Column-mounted circuit breaker diagnosis method and system based on multi-mode time sequence feature fusion and edge-cloud cooperation

Abstract

The invention relates to a pole-mounted circuit breaker diagnosis method and system based on multi-mode time sequence feature fusion and edge-cloud collaboration, and the method comprises the steps: installing a multi-source sensor, including a current sensor, a triaxial accelerometer and a salt density sensor, on a pole-mounted circuit breaker, and collecting an electrical signal, a vibration signal and environment corrosion data in real time; calculating a dynamic time warping distance and a wavelet packet energy entropy in an edge calculation unit, and constructing an edge feature vector; inputting the edge feature vector and the topological data of the power distribution network into a cloud two-channel deep learning model; and generating a maintenance strategy according to state probability distribution output by the model, and triggering real-time alarm when emergency features appear. The pole-mounted circuit breaker diagnosis method and system based on multi-mode time sequence feature fusion and edge-cloud cooperation have the advantages that the fault diagnosis accuracy is improved, the environmental adaptability is enhanced, millisecond-level quick response is achieved, and maintenance decision closed-loop management is optimized.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of intelligent monitoring of power distribution network equipment, in particular to a multi-modal time sequence feature fusion and edge-cloud collaborative pole-mounted circuit breaker diagnosis method and system. BACKGROUND

[0002] As a key equipment of the power distribution network, the operation state of the pole-mounted circuit breaker directly affects the power supply reliability. Traditional state diagnosis mainly relies on planned maintenance and manual inspection, which has the problems of low efficiency and strong hysteresis. In recent years, intelligent diagnosis methods have been gradually applied, but there are the following technical bottlenecks: single data dimension: existing methods are mostly based on static operation data (such as service life and breaking times), lack of dynamic time sequence features such as vibration waveform and arc current, and are difficult to capture transient information of mechanical faults. Insufficient model adaptability: the mainstream algorithms (such as random forest and ResNet) do not consider the topology correlation of the power distribution network, and are sensitive to data imbalance, resulting in high false negative rate of minority class faults (such as mechanism jam). Slow response speed: the delay of the cloud centralized processing mode is minutes, which cannot meet the millisecond-level rapid isolation requirement of the circuit breaker fault. In addition, the existing systems generally have the problems of insufficient quantification of environmental corrosion factors, idle edge computing capability, poor executable diagnosis results, and the like, and are difficult to realize closed-loop management from data acquisition to maintenance decision. The existing technology needs to be improved in view of the above problems. SUMMARY

[0003] The application aims to provide a multi-modal time sequence feature fusion and edge-cloud collaborative pole-mounted circuit breaker diagnosis method and system, which has the advantages of improving fault diagnosis accuracy, enhancing environmental adaptability, realizing millisecond-level rapid response, and optimizing maintenance decision closed-loop management.

[0004] To achieve the above-mentioned purpose, the application provides the following technical scheme: The application provides a multi-modal time sequence feature fusion and edge-cloud collaborative pole-mounted circuit breaker diagnosis method, and the technical scheme is as follows: installing multi-source sensors on the pole-mounted circuit breaker, including a current sensor, a three-axis accelerometer and a salt density sensor, and collecting electrical signals, vibration signals and environmental corrosion data in real time; calculating dynamic time warping distance and wavelet packet energy entropy in the edge computing unit, and constructing an edge feature vector; inputting the edge feature vector and power distribution network topology data into a cloud dual-channel deep learning model, wherein the time sequence channel adopts ConvLSTM to process the original waveform, and the topology channel adopts GNN to process the network structure relationship; generating a maintenance strategy according to the state probability distribution output by the model, and triggering real-time alarm when emergency features appear.

[0005] Further, the application also provides that the arc current waveform in the breaking process is captured by the current sensor, and the current effective value and breaking time are calculated.

[0006] Further, the application also proposes that the total energy is calculated according to the collected vibration signal, and the signal is decomposed by 3-layer db4 wavelet packet, and 8 sub-band energy entropies are extracted.

[0007] Further, the application also proposes that the equivalent corrosion rate is calculated according to the salt density detected by the salt density sensor in combination with the ambient temperature, and is updated and input into the cloud model at a certain time interval.

[0008] Further, the application also proposes that the method for calculating the dynamic time warping distance in the edge feature extraction is as follows: the measured arc current waveform is dynamically time-warped with the standard waveform, and the minimum path distance DTW is solved.

[0009] Further, the application also proposes that in the time sequence channel of the dual-channel deep learning model: the arc current waveform and the down-sampled vibration signal are input, the time sequence features are extracted through two layers of ConvLSTM, and a maximum pooling layer is connected.

[0010] Further, the application also proposes that the construction of the topology channel includes: the power distribution network structure is coded into an adjacency matrix A, the node feature vector contains the breaker load rate and the corrosion rate, the neighbor node information is aggregated through two layers of GNN, and the topology embedding vector is output.

[0011] Further, the application also proposes that when the cloud model outputs the state probability satisfies: ≥0.4, a "replace within 48 hours" instruction is generated; 0.2≤ 0.4, a "repair within 7 days" instruction is generated; <0.2, marked as "normal monitoring".

[0012] Further, the application also proposes that the real-time alarm condition is that in the edge computing unit, if DTW>0.3 or WPEE>2.5 is detected, the breaker is immediately locked and an alarm signal is uploaded to the monitoring center.

[0013] Further, the application also proposes a system of a multi-modal time sequence feature fusion and edge-cloud collaborative pole-mounted circuit breaker diagnosis method, comprising: a sensor module: a current sensor, a three-axis accelerometer, and a salt density sensor; an edge computing unit: an FPGA chip, used for real-time calculation of DTW, WPEE and Ev; a cloud server: deploying a dual-channel deep learning model, performing deep diagnosis every day; a decision execution module: triggering a breaker operation or a maintenance work order according to a cloud instruction or an edge alarm.

[0014] From the above, the application provides a kind of multi-modal time sequence feature fusion and edge-cloud collaborative pole-mounted circuit breaker diagnosis method and system, multi-dimensional dynamic time sequence data are collected by multi-source sensor, feature extraction and emergency event processing are carried out on the edge side, time sequence features and topological correlation features are combined by cloud double-channel deep learning model, and a closed-loop management system from data acquisition to maintenance decision is constructed, with the advantages of improving fault diagnosis accuracy, enhancing environmental adaptability, realizing millisecond-level rapid response and optimizing maintenance decision closed-loop management. BRIEF DESCRIPTION OF DRAWINGS

[0015] The drawings described herein are used to provide further understanding of the present application, and form a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application, and do not constitute an improper limitation on the present application. In the drawings: Figure 1 The flowchart of the embodiment of the present application is shown. DETAILED DESCRIPTION

[0016] The embodiments of the present application will be described in detail below with reference to the drawings and examples, so that the implementation process of the present application can be fully understood and implemented by applying technical means to solve technical problems and achieve technical effects.

[0017] In the prior art, as a key device of power distribution network, the running state of pole-mounted circuit breaker directly affects the reliability of power supply. The traditional state diagnosis mainly relies on planned maintenance and manual inspection, which has the problems of low efficiency and strong lag. In recent years, intelligent diagnosis methods have been gradually applied, but there are technical bottlenecks such as single data dimension, insufficient model adaptability and slow response speed. The existing methods are mostly based on static operation data, lack of dynamic time sequence features, and are difficult to capture transient information of mechanical failure; the mainstream algorithm does not consider the topological correlation of power distribution network, resulting in high false negative rate of a small number of faults; the cloud centralized processing mode has a delay of minutes, which cannot meet the millisecond-level fast isolation demand of circuit breaker fault. For example, a local power outage accident was caused by the failure of a certain power distribution station to timely identify the mechanical jamming fault of the circuit breaker.

[0018] In order to solve the above problems, the inventors found that there are three core contradictions in the prior art: the contradiction between the lack of dynamic features and the demand for capturing transient faults, the contradiction between the lack of topological correlation and the demand for identifying complex faults, and the contradiction between the delay of cloud processing and the demand for real-time response. Through analysis, it is found that the time shift of current waveform and the frequency energy distribution of vibration signal are the key indicators reflecting transient fault, and the load correlation between nodes of power distribution network may affect the fault propagation path. Based on this, feature pre-computation is carried out on the edge side to reduce transmission delay, and a double-channel model integrating time sequence and topological relationship is constructed in the cloud to form a closed-loop diagnosis system of cloud-edge collaboration.

[0019] Therefore, the application provides a pole-mounted circuit breaker diagnosis method based on multi-modal time sequence feature fusion and edge-cloud collaboration, which refers to Figure 1 , and includes the following steps: installing multi-source sensors on the pole-mounted circuit breaker, including a current sensor, a three-axis accelerometer, and a salt density sensor, and collecting electrical signals, vibration signals, and environmental corrosion data in real time; calculating a dynamic time warping distance and a wavelet packet energy entropy in an edge computing unit, and constructing an edge feature vector; inputting the edge feature vector and power distribution network topology data into a cloud double-channel deep learning model, wherein a ConvLSTM is used to process original waveforms in a time sequence channel, and a GNN is used to process network structure relationships in a topology channel; generating a maintenance strategy according to a state probability distribution output by the model, and triggering real-time alarms when emergency features appear.

[0020] Specifically, Step 1: multi-modal data acquisition and preprocessing (edge layer) Electrical signal acquisition: a high-precision current sensor (sampling rate 10 kHz) is installed on the secondary side of the circuit breaker current transformer to collect arc current waveform during opening , and the closing time and opening time of each operation are recorded The effective value of the current in the opening time is calculated according to the arc current waveform, and the calculation formula is as follows: ; T is the opening time; and the closing inrush peak value, ; Meanwhile, vibration signals are collected, and a three-axis accelerometer (sampling rate 20 kHz) is installed on the operating mechanism shell of the circuit breaker to collect vibration signals during closing and opening , , The total vibration energy is calculated ; is the vibration duration. , , is the three-axis acceleration value.

[0021] The time-domain kurtosis is also calculated ; wherein N : the number of sampling points for a single operation; is the acceleration value of the kth sampling point; is the acceleration mean, is the acceleration standard deviation.

[0022] The salt density sensor is installed on the surface of the circuit breaker shell, and the equivalent corrosion rate is calculated according to the data collected by the salt density sensor, and the formula is as follows: ; wherein is the ambient temperature. is the salt content in the air.

[0023] The dynamic time warping distance and wavelet packet energy entropy are calculated in the edge computing unit to construct an edge feature vector, wherein the dynamic time warping distance is calculated as follows: The current arc current waveform is aligned with the standard waveform , the minimum path distance is calculated, and the calculation formula is as follows: ; wherein is the set of all warping paths, and K is the path length; Then, the wavelet packet energy entropy is extracted, the vibration signal is decomposed into 8 sub-band energies by 3-layer wavelet packet decomposition, and the energy entropy is calculated as follows: wherein . is the energy proportion of the jth frequency band.

[0024] The feature vector is formed according to the above content: The feature vector is used for real-time alarm in the edge layer. In addition, the time sequence channel is used as channel one: the time sequence channel inputs the arc current waveform / vibration waveform, and directly obtains the arc current waveform, z-axis vibration waveform and derivative of the z-axis vibration waveform from the edge device, that is, the input tensor of channel one is as follows: ; The input vector at time t contains the current value, z-axis acceleration and its derivative; The topological feature is used as channel two: The feature vector includes: wherein the upstream load rate and the downstream load rate are first obtained by summing the active power of all loads downstream of the circuit breaker through the SCADA system or the smart meter to obtain , then the rated capacity S of the circuit breaker is read from the power grid database, the upstream active power and the rated capacity of the upstream line are obtained through the upper transformer substation or line monitoring point, and the upstream load rate and the downstream load rate are calculated as follows: ; ; then the load rate is normalized to the interval [0, 1] for model processing: The adjacency matrix represents the topological connection relationship of the distribution network.

[0025] According to the time sequence feature of channel one and the topological feature , they are spliced according to the weight, and the formula is as follows: ; λ is a trainable weight coefficient, and the initial value is 0.6. Then, a four-dimensional state probability is generated by inputting a full connection layer (FC Layer), and a loss function is established as follows: ; wherein ; ; : loss function value : batch size (number of samples) : number of classes (circuit breaker status level number, e.g. 4 classes) : weight of the c-th class (for handling sample imbalance) : true label of the i-th sample (1 if belonging to class c, 0 otherwise) : probability of the i-th sample belonging to class c predicted by the model

[0026] wherein the multi-source sensor refers to a combination of sensors capable of synchronously collecting electrical, mechanical and environmental heterogeneous data, which can be specifically implemented by a Hall effect current sensor, a piezoelectric three-axis accelerometer and a conductivity salt density sensor to obtain multi-dimensional information of the circuit breaker operating state. The dynamic time warping distance refers to a similarity measurement method for eliminating the time axis shift of time series signals, which can be specifically implemented by calculating the minimum path distance between the measured waveform and the standard waveform through a dynamic programming algorithm to capture the abnormal deformation characteristics of the current waveform. The wavelet packet energy entropy refers to a complexity index of signal frequency energy distribution, which can be specifically implemented by three-layer decomposition using db4 wavelet basis and calculating the energy entropy of each sub-band to quantify the frequency domain feature abnormality of the vibration signal. The dual-channel deep learning model refers to a neural network architecture for parallel processing of time series signals and topological relationships, which can be specifically implemented by using ConvLSTM network to extract waveform time series features and graph neural network to aggregate node topological information, for fusing spatio-temporal features to improve fault recognition accuracy.

[0027] Specifically, the current sensor captures the arc current waveform during the circuit breaker breaking process, the three-axis accelerometer collects the mechanical vibration signal, and the salt density sensor monitors the salt density on the surface of the insulator. The edge computing unit performs dynamic time warping calculation on the current waveform to eliminate the waveform offset caused by the operation time difference; and performs wavelet packet decomposition on the vibration signal to extract the energy distribution characteristics of each frequency band. The preprocessed edge feature vector and the power grid topology data are jointly input into the cloud model, the time series channel captures the time series correlation of the current and vibration signals through the ConvLSTM layer, and the topological channel analyzes the structural relationship of the circuit breaker in the power grid through the GNN layer. The state probability distribution output by the model triggers the hierarchical maintenance strategy, and when the dynamic time warping distance or the wavelet packet energy entropy exceeds the preset threshold, the edge unit directly triggers the circuit breaker lockout.

[0028] Compared with the prior art, the traditional method only uses a single current feature and static topology data, while the present scheme simultaneously captures electrical, mechanical and environmental corrosion features through multi-modal data fusion, and eliminates time offset interference by combining dynamic time warping technology. The existing model independently processes device state and power grid topology, while the present scheme realizes spatio-temporal feature fusion through a dual-channel model, enhancing the ability to identify associated faults. The existing cloud-based centralized processing mode has a minute-level delay, while the present scheme shortens the response time of critical faults to the millisecond level through edge feature extraction and emergency alarm mechanisms.

[0029] Through the above technical solutions, the present application effectively solves the problem of insufficient transient fault capture ability caused by single data dimension, improves the information completeness of fault features through multi-modal feature fusion; overcomes the problem of false negatives of minority class faults caused by the model not considering topology correlation, and captures the fault propagation relationship between devices using a graph neural network; improves the defect of excessive delay in cloud processing, and through the collaborative mechanism of edge real-time calculation and cloud deep analysis, both diagnostic accuracy and response speed are considered.

[0030] The present application further proposes capturing the arc current waveform during the breaking process according to the current sensor, and calculating the current effective value and breaking time.

[0031] The arc current waveform refers to the dynamic current change curve generated when the circuit breaker contacts separate, which can be realized by a Hall effect sensor combined with a high-speed sampling circuit, for example, the sampling frequency can be set to 10 kHz to record the waveform details completely. The current effective value refers to the thermal effect equivalent value of alternating current, which can be calculated by integrating the root mean square value within the waveform period to quantify the change of contact resistance. The breaking time refers to the time interval from when the circuit breaker receives the tripping instruction to when the arc is completely extinguished, which can be detected by a digital signal processor for zero crossing and time stamping, for example, the accuracy can reach the level of 0.1 ms.

[0032] Specifically, when the circuit breaker performs a tripping operation, the current sensor collects real-time transient waveform data of the arc current. The current effective value calculation module divides the waveform into cycles and performs integration operation in each power frequency cycle to obtain a quantitative index representing the degree of contact ablation. At the same time, the time measurement module detects the duration from when the current waveform drops from the rated value to zero to obtain the breaking time parameter reflecting the action speed of the operating mechanism. When the contact ablation causes the contact resistance to increase, the effective value calculation result will show a gradual upward trend; if the operating mechanism has a stuck defect, the breaking time will significantly exceed the standard threshold range.

[0033] Compared with the prior art, the traditional method only monitors the cumulative breaking number or static resistance value of the circuit breaker, and cannot capture the transient characteristic changes caused by contact ablation and mechanical wear during the breaking process. The scheme realizes the synchronous quantitative evaluation of the electrical performance and mechanical state of the circuit breaker through dynamic waveform analysis.

[0034] Through the above technical scheme, the application solves the problem of insufficient capture of fault transient information caused by the lack of dynamic current characteristics, can extract the time domain characteristics of the arc current waveform, and can identify early fault signs such as abnormal contact ablation and actuator action delay during the breaking process of the circuit breaker, avoiding the lag defect of traditional static parameter monitoring.

[0035] The application further proposes to calculate the total energy according to the collected vibration signal, and perform 3-layer db4 wavelet packet decomposition on the signal to extract 8 sub-band energy entropies.

[0036] The total energy refers to the overall energy integral value of the vibration signal in the time domain, which can be realized by using a numerical integral algorithm, and is used to represent the overall intensity level of the vibration signal. The 3-layer db4 wavelet packet decomposition refers to the use of Daubechies4 wavelet basis function to perform three-layer recursive decomposition on the original vibration signal, which can be realized by using a filter bank. Each layer of decomposition divides the signal into a low-frequency approximation component and a high-frequency detail component. The 8 sub-band energy entropies refer to the calculation of the information entropy value of the energy distribution of the 8 frequency bands generated by the third layer of decomposition, which can be calculated using the Shannon entropy formula, and is used to quantify the complexity difference of energy distribution in different frequency bands.

[0037] Specifically, after the vibration signal is collected by a three-axis accelerometer, the total energy is first calculated to obtain the overall intensity index of the signal. Then, the db4 wavelet basis function is used for three-layer decomposition, the first layer divides the signal into two frequency bands of 0-12.5kHz and 12.5-25kHz, the second layer divides each sub-band again, and the third layer continues to subdivide into 8 frequency bands. After each sub-band obtains the time domain waveform through the reconstruction algorithm, the energy proportion is calculated and input into the Shannon entropy formula, and finally an energy entropy feature vector containing 8 dimensions is formed. This decomposition method allows the transient vibration components in the high frequency band of 6.25-25kHz to be retained, while the traditional Fourier transform cannot effectively separate such non-stationary signal components.

[0038] Compared with the prior art, the traditional mechanical fault diagnosis mostly uses Fourier transform for frequency spectrum analysis, but this method cannot process non-stationary signals and has the problem of frequency spectrum leakage. The wavelet packet decomposition used in the scheme has the characteristics of time-frequency localization, which can effectively separate the high-frequency transient components generated by mechanical impact. At the same time, the energy entropy index can better reflect the complexity change of the energy distribution of the vibration signal in the time-frequency domain than the simple frequency band energy proportion.

[0039] Through the above technical solution, this application can effectively identify non-stationary vibration characteristics caused by mechanical faults such as jamming of the circuit breaker operating mechanism and spring fatigue. By finely dividing the high-frequency transient components and quantifying the energy distribution, the problem of transient feature loss caused by insufficient frequency domain resolution in traditional methods is solved, thereby improving the accuracy of early detection of mechanical faults.

[0040] This application further proposes to calculate the equivalent corrosion rate based on the salt density detected by the salt density sensor and the ambient temperature, and to update and input the data into the cloud model at certain time intervals.

[0041] Salt density detection refers to measuring the amount of salt deposited per unit area on the surface of an equipment using the conductivity method. Specifically, it can be achieved using a contact-type salt density sensor to quantify the degree of pollutant adhesion.

[0042] Among them, ambient temperature refers to the real-time temperature parameters of the environment in which the data acquisition device is located. Specifically, it can be achieved using a digital temperature sensor to correct the effect of temperature on the electrolyte ionization rate.

[0043] The calculation of equivalent corrosion rate refers to establishing a coupling relationship between salt density and ambient temperature. Specifically, it can be achieved using the Arrhenius modified model, which is used to convert the amount of salt deposition into a metal corrosion rate index.

[0044] Among them, time interval update refers to setting a periodic data upload mechanism, which can be implemented by timed triggering, such as performing data synchronization once per hour, to balance the consumption of communication resources and the real-time requirements of status monitoring.

[0045] Specifically, the salt density data output by the salt density sensor and the ambient temperature parameters collected by the temperature sensor are input into the corrosion assessment module. A dynamically changing corrosion rate value is calculated using a preset equivalent conversion formula. This rate value is encapsulated into a time-series data packet every hour and transmitted to the cloud server via a wireless communication module. After receiving the continuous corrosion state sequence, the cloud model performs trend analysis in conjunction with the circuit breaker's operating parameters. An early warning is triggered when a sudden change or sustained exceedance of the corrosion rate is detected.

[0046] Compared to existing technologies, traditional methods rely on quarterly artificial salt density tests and do not consider the impact of temperature fluctuations, leading to static biases in corrosion assessment. This proposed solution uses online sensors to simultaneously collect salt density and environmental parameters, and combines this with a dynamic correction model to generate an equivalent corrosion rate, accurately reflecting the actual corrosion process under different seasons and climatic conditions.

[0047] By the technical solution, the application realizes online monitoring and dynamic evaluation of the corrosion state of the surface of the equipment, solves the technical defect that the traditional static detection method cannot capture the influence of dynamic changes of environmental factors on the corrosion process, and provides continuous and reliable data support for prediction of the corrosion degradation trend of the circuit breaker.

[0048] The application further proposes a method for calculating a dynamic time warping distance in edge feature extraction, which dynamically time warps the measured arc current waveform and the standard waveform, and solves the minimum path distance DTW.

[0049] The dynamic time warping refers to eliminating the time axis difference between the measured waveform and the standard waveform through a nonlinear alignment method, which can be specifically implemented by using a dynamic programming algorithm, and allows elastic matching of waveform segments with different speeds. This feature solves the similarity misjudgment problem caused by the time offset of the traditional Euclidean distance.

[0050] The minimum path distance DTW refers to calculating the total cost of the optimal alignment path through a cumulative distance matrix, which can be specifically implemented by using a recursive search or a constraint window optimization algorithm, and is used to quantify the waveform shape difference. This feature can capture local fluctuation anomalies of the arc current, such as transient signals of arc reignition or premature extinction.

[0051] Specifically, the time axis length of the measured arc current waveform and the standard waveform may not be consistent, and there may be local time stretching or compression. Through dynamic time warping, the waveform segments are elastically mapped in the time dimension, for example, the current rising segment of the measured waveform is non-rigidly aligned with the corresponding segment of the standard waveform. Then, the cumulative distance of all possible alignment paths is calculated, and the minimum path is selected as the DTW value. For example, when the circuit breaker has a mechanical jamming fault, the arc duration of the measured waveform is extended, and the DTW calculation can identify that the minimum path distance of this segment of the waveform and the standard waveform significantly increases, thereby extracting the fault feature.

[0052] Compared with the prior art, the traditional method uses the Euclidean distance calculation of the fixed time window, which cannot handle the nonlinear deformation of the waveform time axis, resulting in that the transient features such as arc reignition are misjudged as noise. However, the present scheme can accurately align the key feature points of the waveform, such as the arc start time and the extinction time, by using the dynamic time warping algorithm, thereby eliminating the influence of time offset on feature extraction.

[0053] By the technical solution, the application realizes dynamic time sequence feature analysis of the arc current waveform during the breaking process of the circuit breaker, solves the transient fault missed detection problem caused by time sequence alignment error of the existing diagnosis method, and can accurately capture the arc abnormal fluctuation feature, thereby providing reliable time sequence feature basis for fault diagnosis.

[0054] The application further proposes inputting the arc current waveform and the down-sampled vibration signal in the time sequence channel of the dual-channel deep learning model, extracting the time sequence features through two layers of ConvLSTM, and connecting a maximum pooling layer.

[0055] The ConvLSTM refers to a structure of embedding a convolution operation in a long short-term memory neural network, which can be specifically implemented by processing time sequence data through a sliding window convolution kernel, extracting local waveform features through the convolution kernel, and modeling time sequence dependency relationships through the memory unit. The down-sampled vibration signal refers to the reduction of the sampling of the original high-frequency vibration data, which can be specifically implemented by retaining one data point every five sampling points, reducing the data dimension to adapt to the processing capability of the edge computing unit. The maximum pooling layer refers to an operation of spatially compressing the feature map, which can be specifically implemented by extracting the regional maximum value through a 2×2 sliding window, and reducing data redundancy by retaining significant features.

[0056] Specifically, the arc current waveform is input into the time sequence channel after being preprocessed by the edge computing unit, directly reflecting the abnormal fluctuation of the current in the breaking process of the circuit breaker. The vibration signal is down-sampled on the edge side, reducing the sampling frequency from 10 kHz to 2 kHz, and reducing the amount of data transmitted to the cloud. Two layers of ConvLSTM network process two types of time sequence data, the first layer of ConvLSTM extracts local waveform features through a 3×1 convolution kernel, and the second layer of ConvLSTM establishes a cross-time step correlation. The output time sequence features are spatially compressed through the maximum pooling layer to eliminate noise interference and retain key feature vectors, and finally generate time sequence feature expressions with a dimension of 1 / 4 of the original data.

[0057] Compared with the prior art, the traditional method cannot effectively capture local waveform features when processing time sequence data with a single LSTM, and it is difficult to model long time sequence dependency relationships by simply using CNN. The present scheme fuses the advantages of convolution and recurrent neural network through ConvLSTM, identifies abnormal current spikes with a duration of less than 10 ms in the arc current waveform, and detects intermittent high-frequency vibration patterns caused by mechanical component loosening in the vibration signal. Direct transmission of the original vibration signal in the prior art causes a cloud processing delay of more than 500 ms, while the present scheme reduces the data transmission amount by 80% through edge side down-sampling, shortening the cloud model processing time to within 200 ms.

[0058] By the technical solution, efficient extraction of dynamic timing characteristics is realized, data dimension is reduced under the premise of retaining transient fault characteristics, and the processing delay problem caused by data redundancy of a traditional method is solved. Through joint modeling of local waveform features and long timing correlation by the ConvLSTM network, the arc duration anomaly caused by the contact wear of the circuit breaker and the vibration energy mutation caused by the mechanism jam are accurately captured, and the transient fault detection rate is increased from 72% of the traditional LSTM to 89%.

[0059] The application further proposes the construction of a topology channel, which includes encoding the power distribution network structure into an adjacency matrix, the node feature vector containing the circuit breaker load rate and the corrosion rate, and aggregating neighbor node information through a two-layer graph neural network to output a topology embedding vector.

[0060] The adjacency matrix refers to a two-dimensional matrix used to represent the connection relationship between devices in the power distribution network, and can be implemented by using binary values to represent the electrical connection state between devices. The matrix can accurately reflect the topology correlation between the circuit breaker and the adjacent devices.

[0061] The node feature vector refers to a multi-dimensional data set describing the state of a single device in the power distribution network, which can be implemented by combining the real-time load rate of the circuit breaker and the environmental corrosion rate into a vector. The feature vector can reflect both the device operating condition and the environmental influence factors.

[0062] The two-layer graph neural network aggregation refers to integrating the feature information of adjacent nodes through two message passing processes, which can be implemented by using graph convolution operations to calculate the weighted average of neighbor node features. This process can capture the fault propagation path and correlation mode between devices.

[0063] Specifically, after the power distribution network structure is converted into an adjacency matrix, each circuit breaker node is assigned a feature vector containing the load rate and the corrosion rate. In the topology channel, the first layer of graph neural network performs aggregation operation on the features of adjacent nodes to extract primary topology correlation features, and the second layer of graph neural network further integrates cross-layer device information based on the previous aggregation to generate a topology embedding vector reflecting the network-level fault propagation characteristics. In this way, the model can learn the nonlinear relationship between the state change of the circuit breaker and the operating parameters of the adjacent devices, thereby identifying the composite fault mode caused by topology correlation.

[0064] Compared with the prior art, the traditional method is only based on the isolated data of a single circuit breaker for diagnosis, and cannot capture the fault correlation characteristics between devices. The present scheme, through graph structure modeling and multi-layer feature aggregation, enables the model to analyze the implicit causal relationship chain in the power distribution network, such as early warning of the mechanism jam risk caused by overload of the current device when the load rate of the adjacent circuit breaker abnormally increases, thereby overcoming the adaptability problem caused by isolated analysis.

[0065] Through the above technical solution, this application achieves effective modeling of the topological correlation of the distribution network, enabling the diagnostic model to identify cascading failure modes caused by network structure relationships, improving the detection sensitivity of a few types of faults such as mechanical jamming, and reducing the false negatives caused by ignoring the interaction between devices.

[0066] This application further proposes maintenance strategy generation rules, when the cloud model outputs state probabilities. satisfy A value ≥0.4 generates a replacement instruction within 48 hours; a value ≤0.2 generates a replacement instruction within 48 hours. When the value is less than 0.4, a maintenance instruction is generated within 7 days. If the value is less than 0.2, it is marked as normal monitoring.

[0067] Here, the state probability p refers to the four-dimensional probability vector output by the cloud-based dual-channel deep learning model. Specifically, it can be achieved by normalizing the output of the last layer of the model using the Softmax function. Each dimension corresponds to the probability that the circuit breaker is in different deterioration levels.

[0068] in, This refers to the probability value of a circuit breaker being at the highest risk level. Specifically, it can be achieved through a comprehensive analysis of features such as arc current distortion and abnormal vibration energy using a model. This parameter is used to characterize the confidence level of a serious equipment failure.

[0069] Among them, the thresholds of 0.4 and 0.2 refer to the trigger boundaries of the maintenance strategy. Specifically, they can be determined by combining historical fault data statistics with engineering experience. Risk classification and control can be achieved by setting response levels for different intervals.

[0070] Specifically, when the cloud model outputs When the probability value exceeds 0.4, it indicates a serious fault risk in the circuit breaker, such as mechanical jamming or insulation breakdown. In this case, a replacement instruction is automatically generated within 48 hours, forcibly triggering the equipment replacement process. When the value is in the range of 0.2 to 0.4, corresponding to repairable defects such as contact wear or spring fatigue, the system generates a maintenance instruction within 7 days and pushes it to the maintenance terminal. When When the value is below 0.2, maintain normal monitoring status, only record data and do not trigger maintenance actions. This rule quantifies the equipment status through probability distribution, replacing traditional manual experience judgment, and enables dynamic adjustment of maintenance strategies.

[0071] Compared to existing technologies, traditional methods rely on fixed-cycle maintenance or manual on-site assessments, which cannot respond to changes in equipment status in real time and result in wasted maintenance resources. This solution classifies maintenance levels using probability thresholds, prioritizing the handling of high-risk equipment while avoiding unnecessary downtime for low-risk equipment.

[0072] Through the technical solution, the application realizes automatic generation and hierarchical response of the circuit breaker maintenance strategy, shortens the response time required by traditional manual decision from hours to milliseconds of model reasoning completion, and concentrates maintenance resources on high-risk equipment through the probability threshold control, so that the number of repair work orders is reduced by about 30% compared with the fixed period maintenance mode.

[0073] The application further proposes to set a real-time alarm condition in the edge computing unit, and when the dynamic time warping distance exceeds 0.3 or the wavelet packet energy entropy exceeds 2.5, the circuit breaker is immediately locked and an alarm signal is uploaded to the monitoring center.

[0074] The dynamic time warping distance refers to the similarity between the measured arc current waveform and the standard waveform calculated by the dynamic time warping algorithm, and can be specifically realized by solving the minimum path distance by the dynamic programming method. This index is used to quantify the abnormality of the current waveform. The wavelet packet energy entropy refers to the entropy value of the energy distribution of each sub-band after the vibration signal is decomposed by the wavelet packet. Specifically, it can be realized by calculating the energy entropy of 8 sub-bands after 3-layer decomposition using db4 wavelet basis. This index is used to represent the abnormal state of the energy distribution of the vibration signal.

[0075] Specifically, the dynamic time warping distance and the wavelet packet energy entropy are calculated in real time in the edge computing unit. When any index exceeds the preset threshold, the edge computing unit directly sends a locking instruction to the circuit breaker and uploads an alarm signal to the monitoring center. This process is completely completed locally and does not need to wait for the diagnosis result of the cloud server, thereby avoiding the time delay caused by network transmission and cloud computing. The threshold of the dynamic time warping distance is set to 0.3, which can capture significant deviations of the current waveform from the standard mode. The threshold of the wavelet packet energy entropy is set to 2.5, which can identify the abnormal concentration phenomenon of the energy distribution in the vibration signal.

[0076] Compared with the prior art, the traditional method relies on cloud centralized processing of all feature data, resulting in a fault response time of minutes. The present scheme preinstalls a key feature threshold judgment mechanism in the edge computing unit, decentralizes the decision-making right of the circuit breaker locking operation to the local device, and shortens the execution time of fault isolation action to milliseconds.

[0077] Through the technical solution, the application realizes localized and rapid response of the circuit breaker fault. When the current waveform distortion or vibration energy anomaly is detected, the circuit breaker locking operation can be completed within 20 milliseconds, effectively preventing the expansion of the fault caused by action delay. At the same time, the parallel execution of the alarm signal and the locking operation ensures that the monitoring center can obtain real-time device state change information.

[0078] The application further provides a pole-mounted circuit breaker diagnosis system based on multi-modal time sequence feature fusion and edge-cloud collaboration, comprising a sensor module, an edge computing unit, a cloud server and a decision execution module. The sensor module is composed of a current sensor, a three-axis accelerometer and a salt density sensor; the edge computing unit uses an FPGA chip to calculate the dynamic time warping distance, wavelet packet energy entropy and equivalent corrosion rate in real time; the cloud server deploys a double-channel deep learning model for daily deep diagnosis; and the decision execution module triggers a circuit breaker operation or a maintenance work order according to the cloud instruction or the edge alarm.

[0079] The FPGA chip refers to a programmable gate array integrated circuit, which can be implemented by using an Xilinx Artix-7 series chip, and the parallel computing architecture of the FPGA chip can meet the real-time calculation requirement of the dynamic time warping distance, and the feature extraction time delay is shortened through hardware acceleration. The double-channel deep learning model refers to a neural network structure comprising a time sequence channel and a topology channel, which can be implemented by using a ConvLSTM layer to process the current waveform time sequence feature, and a graph neural network layer to aggregate the power grid node information, so as to realize joint modeling of the device state and the power grid operating environment. The equivalent corrosion rate refers to a corrosion degree quantitative index calculated based on the salt density and the environmental temperature, which can be implemented by using the linear polarization resistance method in the ASTM G1 standard, and is used to reflect the actual corrosion process of the metal parts of the circuit breaker.

[0080] Specifically, the current signal collected by the sensor module is subjected to dynamic time warping distance calculation by the edge computing unit, the measured arc current waveform is dynamically aligned with the standard waveform, and a waveform similarity index is generated. After the vibration signal is subjected to three-layer wavelet packet decomposition, the energy entropy values of each sub-band are calculated, which are used to represent abnormal vibration patterns of mechanical parts. The salt density sensor data are combined with the temperature parameter to calculate the equivalent corrosion rate, thereby forming the environmental corrosion feature. After the above features are processed in real time by the edge computing unit, the power grid topology data issued by the cloud are input into the double-channel model, the time sequence channel extracts the space-time features of the current and vibration signals by using a convolutional long short-term memory network, the topology channel aggregates the load and corrosion information of adjacent nodes by using a graph neural network, and finally the two types of features are fused to generate a device state probability distribution. The decision execution module triggers a maintenance work order according to the state probability threshold, and directly executes a circuit breaker locking operation when abnormal features are detected at the edge.

[0081] Compared with the prior art, the traditional system uses a general-purpose processor for feature calculation, resulting in a delay of more than one second in processing complex algorithms such as dynamic time warping distance, while the present scheme compresses the feature calculation time to the millisecond level through FPGA hardware acceleration. The existing diagnostic model only relies on single device data and cannot associate the power distribution network topology relationship. The present scheme synchronously processes the device time sequence features and power grid structure features through a dual-channel model, improving the recognition ability of cascading faults in the power grid. The conventional maintenance strategy only relies on periodic cloud analysis. The present scheme realizes multi-time scale coverage of fault response through the cooperative mechanism of edge real-time alarm and periodic cloud diagnosis.

[0082] Through the above technical scheme, the present application solves the problem of minute-level response delay caused by cloud centralized processing in the traditional diagnostic method. The hardware acceleration of the edge computing unit enables anomaly detection to be completed within 20 milliseconds, meeting the requirement of fast isolation of circuit breakers. The multi-modal data fusion mechanism overcomes the limitations of single electrical parameter analysis, and through the joint calculation of vibration energy entropy and equivalent corrosion rate, accurately identifies the composite fault of mechanical wear and environmental corrosion. The structural design of the dual-channel model eliminates the defects of traditional algorithms that ignore the correlation of the power grid topology, and aggregates the state data of adjacent nodes through a graph neural network, so that the diagnostic result can reflect the true health degree of the device in the power grid operating environment.

[0083] As certain terms have been used in the specification and claims, those of ordinary skill in the art will understand that hardware manufacturers can refer to the same component by different names. The specification and claims do not distinguish components based on the difference in name, but rather on the difference in function. As used throughout the specification and claims, "comprise" is an open term that should be interpreted as "comprising but not limited to". "Substantially" means within an acceptable error range, and those skilled in the art can solve the technical problems within a certain error range and basically achieve the technical effects.

[0084] The above description shows and describes several preferred embodiments of the present application, but as mentioned above, it should be understood that the present application is not limited to the forms disclosed herein, and should not be considered as excluding other embodiments, but can be used in various other combinations, modifications and environments, and can be modified by the above teachings or related art or knowledge in the field. Any modification and change made by those skilled in the art without departing from the spirit and scope of the present application shall be within the protection scope of the claims of the present application.

Claims

1. A pole-mounted circuit breaker diagnosis method based on multi-modal temporal feature fusion and edge-cloud collaboration, characterized in that, Comprising the following steps: S100, install multi-source sensors on the circuit breaker, including current sensors, three-axis accelerometers, and salt density sensors, to collect electrical signals, vibration signals, and environmental corrosion data in real time; S200, calculate dynamic time warping distance and wavelet packet energy entropy in the edge computing unit to construct an edge feature vector; S300, input the edge feature vector and power grid topology data into a cloud-based dual-channel deep learning model, where the time series channel uses ConvLSTM to process the original waveform, and the topology channel uses GNN to process the network structure relationship; S400, generate a maintenance strategy based on the state probability distribution output by the model, and trigger real-time alarms when emergency features appear.

2. The pole-mounted circuit breaker diagnostic method of claim 1, wherein, According to the current sensor, the arc current waveform during the breaking process is captured, and the current effective value and breaking time are calculated.

3. The pole-mounted circuit breaker diagnostic method of claim 1, wherein, According to the collected vibration signal, the total energy is calculated, and the signal is decomposed into 3 layers of db4 wavelet packet to extract 8 sub-band energy entropies.

4. The pole-mounted circuit breaker diagnostic method of claim 1, wherein, According to the salt density detected by the salt density sensor, the equivalent corrosion rate is calculated combined with the environmental temperature, and is updated and input into the cloud model at certain time intervals.

5. The pole-mounted circuit breaker diagnostic method of claim 1, wherein, The method for calculating dynamic time warping distance in edge feature extraction is as follows: the measured arc current waveform is dynamically adjusted with the standard waveform, and the minimum path distance DTW is solved.

6. The pole-mounted circuit breaker diagnostic method of claim 1, wherein, In the time series channel of the dual-channel deep learning model: input the arc current waveform and down-sampled vibration signal, extract the time series features through two layers of ConvLSTM, and then connect the maximum pooling layer.

7. The pole-mounted circuit breaker diagnostic method of claim 1, wherein, The construction of the topology channel includes: The power grid structure is encoded as an adjacency matrix A, and the node feature vector contains the breaker load rate and corrosion rate. The neighbor node information is aggregated by GNN to output the topology embedding vector.

8. The pole-mounted circuit breaker diagnostic method of claim 1, wherein, The maintenance strategy generation rule is When the cloud model outputs a state probability satisfies: ≥ 0.4, generate a "replace within 48 hours" instruction; 0.2≤ When < 0.4, generate "7-day service" instruction; <0.2, labeled "normal monitoring." 9. The pole-mounted circuit breaker diagnostic method of claim 1, wherein, The real-time alarm condition is: In the edge computing unit, if DTW > 0.3 or WPEE > 2.5 is detected, the circuit breaker is immediately locked and the alarm signal is uploaded to the monitoring center.

10. A system for implementing the pole-mounted circuit breaker diagnosis method with multi-modal time-series feature fusion and edge-cloud collaboration according to any one of claims 1-9, characterized in that, Including: Sensor module: current sensor, three-axis accelerometer, salt density sensor; Edge computing unit: FPGA chip, used for real-time calculation of DTW, WPEE and Ev ; Cloud server: deploy dual-channel deep learning model, perform deep diagnosis daily; Decision execution module: trigger circuit breaker operation or maintenance work order according to cloud instructions or edge alarms.

Images