Real-time statistical analysis method and system for operation state of electrical equipment

Abstract

The invention provides an electrical equipment operation state real-time statistical analysis method and system, and belongs to the field of electrical equipment operation state monitoring, and the method comprises the steps: obtaining multi-dimensional time sequence sensor data during the operation of electrical equipment, selecting a proper wavelet basis function based on Shannon information entropy to carry out wavelet packet decomposition, and extracting time-frequency domain features; constructing a dynamic relation graph which takes a sensor as a node and takes a Granger causal relationship as a weight, and learning node space features by using a graph convolutional network in combination with time-frequency features; the node features are input into a gating circulation unit, and electrical equipment state sequence codes fused with space-time dependence are generated; and carrying out Viterbi decoding by using the state transition cost matrix, and reasoning an optimal electrical equipment operation state path. According to the method, the uncertainty of a prediction result can be quantified, clear confidence evaluation is provided for a final analysis conclusion, and the reliability of decision making is improved.

Description

Technical Field

[0001] This application belongs to the field of electrical equipment operation status monitoring, and in particular relates to a method and system for real-time statistical analysis of the operation status of electrical equipment. Background Technology

[0002] In fields such as industrial production and energy, real-time and accurate monitoring and analysis of the operating status of large and complex electrical equipment is a key link in ensuring safe production, improving operational efficiency, and implementing predictive maintenance.

[0003] Traditional methods for analyzing the condition of electrical equipment primarily rely on signal processing techniques or classical machine learning models. For example, methods based on Fourier transform or short-time Fourier transform, while capable of analyzing the frequency domain characteristics of signals, struggle to simultaneously address the need for time-frequency analysis. While wavelet analysis offers improvements in time-frequency analysis, the selection of its basis functions often depends on expert experience, lacking adaptability to different signal characteristics and making it difficult to extract optimal features from complex and variable electrical equipment signals. Furthermore, some methods employing deep learning models such as recurrent neural networks (RNNs) and long short-term memory networks (LSTMs), while effectively capturing the temporal dependencies of data, typically treat multiple sensor signals as independent channels, ignoring the complex topological relationships between different sensors in physical space and functional interconnections. This results in an inability to fully utilize the spatial correlation information between components of the electrical equipment, limiting the depth and accuracy of the model's understanding of the overall condition of the electrical equipment.

[0004] To address these issues, some research has begun to explore the integration of Graph Neural Networks (GNNs) and temporal models to simultaneously capture the spatiotemporal dependencies of electrical equipment state data. However, existing technologies still have several shortcomings: When constructing sensor relationship graphs, most methods employ static graph structures based on physical proximity or prior knowledge. This fixed topology fails to reflect the causal relationships between sensors dynamically changing under different operating conditions, reducing the model's representational power. In parsing state sequences, many methods only output instantaneous states through simple classifiers, neglecting the physical logic and temporal constraints of state transitions, easily leading to frequent and unreasonable jumps that do not conform to actual operating patterns. Furthermore, existing models often provide only a deterministic prediction point when giving state analysis results, failing to quantify the reliability or uncertainty of the prediction, which is a significant drawback in industrial scenarios requiring highly reliable decision-making. When encountering unseen operating conditions or subject to noise interference, a reliable confidence assessment is crucial for operator decision-making and risk management. Summary of the Invention

[0005] To address the above problems, this application proposes a method for real-time statistical analysis of the operating status of electrical equipment, comprising the following steps:

[0006] Acquire multidimensional time-series sensor data generated during the operation of electrical equipment; based on the Shannon information entropy of each dimension signal in the multidimensional time-series sensor data, select wavelet basis functions for each dimension signal, and use the wavelet basis functions to perform wavelet packet decomposition on the data to extract time-frequency domain features;

[0007] A dynamic sensor relationship graph is constructed with sensors as nodes and Granger causality index between sensors as weights, with the weight values ​​updated periodically. The time-frequency domain features are used as node attributes, and a graph convolutional network is used to calculate the dynamic sensor relationship graph under each time slice to learn and fuse the node state features of the spatial topology. The node state features are then input into the gated recurrent unit in time sequence to obtain the electrical equipment state sequence encoding that fuses spatiotemporal dependence.

[0008] Using a preset state transition cost matrix of physical constraints for switching between different operating states, the state sequence encoding of the electrical equipment is Viterbi decoded to obtain the optimal operating state path of the electrical equipment. In the inference phase, Monte Carlo dropout is performed in the graph convolutional network and gated recurrent unit, and multiple forward propagations are performed to obtain the probability prediction distribution of the electrical equipment in different operating states at each time. The variance of the probability prediction distribution is calculated and used as the confidence level of the operating state analysis result.

[0009] Preferably, the step of selecting wavelet basis functions for signals in each dimension includes:

[0010] For each dimension signal in the multidimensional time-series sensor data, each wavelet basis function is selected sequentially from the preset wavelet basis function library to perform multi-level wavelet packet decomposition on the signal, and the Shannon information entropy of the energy distribution of each frequency band after decomposition is calculated. The wavelet basis function that minimizes the Shannon information entropy value is selected as the optimal wavelet basis function for the dimension signal.

[0011] Preferably, the extraction of time-frequency domain features includes:

[0012] The energy values ​​of each frequency band obtained after decomposition by the optimal wavelet basis function are used as the time-frequency domain features of the dimensional signal.

[0013] Preferably, the construction of a dynamic sensor relationship graph, with sensors as nodes and Granger causality indices between sensors as weights and the weight values ​​updated periodically, includes:

[0014] Using a sliding time window of length N and a step size of M sampling points, the Granger causality index is calculated by establishing a vector autoregression model for the time series data of any two sensors within the sliding time window.

[0015] The Granger causality indices of all sensor pairs are used to construct an adjacency matrix, which serves as the weight of the dynamic sensor relationship graph under the time slice.

[0016] Preferably, the step of using a graph convolutional network to calculate the dynamic sensor relationship graph for each time slice and learning the node state features that fuse spatial topology includes:

[0017] For any node in the graph, the time-frequency domain features of all its neighboring nodes are weighted and aggregated according to the Granger causality index to obtain aggregated neighborhood features.

[0018] The aggregated neighborhood features are concatenated with the node's own time-frequency domain features, and then nonlinearly transformed through a fully connected layer with a ReLU activation function to obtain node state features that incorporate spatial topological information.

[0019] Preferably, the step of performing Viterbi decoding on the electrical equipment state sequence encoding using a preset state transition cost matrix that utilizes physical constraints for switching between different operating states includes:

[0020] A K×K state transition cost matrix is ​​predefined, where K is the total number of operating states of electrical equipment, and the element C(i,j) of the state transition cost matrix represents the cost of transitioning from state i to state j. Transitions that are physically impossible have an infinite cost.

[0021] In the recursive process of Viterbi decoding, the element C(i,j) of the state transition cost matrix is ​​added as a penalty term to the logarithmic calculation of the state transition probability to constrain the search for the optimal electrical equipment operating state path.

[0022] Preferably, the step of obtaining the probability prediction distribution of the electrical equipment in different operating states at each time by performing Monte Carlo dropout and multiple forward propagations in the graph convolutional network and gated recurrent unit, calculating the variance of the probability prediction distribution, and quantifying the variance as the confidence level of the operating state analysis result includes:

[0023] During the inference phase, the dropout layer in the graph convolutional network and gated recurrent unit is activated, and the same input data is forwarded multiple times to obtain multiple sets of probabilistic prediction results of the operating status of electrical equipment.

[0024] Based on the multiple sets of probability prediction results, the probability distribution of the electrical equipment in different operating states at each time is calculated, and the variance of the probability distribution is calculated. The variance is used as the confidence level of the operating state analysis results.

[0025] Furthermore, this invention also proposes a real-time statistical analysis system for the operating status of electrical equipment, comprising the following units:

[0026] The feature extraction unit is used to acquire multi-dimensional time-series sensor data generated during the operation of electrical equipment; based on the Shannon information entropy of each dimension signal in the multi-dimensional time-series sensor data, a wavelet basis function is selected for each dimension signal, and the wavelet basis function is used to perform wavelet packet decomposition on the data to extract time-frequency domain features;

[0027] The encoding unit is used to construct a dynamic sensor relationship graph with sensors as nodes and Granger causality index between sensors as weights, with the weight values ​​being updated periodically. The time-frequency domain features are used as node attributes, and a graph convolutional network is used to calculate the dynamic sensor relationship graph under each time slice to learn and fuse the node state features of the spatial topology. The node state features are then input into the gated recurrent unit in time sequence to obtain the electrical equipment state sequence encoding that fuses the spatiotemporal dependence.

[0028] The statistical analysis unit is used to perform Viterbi decoding on the electrical equipment state sequence encoding using a preset state transition cost matrix of physical constraints between different operating states, and to parse out the optimal electrical equipment operating state path; in the inference stage, by performing Monte Carlo dropout and multiple forward propagations in the graph convolutional network and gated recurrent unit, the probability prediction distribution of the electrical equipment in different operating states at each time is obtained, and the variance of the probability prediction distribution is calculated, and the variance is used as the confidence level of the operating state analysis result.

[0029] Preferably, the step of selecting wavelet basis functions for signals in each dimension includes:

[0030] For each dimension signal in the multidimensional time-series sensor data, each wavelet basis function is selected sequentially from the preset wavelet basis function library to perform multi-level wavelet packet decomposition on the signal, and the Shannon information entropy of the energy distribution of each frequency band after decomposition is calculated. The wavelet basis function that minimizes the Shannon information entropy value is selected as the optimal wavelet basis function for the dimension signal.

[0031] Preferably, the extraction of time-frequency domain features includes:

[0032] The energy values ​​of each frequency band obtained after decomposition by the optimal wavelet basis function are used as the time-frequency domain features of the dimensional signal.

[0033] Preferably, the construction of a dynamic sensor relationship graph, with sensors as nodes and Granger causality indices between sensors as weights and the weight values ​​updated periodically, includes:

[0034] Using a sliding time window of length N and a step size of M sampling points, the Granger causality index is calculated by establishing a vector autoregression model for the time series data of any two sensors within the sliding time window.

[0035] The Granger causality indices of all sensor pairs are used to construct an adjacency matrix, which serves as the weight of the dynamic sensor relationship graph under the time slice.

[0036] Preferably, the step of using a graph convolutional network to calculate the dynamic sensor relationship graph for each time slice and learning the node state features that fuse spatial topology includes:

[0037] For any node in the graph, the time-frequency domain features of all its neighboring nodes are weighted and aggregated according to the Granger causality index to obtain aggregated neighborhood features.

[0038] The aggregated neighborhood features are concatenated with the node's own time-frequency domain features, and then nonlinearly transformed through a fully connected layer with a ReLU activation function to obtain node state features that incorporate spatial topological information.

[0039] Preferably, the step of performing Viterbi decoding on the electrical equipment state sequence encoding using a preset state transition cost matrix that utilizes physical constraints for switching between different operating states includes:

[0040] A K×K state transition cost matrix is ​​pre-defined, where K is the total number of operating states of electrical equipment, and the element C(i,j) of the state transition cost matrix represents the cost of transitioning from state i to state j. Transitions that are physically impossible have an infinite cost.

[0041] In the recursive process of Viterbi decoding, the state transition cost matrix element C(i,j) is added as a penalty term to the logarithmic calculation of the state transition probability to constrain the search for the optimal electrical equipment operating state path.

[0042] Preferably, the step of obtaining the probability prediction distribution of the electrical equipment in different operating states at each time by performing Monte Carlo dropout and multiple forward propagations in the graph convolutional network and gated recurrent unit, and calculating the variance of the probability prediction distribution, using the variance as the confidence level of the operating state analysis result, includes:

[0043] During the inference phase, the dropout layer in the graph convolutional network and gated recurrent unit is activated, and the same input data is forwarded multiple times to obtain multiple sets of probabilistic prediction results of the operating status of electrical equipment.

[0044] Based on the multiple sets of probability prediction results, the probability distribution of the electrical equipment in different operating states at each time is calculated, and the variance of the probability distribution is calculated. The variance is used as the confidence level of the operating state analysis results.

[0045] Compared with the prior art, the present invention has the following technical effects:

[0046] 1. Select the optimal wavelet basis function based on the information entropy characteristics of the signal itself, so as to extract the key time-frequency domain features reflecting the state of electrical equipment more accurately.

[0047] 2. By constructing a sensor correlation graph using Granger causality and periodically updating the weights, the inherent relationships between various components of electrical equipment that change with operating conditions can be captured, transcending the limitations of fixed physical topology.

[0048] 3. By deeply fusing graph convolutional networks and gated recurrent units, the complex dependencies of electrical equipment states in the spatial and temporal dimensions were comprehensively learned.

[0049] 4. By introducing a state transition cost matrix with physical constraints for Viterbi decoding, the final output state sequence is guaranteed to conform to the physical logic of electrical equipment operation, effectively avoiding disordered jumps in the results.

[0050] This invention can quantify the uncertainty of prediction results, provide a clear confidence assessment for the final analysis conclusion, and greatly enhance the reliability and practical value of the method in actual industrial decision-making. Attached Figure Description

[0051] Figure 1 Flowchart for an embodiment;

[0052] Figure 2 This is a schematic diagram of data from a temperature sensor, pressure sensor, and vibration sensor.

[0053] Figure 3 A schematic diagram of wavelet basis selection;

[0054] Figure 4 This is a schematic diagram of wavelet packet decomposition and time-frequency feature extraction;

[0055] Figure 5 This is a diagram showing the relationship between sensors.

[0056] Figure 6 This is a schematic diagram of the fusion model;

[0057] Figure 7 This is a schematic diagram of state path parsing based on Viterbi decoding;

[0058] Figure 8 A schematic diagram of the Monte Carlo dropout confidence quantification. Detailed Implementation

[0059] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0060] In a specific embodiment, this application proposes a method for real-time statistical analysis of the operating status of electrical equipment, such as... Figure 1 As shown, it includes the following steps:

[0061] S1, acquire multi-dimensional time-series sensor data generated during the operation of electrical equipment; based on the Shannon information entropy of each dimension signal in the multi-dimensional time-series sensor data, select wavelet basis functions for each dimension signal, and use the wavelet basis functions to perform wavelet packet decomposition on the data to extract time-frequency domain features;

[0062] Data is synchronously collected from multiple sensors installed on electrical equipment using a data acquisition system, such as a distributed control system or a data acquisition card. These sensors include, for example, temperature sensors, pressure sensors, and vibration sensors, forming a multi-dimensional time-series data matrix with time as the row and sensor channels as the column. Figure 2 As shown.

[0063] A pre-defined wavelet basis function library is provided, containing various commonly used wavelet bases such as the Daubechies, Coiflets, and Symlets series. For the time-series signal of each sensor channel, its Shannon information entropy value is calculated. Wavelet packet decomposition is performed on the signal using each wavelet base from the library, and the information entropy of the energy distribution in each frequency band after decomposition is calculated. The wavelet base that minimizes the entropy value is selected as the optimal wavelet base for the signal. Figure 3 As shown; wavelet packet decomposition is performed on the signal using the selected optimal wavelet basis, as follows. Figure 4 As shown, the energy, kurtosis and other statistical quantities of each frequency band are extracted as the time-frequency domain features of the signal in the current time window.

[0064] S2, construct a dynamic sensor relationship graph with sensors as nodes and Granger causality index between sensors as weights, with the weight values ​​updated periodically; use the time-frequency domain features as node attributes, use a graph convolutional network to calculate the dynamic sensor relationship graph under each time slice, learn the node state features that fuse spatial topology, and then input the node state features into the gated recurrent unit in time sequence to obtain the electrical equipment state sequence encoding that fuses spatiotemporal dependence;

[0065] Each sensor is considered a node in the graph. Within each preset time window, for the time series data corresponding to any two sensor nodes, a vector autoregression model is used to calculate the F-statistic or p-value of the Granger causality test between them, and this statistic is used as the weight of the edge connecting the two nodes. By sliding the time window, the Granger causality index between all node pairs is calculated periodically, thereby generating a series of time-varying weighted adjacency matrices to represent the dynamic sensor relationship graph, such as... Figure 5 As shown.

[0066] For each time slice, the weighted adjacency matrix of the dynamic sensor relationship graph at that moment, along with the corresponding time-frequency domain feature matrix, is input into a multi-layer graph convolutional network model. Graph convolution operations aggregate neighbor node information to generate high-level node feature representations that include spatial correlations between sensors. Then, the sequence of high-level node feature representations generated sequentially for all time slices is input into a gated recurrent unit network (GRU). Its update and reset gates are used to capture the long-term dependencies of the feature sequence in the time dimension. Finally, the hidden state sequence output by the network is the electrical equipment state sequence encoding that incorporates spatiotemporal dependencies. Figure 6 As shown.

[0067] S3. Using the preset state transition cost matrix of physical constraints between different operating states, the state sequence encoding of the electrical equipment is Viterbi decoded to obtain the optimal operating state path of the electrical equipment. In the inference stage, Monte Carlo dropout is performed in the graph convolutional network and gated recurrent unit and multiple forward propagations are performed to obtain the probability prediction distribution of the electrical equipment in different operating states at each time, and the variance of the probability prediction distribution is calculated. The variance is used as the confidence level of the operating state analysis result.

[0068] The electrical equipment state sequence output by the gated recurrent unit is encoded and passed through a fully connected layer and a Softmax activation function to obtain the transmission probability of the electrical equipment in each predefined operating state at each time step. Simultaneously, a state transition cost matrix is ​​constructed based on expert knowledge or historical data statistics, such as... Figure 7 As shown, the elements of the state transition cost matrix define the cost of transitioning from one operating state to another. For example, the cost of transitions that do not conform to physical laws, such as going directly from normal to severe fault, is set to a maximum value. The emission probability sequence and the state transition cost matrix are used as inputs to the Viterbi algorithm. By solving the problem through dynamic programming, a state sequence with the maximum joint probability is obtained, which is the optimal operating state path of the electrical equipment.

[0069] During model inference, the dropout layers in the graph convolutional network and gated recurrent units are kept active. For the same input data, the complete forward propagation computation is repeated T times. Since the neurons that are randomly deactivated in each propagation are different, T slightly different state prediction probability sequences are obtained. For each time step, the element-wise variance of these T state prediction probability sequences is calculated. The magnitude of this variance directly reflects the uncertainty of the model's prediction result at that time step. The smaller the variance, the higher the confidence level, and vice versa.

[0070] To find the most suitable analysis tool for sensor signals with different characteristics, in a preferred embodiment, the selection of wavelet basis functions for signals in each dimension includes:

[0071] For each dimension signal in the multidimensional time-series sensor data, each wavelet basis function is selected sequentially from the preset wavelet basis function library to perform multi-level wavelet packet decomposition on the signal, and the Shannon information entropy of the energy distribution of each frequency band after decomposition is calculated. The wavelet basis function that minimizes the Shannon information entropy value is selected as the optimal wavelet basis function for the dimension signal.

[0072] Shannon entropy is a metric for measuring signal uncertainty. In wavelet analysis, a lower entropy value indicates that the signal energy is more concentrated and ordered in the decomposed frequency bands, suggesting that the selected wavelet basis function can more effectively capture the main features of the signal. Therefore, by minimizing the entropy, the wavelet basis function with the strongest expressive power can be selected for each signal dimension. For example, suppose we are processing a vibration sensor signal from the spindle of an electrical device. The preset wavelet basis function library includes functions such as db5, sym4, and coif3. First, db5 is used to perform a three-level wavelet packet decomposition of the signal, obtaining the energy of eight frequency bands, and its entropy is calculated to be 2.5. Next, the same operation is performed using sym4, yielding an entropy of 1.9. Finally, coif3 is used to calculate an entropy of 2.8. Comparing these three values, 1.9 is the smallest, thus sym4 is determined to be the optimal wavelet basis function for this vibration signal. Subsequent feature extraction will be based on the decomposition results of sym4.

[0073] In a preferred embodiment, the extraction of time-frequency domain features includes:

[0074] The energy values ​​of each frequency band obtained after decomposition by the optimal wavelet basis function are used as the time-frequency domain features of the dimensional signal.

[0075] To transform complex time-series signals into relatively concise feature vectors, in this embodiment, after determining the optimal wavelet basis function for a sensor signal, such as a temperature signal, using the aforementioned method, multi-level wavelet packet decomposition is performed on the signal using the optimal wavelet basis function. The decomposition process separates the energy of the signal across different frequency ranges, forming a series of frequency bands. Assuming a three-level decomposition is performed on the temperature signal using its optimal wavelet basis function, eight sub-bands are generated. Then, the energy value of the signal within each sub-band is calculated, resulting in a vector composed of eight values, such as [10.5, 3.2, 1.1, 0.5, 8.7, 15.6, 4.3, 0.9]. This eight-dimensional energy vector constitutes the time-frequency domain feature of the temperature sensor within that time period. It not only reflects the overall energy intensity of the signal but also reveals the distribution of energy at different frequencies, which is crucial for identifying early, subtle fault signs in electrical equipment.

[0076] In a preferred embodiment, constructing a dynamic sensor relationship graph with sensors as nodes, Granger causality indices between sensors as weights, and the weight values ​​being updated periodically includes:

[0077] Using a sliding time window of length N and a step size of M sampling points, the Granger causality index is calculated by establishing a vector autoregression model for the time series data of any two sensors within the sliding time window.

[0078] The Granger causality indices of all sensor pairs are used to construct an adjacency matrix, which serves as the weight of the dynamic sensor relationship graph under the time slice.

[0079] The strength and direction of the correlation between sensors also change dynamically. By using a sliding window, the relationship graph can be updated periodically. The Granger causality index can be used to determine whether one time series can effectively predict another time series; its value represents the strength of the predictive ability, i.e., the strength of the causal relationship. Setting a time window of 1000 sampling points with a step size of 100 sampling points, for a system containing vibration sensor A, pressure sensor B, and temperature sensor C, within the first time window, the Granger causality indices for all six directions are calculated. For example, the influence index of A on B is 0.8, the influence index of B on A is 0.2, the influence index of A on C is 0.1, the influence index of C on A is 0.4, the influence index of B on C is 0.7, and the influence index of C on B is 0.6. These values ​​form a 3x3 adjacency matrix. The element in the i-th row and j-th column of the matrix represents the causal relationship strength between sensor i and sensor j. After the time window slides by 100 sampling points, the new data is used to recalculate and obtain a new adjacency matrix, thus realizing the update of the sensor relationship graph.

[0080] To ensure that the features of each sensor not only reflect its own state but also incorporate information from its associated sensors, in a preferred embodiment, the process of using a graph convolutional network to calculate the dynamic sensor relationship graph for each time slice and learning the node state features that fuse spatial topology includes:

[0081] For any node in the graph, the time-frequency domain features of all its neighboring nodes are weighted and aggregated according to the Granger causality index to obtain aggregated neighborhood features.

[0082] The aggregated neighborhood features are concatenated with the node's own time-frequency domain features, and then nonlinearly transformed through a fully connected layer with a ReLU activation function to obtain node state features that incorporate spatial topological information.

[0083] Graph convolutional networks simulate the propagation and fusion of information in a sensor network graph. Granger causality indices are used as weights to ensure that neighboring nodes with greater influence have a higher weight in feature aggregation. Taking pressure sensor B as an example, assuming it is connected to vibration sensor A and temperature sensor C in the graph, the dynamic relationship graph shows that the causality index of vibration sensor A to pressure sensor B is 0.9, and the causality index of temperature sensor C to pressure sensor B is 0.3. First, the time-frequency domain feature vector of vibration sensor A is multiplied by 0.9, and the time-frequency domain feature vector of temperature sensor C is multiplied by 0.3. Then, these two weighted vectors are added together to obtain the aggregated neighborhood feature. This aggregated feature is concatenated with the original time-frequency domain feature vector of pressure sensor B to form a longer feature vector. The concatenated long feature vector is input into a fully connected neural network layer, and after nonlinear transformation, it is output to generate the final node state feature of pressure sensor B. This final node state feature contains not only detailed information about B itself but also incorporates the dynamic influences of vibration sensor A and temperature sensor C on it.

[0084] In a preferred embodiment, the step of performing Viterbi decoding on the electrical equipment state sequence encoding using a preset state transition cost matrix that utilizes physical constraints for switching between different operating states includes:

[0085] A K×K state transition cost matrix is ​​pre-defined, where K is the total number of operating states of electrical equipment, and the element C(i,j) of the state transition cost matrix represents the cost of transitioning from state i to state j. Transitions that are physically impossible have an infinite cost.

[0086] In the recursive process of Viterbi decoding, the state transition cost matrix element C(i,j) is added as a penalty term to the logarithmic calculation of the state transition probability to constrain the search for the optimal electrical equipment operating state path.

[0087] To avoid generating illogical state sequence predictions, the Viterbi algorithm primarily searches for the most probable state sequence based on observed data, but it doesn't consider the physical reality of state transitions. The state transition cost matrix is ​​designed to compensate for this deficiency. For example, an electrical device has four operating states: normal, minor wear, severe fault, and shutdown. A 4x4 cost matrix is ​​defined. A direct transition from the severe fault state to the normal state is impossible; repair or a restart is necessary. Therefore, the cost C(i,j) for this transition can be set to infinity. Conversely, a transition from the normal state to the minor wear state is entirely possible, and its cost can be set to a small value, such as 0.1. In each step of the Viterbi algorithm's recursive calculation, when evaluating the quality of a path from state i in the previous time step to state j in the current time step, in addition to considering the transition probabilities given by the model, this preset cost C(i,j) is added. An infinite cost will result in an extremely low total score for this path, naturally leading to its exclusion in the optimal path search, ensuring that the final output state sequence conforms to the logic of engineering implementation.

[0088] In a preferred embodiment, the step of obtaining the probability prediction distribution of the electrical equipment in different operating states at each time moment by performing Monte Carlo dropout and multiple forward propagations in the graph convolutional network and gated recurrent unit, calculating the variance of the probability prediction distribution, and quantifying the variance as the confidence level of the operating state analysis result includes:

[0089] During the inference phase, the dropout layer in the graph convolutional network and gated recurrent unit is activated, and the same input data is forwarded multiple times to obtain multiple sets of probabilistic prediction results of the operating status of electrical equipment.

[0090] Based on the multiple sets of probability prediction results, the probability distribution of the electrical equipment in different operating states at each time is calculated, and the variance of the probability distribution is calculated. The variance is used as the confidence level of the operating state analysis results.

[0091] To provide a measure of uncertainty in the model's predictions, allowing users to understand the model's confidence in its judgments, Monte Carlo dropout can be used. This involves randomly and temporarily removing some neurons from the neural network during prediction, causing the model to produce several slightly different outputs for the same input. The differences in the outputs reflect the model's degree of uncertainty about the current input data. Figure 8As shown, for example, given sensor data at a certain moment, if the model makes 50 predictions, and the model is highly confident in its judgment, these 50 predictions will be highly consistent, for example, predicting a normal state 48 times with a probability of around 0.95. Calculating the variance of these probability values ​​will yield a very small value, such as 0.001, representing high confidence. Conversely, if the model is confused by the input data, the 50 predictions may be very scattered, for example, predicting normal 20 times, minor wear 20 times, and serious malfunction 10 times, with each probability value fluctuating greatly. Calculating the variance of these probabilities will yield a large value, such as 0.2, representing low confidence, indicating to the user that the analysis results at that moment may be unreliable and require manual intervention or further investigation.

[0092] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention. In addition, various different implementations of the embodiments of the present invention can be arbitrarily combined, as long as they do not violate the spirit of the embodiments of the present invention, and they should also be regarded as the content disclosed in the embodiments of the present invention.

Claims

1. A method for real-time statistical analysis of the operating status of electrical equipment, characterized in that, Includes the following steps: Acquire multidimensional time-series sensor data generated during the operation of electrical equipment; based on the Shannon information entropy of each dimension signal in the multidimensional time-series sensor data, select wavelet basis functions for each dimension signal, and use the wavelet basis functions to perform wavelet packet decomposition on the data to extract time-frequency domain features; A dynamic sensor relationship graph is constructed with sensors as nodes and Granger causality index between sensors as weights, with the weight values ​​updated periodically. The time-frequency domain features are used as node attributes, and a graph convolutional network is used to calculate the dynamic sensor relationship graph under each time slice to learn and fuse the node state features of the spatial topology. The node state features are then input into the gated recurrent unit in time sequence to obtain the electrical equipment state sequence encoding that fuses spatiotemporal dependence. Using a preset state transition cost matrix of physical constraints for switching between different operating states, the state sequence encoding of the electrical equipment is Viterbi decoded to obtain the optimal operating state path of the electrical equipment. In the inference phase, Monte Carlo dropout is performed in the graph convolutional network and gated recurrent unit, and multiple forward propagations are performed to obtain the probability prediction distribution of the electrical equipment in different operating states at each time. The variance of the probability prediction distribution is calculated and used as the confidence level of the operating state analysis result.

2. The method according to claim 1, characterized in that, The selection of wavelet basis functions for signals in each dimension includes: For each dimension signal in the multidimensional time-series sensor data, each wavelet basis function is selected sequentially from the preset wavelet basis function library to perform multi-level wavelet packet decomposition on the signal, and the Shannon information entropy of the energy distribution of each frequency band after decomposition is calculated. The wavelet basis function that minimizes the Shannon information entropy value is selected as the optimal wavelet basis function for the dimension signal.

3. The method according to claim 2, characterized in that, The extraction of time-frequency domain features includes: The energy values ​​of each frequency band obtained after decomposition by the optimal wavelet basis function are used as the time-frequency domain features of the dimensional signal.

4. The method according to claim 1, characterized in that, The construction of a dynamic sensor relationship graph, with sensors as nodes and Granger causality indices between sensors as weights and the weight values ​​updated periodically, includes: Using a sliding time window of length N and a step size of M sampling points, the Granger causality index is calculated by establishing a vector autoregression model for the time series data of any two sensors within the sliding time window. The Granger causality indices of all sensor pairs are used to construct an adjacency matrix, which serves as the weight of the dynamic sensor relationship graph under the time slice.

5. The method according to claim 1, characterized in that, The process of using a graph convolutional network to calculate the dynamic sensor relationship graph for each time slice and learning the node state features that fuse spatial topology includes: For any node in the graph, the time-frequency domain features of all its neighboring nodes are weighted and aggregated according to the Granger causality index to obtain aggregated neighborhood features. The aggregated neighborhood features are concatenated with the node's own time-frequency domain features, and then nonlinearly transformed through a fully connected layer with a ReLU activation function to obtain node state features that incorporate spatial topological information.

6. The method according to claim 1, characterized in that, The step of performing Viterbi decoding on the state sequence encoding of the electrical equipment using a preset state transition cost matrix that controls the switching of different operating states includes: A K×K state transition cost matrix is ​​predefined, where K is the total number of operating states of electrical equipment, and the element C(i,j) of the state transition cost matrix represents the cost of transitioning from state i to state j. The cost of a physically impossible transition is infinite. In the recursive process of Viterbi decoding, the state transition cost matrix element C(i,j) is added as a penalty term to the logarithmic calculation of the state transition probability to constrain the search for the optimal electrical equipment operating state path.

7. The method according to claim 1, characterized in that, The method involves performing Monte Carlo dropout and multiple forward propagations within the graph convolutional network and gated recurrent unit to obtain the probability prediction distribution of the electrical equipment in different operating states at each time point, and calculating the variance of the probability prediction distribution. This variance is then used as the confidence level of the operating state analysis results. The method includes: During the inference phase, the dropout layer in the graph convolutional network and gated recurrent unit is activated, and the same input data is forwarded multiple times to obtain multiple sets of probabilistic prediction results of the operating status of electrical equipment. Based on the probability prediction results of the multiple sets of electrical equipment operating states, the probability distribution of the electrical equipment being in different operating states at each time is calculated, and the variance of the probability distribution is calculated. The variance is used as the confidence level of the operating state analysis results.

8. A real-time statistical analysis system for the operating status of electrical equipment, characterized in that, Includes the following units: The feature extraction unit is used to acquire multi-dimensional time-series sensor data generated during the operation of electrical equipment; based on the Shannon information entropy of each dimension signal in the multi-dimensional time-series sensor data, a wavelet basis function is selected for each dimension signal, and the wavelet basis function is used to perform wavelet packet decomposition on the data to extract time-frequency domain features; The encoding unit is used to construct a dynamic sensor relationship graph with sensors as nodes and Granger causality index between sensors as weights, with the weight values ​​being updated periodically. The time-frequency domain features are used as node attributes, and a graph convolutional network is used to calculate the dynamic sensor relationship graph under each time slice to learn and fuse the node state features of the spatial topology. The node state features are then input into the gated recurrent unit in time sequence to obtain the electrical equipment state sequence encoding that fuses the spatiotemporal dependence. The statistical analysis unit is used to perform Viterbi decoding on the electrical equipment state sequence encoding using a preset state transition cost matrix of physical constraints between different operating states, and to parse out the optimal electrical equipment operating state path; in the inference stage, by performing Monte Carlo dropout and multiple forward propagations in the graph convolutional network and gated recurrent unit, the probability prediction distribution of the electrical equipment in different operating states at each time is obtained, and the variance of the probability prediction distribution is calculated, and the variance is used as the confidence level of the operating state analysis result.

9. The system according to claim 8, characterized in that, The selection of wavelet basis functions for signals in each dimension includes: For each dimension signal in the multidimensional time-series sensor data, each wavelet basis function is selected sequentially from the preset wavelet basis function library to perform multi-level wavelet packet decomposition on the signal, and the Shannon information entropy of the energy distribution of each frequency band after decomposition is calculated. The wavelet basis function that minimizes the Shannon information entropy value is selected as the optimal wavelet basis function for the dimension signal.

10. The system according to claim 9, characterized in that, The extraction of time-frequency domain features includes: The energy values ​​of each frequency band obtained after decomposition by the optimal wavelet basis function are used as the time-frequency domain features of the dimensional signal.

Images