AI-based isolation switch health state management method and system

Abstract

The invention relates to an AI-based isolation switch health state management method and system. The method comprises the following steps: carrying out multi-source data sampling on operation sound, equipment vibration, contact temperature and operation current in the operation process of the isolating switch, and carrying out signal preprocessing to obtain a multi-source characteristic data set; performing time domain parameter calculation, frequency domain component extraction and time-frequency energy decomposition on the multi-source feature data set, and performing polynomial fitting optimization processing to obtain a feature vector sequence; constructing a time sequence dynamic graph network for the feature vector sequence, generating graph nodes based on the time relevance of feature vectors, and constructing directed connection edges according to the similarity between the nodes to obtain a dynamic topological graph; and inputting the dynamic topological graph into a depth map neural network, and performing state feature extraction through multi-layer graph convolution operation, feature pooling compression and full-connection mapping to obtain an equipment state classification result. According to the invention, the method achieves the precise recognition of an abnormal state, and is higher in anti-interference capability.

Description

Technical Field

[0001] The present invention relates to the technical field of isolating switches, and in particular to an AI-based isolating switch health status management method and system. Background Art

[0002] As a key device in the power system, the operating status of the disconnector is directly related to the safe and stable operation of the power grid. With the continuous expansion of the scale of the power grid and the improvement of the level of intelligence, the traditional manual inspection method can no longer meet the health status monitoring needs of large-scale disconnector groups. Although sensors have begun to be used to collect equipment operation data for status monitoring, the multi-source signals such as sound, vibration, temperature, etc. generated during the operation of the disconnector have strong coupling and nonlinear characteristics, which makes it difficult to extract status features and identify faults.

[0003] There are three main problems with the existing disconnector status monitoring methods: the collection and preprocessing of multi-source sensor data lack a unified standard, and different types of signal processing methods are different, making it difficult to achieve effective data fusion; secondly, traditional feature extraction methods often use fixed feature templates, which cannot adaptively capture the dynamic feature changes of disconnectors under different operating conditions; most of the existing state recognition algorithms are based on shallow models, which are prone to underfitting when processing high-dimensional nonlinear data, and it is difficult to mine deep correlations between features. Summary of the invention

[0004] The main purpose of the present invention is to provide an AI-based isolating switch health status management method and system, which realizes accurate identification of abnormal states and has strong anti-interference ability.

[0005] To achieve the above object, the present invention provides an AI-based isolating switch health status management method, comprising the following steps: Multi-source data sampling is performed on the operating sound, equipment vibration, contact temperature and operating current during the operation of the disconnector, and a multi-source feature data set is obtained after signal preprocessing; Performing time domain parameter calculation, frequency domain component extraction and time-frequency energy decomposition on the multi-source feature data set, and performing polynomial fitting optimization processing to obtain a feature vector sequence; Constructing a time-series dynamic graph network for the feature vector sequence, generating graph nodes based on the temporal correlation of the feature vectors, and constructing directed connection edges based on the similarity between the nodes to obtain a dynamic topology graph; The dynamic topology map is input into a deep graph neural network, and state features are extracted through multi-layer graph convolution operations, feature pooling compression and full connection mapping to obtain device state classification results.

[0006] The present invention also provides an AI-based isolating switch health status management system, comprising: The sampling module is used to sample multi-source data of operating sound, equipment vibration, contact temperature and operating current during the operation of the disconnector, and obtain a multi-source feature data set after signal preprocessing; A decomposition module, used to perform time domain parameter calculation, frequency domain component extraction and time-frequency energy decomposition on the multi-source feature data set, and obtain a feature vector sequence through polynomial fitting optimization processing; A construction module is used to construct a time-series dynamic graph network for the feature vector sequence, generate graph nodes based on the time correlation of the feature vectors, and construct directed connection edges based on the similarity between the nodes to obtain a dynamic topology graph; The extraction module is used to input the dynamic topology map into a deep graph neural network, extract state features through multi-layer graph convolution operations, feature pooling compression and full connection mapping, and obtain device state classification results.

[0007] The present invention also provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of any one of the above methods when executing the computer program.

[0008] The present invention also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of any of the above-mentioned methods are implemented.

[0009] In summary, the technical solution provided by the present invention effectively improves the quality of the original data by designing a multi-source data sampling and preprocessing scheme, adopts different sampling frequencies to collect various sensor signals, and performs noise reduction, filtering and standardization processing in a targeted manner, laying the foundation for subsequent feature extraction. The present invention proposes a feature optimization method based on polynomial fitting, and realizes adaptive smoothing of features by constructing a time series polynomial function model and performing least squares optimization, effectively reducing the impact of data noise on feature extraction. The present invention designs a dynamic graph network structure, generates graph nodes through the time correlation of feature vectors, and constructs directed connection edges based on the similarity between nodes, successfully capturing the temporal evolution law of the state characteristics of the disconnector. The present invention adopts a deep model architecture of a three-layer graph convolutional network and a three-layer fully connected neural network, combined with an attention mechanism and a maximum pooling operation, which significantly improves the model's learning ability for high-dimensional nonlinear features. The present invention proposes an anomaly detection method based on kernel density estimation, which realizes accurate identification of abnormal states by calculating the normal state feature distribution and optimizing the detection threshold, and has strong anti-interference ability. The present invention introduces a spatiotemporal feature analysis mechanism to comprehensively evaluate the duration and impact range of abnormal conditions, providing a more comprehensive equipment health status evaluation result. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 It is a schematic diagram of the steps of the isolating switch health status management method based on AI in one embodiment of the present invention; Figure 2 It is a structural block diagram of an AI-based isolating switch health status management system in one embodiment of the present invention; Figure 3 It is a schematic block diagram of the structure of a computer device according to an embodiment of the present invention.

[0011] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0012] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0013] Reference Figure 1 , this embodiment provides an AI-based isolating switch health status management method, comprising the following steps: S1, sampling multi-source data of operating sound, equipment vibration, contact temperature and operating current during the operation of the disconnector, and obtaining a multi-source feature data set after signal preprocessing; Among them, the frequency sampling of the operating sound of the disconnector is carried out to capture the sound wave information generated when the equipment is working. The frequency distribution of the sound signal can reflect the abnormal conditions inside the equipment, such as mechanical friction, electrical breakdown and other faults. In order to ensure the accuracy of sampling, a certain frequency range is selected for sampling, and the sampling frequency is set to 2kHz to 10kHz to ensure that the high-frequency components of the sound signal can be effectively captured. The collected sound data is subjected to wavelet threshold denoising operation, and the noise components of different frequency bands are separated from the effective signal by wavelet transform of the sound signal. Wavelet transform can provide multi-resolution analysis of signals in both time and frequency domains, and use threshold denoising method to suppress high-frequency noise, so that the final sound feature sequence is smoother and more accurate, reflecting a more realistic equipment operation status. In the process of collecting equipment vibration data, frequency sampling is carried out, and the vibration signal is an important indicator of the health status of the equipment. The mechanical part of the disconnector vibrates due to wear, looseness and other factors, and these vibration signals can reflect the state changes of the equipment. The vibration sampling frequency is 1kHz to 2kHz to ensure that sufficiently high-frequency signals are collected to reveal the operating status of the equipment. The acquired vibration data is subjected to bandpass filtering operation. Bandpass filtering can filter out low-frequency environmental noise and high-frequency irrelevant interference, and only retain the relevant frequency components of the equipment fault signal. After bandpass filtering, the obtained vibration feature sequence can accurately reflect the vibration characteristics of the equipment during operation. For the contact temperature data, the acquisition frequency is set to 1Hz, which is sufficient to meet the monitoring needs of temperature changes. Temperature is an important early warning indicator of equipment failure, especially the contact part of the disconnector. Long-term overheating can cause poor contact, arcing or equipment damage. The temperature sampling data will be affected by the ambient temperature. Ambient temperature compensation is performed during acquisition, and the change factor of ambient temperature is removed from the original temperature signal to obtain the temperature feature sequence. Regarding the sampling of operating current, the current signal is directly related to the workload and operating status of the equipment. The change of operating current helps the system determine whether the load condition of the disconnector is normal. In order to ensure the stability and reliability of the current data, frequency sampling is performed in combination with the load condition parameters of the equipment. The collected current data is processed through sliding window averaging operation to smooth current fluctuations, eliminate high-frequency noise, extract the main trend of the current signal, and avoid the impact of instantaneous current fluctuations on fault diagnosis. The obtained current feature sequence reflects the current variation law of the equipment under different loads. The sound feature sequence, vibration feature sequence, temperature feature sequence and current feature sequence are aligned by timestamp and data fused to obtain a multi-source feature data set. Through the aligned data set, the operating status of the equipment is analyzed on the same time scale to reveal the potential correlation between various parameters.

[0014] S2, performing time domain parameter calculation, frequency domain component extraction and time-frequency energy decomposition on the multi-source feature data set, and obtaining a feature vector sequence through polynomial fitting optimization processing; Specifically, the time domain parameters of the multi-source feature data set are calculated to extract the statistical characteristics of the signal in the time dimension, reflecting the basic changes in the operating status of the equipment. Time domain features include peak, variance, skewness, and kurtosis. The peak reveals the extreme fluctuations in the signal, the variance can describe the fluctuation range of the signal, the skewness provides the symmetry information of the signal, and the kurtosis can reflect the sharpness of the signal. These time domain features are helpful to evaluate the stability of equipment operation and determine whether there are sudden failures. By calculating the time domain parameters of each type of data in the multi-source feature data set (such as sound, vibration, temperature, and current), the time domain feature matrix is ​​obtained. The frequency domain components of the multi-source feature data set are extracted, and the spectrum analysis is performed using fast Fourier transform to extract the frequency components in the signal. Frequency domain analysis can effectively reveal the periodicity and frequency characteristics in the signal. For vibration signals, it can help detect whether the equipment has certain specific frequency failure modes. For example, mechanical failures will produce obvious signal components at certain specific frequencies. Through fast Fourier transform, the time domain signal is converted into the frequency domain signal, and the energy value of each frequency band is calculated to obtain the frequency domain feature matrix, which contains the energy distribution information of each frequency band in the signal, which helps to identify potential fault frequencies or abnormal fluctuations. At the same time, in order to improve the accuracy of the analysis, time-frequency energy decomposition is performed, and the signal is multi-resolution analyzed through wavelet packet multi-scale decomposition. Wavelet packet transform can perform fine-grained decomposition of the signal at different scales to obtain more time-frequency information. By calculating the energy coefficients of different frequency bands and their energy distribution ratios, the time-frequency feature matrix is ​​obtained, which reflects the common change law of the signal in time and frequency. The time domain feature matrix, frequency domain feature matrix and time-frequency feature matrix are concatenated and spliced ​​to combine the three different types of features into a multidimensional feature data set. A time series polynomial function model is constructed for the multidimensional feature data, and a fitting function is established by setting an n-order polynomial basis function and its initial parameters. The model captures the trend and law of feature data changing over time through polynomial fitting, and can effectively reduce the complexity of the data when processing high-dimensional data. In order to improve the fitting accuracy, the least squares optimization calculation is performed on the feature fitting function. The fitting error is minimized by iteratively updating the polynomial coefficients to obtain the optimized feature parameters. The fitting residual is evaluated for the optimization result. The accuracy of the feature fitting is evaluated by calculating the comparison result between the fitting error and the preset accuracy threshold. If the fitting error is lower than the set threshold, the feature optimization is considered successful; otherwise, the fitting accuracy is improved by adjusting the optimization process. The principal component dimensionality reduction transformation is performed on the optimized feature parameters. The high-dimensional data is mapped to a lower dimension through linear transformation, while retaining the main information in the data to obtain a feature vector sequence.

[0015] Effectively segment the multidimensional feature data. By setting a fixed time window to segment the data, each time window contains feature data of a certain time period, and obtains a feature segmentation sequence. Based on the feature segmentation sequence, a time series index mapping is established, and the feature data is associated with the timestamp to obtain a feature time correspondence table. According to the feature time correspondence table, an n-order polynomial fitting model is constructed, and the change trend of the feature data over time is described by selecting appropriate polynomial basis functions. Select an appropriate combination of polynomial basis functions, and construct the expression of the fitting function through the combination of these basis functions. The order n of the polynomial is a key parameter that determines the complexity and fitting accuracy of the model. Higher-order polynomials fit the complex changes of data more accurately, but at the same time lead to overfitting problems. Select a suitable order according to the specific application. After constructing the fitting function, set the initial value for the coefficient of the fitting function, and determine the iteration step size according to the characteristics of the data to obtain the optimized initial conditions. Optimize the parameters of the polynomial fitting function by the least squares method. The least squares method solves the optimal parameters of the model by minimizing the fitting error. In this process, the optimization problem is transformed into a square error minimization objective function, that is, the error between the fitting function and the actual data is calculated, and the model is optimized by minimizing this error. This objective function reflects the sum of the fitting errors. The smaller the error, the higher the fitting accuracy of the model. By introducing the least squares iterative solver, the polynomial coefficients are continuously adjusted so that the value of the objective function is gradually reduced to achieve the optimization effect. In the iterative process, the parameter gradient and update direction are calculated. The gradient is the rate of change of the objective function at a certain point in the parameter space, which indicates the sensitivity of the objective function to each parameter. By calculating the parameter gradient, it is determined how to adjust the parameter value to reduce the error. The parameters are updated by adaptive step size adjustment. The adaptive step size can automatically adjust the step size of each iteration according to the size of the gradient, thereby improving the efficiency and stability of the optimization. In each iteration, the polynomial coefficients are gradually updated according to the calculated gradient and update direction until the ideal fitting result is achieved. In order to ensure the stability and effectiveness of the optimization process, the convergence of the parameter iteration sequence is judged to check whether the parameter changes after multiple iterations are small enough to ensure that the optimization process has reached the optimal solution. The parameter change between adjacent iteration steps is calculated, that is, the difference in parameter values ​​in two consecutive iterations. When this change is less than a certain preset threshold, it means that the optimization process has converged and the parameters have stabilized. The convergence judgment index is calculated based on the change in the parameter or the change in the objective function value, and a convergence threshold is set to determine whether the convergence condition is met. When the convergence condition is met, it means that the optimization process of the polynomial coefficients has ended, and the obtained coefficients are the optimized characteristic parameters.

[0016] S3, construct a temporal dynamic graph network for the feature vector sequence, generate graph nodes based on the temporal correlation of the feature vectors, and construct directed connection edges based on the similarity between nodes to obtain a dynamic topology graph; It should be noted that the feature vector sequence is encoded with time point marking. By adding the identification information of the sampling time, a unique time mark is assigned to each feature vector, so that the feature vector is associated with its corresponding time point to obtain time series marking data. A graph node generation operation is performed based on the time series marking data. Each feature vector is mapped to a node in the network, and a unique identifier is assigned to it according to its corresponding timestamp to obtain a node data set. Each node represents the state information of the device at a specific time, and the connection relationship between the nodes reflects their association in the time series. The similarity between the nodes is calculated to determine whether they should be connected through directed edges. In order to measure the similarity between nodes, cosine similarity is used. Cosine similarity measures the similarity of two vectors by calculating the cosine value of the angle between the feature vectors. The value range is between -1 and 1. The larger the value, the more similar the two nodes are. After calculating the cosine similarity between the node pairs, a similarity matrix is ​​obtained, in which each item represents the similarity between the node pairs. By comparing the similarity matrix with the preset connection threshold, it is determined which nodes have a sufficiently high similarity, so as to establish a connection. For node pairs with similarity greater than the set threshold, a connection determination identifier is generated, that is, it is determined which nodes should establish edge connections. Based on the connection determination identifier, an edge connection relationship is constructed. For each pair of nodes with high similarity, a directed connection edge is established to reflect the temporal correlation and state similarity between the nodes. In order to enhance the temporal nature of the graph, a temporal directed assignment is performed on the edge connection set, and a direction is assigned to each connection edge according to the time sequence to ensure that the edges in the graph have a clear time flow direction, and a directed edge data set is obtained. The directed edge data set is merged with the node data set to construct the adjacency matrix of the graph. The adjacency matrix records the connection relationship between all nodes in the graph, represents the connection strength and topological structure between nodes, and obtains the initial network structure. The topological relationship of the initial network structure is optimized so that the network structure can more effectively reflect the real relationship between nodes, while improving the computational efficiency and expression ability of the graph structure. The density and topological characteristics of the graph are evaluated by calculating the node connectivity distribution, that is, the number or strength of connections between nodes and other nodes. If the connection of some nodes is too high or too low, it will lead to an unreasonable network structure. In this case, the network should be adjusted to optimize the network structure by increasing or decreasing the connection, and finally a dynamic topology diagram will be obtained.

[0017] S4, input the dynamic topology map into the deep graph neural network, extract the state features through multi-layer graph convolution operations, feature pooling compression and fully connected mapping, and obtain the device state classification result.

[0018] Specifically, the dynamic topology graph is input into the three-layer graph convolution network in the deep graph neural network for processing. The graph convolution network extracts graph features of different scales layer by layer by performing 64-dimensional, 128-dimensional, and 256-dimensional graph convolution kernel operations in sequence, thereby performing multi-level learning on the local and global features of the nodes in the graph. By continuously enhancing the dimension of the convolution layer, graph features of different scales can be captured, thereby effectively obtaining the structural information of the disconnector at different time points and in different states. Neighborhood information aggregation is performed on multi-scale graph features. In the graph neural network, the characteristics of a node are not only determined by its own attributes, but also affected by the characteristics of its neighboring nodes. When performing neighborhood information aggregation, the local structural features of each node are obtained by fusing the feature representations of each node and its neighboring nodes. Attention is allocated to the local structural features. A feature importance coefficient is assigned to each node by calculating the feature correlation score between nodes, that is, measuring the relative importance of the node feature in the entire graph. Based on the feature importance coefficient, weighted feature aggregation is performed. By performing a weighted feature fusion operation, the features of each node are weighted and synthesized according to their importance to obtain a node representation vector for each node. Perform a maximum pooling operation on the node representation vector to extract the most significant eigenvalue in the node representation vector and generate a graph representation feature that represents the global information of the entire graph. The graph representation feature is input into the three-layer fully connected neural network in the deep graph neural network for processing. The fully connected layer enables the network to capture more complex nonlinear relationships through the ReLU activation function. By performing a 512-dimensional ReLU mapping, and then gradually performing a 256-dimensional ReLU mapping and a 128-dimensional ReLU mapping, further compression and mapping of features are achieved through layer-by-layer mapping. Each layer of the fully connected neural network can effectively perform a nonlinear transformation on the previously extracted features, enhance the model's ability to understand the device status, and obtain a deep feature map. Perform a Softmax classification operation on the deep feature map, and assign a probability score to each possible device state by converting the output eigenvalue into a probability distribution. By calculating the normalized category probability score, the state probability distribution is obtained, which represents the probability of the device being in each health state at the current moment. According to the state probability distribution, the label with the maximum probability is extracted to obtain the state classification result of the device.

[0019] Perform statistical feature analysis on the state probability sequence in the equipment state classification results. Obtain the normal state feature distribution by calculating the probability mean and standard deviation distribution under normal operating conditions. Initialize the threshold of the normal state feature distribution. Set the initial monitoring benchmark threshold by calculating the μ-3σ (mean minus three times the standard deviation) limit value. Determine the maximum fluctuation range of the equipment under normal conditions through the limit value, and all state changes beyond this range indicate abnormalities. Perform abnormal pattern extraction on the historical records of equipment state classification results. Obtain an abnormal probability sample set by counting the probability distribution characteristics of known fault events, representing the state probability distribution of equipment failures in history. Perform kernel density distribution estimation on the abnormal probability sample set. Obtain a continuous abnormal distribution function by using the Gaussian kernel function to calculate the probability density of the samples, describing the continuous change process of the equipment state when it is abnormal. Perform curve feature analysis on the continuous abnormal distribution function. Identify the candidate threshold point set by calculating the slope change point of the curve, representing the critical moment when the abnormal state may occur. By judging whether the trend of state change has fluctuated drastically through the change of the slope, potential abnormal points can be identified. Perform classification performance test on the candidate threshold point set. The detection effect of different thresholds is evaluated by calculating indicators such as precision and recall. Precision measures the proportion of all detected abnormal states that are truly abnormal; while recall measures the proportion of all truly abnormal states that can be correctly detected. Through these indicators, the effectiveness and reliability of different thresholds in anomaly detection are judged, and then evaluated to generate a threshold evaluation matrix. Based on the threshold evaluation matrix, the optimal detection threshold is selected. The selection of the optimal detection threshold is based on the criterion of maximizing the F1 score. The F1 score is the harmonic mean of the precision and recall rates, and is an indicator that comprehensively considers the balance between the two. By maximizing the F1 score, the best balance between accuracy and recall is ensured, and a most suitable anomaly detection threshold is obtained. The anomaly detection threshold is applied to real-time status monitoring to determine whether the device status exceeds the normal range and is marked as an abnormal state. The spatiotemporal feature analysis of the abnormal state mark is performed to calculate the duration of the anomaly and the impact range parameter. The duration of the anomaly reflects the time span of the abnormal state, which helps to analyze the severity of the anomaly; while the impact range parameter reflects the degree of diffusion of the abnormal state in different device parts or systems.

[0020] In one example, multi-source data sampling is performed on the operating sound, equipment vibration, contact temperature, and operating current of the disconnector during operation. After signal preprocessing, a multi-source feature data set is obtained, including: The operating sound of the isolating switch is sampled at a frequency to obtain sound sampling data; the vibration of the isolating switch body is sampled at a frequency to obtain vibration sampling data; the contact temperature of the isolating switch is sampled at a frequency of 1Hz, and the temperature sampling data is obtained by combining the influencing factors of the ambient temperature; the operating current of the motor of the isolating switch operating mechanism is sampled at a frequency to obtain current sampling data by combining the load condition parameters; Perform wavelet threshold noise reduction operation on the sound sampling data to obtain the sound feature sequence; perform bandpass filtering operation on the vibration sampling data to obtain the vibration feature sequence; perform standardization operation on the temperature sampling data to obtain the temperature feature sequence; perform sliding window averaging operation on the current sampling data to obtain the current feature sequence; The sound feature sequence, vibration feature sequence, temperature feature sequence and current feature sequence are aligned according to timestamps and data fusion is performed to obtain a multi-source feature data set.

[0021] In this example, the operating sound of the isolating switch is sampled to obtain sound sampling data. The sound signal is collected by sensors such as microphones installed around the device. The sampling frequency determines the resolution of the sound data. Assume that the sound signal is , whose sampling frequency is , then the sound sampling data is expressed as: ; in, is the sampled sound data, is the sampling time interval, is the sampling frequency. The vibration of the isolating switch body is sampled to obtain vibration sampling data. The vibration signal is obtained by installing an acceleration sensor on the equipment body. The vibration signal is expressed as , and its sampling frequency is also , the corresponding sampling data is: ; Vibration signals are similar to sound signals. They are continuous signals that change over time. The sampled data is used for subsequent signal processing. For the contact temperature of the disconnector, 1Hz sampling is performed, and the temperature sampling data is obtained by combining the ambient temperature factor. The change of the contact temperature is relatively stable, and its sampling frequency is relatively low. 1Hz is selected as the sampling frequency, and the temperature value is sampled once per second. Assume that the temperature signal is , the sampled data is expressed as: Second; In actual measurement, the temperature sampling data is combined with the influence of ambient temperature, such as compensation for temperature changes with ambient temperature. Compensation is performed using the following formula: ; in, is the temperature data after compensation, is the ambient temperature data. For the operating current of the disconnector operating mechanism motor, frequency sampling is performed and combined with the load condition parameters to obtain current sampling data. The current signal is obtained through the current sensor. Assume that the current signal is , whose sampling frequency is , the corresponding current sampling data is: ; The change of current signal is affected by the load condition, and corresponding correction is made in the sampled data. Assume that the load condition parameters are , correct the current data as follows: ; Signal preprocessing is performed on various signals to extract more effective features. For sound sampling data, wavelet threshold denoising is used. Wavelet threshold denoising is a signal denoising method based on wavelet transform. It decomposes the signal into different frequency components through wavelet transform, and then performs threshold processing on the high-frequency noise components to remove noise. Assume that the wavelet transform coefficient of the sound signal is , and its threshold processing is expressed as: ; in, is the threshold value, by adjusting the threshold , controls the noise reduction effect. The processed signal As the sound feature sequence. For the vibration sampling data, a bandpass filter is used to effectively remove the low-frequency and high-frequency noise in the signal. Assume that the frequency domain representation of the vibration signal is , the transfer function of the bandpass filter is , then the filtered signal It is expressed as: ; in, represents the Fourier transform, For inverse Fourier transform, the bandpass filter retains the effective components of the signal within a specific frequency band and removes useless frequency components to obtain the vibration characteristic sequence. For the temperature sampling data, a normalization operation is performed to unify the range and distribution of the data so that data of different dimensions can be compared on the same scale. Assume that the mean of the temperature signal is , the standard deviation is , then the standardized temperature data It is expressed as: ; The obtained temperature characteristic sequence is the temperature data after standardization. For the current sampling data, the sliding window average operation is used. The signal is processed locally to remove the influence of mutation or noise. Assume that the current data is , the window size is , then the current characteristic sequence after sliding window averaging is It is expressed as: ; This operation effectively smoothes the current signal and removes high-frequency noise. Align the sound feature sequence, vibration feature sequence, temperature feature sequence, and current feature sequence according to the timestamp and perform data fusion. Merge data from different signal sources by splicing or weighting. Assuming that all feature sequences are aligned by timestamp, they are fused in the following way: ; in, Represents the fused multi-source feature dataset. Through data fusion, a comprehensive dataset containing all relevant features is obtained.

[0022] In one example, time domain parameter calculation, frequency domain component extraction, and time-frequency energy decomposition are performed on a multi-source feature data set, and polynomial fitting optimization processing is performed to obtain a feature vector sequence, including: The peak, variance, skewness and kurtosis parameters of the multi-source feature data set are extracted to obtain the time domain feature matrix; the multi-source feature data set is subjected to fast Fourier transform and frequency band energy calculation to obtain the frequency domain feature matrix; Perform wavelet packet multi-scale decomposition on multi-source feature data sets, and obtain the time-frequency feature matrix by calculating the energy coefficient and energy distribution ratio of each frequency band; Perform feature cascade splicing on the time domain feature matrix, the frequency domain feature matrix and the time-frequency feature matrix to obtain multi-dimensional feature data; A time series polynomial function model is constructed for multidimensional feature data. By setting the n-order polynomial basis function and initial parameters, a feature fitting function is obtained. The feature fitting function is optimized by least squares calculation. The polynomial coefficients are updated iteratively to obtain optimized feature parameters. The optimized feature parameters are evaluated by fitting residuals, and the feature optimization results are obtained by calculating the comparison results between the fitting error and the preset accuracy threshold. The feature optimization results are transformed into principal component dimension reduction to obtain a feature vector sequence.

[0023] In this example, when extracting time domain features, the indicators include peak, variance, skewness, and kurtosis, which help the system understand the basic characteristics of the signal from the shape and distribution of the signal. The peak value represents the maximum value of the signal, and the formula is: ; in, is the sampling value of the signal, and the peak value reflects the maximum amplitude of the signal within a period of time. The variance is used to describe the degree of fluctuation of the signal, and the calculation formula is: ; in, is the mean of the signal, is the number of sampling points of the signal. Variance reflects the strength of signal fluctuation. Skewness is used to measure the symmetry of the signal. The larger the skewness, the more the signal distribution is biased to one side. The calculation formula is: ; in, is the standard deviation of the signal. Kurtosis is the sharpness of the signal shape, which is used to describe the steepness of the signal distribution. Its calculation formula is: ; By extracting these time domain features, we can obtain the time domain feature matrix, which is the basic feature set of each signal source in the time domain. We can extract frequency domain features from the multi-source feature data set. Frequency domain analysis uses fast Fourier transform to convert the time domain signal to the frequency domain. In the frequency domain, the energy of the signal is mainly concentrated in certain frequency ranges. By calculating the energy of each frequency band, we can obtain the frequency domain features. The calculation formula of fast Fourier transform is: ; in, is the frequency domain signal, is the frequency, is an imaginary unit, is the number of sampling points of the signal. The result after Fourier transform is Used to calculate the energy of each frequency band. Assume that the energy of the signal is in a certain frequency band Internal , then the energy calculation formula is: ; In this way, the frequency domain feature matrix of the signal is obtained, that is, the energy distribution of each signal source in different frequency bands. In time-frequency analysis, the wavelet packet multi-scale decomposition method is used to extract the time-frequency characteristics of the signal. Wavelet packet transform is a multi-scale method for decomposing signals, which decomposes the signal into multiple frequency bands and analyzes each frequency band. Wavelet packet transform is expressed as: ; in, is the wavelet packet basis function, Represents different frequency bands. Through wavelet packet transform, the energy coefficient of each frequency band is obtained, and its energy distribution ratio is calculated, which is expressed as: ; Through this step, the time-frequency feature matrix is ​​obtained, which combines the time domain and frequency domain information and helps to capture the changing characteristics of the signal in time and frequency. The time domain feature matrix, frequency domain feature matrix and time-frequency feature matrix are concatenated and spliced ​​into a multidimensional feature data set in chronological order. Assume that the time domain feature is , the frequency domain characteristics are , the time-frequency characteristics are , then the feature matrix after cascade concatenation is: ; Model the multidimensional feature data set with a time series polynomial function. By constructing an n-order polynomial function model for the multidimensional feature data, the law of feature changes over time is captured. Let the polynomial basis function be , the expression of the model is: ; in, are the polynomial coefficients, is the basis function. These coefficients are fitted by the least squares method, and the objective function that minimizes the fitting error is: ; in, is the total number of time points of the sample. By minimizing the objective function, the optimal polynomial coefficients are obtained. The polynomial coefficients are updated through an iterative optimization algorithm (such as gradient descent method). The update rule for each iteration is: ; in, is the learning rate, The coefficient of the objective function The partial derivative of . By continuously updating the coefficients, the optimized polynomial features are obtained. In order to evaluate the effect of polynomial fitting, the fitting residual is evaluated. The fitting residual refers to the gap between the actual data and the fitting function, and the calculation formula is: ; By evaluating the comparison between the fitting residual and the preset accuracy threshold, it is determined whether the fitting is accurate enough. If the fitting error is less than the threshold, the feature optimization is considered successful. In order to reduce the data dimension, principal component analysis is used for dimensionality reduction to convert the multidimensional features into a more representative feature vector sequence. The dimensionality reduction process is expressed as: ; in, is the eigenvector matrix, is a multidimensional feature dataset, is the feature vector sequence after dimensionality reduction.

[0024] In one example, a time series polynomial function model is constructed for multidimensional feature data, a feature fitting function is obtained by setting an n-order polynomial basis function and initial parameters, and the feature fitting function is optimized by least squares optimization calculation, and the polynomial coefficients are iteratively updated to obtain optimized feature parameters, including: The multi-dimensional feature data is segmented according to a fixed time window to obtain a feature segmentation sequence, and a time series index mapping is established based on the feature segmentation sequence to obtain a feature time correspondence table; An n-order polynomial fitting model is constructed according to the characteristic time correspondence table, and a fitting function expression is obtained by selecting a combination of polynomial basis functions, and the initial value of the coefficient and the iteration step size are set for the fitting function expression to obtain the optimization initial condition; Input the optimization initial conditions into the least squares iterative solver, and obtain the optimization loss function by constructing the square error minimization objective function; The parameter gradient and update direction are calculated based on the optimized loss function, and the coefficients are iteratively calculated through adaptive step size adjustment to obtain the parameter iteration sequence. The convergence of the parameter iteration sequence is judged, and the convergence judgment index is obtained by calculating the parameter changes of adjacent iteration steps. The convergence judgment index is compared with the preset convergence threshold, and the optimized feature parameters are output when the convergence condition is met.

[0025] In this example, the multi-dimensional feature data is segmented according to a fixed time window. Set a time window length of , each time select a length of Then, these time periods are segmented as features. Assume that the total time series of the data set is , through the sliding window method, each time the data is divided into Time period ), each time period contains sampling points. Through this segmentation method, the feature segmentation sequence is obtained, namely: ; Establish a time series index mapping to establish a clear mapping relationship between each feature segment and its corresponding time point. By adding timestamp information to each feature segment sequence, a feature time correspondence table is obtained. Assume that the set of timestamps is , where each Representative feature segment The characteristic time correspondence table is: ; Based on these timestamps, an n-order polynomial fitting model is constructed. By fitting these feature data, more representative regular features are extracted. The expression of the n-order polynomial is expressed as: ; in, are the coefficients of the polynomial, is the time variable. To fit these data, set the initial coefficient values and a suitable iteration step size , which is used to update these coefficients in each iteration. Assuming the initial polynomial model is , then the objective function (i.e. fitting error function) is defined by the least squares method: ; in, Based on the current coefficient About time point The goal is to optimize the coefficients by minimizing this error function. Input into the least squares iterative solver, construct the objective function of minimizing the square error, and solve the optimal coefficient through iterative optimization. In each iteration, the coefficient is updated by calculating the parameter gradient and the update direction. The gradient calculation formula is: ; By gradient descent method, the coefficients are updated according to the gradient information: ; in, is the learning rate, which controls the step size of each iteration. The iteration process will continue until convergence. Convergence is determined by calculating the magnitude of the coefficient change between two consecutive iterations. Assume that the coefficient change between two consecutive iterations is If these changes are less than a certain threshold , then the coefficients are considered to have converged. The convergence determination formula is: Convergence Check: ; When the convergence condition is met, the iteration stops and the final optimized feature parameters are output. ,These optimized parameters reflect the regularity of feature changes over time.

[0026] In one example, a time-series dynamic graph network is constructed for a feature vector sequence, graph nodes are generated based on the temporal correlation of the feature vectors, and directed connection edges are constructed based on the similarity between the nodes to obtain a dynamic topology graph, including: Perform time point mark encoding on the feature vector sequence, and obtain time series mark data by adding sampling moment identification information; Perform graph node generation operation on the time series labeled data, and obtain the node data set by mapping the feature vector at time t to the network node; Perform cosine distance calculation on all node pairs in the node data set, obtain a similarity matrix by calculating the cosine value of the angle between the feature vectors, and compare the similarity matrix with the preset connection threshold to obtain a connection determination mark; The edge connection relationship is constructed based on the connection judgment identifier, and a directed connection is established between high-similarity nodes to obtain an edge connection set, and a time-series directed assignment is performed on the edge connection set to obtain a directed edge data set; The directed edge data set and the node data set are combined to construct an adjacency matrix. The initial network structure is obtained by recording the connection strength relationship between nodes. The topological relationship of the initial network structure is optimized. The dynamic topology graph is obtained by calculating the node connection degree distribution and adjusting the network connection density.

[0027] In this example, the feature vector sequence is encoded with time point tags, and a unique identification information is assigned to each time point to form time series tag data. , add a corresponding timestamp to it ,Right now: ; in, Indicates The sampling time of the feature vector. Based on the time series labeled data, the graph node generation operation is performed. In the graph neural network, the feature vector of each time point is regarded as a node in the graph, and the data of the node is the feature vector. Assume that the node set is , where each node Corresponding to a feature vector and timestamp The node data set is: ; These nodes form the initial structure of the graph. Calculate the similarity between nodes. In order to measure the similarity between feature vectors at different time points, the cosine distance is used. For any two nodes and , the cochord distance calculation formula is: ; in, and The nodes are and The corresponding eigenvector, is their dot product, and are their moduli respectively. The value range of cosine similarity is between [-1,1]. The larger the value, the higher the similarity between the two vectors. By calculating the cosine similarity between all node pairs, we can get the similarity matrix , where each element Representation Node and The similarity between: ; Based on the preset connection threshold , by comparing the similarity matrix with the threshold, the connection judgment mark is obtained If the similarity between two nodes is greater than the threshold , then it is considered that there is a connection between them. That is: ; Connection determination flag Determine which nodes are connected. Based on the connection determination identifier, construct the edge connection relationship. and ,if , then a directed edge is established between these two nodes Each edge Represents a slave node To Node The resulting edge connection set for: ; Since the graph is driven by time series data, each edge has time-series directionality, that is, the direction of the edge is not only determined by the similarity, but also takes into account the order of time. Assume that for each edge ,if , then the direction of the edge is considered to be from arrive , that is, the directedness of the edge conforms to the order of time. With node dataset Combination, build adjacency matrix , used to record the connection strength between nodes in the graph. The elements of the adjacency matrix Representation Node To Node The weight of the edge (if there is an edge connection). That is: ; Optimize the topology of the initial network structure. Adjust the connection relationship of the nodes in the graph to make the graph structure more compact and better reflect the time correlation of the feature vector sequence. In order to optimize the topology, calculate the node connection degree distribution, that is, the number of connections of each node: ; Then, by adjusting the network connection density, the nodes in the graph are connected more evenly or selectively according to certain rules. The optimized graph structure can better describe the temporal relationship between nodes, thus obtaining a dynamic topology graph, namely: .

[0028] In one example, the dynamic topology graph is input into the deep graph neural network, and the state features are extracted through multi-layer graph convolution operations, feature pooling compression and full connection mapping to obtain the device state classification results, including: The dynamic topology graph is input into the three-layer graph convolution network in the deep graph neural network, and multi-scale graph features are obtained by sequentially performing 64-dimensional, 128-dimensional, and 256-dimensional graph convolution kernel operations; Aggregate neighborhood information of multi-scale graph features and obtain local structural features by fusing feature representations of central nodes and adjacent nodes. Calculate the attention allocation weight for local structural features, and obtain the feature importance coefficient through node feature correlation scoring operation; Based on the feature importance coefficient, weighted feature aggregation is performed, and the node representation vector is obtained by performing weighted feature fusion operation; Perform a maximum pooling operation on the node representation vector to obtain the graph representation features by extracting the most significant eigenvalues ​​of each dimension; Input the graph representation features into the three-layer fully connected neural network in the deep graph neural network, and obtain the deep feature map by performing 512-dimensional ReLU mapping, 256-dimensional ReLU mapping, and 128-dimensional ReLU mapping; A Softmax classification operation is performed on the deep feature map. The state probability distribution is obtained by calculating the normalized category probability score. The maximum probability label is extracted from the state probability distribution to obtain the device state classification result.

[0029] In this example, the dynamic topology graph is input into a three-layer graph convolutional network in a deep graph neural network. In the first layer of the graph convolutional network, the initial node features are input. , each node The corresponding eigenvector is ,Right now: ; In graph convolution, the adjacency matrix of the graph Describes the connection relationship between nodes. Its elements 1 indicates a node and nodes There is a connection between them, and 0 means there is no connection. The graph convolution operation is expressed as: ; in, is the normalized adjacency matrix, using , is the degree matrix, is the activation function (e.g. ReLU), is the weight matrix of the first layer of graph convolution. Through graph convolution, we get the first layer feature matrix , where the feature representation of each node has been integrated with the information of its neighboring nodes. In the second layer of graph convolution, the output of the first layer is input , perform similar graph convolution operations to obtain the feature matrix of the second layer : ; same, is the weight matrix of the second layer. Through this layer, the graph convolution network can gradually expand the neighborhood range and capture higher-level graph features. In the third layer of graph convolution, the same operation is applied to obtain the final feature matrix : ; At this point, the node features already contain multi-scale graph features extracted by three layers of graph convolution, and the node representation can fully reflect the structural information of the graph. The graph neural network optimizes the representation of each node by aggregating the feature information of the neighbors around each node. By aggregating neighborhood information, the feature representations of the central node and its adjacent nodes are fused. , a new node representation is obtained by calculating the weighted average of its neighborhood information : ; in, Representation Node The neighbor set of is the number of neighbors, Neighbor node Through aggregation, the representation of a node contains more information about its neighbors, which helps the model understand the complexity of the graph structure. The attention mechanism is applied to calculate the importance of node features. By scoring the relevance of node features, an attention weight is assigned to each neighbor node. , indicating a node For Node The calculation formula is: ; in, is the attention weight vector, In this way, weights are assigned according to the similarity between nodes, so that more important neighbor nodes have a greater impact on the final node representation. Based on the attention weights, weighted feature aggregation is performed to obtain the weighted feature representation of the node. : ; Through weighted aggregation, the final representation of the node The features of the central node and its neighboring nodes are integrated and assigned different weights according to the attention mechanism. The maximum pooling operation is performed on the weighted features of the nodes to extract the most significant feature values ​​in each node dimension. The maximum pooling operation is expressed by the following formula: ; This step helps the network focus on the most important features, reduces redundant information, and improves the accuracy of feature expression. After obtaining the final representation of the node, the features of all nodes are aggregated to obtain the graph representation features, which contain the structural information of the entire graph. The graph representation features are input into the three-layer fully connected neural network in the deep graph neural network. Each layer of the neural network performs the following operations: ; ; in, is the weight matrix of each layer, is the bias term, and ReLU is the activation function. Through the multi-layer fully connected neural network, the high-level features of the graph are extracted, and deeper nonlinear relationships can be captured. Through the Softmax classification operation, the classification result of the device status is obtained. The calculation formula of the Softmax function is: ; By calculating the normalized category probability score, the probability distribution of each category is obtained. According to the maximum probability label, the device status classification result is finally determined, that is, whether the device is currently in a normal state, faulty state, etc.

[0030] In one example, the AI-based isolating switch health status management method further includes: Perform statistical feature analysis on the state probability sequence in the equipment state classification results, and obtain the normal state feature distribution by calculating the probability mean and standard deviation distribution under normal operating conditions; Initialize the threshold of the normal state feature distribution, and obtain the monitoring benchmark threshold by calculating the μ-3σ limit value; Perform abnormal pattern extraction on the historical records of equipment status classification results, and obtain an abnormal probability sample set by statistically analyzing the probability distribution characteristics of known fault events; Perform kernel density distribution estimation on the abnormal probability sample set, and obtain the continuous abnormal distribution function through Gaussian kernel function probability density calculation; The curve characteristics of the continuous abnormal distribution function are analyzed, and the candidate threshold point set is obtained by calculating the change points of the curve slope. The classification performance of the candidate threshold point set is tested, and the threshold evaluation matrix is ​​obtained by calculating the precision and recall indicators. The optimal detection threshold is selected based on the threshold evaluation matrix, and the anomaly detection threshold is obtained by maximizing the F1 score criterion. The anomaly detection threshold is applied to real-time state monitoring to obtain an abnormal state mark. The spatiotemporal characteristics of the abnormal state markers are analyzed, and the abnormal state assessment results are obtained by calculating the abnormal duration and impact range parameters.

[0031] In this example, the state probability sequence in the device state classification result is analyzed for statistical characteristics. The device state classification result will output the state probability value corresponding to each time point, indicating the probability distribution of the device in different states (such as normal or faulty). In order to perform effective anomaly detection, the statistical characteristics of these probability sequences are extracted, especially the probability mean and standard deviation under normal operating conditions, to obtain the normal state feature distribution. The threshold of the normal state feature distribution is initialized. The monitoring benchmark threshold is set based on the mean and standard deviation of the normal state. Calculation , that is, a limit is defined by three multiples of the standard deviation, covering most of the normal state, assuming that the probability of abnormal events is low. The specific formula is as follows: ; This threshold is used in real-time monitoring. When the probability value of the device state is lower than this threshold, it means that the device has entered an abnormal state. The abnormal pattern is extracted from the historical records of the device state classification results. In the historical data, it is assumed that there are known fault events, and the probability distribution characteristics of these known fault events are statistically analyzed to extract the abnormal probability sample set. These samples come from a labeled dataset of equipment failures, extracted in the following way: ; After obtaining the abnormal samples, kernel density estimation is performed to smooth them and obtain a continuous abnormal distribution function. . Use the Gaussian kernel function to estimate the probability density: ; in, is the sample size, is the bandwidth parameter, which controls the degree of smoothing. Through kernel density estimation, a continuous anomaly distribution function is obtained. The curve characteristics of the continuous anomaly distribution function are analyzed. By analyzing the slope change of the curve, potential candidate threshold point sets are identified. Assume is the abnormal distribution function, by calculating its first-order derivative (i.e. slope): ; Determine the points where the slope changes significantly. These points represent the turning points or change points of the abnormal distribution, and then determine the candidate threshold point set. Among these candidate points, perform classification performance tests and use classification evaluation indicators such as precision and recall to evaluate the performance of each candidate threshold point. Precision and recall are calculated using the following formulas: ; ; in, It is a real example. It is a false positive example. is a false negative example. By calculating the precision and recall of each candidate threshold point, we get the threshold evaluation matrix, which records the classification performance of each candidate threshold. Based on the evaluation matrix, we select the optimal detection threshold to maximize the F1 score, which is the harmonic mean of the precision and recall, and the formula is as follows: ; By maximizing the F1 score, the optimal anomaly detection threshold is obtained. When the state probability of the device is lower than this threshold, the system will determine that it has entered an abnormal state. The optimal anomaly detection threshold is applied to real-time state monitoring and the abnormal state is marked. During the monitoring process, once the state probability of the device is lower than the threshold, the system will trigger the abnormal state marking. The spatiotemporal feature analysis of the abnormal state marking is performed to evaluate the duration of the anomaly and the scope of the impact of the anomaly on the device. The duration of the anomaly is obtained by calculating the time difference between the start and end of the marking, and the scope of impact is measured by the degree of functional limitation of the device in the abnormal state. Through these spatiotemporal feature analyses, the abnormal state evaluation results are obtained.

[0032] Reference Figure 2 , this embodiment provides an AI-based isolating switch health status management system, including: Sampling module 1 is used to perform multi-source data sampling on the operating sound, equipment vibration, contact temperature and operating current during the operation of the disconnector, and obtain a multi-source feature data set after signal preprocessing; Decomposition module 2 is used to calculate time domain parameters, extract frequency domain components and decompose time-frequency energy for multi-source feature data sets, and obtain feature vector sequences through polynomial fitting optimization processing; Construction module 3 is used to construct a time-series dynamic graph network for the feature vector sequence, generate graph nodes based on the time correlation of the feature vectors, and construct directed connection edges based on the similarity between the nodes to obtain a dynamic topology graph; Extraction module 4 is used to input the dynamic topology map into the deep graph neural network, extract state features through multi-layer graph convolution operations, feature pooling compression and full connection mapping, and obtain the device state classification result.

[0033] In this embodiment, for the specific implementation of each unit in the above system embodiment, please refer to the above method embodiment, which will not be repeated here.

[0034] Reference Figure 3 In an embodiment of the present invention, a computer device is also provided. The computer device may be a server, and its internal structure may be as follows: Figure 3 As shown. The computer device includes a processor, a memory, a display screen, an input device, a network interface and a database connected through a system bus. Among them, the processor designed by the computer is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the corresponding data in this embodiment. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the above method is implemented.

[0035] Those skilled in the art will understand that Figure 3 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present invention, and does not constitute a limitation on the computer device to which the solution of the present invention is applied.

[0036] An embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the above method is implemented. It can be understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0037] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media provided by the present invention and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double-speed data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM.

[0038] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, device, article or method including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, device, article or method. In the absence of further restrictions, an element defined by the sentence "includes a ..." does not exclude the presence of other identical elements in the process, device, article or method including the element.

[0039] The above description is only a preferred embodiment of the present invention, and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. An AI-based isolating switch health status management method, characterized in that: The following steps are involved: Multi-source data sampling is performed on the operating sound, equipment vibration, contact temperature and operating current during the operation of the disconnector, and a multi-source feature data set is obtained after signal preprocessing; Performing time domain parameter calculation, frequency domain component extraction and time-frequency energy decomposition on the multi-source feature data set, and performing polynomial fitting optimization processing to obtain a feature vector sequence; Constructing a time-series dynamic graph network for the feature vector sequence, generating graph nodes based on the temporal correlation of the feature vectors, and constructing directed connection edges based on the similarity between the nodes to obtain a dynamic topology graph; The dynamic topology map is input into a deep graph neural network, and state features are extracted through multi-layer graph convolution operations, feature pooling compression and full connection mapping to obtain device state classification results.

2. The AI-based isolating switch health status management method according to claim 1 is characterized in that: The multi-source data sampling of the operating sound, equipment vibration, contact temperature and operating current during the operation of the disconnector is performed, and a multi-source feature data set is obtained after signal preprocessing, including: The operating sound of the isolating switch is sampled at a frequency to obtain sound sampling data; the vibration of the isolating switch body is sampled at a frequency to obtain vibration sampling data; the contact temperature of the isolating switch is sampled at a frequency of 1Hz, and the temperature sampling data is obtained by combining the influencing factors of the ambient temperature; the operating current of the motor of the isolating switch operating mechanism is sampled at a frequency to obtain current sampling data by combining the load condition parameters; Performing a wavelet threshold noise reduction operation on the sound sampling data to obtain a sound feature sequence; performing a bandpass filtering operation on the vibration sampling data to obtain a vibration feature sequence; performing a normalization operation on the temperature sampling data to obtain a temperature feature sequence; performing a sliding window average operation on the current sampling data to obtain a current feature sequence; The sound feature sequence, the vibration feature sequence, the temperature feature sequence and the current feature sequence are aligned according to timestamps and data fusion is performed to obtain a multi-source feature data set.

3. The AI-based isolating switch health status management method according to claim 2 is characterized in that: The multi-source feature data set is subjected to time domain parameter calculation, frequency domain component extraction and time-frequency energy decomposition, and polynomial fitting optimization processing, Get the feature vector sequence, including: Extracting peak, variance, skewness and kurtosis parameters of the multi-source feature data set to obtain a time domain feature matrix; performing fast Fourier transform and frequency band energy calculation on the multi-source feature data set to obtain a frequency domain feature matrix; Performing wavelet packet multi-scale decomposition on the multi-source feature data set, and obtaining a time-frequency feature matrix by calculating the energy coefficient and energy distribution ratio of each frequency band; Performing feature cascade splicing on the time domain feature matrix, the frequency domain feature matrix and the time-frequency feature matrix to obtain multi-dimensional feature data; Constructing a time series polynomial function model for the multidimensional feature data, obtaining a feature fitting function by setting an n-order polynomial basis function and initial parameters, performing least squares optimization calculation on the feature fitting function, and obtaining optimized feature parameters by iteratively updating polynomial coefficients; The optimized feature parameters are evaluated for fitting residuals, and feature optimization results are obtained by calculating the comparison result between the fitting error and a preset accuracy threshold, and the feature optimization results are transformed into principal components for dimensionality reduction to obtain a feature vector sequence.

4. The AI-based isolating switch health status management method according to claim 3 is characterized in that: The method of constructing a time series polynomial function model for the multidimensional feature data, obtaining a feature fitting function by setting an n-order polynomial basis function and initial parameters, performing least squares optimization calculation on the feature fitting function, and obtaining optimized feature parameters by iteratively updating polynomial coefficients includes: Segmenting the multidimensional feature data according to a fixed time window to obtain a feature segmentation sequence, and establishing a time series index mapping based on the feature segmentation sequence to obtain a feature time correspondence table; Constructing an n-order polynomial fitting model according to the characteristic time correspondence table, obtaining a fitting function expression by selecting a combination of polynomial basis functions, and setting coefficient initial values ​​and iteration steps for the fitting function expression to obtain an optimized initial condition; Inputting the optimization initial condition into a least squares iterative solver, and obtaining an optimization loss function by constructing a square error minimization objective function; Calculating the parameter gradient and update direction based on the optimization loss function, performing coefficient iterative calculation through adaptive step size adjustment to obtain a parameter iteration sequence, and performing convergence judgment on the parameter iteration sequence, and obtaining a convergence judgment index by calculating the parameter change amount of adjacent iteration steps; The convergence judgment index is compared with a preset convergence threshold, and when the convergence condition is met, the optimized characteristic parameters are output.

5. The AI-based isolating switch health status management method according to claim 4 is characterized in that: The step of constructing a time-series dynamic graph network for the feature vector sequence, generating graph nodes based on the temporal correlation of the feature vectors, and constructing directed connection edges based on the similarity between the nodes to obtain a dynamic topology graph includes: Performing time point mark encoding on the feature vector sequence, and obtaining time series mark data by adding sampling moment identification information; Performing a graph node generation operation on the time series labeled data, and obtaining a node data set by mapping the feature vector at time t to a network node; Performing cosine distance calculation on all node pairs in the node data set, obtaining a similarity matrix by calculating the cosine value of the angle between the feature vectors, and comparing the similarity matrix with a preset connection threshold to obtain a connection determination identifier; Building an edge connection relationship based on the connection determination identifier, obtaining an edge connection set by establishing directed connections between high-similarity nodes, and performing temporal directed assignment on the edge connection set to obtain a directed edge data set; The directed edge data set and the node data set are combined to construct an adjacency matrix, and the initial network structure is obtained by recording the connection strength relationship between nodes. The topological relationship of the initial network structure is optimized, and the dynamic topology graph is obtained by calculating the node connection degree distribution and adjusting the network connection density.

6. The AI-based isolating switch health status management method according to claim 5, characterized in that: The dynamic topology graph is input into a deep graph neural network, and state feature extraction is performed through multi-layer graph convolution operation, feature pooling compression and full connection mapping to obtain a device state classification result, including: The dynamic topology graph is input into a three-layer graph convolutional network in a deep graph neural network, and multi-scale graph features are obtained by sequentially performing 64-dimensional, 128-dimensional, and 256-dimensional graph convolution kernel operations; Aggregating neighborhood information of the multi-scale graph features, and obtaining local structural features by fusing feature representations of the central node and adjacent nodes; Calculating the attention allocation weight for the local structural features, and obtaining the feature importance coefficient through node feature correlation scoring operation; Performing weighted feature aggregation based on the feature importance coefficients, and obtaining a node representation vector by performing a weighted feature fusion operation; Performing a maximum pooling operation on the node representation vector to obtain a graph representation feature by extracting the most significant eigenvalue of each dimension; Inputting the graph representation features into a three-layer fully connected neural network in the deep graph neural network, and obtaining a deep feature map by performing 512-dimensional ReLU mapping, 256-dimensional ReLU mapping, and 128-dimensional ReLU mapping; A Softmax classification operation is performed on the deep feature map, a state probability distribution is obtained by calculating the normalized category probability score, and a maximum probability label is extracted from the state probability distribution to obtain a device state classification result.

7. The AI-based isolating switch health status management method according to claim 6, characterized in that: The AI-based isolating switch health status management method also includes: Performing statistical characteristic analysis on the state probability sequence in the equipment state classification result, and obtaining the normal state characteristic distribution by calculating the probability mean and standard deviation distribution under the normal operating state; Initializing the threshold of the normal state characteristic distribution, and obtaining the monitoring reference threshold by calculating the μ-3σ limit value; Performing abnormal pattern extraction on the historical records of the equipment status classification results, and obtaining an abnormal probability sample set by statistically analyzing the probability distribution characteristics of known fault events; Performing kernel density distribution estimation on the abnormal probability sample set, and obtaining a continuous abnormal distribution function by Gaussian kernel function probability density calculation; Performing curve feature analysis on the continuous abnormal distribution function, obtaining a candidate threshold point set by calculating the curve slope change point, and performing classification performance test on the candidate threshold point set, obtaining a threshold evaluation matrix by calculating precision and recall indicators; Selecting an optimal detection threshold based on the threshold evaluation matrix, obtaining an abnormality detection threshold by maximizing the F1 score criterion, and applying the abnormality detection threshold to real-time state monitoring to obtain an abnormal state mark; The spatiotemporal characteristics of the abnormal state mark are analyzed, and the abnormal state evaluation result is obtained by calculating the abnormal duration and the impact range parameters.

8. An AI-based isolating switch health status management system, characterized in that: For implementing the steps of the method according to any one of claims 1 to 7, the system comprises: The sampling module is used to sample multi-source data of operating sound, equipment vibration, contact temperature and operating current during the operation of the disconnector, and obtain a multi-source feature data set after signal preprocessing; A decomposition module, used to perform time domain parameter calculation, frequency domain component extraction and time-frequency energy decomposition on the multi-source feature data set, and obtain a feature vector sequence through polynomial fitting optimization processing; A construction module is used to construct a time-series dynamic graph network for the feature vector sequence, generate graph nodes based on the time correlation of the feature vectors, and construct directed connection edges based on the similarity between the nodes to obtain a dynamic topology graph; The extraction module is used to input the dynamic topology map into a deep graph neural network, extract state features through multi-layer graph convolution operations, feature pooling compression and full connection mapping, and obtain device state classification results.

9. A computer device comprising a memory and a processor, wherein a computer program is stored in the memory, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images