A method and system for real-time statistical analysis of the operating state of an electrical device

Abstract

The application provides a kind of electrical equipment operating state real-time statistical analysis method and system, belongs to electrical equipment operating state monitoring field, including obtaining the multidimensional time sequence sensor data of electrical equipment operation, based on Shannon information entropy selects appropriate wavelet base function to carry out wavelet packet decomposition, extracts time-frequency domain feature;Sensor is constructed as node, and the dynamic relationship diagram with granger causality is as weight, and the node spatial feature is learned by graph convolution network combined with time-frequency feature;The node feature is input into the gated recurrent unit, and the electrical equipment state sequence coding that fuses space-time dependence is generated;The state transition cost matrix is used for viterbi decoding, and the optimal electrical equipment operating state path is inferred.The application can quantify the uncertainty of prediction result, provide clear confidence evaluation for the final analysis conclusion, and improve the reliability of decision.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of electrical equipment operation state monitoring, and particularly relates to an electrical equipment operation state real-time statistical analysis method and system. BACKGROUND

[0002] In the fields of industrial production, energy, etc., real-time and accurate monitoring and analysis of the operation state of large and complex electrical equipment is a key link for ensuring safe production, improving operation efficiency and implementing predictive maintenance.

[0003] Traditional electrical equipment state analysis methods mainly rely on signal processing techniques or classic machine learning models. For example, methods based on Fourier transform or short-time Fourier transform can analyze the frequency domain characteristics of signals, but cannot simultaneously analyze time and frequency; wavelet analysis has improved in time-frequency analysis, but the selection of its basis functions usually depends on expert experience and lacks adaptive ability to different signal characteristics, making it difficult to optimally extract features from complex and variable electrical equipment signals. In addition, some methods that apply recurrent neural networks (RNN), long short-term memory networks (LSTM) and other deep learning models can effectively capture the dependence of data in the time dimension, but usually treat multi-channel sensor signals as independent channels for processing, ignoring the complex topological relationships between different sensors in physical space and functional association, resulting in a failure to fully utilize the spatial correlation information between components of electrical equipment, limiting the depth and accuracy of the model's understanding of the overall state of the electrical equipment.

[0004] To solve the above problems, some research has begun to try to combine graph neural networks (GNN) with time series models to capture the spatio-temporal dependence of electrical equipment state data. However, existing technologies still have several shortcomings: when constructing a sensor relationship graph, a static graph structure based on physical proximity or prior knowledge is usually used, which cannot reflect the dynamic causal relationships between sensors under different operating conditions of the electrical equipment, reducing the model's representation ability. In the analysis of state sequences, many methods only output instantaneous states through simple classifiers, without considering the physical logic and time constraints of state switching, which can easily produce frequent and unreasonable jump results that do not conform to actual operating rules. In addition, existing models often only provide a deterministic prediction point when giving state analysis results, and cannot quantify the reliability or uncertainty of the prediction results, which is a major defect in industrial scenarios that require high reliability decisions. When encountering unseen conditions or being disturbed by noise, a reliable confidence assessment is crucial for the decision-making and risk management of operating personnel. SUMMARY

[0005] To solve the above problems, the present application proposes an electrical equipment operation state real-time statistical analysis method, comprising the following steps:

[0006] acquiring multi-dimensional time-series sensor data generated by the electrical equipment in operation; selecting a wavelet basis function for each dimension signal based on a Shannon information entropy of each dimension signal in the multi-dimensional time-series sensor data, and performing wavelet packet decomposition on the data using the wavelet basis function to extract time-frequency domain features;

[0007] constructing a dynamic sensor relationship graph taking sensors as nodes, taking a Granger causality index between sensors as weights, and periodically updating the weight values; taking the time-frequency domain features as node attributes, calculating the dynamic sensor relationship graph at each time slice using a graph convolution network, learning node state features fused with spatial topological structures, and then inputting the node state features in sequence to a gated recurrent unit to obtain an electrical equipment state sequence encoding fused with space-time dependencies;

[0008] performing Viterbi decoding on the electrical equipment state sequence encoding using a state transition cost matrix of preset physical constraints between different operating states to obtain an optimal electrical equipment operating state path; in the inference stage, performing Monte Carlo dropout and multiple forward propagations in the graph convolution network and the gated recurrent unit to obtain a probability prediction distribution of the electrical equipment being in different operating states at each time, and calculating a variance of the probability prediction distribution as a confidence of the operating state analysis result.

[0009] Preferably, the wavelet basis function selection for each dimension signal comprises:

[0010] for each dimension signal in the multi-dimensional time-series sensor data, selecting each wavelet basis function from a preset wavelet basis function library to perform multi-layer wavelet packet decomposition on the signal, and calculating a Shannon information entropy of the energy distribution of each frequency band after decomposition to select a wavelet basis function with the minimum Shannon information entropy as the optimal wavelet basis function for the dimension signal.

[0011] Preferably, the time-frequency domain feature extraction comprises:

[0012] using the energy values of each frequency band obtained after decomposition by the optimal wavelet basis function as the time-frequency domain features of the dimension signal.

[0013] Preferably, the dynamic sensor relationship graph taking sensors as nodes and taking a Granger causality index between sensors as weights with periodically updated weight values comprises:

[0014] using a sliding time window with a length of N and a step of M sampling points to calculate the Granger causality index between the time series data of any two sensors in the sliding time window by establishing a vector autoregressive model;

[0015] The Granger causality indexes of all sensor pairs are constituted into an adjacency matrix as weights of the dynamic sensor relation graph under the time slice.

[0016] Preferably, the dynamic sensor relation graph under each time slice is calculated by using the graph convolution network to learn the node state features fused with spatial topology structure, including:

[0017] For any node in the graph, the time-frequency domain features of all adjacent nodes of the node are aggregated by weighting according to the Granger causality index to obtain aggregated neighborhood features;

[0018] The aggregated neighborhood features and the time-frequency domain features of the node itself are spliced, and a fully connected layer with a ReLU activation function is used for nonlinear transformation to obtain node state features fused with spatial topology information.

[0019] Preferably, the state sequence coding of the electrical equipment is decoded by using a state transition cost matrix of preset different operating states and switching physical constraints, including:

[0020] A K×K state transition cost matrix is preset, where K is the total number of operating states of the electrical equipment, and an element C(i,j) of the state transition cost matrix represents a cost of transition from state i to state j, where a physically impossible transition has an infinite cost;

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

[0022] Preferably, the probability prediction distribution of the electrical equipment in different operating states at each time is obtained by performing Monte Carlo dropout and multiple forward propagations in the graph convolution network and the gated recurrent unit, and the variance of the probability prediction distribution is calculated to quantify the confidence of the operating state analysis result, including:

[0023] In the inference phase, the dropout layer in the graph convolution network and the gated recurrent unit is activated, and multiple forward propagations are performed on the same input data to obtain multiple sets of probability prediction results of the electrical equipment operating state;

[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 as the confidence of the operating state analysis result.

[0025] In addition, the present application also proposes an electrical equipment operating state real-time statistical analysis system, including the following units:

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

[0027] an encoding unit configured to construct a dynamic sensor relationship graph with sensors as nodes and a Granger causality index between sensors as weights, periodically update the weight values, and calculate the dynamic sensor relationship graph at each time slice using a graph convolution network, learn node state features fused with spatial topological structures, and input the node state features in sequence to a gated recurrent unit to obtain an electrical equipment state sequence encoding fused with spatio-temporal dependencies;

[0028] a statistical analysis unit configured to perform Viterbi decoding on the electrical equipment state sequence encoding using a state transition cost matrix of preset physical constraints between different operating states, analyze to obtain an optimal electrical equipment operating state path, perform Monte Carlo dropout and multiple forward propagations in the graph convolution network and the gated recurrent unit in an inference stage to obtain a probability prediction distribution of the electrical equipment being in different operating states at each time, and calculate a variance of the probability prediction distribution as a confidence level of an operating state analysis result.

[0029] Preferably, the wavelet basis function selection for each dimension signal comprises:

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

[0031] Preferably, the time-frequency domain feature extraction comprises:

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

[0033] Preferably, the dynamic sensor relationship graph with sensors as nodes and the Granger causality index between sensors as weights, and the weight values being periodically updated comprises:

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

[0035] The Granger causality indexes of all sensor pairs form an adjacency matrix as the weight of the dynamic sensor relation graph under the time slice.

[0036] Preferably, the dynamic sensor relation graph under each time slice is calculated by using the graph convolution network to learn the node state features fused with spatial topology structure, including:

[0037] For any node in the graph, the time-frequency domain features of all adjacent nodes of the node are aggregated by weighting according to the Granger causality index to obtain aggregated neighborhood features;

[0038] The aggregated neighborhood features and the time-frequency domain features of the node itself are spliced, and a fully connected layer with a ReLU activation function is used for nonlinear transformation to obtain node state features fused with spatial topology information.

[0039] Preferably, the state sequence coding of the electrical equipment is decoded by using a state transition cost matrix of preset different operating states and switching physical constraints, including:

[0040] A K×K state transition cost matrix is preset, where K is the total number of electrical equipment operating states, and the state transition cost matrix element C(i,j) represents the cost of transition from state i to state j, where the transition that is physically impossible has an infinite cost;

[0041] In the recursive process of the 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 of the optimal electrical equipment operating state path.

[0042] Preferably, the probability prediction distribution of the electrical equipment in different operating states at each time is obtained by performing Monte Carlo dropout and multiple forward propagations in the graph convolution network and the gated recurrent unit, and the variance of the probability prediction distribution is calculated, and the variance is taken as the confidence of the operating state analysis result, including:

[0043] In the inference phase, the dropout layer in the graph convolution network and the gated recurrent unit is activated, and multiple forward propagations are performed on the same input data to obtain multiple sets of probability prediction results of the electrical equipment operating state;

[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, and the variance is taken as the confidence of the operating state analysis result.

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

[0046] 1. According to the information entropy characteristics of the signal itself, the optimal wavelet basis function is selected, so that the key time-frequency domain features reflecting the state of the electrical equipment are more accurately extracted.

[0047] 2. The sensor correlation graph is constructed by using the Granger causality, and the weight is periodically updated, so that the internal relationship between the components of the electrical equipment changing with the working condition can be captured, and the limitation of the fixed physical topology is overcome.

[0048] 3. The complex dependence relationship of the electrical equipment state in the spatial and time dimensions is comprehensively learned by deeply fusing the graph convolution network and the gate recurrent unit.

[0049] 4. The state transition cost matrix with physical constraints is introduced for Viterbi decoding, so that the state sequence of the final output conforms to the physical logic of the electrical equipment operation, and the disorder jump of the result is effectively avoided.

[0050] The present application can quantitatively predict the uncertainty of the result, provide clear confidence evaluation for the final analysis conclusion, and greatly enhance the reliability and practical value of the method in actual industrial decision-making. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1 It is a flowchart of the embodiment;

[0052] Figure 2 It is a data schematic diagram of the temperature sensor, the pressure sensor and the vibration sensor;

[0053] Figure 3 It is a schematic diagram of wavelet basis selection;

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

[0055] Figure 5 It is a sensor relationship graph;

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

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

[0058] Figure 8 It is a schematic diagram of Monte Carlo discard confidence quantification. DETAILED DESCRIPTION

[0059] With reference to the drawings and embodiments of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0060] In specific embodiments, the present application provides a real-time statistical analysis method for the running state of electrical equipment, as shown in Figure 1 The method comprises the following steps:

[0061] S1, acquiring multi-dimensional time sequence sensor data generated during the running of electrical equipment; based on the Shannon information entropy of each dimension signal in the multi-dimensional time sequence sensor data, selecting a wavelet base function for each dimension signal, and using the wavelet base function to perform wavelet packet decomposition on the data to extract time-frequency domain features;

[0062] Through a data acquisition system, such as a distributed control system or a data acquisition card, data is synchronously collected from multiple sensors installed on the electrical equipment, such as temperature sensors, pressure sensors, and vibration sensors, to form a multi-dimensional time sequence data matrix with time as the row and sensor channels as the column, as shown in Figure 2 .

[0063] A wavelet base function library containing multiple commonly used wavelet bases, such as Daubechies series, Coiflets series, and Symlets series, is preset; for the time sequence signal of each sensor channel, the Shannon information entropy value is calculated; each wavelet base in the wavelet base function library is used to perform wavelet packet decomposition on the signal, and the information entropy of the energy distribution of each frequency band after decomposition is calculated, and the wavelet base with the minimum entropy value is selected as the optimal wavelet base for the signal, as shown in Figure 3 ; the selected optimal wavelet base is used to perform wavelet packet decomposition on the signal, as shown in Figure 4 , and the energy and kurtosis of each frequency band are extracted as the time-frequency domain features of the signal under the current time window.

[0064] S2, constructing a dynamic sensor relationship graph with sensors as nodes and the Granger causality index between sensors as weights, and periodically updating the weight values; using the time-frequency domain features as node attributes, using a graph convolution network to calculate the dynamic sensor relationship graph under each time slice, learning the node state features fused with the spatial topology structure, and then inputting the node state features in sequence to a gated recurrent unit to obtain an electrical equipment state sequence encoding fused with spatial and temporal dependencies;

[0065] Each sensor is regarded as a node in the graph; in each preset time window, the F statistic or p value of Granger causality test between any two sensor nodes is calculated by using vector autoregressive model, and the statistic is taken as the weight of the edge connecting the two nodes; by sliding the time window, the Granger causality index between all node pairs is repeatedly calculated periodically, so as to generate a series of time-varying weighted adjacency matrices to represent the dynamic sensor relationship graph, as shown in Figure 5 .

[0066] For each time slice, the weighted adjacency matrix of the dynamic sensor relationship graph at the moment is input into a multi-layer graph convolution network model together with the corresponding time-frequency domain feature matrix, neighbor node information is aggregated by graph convolution operation to generate a node high-level feature representation containing spatial correlation between sensors; then, the node high-level feature representation sequence generated in sequence by all time slices is input into a gated recurrent unit network, and the long-term dependence of the feature sequence in the time dimension is captured by using the update gate and reset gate inside the network, and the hidden state sequence output by the network is the electrical equipment state sequence encoding fused with space-time dependence, as shown in Figure 6 .

[0067] S3, using a state transition cost matrix of preset different running states, the electrical equipment state sequence encoding is subjected to Vitbebi decoding, and the optimal electrical equipment running state path is obtained; in the inference stage, the probability prediction distribution of the electrical equipment in different running states at each moment is obtained by performing Monte Carlo dropout and multiple forward propagation in the graph convolution network and the gated recurrent unit, and the variance of the probability prediction distribution is calculated, and the variance is taken as the confidence of the running state analysis result.

[0068] The electrical equipment state sequence encoding output by the gated recurrent unit is passed through a fully connected layer and a Softmax activation function to obtain the emission probability of the electrical equipment in each predefined running state at each moment; at the same time, according to expert knowledge or historical data statistics, a state transition cost matrix is constructed, as shown in Figure 7 , wherein the state transition cost matrix element defines the cost of transition from one running state to another, for example, the cost of transition that does not conform to the physical law, such as directly from normal to serious fault, is set to a maximum value; the emission probability sequence and the state transition cost matrix are taken as the input of the Vitbebi algorithm, and a state sequence with the maximum joint probability is obtained by dynamic programming, which is the optimal electrical equipment running state path.

[0069] During model inference, the dropout layers in the graph convolution network and the gated recurrent unit are kept active; for the same segment of input data, the complete forward propagation calculation is repeatedly performed T times, and since the randomly inactivated neurons are different in each propagation, T slightly different state prediction probability sequences are obtained; for each time, the element-wise variance of the T state prediction probability sequences is calculated, and the size of the variance directly reflects the degree of uncertainty of the model's prediction at that time; the smaller the variance, the higher the confidence, and vice versa.

[0070] In order to find the most suitable analysis tool for sensor signals with different characteristics, in a preferred embodiment, the wavelet basis function is selected for each dimensional signal, including:

[0071] For each dimensional signal in the multi-dimensional time series sensor data, each wavelet basis function is sequentially selected from the preset wavelet basis function library to perform multi-layer 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 with the minimum Shannon information entropy value is selected as the optimal wavelet basis function for the dimensional signal.

[0072] Shannon information entropy is an indicator of signal uncertainty. In wavelet analysis, a lower information entropy value means that the energy of the signal is more concentrated and ordered in the frequency bands after decomposition, which indicates that the selected wavelet basis function can more effectively capture the main characteristics of the signal. Therefore, by minimizing the information entropy, the strongest wavelet basis function can be selected for each signal dimension. For example, assume that a vibration sensor signal from the main shaft of an electrical equipment is processed, and the preset wavelet basis function library contains db5, sym4 and coif3 functions. First, use db5 to perform three-layer wavelet packet decomposition on the signal to obtain the energy of eight frequency bands, and calculate the information entropy as 2.5. Next, use sym4 to perform the same operation to calculate the information entropy as 1.9. Finally, use coif3 to calculate the information entropy as 2.8. Comparing these three values, 1.9 is the smallest, and sym4 is finally determined as the optimal wavelet basis function for the vibration signal, and subsequent feature extraction will be based on the decomposition results of sym4.

[0073] In a preferred embodiment, the extracted time-frequency domain features include:

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

[0075] To convert the complex time series into a relatively simple feature vector, in this embodiment, after the optimal wavelet basis function is determined for a certain sensor signal, such as temperature signal, by the aforementioned method, the signal is decomposed into multiple layers using the optimal wavelet basis function. The decomposition process separates the energy of the signal in different frequency ranges and forms a series of frequency bands. Assuming that the temperature signal is decomposed into three layers using its optimal wavelet basis function, eight sub-bands are generated, and then the energy values of the signal in each sub-band are calculated to obtain an eight-dimensional vector, 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 in this time period. It not only reflects the overall energy intensity of the signal, but also reveals the energy distribution at different frequencies, which is crucial for identifying early and weak fault signs of electrical equipment.

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

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

[0078] The Granger causality index of all sensor pairs forms an adjacency matrix, which is used as the weight of the dynamic sensor relationship graph at the time slice.

[0079] The correlation strength and direction between sensors also dynamically change, and by using a sliding window, the relationship graph can be periodically updated. The Granger causality index can be used to determine whether one time series can effectively predict another time series, and the numerical value represents the strength of the prediction ability, i.e., the strength of the causal relationship. Assuming a time window with a length of 1000 sampling points and a step of 100 sampling points, for a system containing vibration sensor A, pressure sensor B, and temperature sensor C, in the first time window, the Granger causality index in all six directions is calculated, such as 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, and the element value in the ith row and jth column of the matrix is the causal relationship strength of sensor i on sensor j. When the time window slides by 100 sampling points, the data is calculated again to obtain a new adjacency matrix, thereby updating the sensor relationship graph.

[0080] In order to let the characteristics of each sensor not only reflect its own state, but also integrate the information of its associated sensors, in a preferred embodiment, the dynamic sensor relation graph at each time slice is calculated using a graph convolution network to learn the node state features that integrate the spatial topology, including:

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

[0082] The aggregated neighborhood features are concatenated with the time-frequency domain features of the node itself, and a fully connected layer with a ReLU activation function is used for nonlinear transformation to obtain the node state features that integrate the spatial topology information.

[0083] The graph convolution network simulates the propagation and integration process of information on the sensor network graph. The Granger causality index is used as the weight to ensure that the adjacent nodes with greater influence have a higher proportion 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, it is known from the dynamic relation graph that the causality index of vibration sensor A on pressure sensor B is 0.9, and the causality index of temperature sensor C on pressure sensor B is 0.3. First, multiply the time-frequency domain feature vector of vibration sensor A by 0.9, and multiply the time-frequency domain feature vector of temperature sensor C by 0.3, then add the two weighted vectors to obtain the aggregated neighborhood features. The aggregated features are concatenated with the original time-frequency domain feature vector of pressure sensor B to form a longer feature vector. The long feature vector after concatenation is input into a fully connected neural network layer, and after nonlinear transformation, the final node state feature of pressure sensor B is generated, which contains not only the detailed information of B itself, but also the dynamic influence of vibration sensor A and temperature sensor C on it.

[0084] In a preferred embodiment, the state transition cost matrix with preset different running state switching physical constraints is used to perform Viterbi decoding on the electrical equipment state sequence encoding, including:

[0085] A KxK state transition cost matrix is preset, where K is the total number of electrical equipment running states, and the state transition cost matrix element C(i,j) represents the cost of transitioning from state i to state j, where physically impossible transitions 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 state transition probability to constrain the search of the optimal electrical equipment running state path.

[0087] To avoid generating an unreasonable state sequence prediction, the Viterbi algorithm itself mainly looks for the most likely state sequence based on observation data, but it does not consider the physical reality of state transition. The state transition cost matrix is used to make up for this deficiency. For example, an electrical device has four operating states, namely normal, slight wear, serious failure and shutdown. A four-by-four cost matrix is defined. It is impossible to directly jump from a serious failure state to a normal state, and it must go through maintenance or shutdown and restart. Therefore, the cost C(i,j) of the transition from serious failure to normal can be set to infinity. On the contrary, it is entirely possible to transition from a normal state to a slightly worn state, and the cost can be set to a small value, such as 0.1. In each step of the recursive calculation of the Viterbi algorithm, when evaluating the pros and cons of the path from the previous state i to the current state j, in addition to considering the transition probability given by the model, the preset cost C(i,j) is also added. An infinite cost will make the total score of this path very low, so it will be naturally excluded in the optimal path search, ensuring that the final output state sequence meets the logic of engineering implementation.

[0088] In a preferred embodiment, the probability prediction distribution of the electrical equipment in different operating states at each time is obtained by performing Monte Carlo dropout and multiple forward propagation in the graph convolution network and the gated recurrent unit, and the variance of the probability prediction distribution is calculated, and the variance is quantified as the confidence of the operating state analysis result, comprising:

[0089] In the inference phase, the dropout layer in the graph convolution network and the gated recurrent unit is activated, and multiple forward propagations are performed on the same input data to obtain multiple sets of probability prediction results of the electrical equipment operating state;

[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 as the confidence of the operating state analysis result.

[0091] In order to provide a measure of uncertainty for the prediction results of the model, so that the user knows how confident the model is in its judgment, the Monte Carlo dropout method can be used. By temporarily removing some neurons in the neural network at random during prediction, the model can produce multiple slightly different output results for the same input. The difference between the output results reflects the degree of uncertainty of the model for the current input data, such as Figure 8As shown, for example, for sensor data at a certain time, let the model make 50 predictions, if the model is very confident in its judgment, the 50 predictions will be highly consistent, for example, 48 times predict normal state, and the probability is about 0.95, calculate the variance of these probability values, a very small value, such as 0.001, which represents high confidence. Conversely, if the model is confused by the input data, the 50 predictions may be very scattered, for example, 20 times predict normal, 20 times predict slight wear, 10 times predict serious failure, and the respective probability values fluctuate greatly, calculating the variance of these probabilities will get a larger value, such as 0.2, which represents low confidence, prompting the user that the analysis result at this time may not be reliable, and manual intervention or further inspection is required.

[0092] The above examples are only used to illustrate the technical solutions of the present application, but not to limit it; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that the technical solutions recorded in the foregoing examples can still be modified, or some technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application. In addition, the various different embodiments of the embodiments of the present application can also be combined arbitrarily, as long as it does not deviate from the idea of the embodiments of the present application, it should also be considered as the disclosed content of the embodiments of the present application.

Claims

1. A method for real-time statistical analysis of the operating state of an electrical device, characterized in that, The method comprises the following steps: acquiring multi-dimensional time sequence sensor data generated during operation of an electrical device; selecting a wavelet basis function for each dimension signal based on the Shannon information entropy of each dimension signal in the multi-dimensional time sequence sensor data, and performing wavelet packet decomposition on the data using the wavelet basis function to extract time-frequency domain features; constructing a dynamic sensor relationship graph with sensors as nodes and the Granger causality index between sensors as weights, and periodically updating the weight values; using the time-frequency domain features as node attributes, calculating the dynamic sensor relationship graph at each time slice using a graph convolution network, learning node state features that fuse spatial topological structures, and then inputting the node state features in sequence into a gated recurrent unit to obtain an electrical device state sequence encoding that fuses spatial and temporal dependencies; performing Viterbi decoding on the electrical device state sequence encoding using a state transition cost matrix that is preset for switching between different operating states, and analyzing to obtain an optimal electrical device operating state path; in the reasoning phase, performing Monte Carlo dropout and multiple forward propagations in the graph convolution network and the gated recurrent unit to obtain a probability prediction distribution of the electrical device being in different operating states at each time, and calculating the variance of the probability prediction distribution as the confidence of the operating state analysis result; the calculation of the dynamic sensor relationship graph at each time slice using the graph convolution network to learn node state features that fuse spatial topological structures comprises: for any node in the graph, weighting and aggregating the time-frequency domain features of all adjacent nodes according to the Granger causality index to obtain aggregated neighborhood features; concatenating the aggregated neighborhood features with the time-frequency domain features of the node itself, and performing nonlinear transformation through a fully connected layer with a ReLU activation function to obtain node state features that fuse spatial topological information.

2. The method of claim 1, wherein, the selection of a wavelet basis function for each dimension signal comprises: for each dimension signal in the multi-dimensional time sequence sensor data, sequentially selecting each wavelet basis function from a preset wavelet basis function library to perform multi-layer wavelet packet decomposition on the each dimension signal, and calculating the Shannon information entropy of the energy distribution of each frequency band after decomposition to select a wavelet basis function with the minimum Shannon information entropy value as the optimal wavelet basis function for the dimension signal.

3. The method of claim 2, wherein, the extraction of time-frequency domain features comprises: using the energy values of each frequency band obtained after decomposition by the optimal wavelet basis function as the time-frequency domain features of the dimension signal.

4. The method of claim 1, wherein, the construction of a dynamic sensor relationship graph with sensors as nodes and the Granger causality index between sensors as weights, and periodically updating the weight values, comprises: using a sliding time window with a length of N and a step of M sampling points to calculate the Granger causality index between the time series data of any two sensors in the sliding time window by establishing a vector autoregressive model; constructing an adjacency matrix of the Granger causality index of all sensor pairs as the weight of the dynamic sensor relationship graph at the time slice.

5. The method of claim 1, wherein, The state transition cost matrix using preset different running state inter-switch physical constraints is used for Viterbi decoding of the electrical equipment state sequence coding, and the Viterbi decoding includes the following steps: A K*K state transition cost matrix is preset, wherein K is the total number of electrical equipment running states, and a state transition cost matrix element C(i,j) represents a cost of transition from a state i to a state j, wherein a physically impossible transition has an infinite cost; In the recursive process of Viterbi decoding, the state transition cost matrix element C(i,j) is added as a penalty term to the calculation of the logarithm of the state transition probability, so as to constrain the search of the optimal electrical equipment running state path.

6. The method of claim 1, wherein, The probability prediction distribution of the electrical equipment in different running states at each time is obtained by performing Monte Carlo dropout and multiple forward propagation in the graph convolution network and the gated recurrent unit, and the variance of the probability prediction distribution is calculated, and the variance is used as the confidence of the running state analysis result. In the inference stage, the dropout layer in the graph convolution network and the gated recurrent unit is activated, multiple forward propagations are performed on the same input data, and multiple sets of probability prediction results of the electrical equipment running state are obtained. Based on the multiple sets of probability prediction results of the electrical equipment running state, the probability distribution of the electrical equipment in different running states at each time is calculated, and the variance of the probability distribution is calculated, and the variance is used as the confidence of the running state analysis result.

7. An electric equipment operation state real-time statistical analysis system, characterized by comprising: an electric equipment operation state real-time statistical analysis device; and an electric equipment operation state real-time statistical analysis server. The unit includes the following units: The feature extraction unit is configured to obtain multi-dimensional time sequence sensor data generated during electrical equipment operation, select a wavelet basis function for each dimension signal based on the Shannon information entropy of each dimension signal in the multi-dimensional time sequence sensor data, and perform wavelet packet decomposition on the data using the wavelet basis function to extract time-frequency domain features; The encoding unit is configured to construct a dynamic sensor relationship graph taking sensors as nodes and a Granger causality index between sensors as weights, periodically update the weight values, and calculate the dynamic sensor relationship graph at each time slice using a graph convolution network, learn node state features fused with spatial topological structures, and then input the node state features in sequence to a gated recurrent unit to obtain electrical equipment state sequence coding fused with spatial and temporal dependencies; The statistical analysis unit is configured to use a state transition cost matrix with preset different running state inter-switch physical constraints to perform Viterbi decoding on the electrical equipment state sequence coding to obtain an optimal electrical equipment running state path, and in the inference stage, obtain a probability prediction distribution of the electrical equipment in different running states at each time by performing Monte Carlo dropout and multiple forward propagation in the graph convolution network and the gated recurrent unit, and calculate the variance of the probability prediction distribution, and use the variance as the confidence of the running state analysis result. The graph convolution network is used to calculate the dynamic sensor relationship graph at each time slice, and learn node state features fused with spatial topological structures, and the learning includes the following steps: For any node in the figure, time-frequency domain features of all adjacent nodes of the node are aggregated according to the Granger causality index to obtain aggregated neighborhood features; The aggregated neighborhood features are spliced with the time-frequency domain features of the node itself, and are nonlinearly transformed through a fully connected layer with a ReLU activation function to obtain node state features fused with spatial topology information.

8. The system of claim 7, wherein, The wavelet basis function is selected for each dimension signal, including: For each dimension signal in the multi-dimensional time series sensor data, each wavelet basis function is sequentially selected from a preset wavelet basis function library to perform multi-layer wavelet packet decomposition on each dimension signal, and the Shannon information entropy of the energy distribution of each frequency band after decomposition is calculated, and the wavelet basis function with the minimum Shannon information entropy value is selected as the optimal wavelet basis function of the dimension signal.

9. The system of claim 8, wherein, The time-frequency domain features are extracted, including: The energy value of each frequency band obtained after decomposition by the optimal wavelet basis function is used as the time-frequency domain feature of the dimension signal.

Images