Pole-mounted circuit breaker operation state monitoring method based on multi-source information fusion

Abstract

The present application relates to the field of control adjustment system, specifically to a pole-mounted circuit breaker operation state monitoring method based on multi-source information fusion, comprising obtaining multi-source heterogeneous sensing data; constructing a heterogeneous time sequence diagram according to the sampling rate and the communication delay, the nodes correspond to the original time sequence sequence and retain the original sampling resolution, and the edge weight is dynamically updated based on the physical coupling delay; using a dynamic graph convolution network to aggregate information on the heterogeneous time sequence diagram, and extracting a cross-modal asynchronous correlation feature vector; inputting the asynchronous correlation feature vector into a long short-term memory network to output a latent fault probability value and a deterioration trend sequence. The monitoring method overcomes the problem of missing high-frequency transient characteristics caused by forced synchronization interpolation, retains the original time sequence resolution of multi-source data, and the asynchronous correlation feature reflects the dynamic coupling relationship of mechanical and electrical states, thereby improving the latent fault recognition accuracy and the reliability of the deterioration trend prediction.

Description

Technical Field

[0001] This invention relates to the field of control and regulation systems, specifically to a method for monitoring the operating status of pole-mounted circuit breakers based on multi-source information fusion. Background Technology

[0002] Pole-mounted circuit breakers (PBBs) play a crucial role in control and protection within power distribution networks, and their operational status monitoring is a key aspect of ensuring power supply reliability. Current technologies typically deploy various sensors on the PBB itself, including vibration sensors to collect mechanical characteristic signals, current transformers to collect opening and closing coil current signals, and temperature and humidity sensors to collect environmental data. After data acquisition, conventional monitoring systems transmit these sensor data from different sources and with different sampling frequencies to a centralized processing system. Because the sampling frequency of mechanical vibration signals reaches the kilohertz level, the coil current signal is at the hundred hertz level, and the environmental temperature and humidity signal sampling frequency is less than one hertz, current technologies forcibly perform resampling during the data preprocessing stage to facilitate subsequent data alignment and matrix operations.

[0003] The specific implementation of this forced resampling involves downsampling high-frequency mechanical vibration data or upsampling low-frequency temperature and humidity data using the lowest sampling frequency among the multi-source sensors as a benchmark, so that all sensor data form a uniform sampling interval on the time axis. Under the unified time benchmark, the mechanical vibration characteristics, coil current characteristics, and environmental temperature and humidity characteristics at the same moment are vector-concatenated to form a feature vector of equal dimension. Subsequently, the concatenated feature vector is input into a classification model or neural network to perform state assessment and fault identification.

[0004] The aforementioned forced resampling and feature stitching methods suffer from a core technical problem: forced time alignment of multi-source data leads to the loss of high-frequency features. High-frequency mechanical vibration data contains key transient impact features at the moment of circuit breaker operation. Downsampling directly filters out these high-frequency components, and the data points generated by interpolation of low-frequency data do not reflect the true physical changes. Simple feature stitching severs the asynchronous coupling relationship between electromechanical and thermal multi-physics fields. The weak feature shifts caused by early latent faults in the circuit breaker cannot be extracted through cross-modal correlation, making latent fault identification difficult. Summary of the Invention

[0005] The purpose of this invention is to provide a method for monitoring the operating status of pole-mounted circuit breakers based on multi-source information fusion, which can effectively solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A method for monitoring the operational status of pole-mounted circuit breakers based on multi-source information fusion includes:

[0008] Acquire multi-source heterogeneous sensing data of the pole-mounted circuit breaker, including mechanical vibration data, coil current data, and ambient temperature and humidity data;

[0009] A heterogeneous time series graph is constructed based on the sampling rate and communication delay of each sensor in the multi-source heterogeneous sensing data. The nodes in the heterogeneous time series graph correspond to the original time series sequences of each sensor. The attributes of the nodes retain the original sampling resolution. The edge weights between nodes in the heterogeneous time series graph are dynamically updated based on the physical coupling delay between each information source.

[0010] A dynamic graph convolutional network is used to perform information aggregation operations on the heterogeneous time series graph to extract asynchronous correlation feature vectors across modalities;

[0011] The asynchronous correlation feature vector is input into a long short-term memory network, which outputs the latent fault probability value and deterioration trend sequence of the pole-mounted circuit breaker.

[0012] Preferably, the step of dynamically updating the edge weights between nodes in the heterogeneous time series graph based on the physical coupling delay between each information source includes:

[0013] The electromagnetic force transmission delay between the node corresponding to the mechanical vibration data and the node corresponding to the coil current data is calculated as the weight of the first side.

[0014] The thermal expansion and contraction response delay between the node corresponding to the environmental temperature and humidity data and the node corresponding to the mechanical vibration data is calculated as the weight of the second side.

[0015] An adjacency matrix is ​​constructed based on the first edge weight and the second edge weight. The off-diagonal elements in the adjacency matrix are attenuated or enhanced according to the real-time operating conditions of the pole-mounted circuit breaker, and the diagonal elements are the retention coefficients of the original sampling resolution.

[0016] Preferably, the step of preserving the original sampling resolution of the node attributes in the heterogeneous timing graph corresponding to the original timing sequence of each sensor includes:

[0017] The mechanical vibration data, the coil current data, and the ambient temperature and humidity data are continuously extracted without interpolation according to their respective inherent sampling frequencies to form multi-granularity time slices;

[0018] The multi-granularity time-series slices are mapped to the initial feature matrices of the corresponding nodes in the heterogeneous time-series graph. The number of rows in the initial feature matrix is ​​positively correlated with the time length of the multi-granularity time-series slices, and the number of columns is consistent with the measurement dimension of the sensor.

[0019] The dimensionality heterogeneity of the initial feature matrix is ​​maintained in the topology of the heterogeneous time series graph, and dimensionality reduction alignment to the lowest sampling resolution is rejected.

[0020] Preferably, the step of using a dynamic graph convolutional network to perform information aggregation operations on the heterogeneous time series graph and extract cross-modal asynchronous correlation feature vectors includes:

[0021] For the target node in the heterogeneous time series graph, the feature projection dimension is determined based on the original sampling resolution of the target node;

[0022] The feature matrices of the neighboring nodes of the target node are linearly transformed according to the feature projection dimension to eliminate the dimensional difference between the neighboring nodes and the target node;

[0023] The feature matrices of the adjacent nodes after linear transformation are weighted and summed based on the dynamically updated edge weights. The summation result is then concatenated with the original feature matrix of the target node and nonlinearly activated to generate the fused feature representation of the target node. The asynchronous association feature vector is then extracted by traversing all nodes in the heterogeneous time series graph.

[0024] Preferably, the step of inputting the asynchronous correlation feature vector into a long short-term memory network and outputting the latent fault probability value and degradation trend sequence of the pole-mounted circuit breaker includes:

[0025] The continuously acquired asynchronous correlation feature vectors are constructed into an asynchronous feature sequence according to the time step;

[0026] The asynchronous feature sequence is sequentially input into the forget gate and input gate of the long short-term memory network. The forget gate outputs the forgetting coefficient based on the asynchronous correlation feature vector at the current time and the hidden state at the previous time. The input gate generates candidate memory cell states based on the asynchronous correlation feature vector at the current time and the hidden state at the previous time.

[0027] The current memory cell state is updated based on the forgetting coefficient and the candidate memory cell state, and the current memory cell state is mapped to the latent failure probability value and the corresponding degradation trend sequence for multiple future time steps through the output gate.

[0028] Preferably, the step of acquiring multi-source heterogeneous sensor data of the pole-mounted circuit breaker includes:

[0029] Raw data packets from different sensors are received through a heterogeneous communication interface. The raw data packets carry a timestamp of the acquisition time and a sensor identifier.

[0030] The original data packet is parsed to extract the mechanical vibration data, the coil current data, and the ambient temperature and humidity data.

[0031] The arrival order of the mechanical vibration data, the coil current data, and the ambient temperature and humidity data is verified based on the timestamp, and late data exceeding the delay threshold due to communication congestion is identified and eliminated.

[0032] The retained mechanical vibration data, coil current data, and ambient temperature and humidity data are arranged according to the timestamps to form an asynchronously arriving multi-source data stream, which maintains the independence of the original sampling clock of each sensor.

[0033] Preferably, the step of adjusting the attenuation or enhancement of the off-diagonal elements in the adjacency matrix according to the real-time operating conditions of the pole-mounted circuit breaker includes:

[0034] Extract the real-time load current value and contact temperature value of the pole-mounted circuit breaker;

[0035] When the real-time load current value exceeds the preset current threshold, the value of the adjacent matrix element corresponding to the electromagnetic force transmission delay in the first side weight is increased to enhance the driving effect of the coil current data on the mechanical vibration data.

[0036] When the contact temperature exceeds a preset temperature threshold, the adjacent matrix element value corresponding to the thermal expansion and contraction response delay in the second side weight is increased to enhance the modulation effect of the ambient temperature and humidity data on the mechanical vibration data.

[0037] When the pole-mounted circuit breaker is in the open holding state, the adjacent matrix element values ​​corresponding to the first side weight and the second side weight are decayed to the reference value.

[0038] Preferably, the step of mapping the multi-granularity temporal slices to the initial feature matrix of the corresponding nodes in the heterogeneous temporal graph includes:

[0039] Within the multi-granularity time slice, local peaks and local troughs of the mechanical vibration data are retrieved, and the zero-crossing time interval and amplitude difference between the local peaks and local troughs are calculated.

[0040] The zero-crossing time interval and the amplitude difference are added as high-frequency enhancement features to the extended column of the initial feature matrix corresponding to the mechanical vibration data;

[0041] For the multi-granularity time slice corresponding to the coil current data, the rising edge steepness and peak duration of the coil current data are calculated. The rising edge steepness and peak duration are added as transient enhancement features to the extended column of the initial feature matrix corresponding to the coil current data. The extended column is updated synchronously with the multi-granularity time slice.

[0042] Preferably, the step of weighted summation of the feature matrices of the adjacent nodes after linear transformation based on the dynamically updated edge weights includes:

[0043] Calculate the dot product of the query vector of the target node and the key vector of the adjacent node, and add the dot product to the logarithm of the dynamically updated edge weights to obtain the joint attention score;

[0044] The joint attention score is normalized to generate the attention coefficients of the neighboring nodes relative to the target node;

[0045] The attention coefficient is multiplied as a weight by the value vector in the feature matrix of the neighboring node after linear transformation. The product results of all neighboring nodes of the target node are summed to complete one graph convolution information aggregation operation. The attention coefficient is redistributed as the edge weight is updated to suppress feature interference from neighboring nodes that are irrelevant to the current operating conditions.

[0046] Preferably, the steps of the forgetting gate outputting the forgetting coefficient based on the asynchronous correlation feature vector at the current time and the hidden state at the previous time, and the input gate generating candidate memory cell states based on the asynchronous correlation feature vector at the current time and the hidden state at the previous time, include:

[0047] Detect the sensor source identifier corresponding to the asynchronous correlation feature vector input at the current moment;

[0048] If the sensor source identifier indicates that the current input is a feature corresponding to the mechanical vibration data, then a first bias vector is injected into the forget gate, and the first bias vector guides the forget gate to retain long-term mechanical wear memory;

[0049] If the sensor source identifier indicates that the current input is a feature corresponding to the coil current data, then a second bias vector is injected into the input gate. The second bias vector guides the input gate to strengthen the current transient operation memory. The first bias vector and the second bias vector are dynamically adjusted according to the arrival frequency of the asynchronous associated feature vector.

[0050] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0051] 1. By constructing a heterogeneous time series graph and using the original time series sequences of each sensor as graph nodes while preserving the original sampling resolution, downsampling and interpolation operations are avoided. The node attributes retain the high-frequency sampling resolution of mechanical vibration data and the low-frequency sampling resolution of temperature and humidity data, maintaining the inherent physical characteristics of the original time series sequences. This processing method overcomes the problem of missing high-frequency transient features caused by forced synchronous interpolation, enabling the complete preservation of transient impact components in mechanical vibration data and providing an undistorted underlying data foundation for subsequent condition assessment.

[0052] 2. By dynamically updating the edge weights of a heterogeneous time-series graph based on physical coupling delays, and employing a dynamic graph convolutional network for information aggregation on the heterogeneous time-series graph, asynchronous correlation feature vectors across modes are extracted. The edge weights are dynamically updated based on the electromagnetic force transmission delay and the thermal expansion and contraction response delay. The graph convolutional operation performs feature fusion along the graph topology while preserving the original resolution, reflecting the deep dynamic coupling relationship between mechanical and electrical states. This mechanism solves the problem of simple feature splicing severing the physical field correlation, improving the system's ability to identify latent faults and the reliability of degradation trend prediction.

[0053] 3. By injecting different bias vectors into the Long Short-Term Memory (LSTM) network based on the sensor source identifiers corresponding to the asynchronous associated feature vectors, differentiated updates of memory cell states are achieved. When the input is a mechanical vibration feature, the forget gate is guided to retain long-term mechanical wear memory; when the input is a coil current feature, the input gate is guided to strengthen the current transient operation memory. This mechanism enables the network to adaptively adjust memory according to the dynamic characteristics of different physical quantities, suppressing interference from irrelevant states and enhancing the output stability of the degradation trend sequence. Attached Figure Description

[0054] Figure 1 This is a flowchart of the multi-source heterogeneous sensing data acquisition and preprocessing process of the present invention;

[0055] Figure 2 This is a flowchart of the heterogeneous temporal graph construction and edge weight dynamic update process of the present invention;

[0056] Figure 3 This is a flowchart of the multi-granularity temporal slice and initial feature matrix generation process of the present invention;

[0057] Figure 4 This is a flowchart of the cross-modal feature extraction process of the dynamic graph convolutional network of the present invention;

[0058] Figure 5 This is a flowchart of the Long Short-Term Memory Network State Evaluation and Trend Prediction of the present invention;

[0059] Figure 6 This is a flowchart of the LSTM bias vector dynamic injection and memory update process of the present invention. Detailed Implementation

[0060] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0061] Please refer to Figure 1 This embodiment provides a pole-mounted circuit breaker operation status monitoring system based on multi-source information fusion, deployed at the edge computing node of the distribution network. It connects to vibration sensors, coil current sensors, and ambient temperature and humidity sensors installed on the pole-mounted circuit breaker body via heterogeneous communication interfaces. The edge computing node possesses complete computing capabilities for data reception, preprocessing, graph construction, feature extraction, and status assessment, eliminating the need to transmit raw sensor data to the cloud for processing. The system's workflow sequentially includes multi-source heterogeneous sensor data acquisition, heterogeneous time-series graph construction, cross-modal asynchronous correlation feature extraction, and operation status assessment and trend prediction.

[0062] Specifically, edge computing nodes receive raw data packets from different sensors via heterogeneous communication interfaces. Vibration sensors, employing piezoelectric sensors, collect mechanical vibration signals from the pole-mounted circuit breaker's operating mechanism during opening and closing processes and operation. Coil current sensors, using Rogowski coils, collect the current waveforms of the opening and closing coils. Ambient temperature and humidity sensors, employing capacitive sensors, collect the ambient temperature and relative humidity at the pole-mounted circuit breaker's installation location. Each raw data packet carries a high-precision timestamp of the acquisition time and a unique sensor identifier. The timestamp's accuracy is at the microsecond level, and the sensor identifier distinguishes different types of sensor data.

[0063] After receiving the raw data packet, the edge computing node first parses the packet, extracting the timestamp, sensor identifier, and corresponding sensor data value. After parsing, the arrival order of each sensor data point is verified based on the timestamp. The system presets a delay threshold, which is determined according to the communication protocol and transmission distance of different sensors. Late data exceeding the delay threshold is identified as abnormal data caused by communication congestion and discarded. The retained sensor data is arranged in chronological order of its timestamps, forming an asynchronously arriving multi-source data stream. In this multi-source data stream, the original sampling clocks of each sensor remain independent, without any form of clock synchronization.

[0064] In this embodiment, reference Figure 2 A heterogeneous time series graph is defined as a dynamic directed graph, and its mathematical expression is: .

[0065] in, Indicates time Heterogeneous timing diagram, Indicates time The set of nodes, Indicates time The set of edges, Indicates time The adjacency matrix.

[0066] It contains three nodes, corresponding to mechanical vibration data, coil current data, and ambient temperature and humidity data, respectively. Each node corresponds to an original time series, and the node attributes retain the original sampling resolution. Specifically, the node corresponding to the mechanical vibration data... The attribute is a two-dimensional matrix, where the number of rows equals the number of sampling points within the time window corresponding to that node, and the number of columns equals the measurement dimension of the vibration sensor; the node corresponding to the coil current data The attributes are represented by a two-dimensional matrix, where the number of rows equals the number of sampling points within the time window corresponding to that node, and the number of columns equals the measurement dimension of the coil current sensor; the nodes corresponding to the environmental temperature and humidity data... The attribute is a two-dimensional matrix, the number of rows of which is equal to the number of sampling points in the time window corresponding to the node, and the number of columns is equal to the measurement dimension of the temperature and humidity sensor.

[0067] It contains three directed edges, each representing a node from which a branch node is located. To the node edge From node To the node edge and from the node To its own self-loop edge . The weights are determined based on the electromagnetic force transmission delay between the coil current and mechanical vibration. The weights are determined based on the thermal expansion and contraction response delay between ambient temperature and humidity and mechanical vibration, from the ring edge. The weights are the retention coefficients of the original sampling resolution.

[0068] It is a 3×3 matrix with elements Indicates from node To the node The edge weights are defined in the adjacency matrix. The diagonal elements of the adjacency matrix are the original sampling resolution retention coefficients for each node, while the off-diagonal elements are the weights of the corresponding edges. The off-diagonal elements in the adjacency matrix are dynamically updated based on the real-time operating conditions of the pole-mounted circuit breaker.

[0069] In this embodiment, reference Figure 4 The dynamic graph convolutional network consists of two graph convolutional layers and one fully connected layer. The first graph convolutional layer extracts local features within nodes, and the second graph convolutional layer extracts global relational features across nodes. The fully connected layer maps the features output by the graph convolutional layers into asynchronous relational feature vectors of uniform dimension.

[0070] Specifically, for target nodes in heterogeneous time series graphs First, the feature projection dimension is determined based on the original sampling resolution of the target node. . Sampling frequency of the target node It is directly proportional, and its calculation formula is:

[0071] in, This is the proportionality coefficient.

[0072] Then, the target node All neighboring nodes Feature matrix according to Perform a linear transformation. The formula for calculating the linear transformation is: .

[0073] in, The feature matrix of adjacent nodes after linear transformation. Let be a learnable linear transformation matrix with dimension . , Adjacent nodes The original feature dimensions.

[0074] After the linear transformation is completed, based on right Perform a weighted summation. The formula for calculating the weighted summation is: .

[0075] in, This is the weighted sum of the features of adjacent nodes. For the target node The set of all adjacent nodes.

[0076] Will With the original feature matrix of the target node The features are concatenated to obtain the concatenated feature matrix. . After processing with a nonlinear activation function, a fused feature representation of the target node is generated. The nonlinear activation function used is the ReLU function, and its calculation formula is as follows: .

[0077] in, The weight matrix is ​​a learnable matrix. This is a learnable bias vector.

[0078] Traverse all nodes in the heterogeneous time series graph and generate a fused feature representation for all nodes. Concatenate the fused feature representations of all nodes to obtain the cross-modal asynchronous correlation feature vector. .

[0079] In this embodiment, reference Figure 5 A Long Short-Term Memory (LSTM) network consists of an input layer, two hidden layers, and an output layer. The input layer receives asynchronously correlated feature vectors, the hidden layers process temporal features, and the output layer outputs the probability values ​​of latent faults and the degradation trend sequence.

[0080] Specifically, the asynchronously acquired asynchronous feature vectors are constructed into an asynchronous feature sequence according to a preset time step. ,in, for The Middle One element, The time step is defined as follows: The asynchronous feature sequence is sequentially input into each time step of the Long Short-Term Memory network.

[0081] At each time step Long Short-Term Memory (LSTM) networks receive asynchronous correlation feature vectors at the current time step. Compared to the previous hidden state The hidden state at the current moment is updated through the coordinated action of the forget gate, input gate, memory cells, and output gate. With memory cell state .

[0082] The formula for calculating the forgetting gate is: .

[0083] in, The forgetting coefficient is the output of the forgetting gate. It is the sigmoid activation function. Here is the weight matrix for the forget gate. This is the bias vector for the forget gate.

[0084] The formula for calculating the input gate is: .

[0085] in, The output of the input gate, Candidate memory cell states, The weight matrix of the input gate, The weight matrix for candidate memory cells, Let be the bias vector of the input gate. is the bias vector of the candidate memory cells.

[0086] The formula for updating the state of memory cells is: .

[0087] in, This represents element-wise multiplication. This refers to the state of memory cells before the update. This represents the updated state of memory cells.

[0088] The formula for calculating the output gate is: .

[0089] in, For the output of the output gate, Here is the weight matrix of the output gate. This is the bias vector for the output gate.

[0090] Finally, the hidden state at the current moment. The latent fault probability value of the fully connected network of the input and output layers and the output pole-mounted circuit breaker. And corresponding to the future Deterioration trend sequence at each time step , for The Element, probability value of latent failure The value range is [0,1], and a larger value indicates a higher probability of latent failure. Deterioration trend sequence. Each element in the table represents the health status score of the pole-mounted circuit breaker at the corresponding future time step, with a score range of [0, 100]. The smaller the value, the worse the health status. The basic parameters of the multi-source heterogeneous sensor data are shown in Table 1.

[0091] Table 1 Basic Parameters of Multi-Source Heterogeneous Sensing Data

[0092] Vibration sensor 2000 3 LoRa 500 16-bit Coil current sensor 200 1 RS-485 200 12-digit Ambient temperature and humidity sensor 0.5 2 NB-IoT 2000 8-bit

[0093] Table 1 shows the basic parameters of the three sensors used in this embodiment. The vibration sensor uses a high sampling frequency to capture the high-frequency transient characteristics of mechanical vibration, the coil current sensor uses a medium sampling frequency to record the current waveform of the opening and closing coils, and the ambient temperature and humidity sensor uses a low sampling frequency to monitor slow changes in environmental parameters. Different sensors use different communication protocols to adapt to their respective data transmission requirements, and the delay threshold is set according to the characteristics of the communication protocol to ensure timely reception of sensor data under normal communication conditions.

[0094] This embodiment avoids the loss of high-frequency features caused by forced resampling by constructing a heterogeneous time series graph and preserving the original sampling resolution of each sensor. The dynamic graph convolutional network can aggregate cross-modal information on the heterogeneous time series graph, extracting asynchronous correlation features between different physical quantities. The long short-term memory network can process time series features and output latent fault probabilities and degradation trends, achieving comprehensive monitoring of the pole-mounted circuit breaker's operating status.

[0095] In a preferred embodiment, the edge weights between nodes in the heterogeneous timing graph are dynamically updated based on the physical coupling delay between each information source. Specifically, the electromagnetic force transmission delay between the node corresponding to the mechanical vibration data and the node corresponding to the coil current data is first calculated as the first edge weight. The electromagnetic force transmission delay refers to the time difference between the electromagnetic force generated by the coil current acting on the operating mechanism and causing mechanical vibration. This delay is related to the mechanical structure parameters of the operating mechanism and the magnitude of the electromagnetic force.

[0096] Then, the thermal expansion and contraction response delay between the nodes corresponding to the ambient temperature and humidity data and the nodes corresponding to the mechanical vibration data is calculated as the weight of the second side. The thermal expansion and contraction response delay refers to the time difference between the change in ambient temperature causing the operating mechanism components to expand and contract and affecting the mechanical vibration characteristics. This delay is related to the component's material thermal expansion coefficient, mass, and heat conduction rate.

[0097] An adjacency matrix is ​​constructed based on the weights of the first and second edges. The off-diagonal elements in the adjacency matrix are adjusted for attenuation or enhancement according to the real-time operating conditions of the pole-mounted circuit breaker, while the diagonal elements are the retention coefficients of the original sampling resolution. These retention coefficients control the weight of node features during information aggregation, ensuring that the original features are not excessively diluted.

[0098] Furthermore, edge computing nodes receive raw data packets from different sensors via heterogeneous communication interfaces. These interfaces include LoRa, RS-485, and NB-IoT interfaces, used to receive data from vibration sensors, coil current sensors, and ambient temperature and humidity sensors, respectively. Each raw data packet is formatted with a preamble, synchronization word, sensor identifier, timestamp, data length, data payload, and checksum. The preamble and synchronization word are used for data packet synchronization detection, the sensor identifier distinguishes different sensors, the timestamp records the time of data acquisition, the data length indicates the number of bytes in the data payload, the data payload contains the actual sensor data value, and the checksum is used for data packet error detection.

[0099] After receiving the raw data packet, the edge computing node first performs a checksum verification. If the checksum verification fails, the data packet is discarded. If the checksum verification passes, the data packet is parsed to extract the sensor identifier, timestamp, and data payload. Based on the sensor identifier, the data type is determined, and the data payload is converted into the corresponding physical quantity value.

[0100] After parsing, the arrival order of each sensor data point is verified based on its timestamp. The system maintains a data buffer to store the most recently received sensor data. When new sensor data arrives, its timestamp is compared with the timestamps of existing data in the buffer, and it is inserted into the corresponding position in the buffer according to the timestamp order. For late data exceeding the delay threshold, the system removes it from the buffer. The retained sensor data is arranged by timestamp, forming an asynchronously arriving multi-source data stream. In the multi-source data stream, the raw sampling clocks of each sensor remain independent, without any form of clock synchronization or interpolation.

[0101] In this embodiment, the off-diagonal elements in the adjacency matrix are adjusted for attenuation or enhancement based on the real-time operating conditions of the pole-mounted circuit breaker. Specifically, the system extracts the real-time load current and contact temperature values ​​of the pole-mounted circuit breaker. The real-time load current value is obtained through a current transformer installed on the incoming side of the pole-mounted circuit breaker, and the contact temperature value is obtained through a temperature sensor installed near the circuit breaker contacts.

[0102] When the real-time load current exceeds the preset current threshold, the value of the adjacent matrix element corresponding to the electromagnetic force transmission delay in the first edge weight is increased. This is because as the load current increases, the influence of the electromagnetic force generated by the coil current on the operating mechanism strengthens, and the coupling relationship between the coil current data and the mechanical vibration data becomes tighter. Increasing the corresponding adjacent matrix element value can enhance the driving influence of the coil current data on the mechanical vibration data, making the graph convolutional network pay more attention to the correlation features between the two.

[0103] When the contact temperature exceeds a preset temperature threshold, the value of the adjacent matrix element corresponding to the thermal expansion and contraction response delay in the second side weights is increased. This is because as the contact temperature rises, the influence of ambient temperature on the thermal expansion and contraction of the operating mechanism components intensifies, and the coupling relationship between ambient temperature and humidity data and mechanical vibration data becomes more pronounced. Increasing the corresponding adjacent matrix element value can strengthen the modulation effect of ambient temperature and humidity data on mechanical vibration data, enabling the graph convolutional network to better capture the impact of temperature changes on mechanical characteristics.

[0104] When the pole-mounted circuit breaker is in the open-hold state, the values ​​of the adjacent matrix elements corresponding to the first and second side weights are attenuated to the reference values. This is because in the open-hold state, the coil current is zero, there is no electromagnetic force transmission delay, and the influence of ambient temperature and humidity on mechanical vibration is relatively small. Attenuating the adjacent matrix element values ​​to the reference values ​​can reduce interference from irrelevant information and improve the accuracy of condition monitoring. The dynamic adjustment rules for the off-diagonal elements of the adjacent matrix are shown in Table 2.

[0105] Table 2. Rules for Dynamic Adjustment of Off-Diagonal Elements in the Adjacency Matrix

[0106] Normal operating conditions 1.0 1.0 none Overload conditions 1.5 1.0 Load current > 1.2 times rated current High temperature working conditions 1.0 1.3 Contact temperature >80℃ Overload + High Temperature Conditions 1.5 1.3 Both of the above conditions must be met. Trip holding state 0.2 0.5 The circuit breaker is in the open position and there is no operating command.

[0107] Table 2 illustrates the dynamic adjustment rules for the off-diagonal elements of the adjacency matrix. The adjustment coefficient is used to multiply the baseline edge weight value to obtain the adjusted edge weight value. Different adjustment coefficients are used under different operating conditions to adapt to changes in the physical coupling relationship. In the open-hold state, the adjustment coefficient for the first edge weight is set to a smaller value because the coil current is zero and the electromagnetic coupling relationship does not exist; the adjustment coefficient for the second edge weight is set to a medium value because ambient temperature and humidity still have a certain impact on the mechanical structure.

[0108] This embodiment dynamically updates the edge weights of the heterogeneous time-series graph based on physical coupling delay, enabling the graph's topology to reflect the physical coupling relationships under real-time operating conditions of the pole-mounted circuit breaker. The dynamic adjustment mechanism of the adjacency matrix elements strengthens the association of features relevant to the current operating condition and suppresses interference from irrelevant information. The acquisition and preprocessing of multi-source heterogeneous sensor data retains the asynchronous characteristics of the original data, providing an accurate data foundation for subsequent cross-modal information fusion.

[0109] In a preferred embodiment, reference Figure 3 In the heterogeneous time series graph, nodes correspond to the original time series sequences of each sensor, and the node attributes retain the original sampling resolution. Specifically, mechanical vibration data, coil current data, and ambient temperature and humidity data are continuously extracted without interpolation according to their respective inherent sampling frequencies to form multi-granularity time series slices. The time window length of the time series slice is set according to the characteristics of different sensors to ensure that each time series slice contains sufficient feature information.

[0110] Multi-granularity time-series slices are mapped to the initial feature matrices of corresponding nodes in the heterogeneous time-series graph. The number of rows in the initial feature matrix is ​​positively correlated with the time length of the multi-granularity time-series slice, and the number of columns is consistent with the measurement dimension of the sensor. The dimensionality heterogeneity of the initial feature matrix is ​​maintained within the topology of the heterogeneous time-series graph, rejecting dimensionality reduction alignment to the lowest sampling resolution. This approach preserves the temporal resolution characteristics of the original data from each sensor, avoiding the loss of high-frequency features caused by downsampling and the generation of spurious data due to interpolation.

[0111] Furthermore, when mapping multi-granularity time-series slices to the initial feature matrix of corresponding nodes in the heterogeneous time-series graph, targeted feature enhancement processing is performed on different types of sensor data. For multi-granularity time-series slices corresponding to mechanical vibration data, local peaks and local troughs are retrieved within the slice. The detection method for local peaks is as follows: for each data point in the time-series, if its value is greater than the values ​​of the two adjacent data points, it is marked as a local peak. The detection method for local troughs is as follows: for each data point in the time-series, if its value is less than the values ​​of the two adjacent data points, it is marked as a local trough.

[0112] The zero-crossing time interval and amplitude difference between each local peak and its adjacent local trough are calculated. The zero-crossing time interval refers to the time difference between the signal crossing the zero point from the local peak to the next local trough. The amplitude difference refers to the difference between the amplitude of the local peak and the amplitude of the adjacent local trough. The zero-crossing time interval and amplitude difference are added as high-frequency enhancement features to the extended columns of the initial feature matrix corresponding to the mechanical vibration data. The number of extended columns is consistent with the number of detected local peak-trough pairs and is updated synchronously with multi-granularity time slices.

[0113] For multi-granularity timing slices corresponding to coil current data, calculate the rise edge steepness and peak duration of the coil current data. Rise edge steepness refers to the slope of the coil current as it rises from the 10% peak value to the 90% peak value; its calculation formula is as follows: .

[0114] in, For the steepness of the rising edge, This represents the peak value of the coil current. The moment when the current reaches 10% of its peak value. This is the moment when the current reaches 90% of its peak value.

[0115] Peak sustain duration refers to the length of time the coil current remains above 90% of its peak value. Rise edge steepness and peak sustain duration are added as transient enhancement features to the extended columns of the initial feature matrix corresponding to the coil current data. The extended columns are two in number, corresponding to rise edge steepness and peak sustain duration respectively, and are updated synchronously with multi-granularity time slices.

[0116] For multi-granularity time-series slices corresponding to environmental temperature and humidity data, due to their low sampling frequency, no additional feature enhancement processing is performed; the original data is directly used as elements of the initial feature matrix. The parameter configuration for multi-granularity time-series slices is shown in Table 3.

[0117] Table 3. Parameter Configuration Table for Multi-Grained Temporal Slices

[0118] Vibration sensor 0.1 50 200×3 variable Coil current sensor 1.0 30 200×1 2 Ambient temperature and humidity sensor 60.0 0 30×2 0

[0119] Table 3 shows the parameter configuration for multi-granularity time-series slicing. The time window length for the vibration sensor is set to 0.1 seconds to capture transient vibration characteristics during the opening and closing process; the slice overlap rate is set to 50% to ensure continuous feature coverage. The time window length for the coil current sensor is set to 1.0 second to completely record the current waveform of the opening and closing coil; the slice overlap rate is set to 30% to balance computational load with feature continuity. The time window length for the ambient temperature and humidity sensor is set to 60 seconds to match the sampling frequency; the slice overlap rate is set to 0 because environmental parameters change slowly, eliminating the need for overlapping slices.

[0120] This embodiment forms multi-granularity time-series slices through continuous truncation without interpolation, preserving the temporal resolution characteristics of the original data from each sensor. Targeted feature enhancement processing can extract key features from different types of sensor data, improving the representational capability of the initial feature matrix. Maintaining the dimensionality heterogeneity of the initial feature matrix provides a rich feature foundation for cross-modal information aggregation in dynamic graph convolutional networks.

[0121] In a preferred embodiment, a dynamic graph convolutional network is used to perform information aggregation operations on a heterogeneous temporal graph to extract asynchronous correlation feature vectors across modalities. Specifically, for a target node in the heterogeneous temporal graph, the feature projection dimension is determined based on the original sampling resolution of the target node. The feature projection dimension is proportional to the sampling frequency of the target node, ensuring that nodes sampled at high frequencies have higher feature dimensions to retain more detailed information.

[0122] The feature matrices of the target node's neighboring nodes are linearly transformed according to their feature projection dimensions to eliminate dimensional differences between the neighboring nodes and the target node. The linear transformation matrix is ​​learned through training and can map features of different dimensions to a unified feature space. Based on dynamically updated edge weights, the feature matrices of the linearly transformed neighboring nodes are weighted and summed. The sum is then concatenated with the original feature matrix of the target node and nonlinearly activated to generate a fused feature representation of the target node. Asynchronous association feature vectors are extracted by traversing all nodes in the heterogeneous temporal graph.

[0123] Furthermore, when performing a weighted summation of the feature matrices of neighboring nodes after linear transformation based on dynamically updated edge weights, an attention mechanism is introduced to adaptively adjust the contribution weights of neighboring nodes. Specifically, the dot product of the query vector of the target node and the key vectors of neighboring nodes is calculated, and the dot product is added to the logarithm of the dynamically updated edge weights to obtain the joint attention score. The query vector and key vector are obtained by performing a linear transformation on the node feature matrices, and the calculation formula is as follows: .

[0124] in, For the target node The query vector, Adjacent nodes The key vector, To find the learnable linear transformation matrix of the query vector, Let be the learnable linear transformation matrix of the key vector.

[0125] The formula for calculating the joint attention score is: .

[0126] in, Adjacent nodes Relative to the target node Joint attention score, The edge weights are dynamically updated.

[0127] The joint attention score is normalized to generate attention coefficients between neighboring nodes and the target node. The normalization process uses the softmax function, calculated as follows: .

[0128] in, Adjacent nodes Relative to the target node Attention coefficient.

[0129] The attention coefficients are used as weights to multiply the value vectors in the feature matrices of neighboring nodes after linear transformation. The sum of these products over all neighboring nodes of the target node completes one graph convolution information aggregation operation. The value vectors are obtained by linearly transforming the feature matrices of neighboring nodes; the calculation formula is as follows: .

[0130] in, Adjacent nodes The value vector, It is a learnable linear transformation matrix.

[0131] The formula for weighted summation is: .

[0132] in, This is a weighted summation of the features of adjacent nodes.

[0133] The attention coefficients are redistributed as the edge weights are updated, which can suppress the feature interference of neighboring nodes that are irrelevant to the current operating condition and enhance the feature contribution of neighboring nodes that are relevant to the current operating condition.

[0134] In this embodiment, asynchronous correlated feature vectors are input into a long short-term memory network, which outputs the latent fault probability value and degradation trend sequence of the pole-mounted circuit breaker. Specifically, the continuously acquired asynchronous correlated feature vectors are constructed into an asynchronous feature sequence according to the time step. Since the asynchronous correlated feature vectors are generated at different times and their time intervals are not fixed, the system converts the asynchronous feature sequence into a sequence with a fixed time step through interpolation. Linear interpolation is used to ensure the continuity of the time-series features.

[0135] The asynchronous feature sequence is sequentially input into the forget gate and input gate of the Long Short-Term Memory (LSTM) network. The forget gate outputs a forgetting coefficient based on the asynchronous correlation feature vector at the current time step and the hidden state at the previous time step. The input gate generates candidate memory cell states based on the asynchronous correlation feature vector at the current time step and the hidden state at the previous time step. The current memory cell state is updated based on the forgetting coefficient and the candidate memory cell states, and the current memory cell state is mapped to a latent failure probability value and a corresponding degradation trend sequence for multiple future time steps through the output gate.

[0136] Further, refer to Figure 6 The bias vectors of the forget gate and the input gate are dynamically adjusted based on the sensor source identifier corresponding to the asynchronous correlation feature vector of the input. Specifically, the sensor source identifier corresponding to the asynchronous correlation feature vector of the current time step is detected. The sensor source identifier is a one-dimensional feature added during the generation of the asynchronous correlation feature vector, used to indicate which sensor's data the feature vector mainly originates from.

[0137] If the sensor source identifier indicates that the current input is a feature corresponding to mechanical vibration data, a first bias vector is injected into the forget gate. All elements of the first bias vector are positive, and its function is to increase the output of the forget gate, thereby guiding the forget gate to retain more long-term memory cell states. This is because mechanical vibration data mainly reflects the long-term mechanical wear of the pole-mounted circuit breaker operating mechanism, and long-term historical memory needs to be retained to accurately assess the deterioration trend.

[0138] If the sensor source identifier indicates that the current input corresponds to the coil current data feature, a second bias vector is injected into the input gate. All elements of the second bias vector are positive, and its function is to increase the output of the input gate, thereby guiding the input gate to strengthen the update of the current candidate memory cell state. This is because the coil current data mainly reflects the transient electrical characteristics of the pole-mounted circuit breaker's opening and closing operations, and it is necessary to strengthen the current transient memory in order to accurately identify abnormal situations during the operation.

[0139] The first and second bias vectors are dynamically adjusted based on the arrival frequency of the asynchronously correlated feature vectors. A higher arrival frequency indicates more frequent updates of this type of sensor data, and the corresponding bias vector value is larger. The calculation formula is as follows: .

[0140] in, This is the first bias vector. This is the second bias vector. and This is the proportionality coefficient. The arrival frequency of the asynchronous correlated feature vector corresponding to the mechanical vibration data. This represents the arrival frequency of the asynchronous correlated feature vector corresponding to the coil current data. The dynamic adjustment parameters of the LSTM bias vector are shown in Table 4.

[0141] Table 4. LSTM Bias Vector Dynamic Adjustment Parameters

[0142] Mechanical vibration data Forgotten Gate [0.2,0.2,...,0.2] 0.001 [0.2,0.5] Coil current data Input gate [0.3,0.3,...,0.3] 0.002 [0.3,0.7] Ambient temperature and humidity data none none none none

[0143] Table 4 shows the dynamic adjustment parameters of the LSTM bias vector. The initial value of the first bias vector is set to 0.2, the scaling factor is set to 0.001, and the adjustment range is [0.2, 0.5]. The initial value of the second bias vector is set to 0.3, the scaling factor is set to 0.002, and the adjustment range is [0.3, 0.7]. No additional bias vector is injected into the features corresponding to the environmental temperature and humidity data because environmental parameters change slowly and have a relatively small impact on the state of memory cells.

[0144] This embodiment introduces an attention mechanism into a dynamic graph convolutional network, enabling adaptive adjustment of the feature contribution weights of adjacent nodes and improving the accuracy of cross-modal information aggregation. By injecting different bias vectors into the long short-term memory network based on sensor source identifiers, differentiated updates of memory cell states are achieved, allowing the network to adaptively adjust its memory according to the dynamic characteristics of different physical quantities. This mechanism can retain long-term mechanical wear memory, strengthen current transient operational memory, and improve the reliability of latent fault identification and degradation trend prediction.

Claims

1. A method for monitoring the operating status of pole-mounted circuit breakers based on multi-source information fusion, characterized in that, include: Acquire multi-source heterogeneous sensing data of the pole-mounted circuit breaker, including mechanical vibration data, coil current data, and ambient temperature and humidity data; A heterogeneous time series graph is constructed based on the sampling rate and communication delay of each sensor in the multi-source heterogeneous sensing data. The nodes in the heterogeneous time series graph correspond to the original time series sequences of each sensor. The attributes of the nodes retain the original sampling resolution. The edge weights between nodes in the heterogeneous time series graph are dynamically updated based on the physical coupling delay between each information source. A dynamic graph convolutional network is used to perform information aggregation operations on the heterogeneous time series graph to extract asynchronous correlation feature vectors across modalities; The asynchronous correlation feature vector is input into a long short-term memory network to output the latent fault probability value and deterioration trend sequence of the pole-mounted circuit breaker. The step of dynamically updating the edge weights between nodes in the heterogeneous time series graph based on the physical coupling delay between each information source includes: The electromagnetic force transmission delay between the node corresponding to the mechanical vibration data and the node corresponding to the coil current data is calculated as the weight of the first side. The thermal expansion and contraction response delay between the node corresponding to the environmental temperature and humidity data and the node corresponding to the mechanical vibration data is calculated as the weight of the second side. An adjacency matrix is ​​constructed based on the first edge weight and the second edge weight. The off-diagonal elements in the adjacency matrix are attenuated or enhanced according to the real-time operating conditions of the pole-mounted circuit breaker, and the diagonal elements are the retention coefficients of the original sampling resolution. The steps for preserving the original sampling resolution of the node attributes in the heterogeneous timing graph, where the nodes correspond to the original timing sequences of each sensor, include: The mechanical vibration data, the coil current data, and the ambient temperature and humidity data are continuously extracted without interpolation according to their respective inherent sampling frequencies to form multi-granularity time slices; The multi-granularity time-series slices are mapped to the initial feature matrices of the corresponding nodes in the heterogeneous time-series graph. The number of rows in the initial feature matrix is ​​positively correlated with the time length of the multi-granularity time-series slices, and the number of columns is consistent with the measurement dimension of the sensor. The dimensionality heterogeneity of the initial feature matrix is ​​maintained in the topology of the heterogeneous time series graph, and dimensionality reduction alignment to the lowest sampling resolution is rejected.

2. The method for monitoring the operating status of pole-mounted circuit breakers based on multi-source information fusion according to claim 1, characterized in that, The step of using a dynamic graph convolutional network to perform information aggregation operations on the heterogeneous time series graph and extract cross-modal asynchronous correlation feature vectors includes: For the target node in the heterogeneous time series graph, the feature projection dimension is determined based on the original sampling resolution of the target node; The feature matrices of the neighboring nodes of the target node are linearly transformed according to the feature projection dimension; The feature matrices of the adjacent nodes after linear transformation are weighted and summed based on the dynamically updated edge weights. The summation result is then concatenated with the original feature matrix of the target node and nonlinearly activated to generate the fused feature representation of the target node. The asynchronous association feature vector is then extracted by traversing all nodes in the heterogeneous time series graph.

3. The method for monitoring the operating status of pole-mounted circuit breakers based on multi-source information fusion according to claim 1, characterized in that, The steps of inputting the asynchronous correlation feature vector into a long short-term memory network and outputting the latent fault probability value and degradation trend sequence of the pole-mounted circuit breaker include: The continuously acquired asynchronous correlation feature vectors are constructed into an asynchronous feature sequence according to the time step; The asynchronous feature sequence is sequentially input into the forget gate and input gate of the long short-term memory network. The forget gate outputs the forgetting coefficient based on the asynchronous correlation feature vector at the current time and the hidden state at the previous time. The input gate generates candidate memory cell states based on the asynchronous correlation feature vector at the current time and the hidden state at the previous time. The current memory cell state is updated based on the forgetting coefficient and the candidate memory cell state, and the current memory cell state is mapped to the latent failure probability value and the corresponding degradation trend sequence for multiple future time steps through the output gate.

4. The method for monitoring the operating status of pole-mounted circuit breakers based on multi-source information fusion according to claim 1, characterized in that, The step of acquiring multi-source heterogeneous sensor data of the pole-mounted circuit breaker includes: Raw data packets from different sensors are received through a heterogeneous communication interface. The raw data packets carry a timestamp of the acquisition time and a sensor identifier. The original data packet is parsed to extract the mechanical vibration data, the coil current data, and the ambient temperature and humidity data. The arrival order of the mechanical vibration data, the coil current data, and the ambient temperature and humidity data is verified based on the timestamp, and late data exceeding the delay threshold due to communication congestion is identified and eliminated. The retained mechanical vibration data, coil current data, and ambient temperature and humidity data are arranged according to the timestamps to form an asynchronously arriving multi-source data stream, which maintains the independence of the original sampling clock of each sensor.

5. The method for monitoring the operating status of pole-mounted circuit breakers based on multi-source information fusion according to claim 1, characterized in that, The step of adjusting the attenuation or enhancement of the off-diagonal elements in the adjacency matrix according to the real-time operating conditions of the pole-mounted circuit breaker includes: Extract the real-time load current value and contact temperature value of the pole-mounted circuit breaker; When the real-time load current value exceeds the preset current threshold, the value of the adjacent matrix element corresponding to the electromagnetic force transmission delay in the first side weight is increased to enhance the driving effect of the coil current data on the mechanical vibration data. When the contact temperature exceeds a preset temperature threshold, the adjacent matrix element value corresponding to the thermal expansion and contraction response delay in the second side weight is increased to enhance the modulation effect of the ambient temperature and humidity data on the mechanical vibration data. When the pole-mounted circuit breaker is in the open holding state, the adjacent matrix element values ​​corresponding to the first side weight and the second side weight are decayed to the reference value.

6. The method for monitoring the operating status of pole-mounted circuit breakers based on multi-source information fusion according to claim 1, characterized in that, The step of mapping the multi-granularity temporal slices to the initial feature matrix of the corresponding nodes in the heterogeneous temporal graph includes: Within the multi-granularity time slice, local peaks and local troughs of the mechanical vibration data are retrieved, and the zero-crossing time interval and amplitude difference between the local peaks and local troughs are calculated. The zero-crossing time interval and the amplitude difference are added as high-frequency enhancement features to the extended column of the initial feature matrix corresponding to the mechanical vibration data; For the multi-granularity time slice corresponding to the coil current data, the rising edge steepness and peak duration of the coil current data are calculated. The rising edge steepness and peak duration are added as transient enhancement features to the extended column of the initial feature matrix corresponding to the coil current data. The extended column is updated synchronously with the multi-granularity time slice.

7. The method for monitoring the operating status of pole-mounted circuit breakers based on multi-source information fusion according to claim 2, characterized in that, The step of weighted summation of the feature matrices of adjacent nodes after linear transformation based on the dynamically updated edge weights includes: Calculate the dot product of the query vector of the target node and the key vector of the adjacent node, and add the dot product to the logarithm of the dynamically updated edge weights to obtain the joint attention score; The joint attention score is normalized to generate the attention coefficients of the neighboring nodes relative to the target node; The attention coefficient is multiplied as a weight by the value vector in the feature matrix of the neighboring node after linear transformation. The product results of all neighboring nodes of the target node are summed to complete one graph convolution information aggregation operation. The attention coefficient is redistributed as the edge weight is updated to suppress feature interference from neighboring nodes that are irrelevant to the current operating conditions.

8. The method for monitoring the operating status of pole-mounted circuit breakers based on multi-source information fusion according to claim 3, characterized in that, The steps of the forgetting gate outputting the forgetting coefficient based on the asynchronous correlation feature vector at the current time and the hidden state at the previous time, and the input gate generating candidate memory cell states based on the asynchronous correlation feature vector at the current time and the hidden state at the previous time, include: Detect the sensor source identifier corresponding to the asynchronous correlation feature vector input at the current moment; If the sensor source identifier indicates that the current input is a feature corresponding to the mechanical vibration data, then a first bias vector is injected into the forget gate, and the first bias vector guides the forget gate to retain long-term mechanical wear memory; If the sensor source identifier indicates that the current input is a feature corresponding to the coil current data, then a second bias vector is injected into the input gate. The second bias vector guides the input gate to strengthen the current transient operation memory. The first bias vector and the second bias vector are dynamically adjusted according to the arrival frequency of the asynchronous associated feature vector.

Images