a logical comparison result after a continuity test link; S1604, if the logical comparison result meets the preset logical threshold, extracting a feature vector of the candidate early warning sequence, and obtaining a refined early warning sequence after filtering processing.

[0087] In step S1601, the preliminary early warning signal is mapped to a preset time window to form an original early warning sequence.

[0088] It is to be noted that the one or more preliminary warning signals generated are mapped into preset consecutive time windows according to their triggering time to form a raw warning sequence for trend analysis. A fixed time window length, for example, 1 hour, is first set based on prior knowledge of the evolution rate of equipment failure or analysis requirements; then, all warning signals are assigned to the corresponding time window according to their time stamps; then, all warning signals in each time window are aggregated to calculate the warning intensity index of the window, for example, which can be defined as the sum of the risk level values of all warning signals in the window; finally, the warning intensity indexes are arranged in time window order to form a time-sequential raw warning sequence. The consecutive inspection window is used for warning event aggregation.

[0089] In an embodiment, the system generates 15 preliminary warning signals within 24 hours. The signals are assigned to the corresponding 24 windows with a fixed window length of 1 hour. For example, 3 signals are collected in the window from 10:00 to 11:00 am, with risk level values of 1, 2, and 1 respectively, and the warning intensity index of the window is calculated as 4. After processing all the windows in this way, a raw warning sequence containing 24 warning intensity index values is obtained, reflecting the concentration trend of warnings within a day.

[0090] In step S1602, the sequence density of the raw warning sequence is calculated, and noise interference points of discrete distribution are identified according to a preset discrimination criterion.

[0091] It is to be noted that the sliding window method is used to calculate the sequence density, and an analysis sub-window, for example, 3 consecutive time windows, is set to calculate the average value of the warning intensity index in each sub-window as the local density value of the center window. Then, noise interference is identified by comparing the warning intensity index of each window with the overall statistical characteristics of the sequence. First, the mean μ and standard deviation σ of the warning intensity index of the entire raw warning sequence are calculated; if the warning intensity index value of a time window satisfies the following two discrimination criteria at the same time, it is determined to be a discrete noise interference point. One is that the value is lower than μ-1.5σ; the other is that the difference between the value and the average value of the warning intensity indexes of the two time windows directly adjacent to it is greater than σ. The discrimination criterion is based on the principle of statistical distribution, aiming to capture abnormal points with significantly lower intensity and sudden changes in adjacent time signals.

[0092] In one embodiment, the original 24-hour warning sequence is analyzed, and the average value of the warning intensity index (dimensionless) is 4.2. In the window from 2:00 to 3:00, the warning intensity index is 1, but the values of the adjacent windows before and after it are 5 and 6, respectively. The local density is calculated as 4 (taking the average of the three windows before, the current window, and the window after), which is close to the global average. However, the mutation rate of the value (1) of this window relative to the values (5, 6) of the adjacent windows exceeds the preset difference threshold, so it is identified as a discrete noise interference point.

[0093] In step S1603, the noise interference points are removed to obtain a candidate warning sequence, and a logical comparison result is obtained after a continuity verification link.

[0094] It should be noted that the continuity verification of the candidate warning sequence can be performed by time window sliding matching. A continuity determination standard is set, for example, an effective continuous warning event must satisfy that it starts from a time window in which the warning intensity index is greater than zero, and the warning intensity index is continuously greater than zero in at least N-1 consecutive time windows from the starting window. The parameter N is the minimum continuous window number set according to the typical duration of the device failure precursor and the typical duration of the transient interference by analyzing historical data, for example, N=3. The system will traverse the entire candidate sequence, identify and mark all event segments that satisfy the continuity rule. The logical comparison result is a structured list that lists the starting time window, the total duration, and the average warning intensity of each continuous warning event that passes the verification.

[0095] In one embodiment, after noise removal, the warning signals in 24 hours are reduced from 50 windows to a candidate sequence involving 35 windows. After continuity verification (requiring at least 3 consecutive windows to have signals), the system identifies 2 effective continuous events, one lasting for 5 hours with an average intensity of 4.5, and the other lasting for 3 hours with an average intensity of 3.8.

[0096] In step S1604, if the logical comparison result meets the preset logical threshold, the feature vector of the candidate warning sequence is extracted, and a refined warning sequence is obtained after filtering.

[0097] It should be noted that a logical threshold for judging whether the early warning event has persistence and significance is set, and the minimum persistence time length such as 4 hours and the minimum average intensity such as 3.5 in the logical threshold are determined based on statistical analysis of the characteristics of historical effective early warning events, i.e. events finally confirmed to be related to real faults. The distribution of the persistence time length and the average intensity of all historical effective early warning events is calculated, and the lower quantile such as the 10th percentile is taken as the minimum requirement to ensure that most effective events are not filtered out, while most signals caused by non-fault factors, weak intensity or short duration are filtered out. The obtained logical comparison result is compared with the threshold, and if the persistence time length and the average intensity of a certain persistent early warning event meet the preset minimum requirement at the same time, it is determined that the event meets the threshold, and the system extracts the feature vector of the candidate early warning sequence including but not limited to the event duration, the maximum and average values of the early warning intensity within the event, and the occurrence frequency distribution of the early warning signal within the event time window. Then, all the extracted feature vectors are filtered based on the threshold of the feature value. After this filtering process, the finally retained event sequence constitutes the refined early warning sequence.

[0098] It should be noted that the feature vector of the refined early warning sequence needs to be zero-padded or truncated to a fixed length to adapt to the input dimension of the LSTM. The model structure uses a single-layer LSTM (hidden layer 128 units) followed by a Dropout layer (dropout rate 0.2), and the output layer uses a Sigmoid activation function.

[0099] In one embodiment, the preset logical threshold requires that the event lasts at least 4 hours and the average intensity is greater than 3.5. Three events are identified in the candidate sequence, one of which lasts 5 hours and has an average intensity of 4.5, and another of which lasts 4 hours and has an average intensity of 3.8. The system extracts the feature vectors of these two events, filters them (set the average intensity lower limit to 3.8), and finally only retains the event with an average intensity of 4.5 to form the refined early warning sequence.

[0100] In step S17, according to the refined early warning sequence, a final risk prediction output is generated in combination with a preset long short-term memory network model and distributed to a response system for execution, including: S1701, extracting multi-dimensional risk features in the refined early warning sequence and inputting them into the preset long short-term memory network model to obtain an original prediction value; S1702, calculating a risk probability distribution according to the original prediction value, and matching a corresponding risk level and generating confidence interval data if the risk probability distribution falls in a high-risk interval; S1703, encapsulating the risk level and the confidence interval data into an output message and generating a corresponding verification fingerprint; S1704, parsing the output message to identify the responding node and match the distribution protocol with the routing address; S1705, sending the output message carrying the check fingerprint to the responding system corresponding to the routing address through the distribution protocol, and executing the final instruction by the responding system.

[0101] In step S1701, the multi-dimensional risk features in the refined early warning sequence are extracted and input into a pre-set long short-term memory network model to obtain an original prediction value.

[0102] It should be noted that after obtaining the refined early warning sequence, a set of multi-dimensional risk features is extracted from the time window sequence of each persistent early warning event that passes the test, including signal intensity peak, fluctuation amplitude (variance), duration (window number), and frequency change rate (standard deviation of intensity change between adjacent windows). Each event corresponds to a fixed-dimensional feature vector.

[0103] The long short-term memory network (LSTM) model is constructed and trained. The input of the model is a sequence of the above-mentioned multi-dimensional risk features arranged in chronological order, and the output is the probability of failure occurring within 12 hours in the future. The model uses a single-layer LSTM structure, the number of hidden layer neurons is determined by grid search based on the performance of the validation set (for example, search in the set {64, 128, 256}, and finally select 128), and a Dropout layer (dropout rate set to 0.2) is added after the LSTM layer to alleviate overfitting, and finally a fully connected output layer using the Sigmoid activation function is connected. The training data of the model is composed of historical refined early warning sequences and their corresponding binary labels (1 for occurrence, 0 for non-occurrence) of whether a failure actually occurs within the next 12 hours. During training, the Adam optimizer is used, the learning rate is set to 0.001, and the binary cross-entropy is used as the loss function, and the training is stopped until the prediction performance (such as AUC) of the model on the validation set tends to be stable. The feature sequence extracted from the real-time refined early warning sequence to be analyzed is input into the trained LSTM model. After the model processes the entire sequence, an original prediction value between 0 and 1 is output, which represents the probability of failure of the device within the next 12 hours based on historical patterns.

[0104] In one embodiment, the refined early warning sequence contains a persistent early warning event lasting 8 hours, and the calculation obtains an intensity peak of 85 (normalized units), a fluctuation amplitude of 15%, a duration of 8 hours, and a frequency change rate of 10%. The 4-dimensional features of this event are input as a time step. If the current refined early warning sequence contains 3 events occurring in chronological order, a 3x4 feature matrix is formed and input into the model. The model outputs an original prediction value of 0.72, indicating that the device has a high probability of failure within the next 12 hours.

[0105] In step S1702, a risk probability distribution is calculated according to the original prediction value, and if the risk probability distribution falls in a high-risk interval, a corresponding risk level is matched and confidence interval data is generated.

[0106] It should be noted that the system can convert the original prediction value into a probability distribution through a softmax layer, and each dimension respectively represents the probability of low risk, medium risk and high risk. Secondly, the probability value corresponding to the "high risk" dimension in the distribution is compared with a preset high-risk interval threshold; the threshold is determined by analyzing the risk probability distribution characteristics of the early warning signals in historical failure cases. Statistics of the high-risk probability value distribution output by the model before all historical failure events occur, and a quantile that can cover most failure cases such as 90% and can effectively avoid false positives of normal fluctuations is selected as the threshold. For example, it is found that 90% of the failure cases have a high-risk probability value exceeding 0.6, and the probability of the value exceeding 0.6 under normal conditions is less than 5%, so 0.6 is set as the threshold. If the high-risk probability value exceeds this threshold, the result is matched to a specific risk level (for example, "level three risk") according to the preset mapping rule. Finally, based on the original prediction value and its historical fluctuation, the Bootstrap repeated sampling method is used to repeat sampling 1000 times, and the 95% confidence interval is taken.

[0107] In one embodiment, the original prediction value is 0.72, and the risk probability distribution obtained after Softmax conversion is low risk 0.10, medium risk 0.25, and high risk 0.65. Since the high-risk probability 0.65 exceeds the threshold 0.6, the system matches it to "level three risk", and the 95% confidence interval is calculated to be [0.58, 0.72] through repeated sampling. The confidence interval data is used to evaluate the credibility of the prediction result.

[0108] In step S1703, the risk level and the confidence interval data are packaged into an output message and a corresponding verification fingerprint is generated.

[0109] It should be noted that the determined risk level and the calculated confidence interval data are packaged into a structured output message, and a verification fingerprint for verifying the integrity of the message is generated. First, a standard format data message is created, for example, using a JSON structure, which at least includes "risk level" (such as "level three"), "confidence interval lower limit" (such as 0.58) and "confidence interval upper limit" (such as 0.72). Then, all the contents (field names and values) of the message are normalized and concatenated, and a hash algorithm such as SHA-256 is used to encrypt and calculate the entire concatenated string to generate a fixed-length unique hash value as the verification fingerprint of the message. Finally, the fingerprint is attached to the end of the message or a separate "verification code" field.

[0110] In an embodiment, the system generates a message content of {"risk_level":"level 3", "ci_lower":0.58, "ci_upper":0.72}, calculates the SHA-256 hash value thereof, and obtains a verification fingerprint of a1b2c3d4... The fingerprint is sent or stored together with the message; the receiving party can verify whether the data has been tampered with in the transmission process by recalculating the hash of the message and comparing it with the attached fingerprint.

[0111] In step S1704, the output message is parsed to identify the responding node and match the distribution protocol and the routing address.

[0112] It should be noted that the system parses the message content and extracts the "risk level" and other key fields therein; according to the pre-configured rule base, different risk levels are mapped to corresponding responsible nodes (for example, "level 3 risk" is mapped to the maintenance management system, and "level 1 risk" is mapped to the on-site mobile device). Subsequently, according to the type of the identified responding node, a corresponding network distribution protocol (for example, the MQTT protocol is used for mobile devices that need real-time pushing, and the HTTPS protocol is used for management systems that need secure transmission) and its corresponding routing address (such as IP address and port number) are matched. The address information is obtained from the pre-configured node address table.

[0113] In an embodiment, the parsed message has a risk level of "level 3", and the system identifies the responding node as "maintenance management system" accordingly. After querying the rule base, it matches the MQTT protocol and the corresponding routing address 192.168.1.100:1883. Before final sending, the system verifies the integrity of the message using the verification fingerprint attached in the message, and confirms that it is correct. After confirmation, the message is distributed to the target node according to the protocol and address determined above.

[0114] In step S1705, the output message carrying the verification fingerprint is sent to the responding system corresponding to the routing address through the distribution protocol, and the final instruction is executed by the responding system.

[0115] It should be noted that the output message carrying the check fingerprint is sent to the response system corresponding to the routing address through the distribution protocol. After receiving the message, the response system first recalculates the hash value of the message and compares it with the check fingerprint attached in the message to verify the integrity of the data during transmission. After verification, the response system parses the message content, executes the preset final instruction corresponding to the risk level according to the "risk level" field, and completes the early warning process. The instruction is usually an automatic operation. If the risk level is three, a maintenance work order is automatically generated and sent to the maintenance personnel terminal. If it is level one, the sound and light alarm is triggered and the equipment shutdown program is started. For example, triggering the sound and light alarm, sending a notification message to the designated maintenance personnel, or issuing a command to shut down the equipment for inspection.

[0116] In an embodiment, the system successfully sends the message to the address of the maintenance management system through the MQTT protocol. After verifying that the fingerprint is correct, the maintenance management system parses the risk level as "three", and immediately generates and distributes a "immediately arrange on-site inspection" work order to the mobile terminal of the relevant personnel, and the yellow early warning indicator light on the large screen of the monitoring center is lit, thereby realizing timely intervention on potential faults.

[0117] In summary, the present application discloses a regional fire warning method based on multiple sensors, which includes obtaining current, temperature and humidity multi-source data streams through a sensor network, and obtaining time sequence feature vectors after preprocessing; a dynamic contact graph between nodes is constructed and analyzed using a graph neural network to determine the hidden node correlation strength; if the correlation strength exceeds the threshold, the time correlation sequence of the strongly correlated node is extracted, and a comprehensive risk score is constructed and weighted accordingly; the score threshold is dynamically corrected based on environmental humidity and insulation thermal aging risk coefficient, and a warning signal is generated when the risk score exceeds the threshold; the warning signal is subjected to continuity test and noise filtering to obtain a refined warning sequence; finally, a risk prediction result is generated through a long short-term memory network model, and after encapsulation and verification, it is distributed to the response system to complete the warning. The present application realizes early and accurate warning of complex electrical fire hazards by mining the time sequence correlation and dynamic coupling of multiple source signals, effectively reducing the false positive and false negative rates.

[0118] Referring to Figure 2 , the second embodiment of the present application provides a regional fire warning system based on multiple sensors, comprising: a multi-source data acquisition and preprocessing module for obtaining multi-source information data streams and obtaining time sequence feature vectors after preprocessing; the multi-source information data streams include current fluctuation data, temperature signal data and environmental humidity data; a dynamic correlation graph construction and analysis module for constructing a dynamic contact graph between nodes based on the time sequence feature vectors and the physical connection relationship between sensors, and determining the correlation strength between nodes in the hidden mode through analysis using a graph neural network. The correlation strength threshold judgment and feature extraction module extracts the time correlation characteristic sequence of the corresponding node from the inter-node dynamic contact graph if the inter-node correlation strength exceeds a preset correlation strength threshold. The risk feature fusion and evaluation module constructs a risk prediction vector according to the time correlation characteristic sequence, constructs a weight distribution matrix according to the node topology centrality corresponding to the risk prediction vector, and performs weighted summation on the risk prediction vector and the weight distribution matrix to generate a comprehensive risk score. The adaptive threshold adjustment and preliminary warning module 5. modifies the comprehensive risk score based on the environmental humidity data, and calculates an insulation thermal aging risk coefficient based on the temperature signal data. If the insulation thermal aging risk coefficient exceeds the corresponding preset risk threshold, a preliminary warning signal is obtained. The warning signal purification and refining module removes noise interference in the preliminary warning signal through time sequence processing to obtain a candidate warning sequence, and performs continuity test and logical comparison. If the logical comparison result meets the preset logical threshold, a refined warning sequence is obtained. The risk prediction output and response distribution module generates a final risk prediction output based on the refined warning sequence in combination with a preloaded long short-term memory network model, and distributes it to a response system for execution.

[0119] It should be noted that the region fire warning system based on multiple sensors provided by the embodiments of the present application is used to execute all process steps of the region fire warning method based on multiple sensors provided by the embodiments described above. The working principles and beneficial effects of the two are one-to-one correspondence, and thus will not be repeated.

[0120] The embodiments of the present application also provide an electronic device. The electronic device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, such as a region fire warning program based on multiple sensors. The processor implements the steps in each of the above region fire warning methods based on multiple sensors when executing the computer program, such as Figure 1 The steps S11 shown. Alternatively, the processor implements the functions of each module / unit in each of the above system embodiments when executing the computer program, such as the risk feature fusion and evaluation module.

[0121] It should be noted that the system embodiments described above are merely illustrative, wherein the units described as separate components can or can not be physically separated, and the components displayed as units can or can not be physical units, i.e. can be located in one place or distributed to multiple network units. Part or all of the modules can be selected to achieve the purpose of the embodiment according to actual needs. In addition, the connection relationship between the modules in the system embodiment provided by the present application indicates that there is a communication connection between them, which can be implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement it without creative labor.

[0122] The above specific embodiments further illustrate the purpose, technical scheme and beneficial effects of the present application. It should be understood that the above description is only a specific embodiment of the present application and is not intended to limit the protection scope of the present application. It is particularly pointed out that any modification, equivalent replacement, improvement, etc. made by those skilled in the art within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A regional fire early warning method based on multiple sensors, characterized in that, include: The multi-source information data stream is acquired, and the time series feature vector is obtained after preprocessing. The multi-source information data stream includes current fluctuation data, temperature signal data, and environmental humidity data; Based on the time series feature vectors and the physical connection relationships between sensors, a graph neural network is used to construct a dynamic connection graph between nodes, and the correlation strength between nodes in the hidden mode is determined through parsing. If the correlation strength between the nodes exceeds a preset correlation strength threshold, then the time correlation characteristic sequence of the corresponding node is extracted from the dynamic connection graph between the nodes. A risk prediction vector is constructed based on the time-related characteristic sequence, and a weight allocation matrix is ​​constructed based on the node topological centrality corresponding to the risk prediction vector. The risk prediction vector and the weight allocation matrix are then weighted and summed to generate a comprehensive risk score. The comprehensive risk score is corrected based on the environmental humidity data, and the insulation thermal aging risk coefficient is calculated in combination with the temperature signal data. If the insulation thermal aging risk coefficient exceeds the corresponding preset risk threshold, a preliminary warning signal is generated. Noise interference in the preliminary warning signal is removed by time-series processing to obtain a candidate warning sequence. Continuity checks and logical comparisons are then performed. If the logical comparison result meets a preset logical threshold, a refined warning sequence is obtained. Based on the refined early warning sequence, a final risk prediction output is generated by combining it with a pre-set long short-term memory network model, and then distributed to the response system for execution.

2. The multi-sensor-based regional fire early warning method according to claim 1, characterized in that, The multi-source information data stream is acquired, and the time series feature vector is obtained after preprocessing. The multi-source information data stream includes current fluctuation data, temperature signal data, and environmental humidity data, including: The sensor acquires a multi-source information data stream containing current fluctuation data, temperature signal data, and ambient humidity data. The multi-source information data stream is subjected to low-pass filtering to remove noise and signal alignment to obtain a standardized data stream; The standardized data stream is segmented and truncated using a sliding window to obtain the time series segments to be analyzed. Extract the frequency domain components and time domain statistics of the time segment to be analyzed to obtain a multidimensional feature set; Principal component analysis and dimensionality compression are performed on the multidimensional feature set to obtain the time series feature vector.

3. The regional fire early warning method based on multiple sensors according to claim 1, characterized in that, Based on the time-series feature vectors and the physical connections between sensors, a graph neural network is used to construct a dynamic connection graph between nodes, and the strength of the association between nodes in the hidden mode is determined through parsing, including: The time series feature vectors and the physical connection relationships between sensors are combined to generate an initial adjacency matrix; Calculate the feature similarity between nodes in the initial adjacency matrix and construct a weighted graph structure; The weighted graph structure is input into the graph neural network layer to update the hidden state vector corresponding to each node. Adjust the dynamic weights on the propagation path in the weighted graph structure according to the difference values ​​between the hidden layer state vectors, and reconstruct and generate a dynamic connection graph between nodes; The connection tightness of the dynamic relationship graph between the nodes is analyzed to determine the strength of the association between nodes in the hidden mode.

4. The regional fire early warning method based on multiple sensors according to claim 1, characterized in that, If the correlation strength between the nodes exceeds a preset correlation strength threshold, then the temporal correlation characteristic sequence of the corresponding nodes is extracted from the dynamic connection graph between the nodes, including: If the correlation strength between the nodes does not exceed the preset correlation strength threshold, then a supplementary signal is obtained from the multi-source information data stream; The supplementary signal is subjected to feature alignment processing, and the updated topology result of the dynamic connection graph between nodes is obtained based on the alignment processing result; Based on the updated topology results, the association strength between nodes is recalculated. If the updated association strength exceeds the preset association strength threshold, the temporal feature data of the corresponding nodes is extracted from the dynamic connection graph between nodes, and arranged and spliced ​​in chronological order to generate a temporal association feature sequence.

5. The multi-sensor-based regional fire early warning method according to claim 1, characterized in that, A risk prediction vector is constructed based on the time-related characteristic sequence, and a weight allocation matrix is ​​constructed based on the node topological centrality corresponding to the risk prediction vector. The risk prediction vector and the weight allocation matrix are then weighted and summed to generate a comprehensive risk score, including: A multidimensional fault feature matrix is ​​constructed based on the time-related characteristic sequence; The multidimensional fault feature matrix is ​​transformed by feature space mapping to generate a risk prediction vector; Based on the topological centrality of the nodes corresponding to the risk prediction vector, the weight coefficients are obtained by normalization using the Softmax function, and the main diagonal elements are determined to construct a weight allocation matrix. The risk feature set is obtained by performing a weighted operation between the risk prediction vector and the weight allocation matrix. The comprehensive risk score is obtained by additive aggregation of the risk feature set.

6. The regional fire early warning method based on multiple sensors according to claim 1, characterized in that, The comprehensive risk score is corrected based on the environmental humidity data, and the insulation thermal aging risk coefficient is calculated in conjunction with the temperature signal data. If the insulation thermal aging risk coefficient exceeds the corresponding preset risk threshold, a preliminary warning signal is generated, including: Environmental humidity data is obtained from the multi-source information data stream and mapped into the comprehensive risk score to generate an environmentally corrected comprehensive risk score. The insulation strength is calculated by correlating the environmentally corrected comprehensive risk score with the temperature signal data, and the insulation strength is then fused with the harmonic distortion rate to obtain the insulation thermal aging risk coefficient. If the insulation thermal aging risk coefficient exceeds the preset insulation thermal aging risk threshold, a preliminary warning signal will be generated.

7. The regional fire early warning method based on multiple sensors according to claim 1, characterized in that, Noise interference in the preliminary warning signal is removed through time-series processing to obtain a candidate warning sequence. Continuity checks and logical comparisons are then performed. If the logical comparison result meets a preset logical threshold, a refined warning sequence is obtained, including: The initial warning signal is mapped to a preset time window to form the original warning sequence; Calculate the sequence density of the original warning sequence, and identify discrete noise interference points based on preset discrimination criteria; After removing the noise interference points, a candidate warning sequence is obtained, and after a continuity test, a logical comparison result is obtained. If the logical comparison result meets the preset logical threshold, the feature vector of the candidate warning sequence is extracted, and a refined warning sequence is obtained after filtering.

8. The multi-sensor-based regional fire early warning method according to claim 1, characterized in that, Based on the refined early warning sequence, a final risk prediction output is generated using a pre-set long short-term memory network model and distributed to the response system for execution, including: Extract the multidimensional risk features from the refined early warning sequence and input them into a pre-set long short-term memory network model to obtain the original predicted value; The risk probability distribution is calculated based on the original predicted value. If the risk probability distribution falls in the high-risk interval, the corresponding risk level is matched and confidence interval data is generated. The risk level and the confidence interval data are encapsulated into an output message and a corresponding verification fingerprint is generated. The output message is parsed to identify the responding node and match the distribution protocol and routing address; The output message carrying the verification fingerprint is sent to the response system corresponding to the routing address through the distribution protocol, and the response system executes the final instruction.

9. A regional fire early warning system based on multiple sensors, characterized in that, include: The multi-source data acquisition and preprocessing module acquires multi-source information data streams and obtains time series feature vectors after preprocessing. The multi-source information data stream includes current fluctuation data, temperature signal data, and environmental humidity data; The dynamic association graph construction and parsing module, based on the time series feature vector and the physical connection relationship between sensors, uses a graph neural network to construct a dynamic association graph between nodes, and determines the association strength between nodes in the hidden mode through parsing; The association strength threshold judgment and feature extraction module extracts the temporal association characteristic sequence of the corresponding node from the dynamic connection graph between the nodes if the association strength between the nodes exceeds the preset association strength threshold. The risk feature fusion and assessment module constructs a risk prediction vector based on the time-related characteristic sequence, and constructs a weight allocation matrix based on the node topological centrality corresponding to the risk prediction vector. The risk prediction vector and the weight allocation matrix are weighted and summed to generate a comprehensive risk score. The adaptive threshold adjustment and preliminary warning module corrects the comprehensive risk score based on the environmental humidity data and calculates the insulation thermal aging risk coefficient in combination with the temperature signal data. If the insulation thermal aging risk coefficient exceeds the corresponding preset risk threshold, a preliminary warning signal is generated. The warning signal purification and refinement module removes noise interference from the preliminary warning signal through time-series processing to obtain a candidate warning sequence, and performs continuity checks and logical comparisons. If the logical comparison result meets the preset logical threshold, a refined warning sequence is obtained. The risk prediction output and response distribution module generates the final risk prediction output based on the refined early warning sequence and a pre-set long short-term memory network model, and distributes it to the response system for execution.

Images