Intelligent diagnosis method and system for high-voltage cable insulation fault based on multi-information fusion

Abstract

The application relates to the field of high-voltage cable fault diagnosis, and discloses a high-voltage cable insulation fault intelligent diagnosis method and system based on multi-information fusion, which comprises the following steps: S1: acquiring operation multi-source data of a high-voltage cable, and constructing a multi-source original data matrix; S2: preprocessing to obtain a time-space aligned standardized data matrix; S3: performing feature extraction, acquiring a multi-dimensional feature vector, and learning the internal correlation among features by using a multi-modal deep network to obtain joint multi-modal features; S4: constructing a fault type recognition model based on Bayesian inference and Monte Carlo sampling, and acquiring a fault type classification result; S5: combining deep learning and a physical model to acquire interval and positioning information of fault occurrence and a fault grade; and S6: acquiring the fault grade based on the fault type classification result and the interval and positioning information of fault occurrence, and generating a diagnosis report for early warning pushing. The application realizes high-precision identification, positioning and risk assessment of high-voltage cable insulation faults.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of high-voltage cable fault diagnosis, and in particular to a high-voltage cable insulation fault intelligent diagnosis method and system based on multi-information fusion. BACKGROUND

[0002] With the continuous rise of the demand for stability and reliability of power supply in modern society, the operation safety of high-voltage cables, as the core carrier of urban power grids and cross-regional power transmission networks, directly affects the social production and life order. In recent years, the laying length and voltage level of high-voltage cables have been continuously improved, and the working environment has become more complex. Long-term exposure to electrical stress, thermal stress, mechanical stress, and chemical corrosion and other multiple factors has led to an increasing number of insulation aging, partial discharge, overheating and other fault risks. According to statistics, insulation faults have become the primary cause of high-voltage cable failure. Their suddenness and concealment characteristics not only may cause large-scale power outages, but also pose safety risks such as fires, causing great threats to power grid safety and the economy.

[0003] Traditional high-voltage cable insulation fault diagnosis methods mainly rely on single sensor data or manual inspection. For example, the method based on partial discharge detection can only capture electrical signal characteristics and is difficult to reflect insulation degradation caused by mechanical damage and environmental factors. Although infrared temperature measurement technology can detect overheating anomalies, it cannot accurately locate the root cause of the fault. Manual inspection is limited by subjective experience and work intensity, and has problems such as low efficiency, high missed detection rate, etc. These methods lack diagnostic accuracy and timeliness when faced with complex operating conditions and early weak faults, and are difficult to meet the needs of intelligent power grids for fault prediction and rapid disposal. SUMMARY

[0004] To solve the above problems, the purpose of the present application is to provide a high-voltage cable insulation fault intelligent diagnosis method based on multi-information fusion, which effectively realizes high-precision identification, positioning and risk assessment of high-voltage cable insulation faults.

[0005] To achieve the above purpose, the present application adopts the following technical solutions:

[0006] A high-voltage cable insulation fault intelligent diagnosis method based on multi-information fusion, comprising the following steps:

[0007] S1: Obtain the operating multi-source data of the high-voltage cable, and construct a multi-source original data matrix;

[0008] S2: Preprocess the obtained operating multi-source data of the high-voltage cable to obtain a spatio-temporally aligned standardized data matrix;

[0009] S3: According to the spatio-temporally aligned standardized data matrix, feature extraction is performed to obtain a multi-dimensional feature vector, and a multi-modal deep network is used to learn the inherent correlation between the features to obtain a joint multi-modal feature;

[0010] S4: Constructing a fault type identification model based on Bayesian inference and Monte Carlo sampling, and discriminating the fault type according to the joint multi-modal features to obtain a fault type classification result;

[0011] S5: Obtaining the interval and positioning information of fault occurrence according to the joint multi-modal features combined with deep learning and physical model;

[0012] S6: Obtaining the fault level based on the fault type classification result and the interval and positioning information of fault occurrence, and generating a diagnosis report for early warning push.

[0013] Further, the operation multi-source data of the high-voltage cable is obtained, including partial discharge signals, dielectric loss factor, cable temperature and humidity, environmental parameters, operation and maintenance historical data and fault labels, which are as follows:

[0014] A high-frequency current transformer is deployed at the cable joint to detect the electric pulse signal, and a piezoelectric ceramic sensor is installed along the cable body to capture the ultrasonic signal; an optical fiber Bragg grating array is embedded in the cable shielding layer to monitor the optical pulse signal in real time;

[0015] A digital dielectric loss tester is installed on the cable grounding wire to obtain the dielectric loss factor and the capacitance change rate time series data by injecting test signals through the low-frequency superposition method;

[0016] A distributed optical fiber temperature measurement system is spirally wound along the cable outer sheath to obtain the axial temperature distribution matrix; a three-axis MEMS vibration sensor is installed at the cable joint to monitor the X / Y / Z three-axis vibration acceleration time series signal; and the environmental parameter time series data and gas concentration spectrum are obtained through the environmental sensor;

[0017] The cable load curve, operation and maintenance historical data and fault labels are extracted by accessing the SCADA system through the OPC UA protocol;

[0018] The above data is collected through the main control cabinet node, the main control node is equipped with GPS and Beidou dual-mode clock source, and all sensor clocks are synchronized through PTP protocol; when any partial discharge channel detects a >50pC pulse, all sensors are triggered to record 1ms time window data synchronously;

[0019] A three-dimensional coordinate system of the cable is established, and the installation position of each sensor is labeled, and the temperature data is automatically mapped to the corresponding coordinate point;

[0020] Further, the preprocessing includes filtering and denoising, multi-channel synchronization processing distortion correction and abnormal value screening, which are as follows:

[0021] According to S1, the original multi-source data matrix X is obtained raw :

[0022] X raw ={x1(t),x2(t),...,x a (t),...,x n (t)};

[0023] where x a (t) represents the a-th sensor signal, including electrical signal, acoustic signal, temperature field, and t is the time sequence;

[0024] Corresponding filtering and denoising processing is adopted for different sensor signals, elliptical filtering is adopted for electrical signal, EMD denoising is adopted for acoustic signal, and wavelet threshold is adopted for optical signal, to obtain the denoised signal data X denoised ;

[0025] The denoised signal data X denoised is subjected to multi-channel synchronous processing distortion correction, the electrical signal amplitude is corrected through the transmission line model, and the temperature field is compensated through heat conduction:

[0026] to obtain the physical quantity real value signal data X corrected ;

[0027] Finally, combined with the local outlier factor and Mahalanobis distance, abnormal data caused by hardware failure or environmental interference is removed, to obtain the standardized space-time alignment matrix X proc .

[0028] Further, according to the standardized space-time alignment matrix, feature extraction is performed, multi-dimensional feature vectors are obtained and fused, to obtain joint multi-modal feature representation, specifically as follows:

[0029] According to the standardized space-time alignment matrix X proc obtained by S2, features are extracted from each modal data, including partial discharge signal feature, acoustic signal feature and temperature field feature; the partial discharge signal adopts ResNet-18 to extract spatial feature as the partial discharge signal feature; the ultrasonic signal adopts bidirectional LSTM to capture evolution trend as the acoustic signal feature; the temperature field adopts Transformer encoder to model spatial dependence as the temperature field feature;

[0030] Based on the extracted modal features, Z-score standardization is adopted for standardization processing, and mutual information I i between each feature f and fault label y is calculated:

[0031] ;

[0032] where, is the joint probability density; is the feature fi the independent probability distribution of y; p(y) is the prior probability of the fault label y;

[0033] reserved Top-K high mutual information features F MI , delete redundant features that satisfy ∣p(f i ,f j )∣>0.7, wherein p(f i ,f j ) is the joint probability distribution of f i and f j ;

[0034] The multi-modal deep network is used to learn the internal association between the high mutual information features, obtain a high expression representation of the fusion features, and construct a joint multi-modal feature h fusion .

[0035] Further, the multi-modal deep network comprises a feature embedding layer, a feature interaction layer, an attention mechanism and a feature fusion layer, and specifically as follows:

[0036] The feature embedding layer maps the features of different modalities to a unified feature space:

[0037] h pd =W pd ·f pd +b pd ;

[0038] h sound =W sound ·f sound +b sound ;

[0039] h temp =W temp ·f temp +b temp ;

[0040] Wherein: f pd , f sound , f temp are feature vectors of partial discharge, acoustic signal and temperature field respectively; W pd , W sound , W temp are learnable weight matrices; b pd , b sound , b temp are bias vectors; h pd , h sound , h temp are feature vectors embedded by partial discharge, acoustic signal and temperature field respectively;

[0041] The feature interaction layer models the complex correlation between different modalities, and learns the complex correlation between features by using a multi-modal interaction network Z:

[0042] Z = σ(W v · [h pd ; h sound ; h temp ] + b v );

[0043] where [; ] represents the concatenation operation of features; W v and b v are the weights and biases of the feature interaction layer; and σ is an activation function.

[0044] The self-attention mechanism MultiHead is used to capture long-distance dependencies between features:

[0045] Q = W Q · Z, K = W K · Z, V = W V · Z;

[0046] ;

[0047] ;

[0048] where Q, K, V represent the table query, key and value matrices respectively, W Q , W K , W V are the parameter matrices corresponding to Q, K, V; , , are the learned projection matrices corresponding to Q, K, V of the hth attention head; head h is the hth attention head; H is the number of attention heads; Attention represents the attention mechanism calculation function; W O is the output weight matrix; and Concat represents the concatenation function.

[0049] The importance of different modalities is adaptively weighted:

[0050] ;

[0051] where is the attention weight of each modality; W a1 , W a2 , W a3 and b a1 , b a2 , b a3 are learned parameters.

[0052] The feature fusion layer combines the results of the multi-head attention and the modality attention to generate a final joint multi-modal feature representation h fusion :

[0053] ;

[0054] ;

[0055] ;

[0056] wherein h attn is the output of the self-attention mechanism; h modal is the output of the modality attention mechanism; W f and b f are the weights and biases of the feature fusion layer.

[0057] Further, a fault type recognition model is constructed based on Bayesian inference and Monte Carlo sampling, and according to the joint multi-modal feature, fault type discrimination is performed to obtain a fault type classification result, specifically as follows:

[0058] First, the joint multi-modal feature is subjected to nonlinear transformation and feature enhancement:

[0059] F enhanced =σ(W1·h fusion +b1)+h fusion ;

[0060] F attention =Softmax(W2·F enhanced +b2)⊙F enhanced ;

[0061] wherein represents element-wise multiplication; F enhanced represents the output after nonlinear transformation; F attention represents the output after feature enhancement; W1 and b1 are respectively the weights and biases of the nonlinear transformation; W2 and b2 are respectively the weights and biases of the feature enhancement;

[0062] A multi-scale feature extraction module is constructed, which is processed by maximum pooling MaxPool and average pooling AvgPool:

[0063] F local =MaxPool(W3·F attention +b3);

[0064] F global =AvgPool(W4·F attention +b4);

[0065] F multi-scale =Concat(F localF global );

[0066] where F local represents the output after max-pooling; F global represents the output after average-pooling; W3 and b3 are the weight and bias of the max-pooling MaxPool, respectively; W4 and b4 are the weight and bias of the average-pooling AvgPool, respectively; F multi-scale represents the extracted multi-scale feature;

[0067] Based on Bayesian inference, a classifier is constructed, and a probability distribution is output :

[0068] ;

[0069] ;

[0070] ;

[0071] where, is random noise, is the probability density function of the noise; and are the mean vector and standard deviation vector of the class y, respectively; W σ and b σ are the weight and bias of the standard deviation prediction, respectively; and are the weight and bias of the mean prediction, respectively;

[0072] Estimate the classification uncertainty by Monte Carlo sampling:

[0073] ;

[0074] ;

[0075] where, is the expected probability of the average; T is the number of samplings; is the random noise of the t-th sampling; Uncertainty is the information entropy of the model prediction distribution; c represents the class c;

[0076] Simultaneously optimize the classification and uncertainty estimation, and the objective function is:

[0077] ;

[0078] where, is the cross-entropy loss; is the regularization loss; is the uncertainty constraint loss, which adopts entropy regularization; , is a weight coefficient;

[0079] Final classification decision, predict fault type:

[0080] ;

[0081] wherein, is the probability of predicting the fault type as c.

[0082] Further, according to the joint multi-modal feature, combining deep learning and physical model, the interval and positioning information of fault occurrence are obtained, and the specific process is as follows:

[0083] The spatio-temporal feature is extracted from the joint multi-modal feature h fusion :

[0084] F temp =TCN(h fusion );

[0085] F space =GCN([F temp ,A]);

[0086] Wherein, TCN is a time convolution network, capturing the time sequence feature F temp ; GCN is a graph convolution network, modeling the spatial relationship F space ;

[0087] The cable physical parameters are fused with the spatio-temporal feature to embed the positioning process:

[0088] F phys =Embed(R,L,C,G);

[0089] F joint =a’·F space +(1-a’)·F phys ;

[0090] Wherein, R, L, C, G are the physical parameters of the cable, respectively, resistance, inductance, capacitance, conductance; Embed is a physical parameter embedding function; a' is a learnable fusion weight; F phys is a cable physical feature vector; F joint is a fusion feature;

[0091] The probability mapping from the feature to the position is constructed:

[0092] ;

[0093] ;

[0094] Wherein, p(l∣F joint) is the probability distribution of position l; l i is the predefined i-th candidate position; is the predicted fault position; W loc and b loc are the weight and bias of the probability mapping respectively; N is the number of candidate positions.

[0095] Further, based on the fault type classification result and the interval and positioning information of the fault occurrence, the fault level is obtained, and a diagnosis report is generated for early warning push, specifically as follows:

[0096] In combination with historical data and expert experience, a fault level evaluation model is constructed, and the fault level is obtained according to the fault type classification result and the interval and positioning information of the fault occurrence;

[0097] According to the fault type classification result, the interval and positioning information of the fault occurrence, and the fault level, a detailed diagnosis report is generated; the diagnosis report content includes a multi-source evidence chain, and the original information and processing process of the multi-source data such as partial discharge signal, sound signal, temperature field used for diagnosis are displayed; at the same time, according to the fault condition and historical maintenance experience, disposal suggestions are provided; different early warning mechanisms are set according to the fault level, for emergency level fault, strong reminder notice is immediately sent to the operation and maintenance management personnel through the Web system and the mobile terminal APP; for serious level fault, detailed fault information and disposal suggestions are pushed within the preset time; for general level fault, messages can be pushed within the preset time, and they are included in the daily inspection task reminder.

[0098] A high-voltage cable insulation fault intelligent diagnosis system based on multi-information fusion, comprising a processor, a memory and a computer program stored in the memory, wherein the processor executes the computer program to specifically execute the steps of the high-voltage cable insulation fault intelligent diagnosis method based on multi-information fusion as described above.

[0099] A computer storage medium stores a plurality of instructions, the instructions are suitable for being loaded and executed by a processor to execute the steps of the high-voltage cable insulation fault intelligent diagnosis method based on multi-information fusion as described above.

[0100] The present application has the following beneficial effects:

[0101] 1、The hierarchical and multi-scale feature fusion method can fully exploit the complementary information between different modal features, generate a joint feature representation with higher expression ability, and thus improve the accuracy and reliability of fault diagnosis;

[0102] 2、The application utilizes the complementarity of multi-modal features, realizes the probability prediction and uncertainty quantification of fault types through the Bayesian inference framework, can effectively improve the accuracy and reliability of fault classification, and combines the representation ability of deep learning and the explainability of physical models, realizes high-precision fault positioning, and provides uncertainty quantification, providing a reliable basis for maintenance decision-making;

[0103] 3、The application realizes the intelligentization of the whole process from fault diagnosis to decision execution, reduces the cost of manual analysis, improves the efficiency and safety of cable operation and maintenance, and effectively reduces power outages and economic losses caused by cable faults. BRIEF DESCRIPTION OF DRAWINGS

[0104] Figure 1 The method flowchart of the application. DETAILED DESCRIPTION

[0105] The application will be further described in detail below in combination with the drawings and specific embodiments.

[0106] Embodiment 1

[0107] Reference Figure 1 In this embodiment, a high-voltage cable insulation fault intelligent diagnosis method based on multi-information fusion is improved, which includes the following steps:

[0108] S1: Obtain the operation multi-source data of the high-voltage cable, and construct a multi-source original data matrix;

[0109] S2: Preprocess the obtained operation multi-source data of the high-voltage cable to obtain a spatio-temporally aligned standardized data matrix;

[0110] S3: According to the spatio-temporally aligned standardized data matrix, feature extraction is performed to obtain a multi-dimensional feature vector, and a multi-modal deep network is used to learn the internal correlation between features to obtain joint multi-modal features;

[0111] S4: Construct a fault type recognition model based on Bayesian inference and Monte Carlo sampling, and perform fault type discrimination according to the joint multi-modal features to obtain fault type classification results;

[0112] S5: According to the joint multi-modal features, combine deep learning and physical models to obtain fault occurrence interval and positioning information;

[0113] S6: Based on the fault type classification results and the fault occurrence interval and positioning information, obtain the fault level, and generate a diagnosis report for early warning push.

[0114] In this embodiment, the operation multi-source data of high-voltage cable is acquired, including partial discharge signals (electric, acoustic, ultrasonic, light, etc. multi-channel), dielectric loss factor (using dielectric loss tester), cable temperature and humidity, environmental parameters (installing environmental sensor), operation and maintenance history data and fault label (database access), as follows:

[0115] A high-frequency current transformer is deployed at the cable joint to detect electric pulse signals, piezoelectric ceramic sensors are installed along the cable body to capture ultrasonic signals, and an optical fiber Bragg grating (FBG) array is embedded in the cable shielding layer to monitor light pulse signals in real time.

[0116] A digital dielectric loss tester is installed on the cable grounding wire to inject test signals through low-frequency superposition method (0.1 Hz / 10 kV) to obtain time series data of dielectric loss factor and capacitance change rate.

[0117] A distributed optical fiber temperature measurement system is used to spiral wrap along the cable outer sheath to obtain the axial temperature distribution matrix, a three-axis MEMS vibration sensor is installed at the cable joint to monitor X / Y / Z three-axis vibration acceleration time series signals, and environmental parameter time series data and gas concentration spectrum are obtained through environmental sensors.

[0118] The OPC UA protocol is used to access the SCADA system to extract the cable load curve, operation and maintenance history data and fault label.

[0119] The main control cabinet node collects the above data, the main control node is equipped with GPS and Beidou dual-mode clock source, and all sensor clocks are synchronized through PTP protocol. When any partial discharge channel detects a pulse greater than 50pC, all sensors are triggered to record 1ms time window data simultaneously.

[0120] A three-dimensional coordinate system of the cable is established to mark the installation position of each sensor (such as "HFCT@X=32.5m"). Temperature data is automatically mapped to the corresponding coordinate point.

[0121] In this embodiment, preprocessing includes filtering and denoising, multi-channel synchronization processing, distortion correction and outlier removal, as follows:

[0122] According to S1, the original multi-source data matrix X is obtained raw :

[0123] X raw ={x1(t),x2(t),...,x a (t),...,x n (t)};

[0124] Where x a (t) represents the signal of the a-th type of sensor, including electric signal, acoustic signal, temperature field, etc., and t is the time series.

[0125] Corresponding filtering and denoising processing is adopted for different sensor signals, elliptical filtering is adopted for electric signals, EMD denoising is adopted for acoustic signals, and wavelet threshold is adopted for optical signals to obtain denoised signal data X denoised ;

[0126] The denoised signal data X denoised is subjected to multi-channel synchronous processing distortion correction, the electric signal amplitude is corrected through a transmission line model, and the temperature field is compensated through heat conduction:

[0127] The physical quantity real value signal data X corrected is obtained.

[0128] Finally, combined with the local outlier factor and Mahalanobis distance, abnormal data caused by hardware failure or environmental interference is removed, and a standardized space-time alignment matrix X proc is obtained, including partial discharge (electric, acoustic, and optical signals), dielectric loss factor (tan δ), temperature field distribution, vibration spectrum, and environmental parameters.

[0129] In this embodiment, according to the standardized data matrix of space-time alignment, feature extraction is performed, multi-dimensional feature vectors are obtained and fused, and a joint multi-modal feature representation is obtained, as follows:

[0130] According to the standardized space-time alignment matrix X proc obtained in S2, features are extracted from each modal data, including partial discharge signal features, acoustic signal features, and temperature field features; the partial discharge signal adopts ResNet-18 to extract spatial features as partial discharge signal features; the ultrasonic signal adopts bidirectional LSTM to capture the evolution trend as acoustic signal features; and the temperature field adopts a Transformer encoder to model spatial dependence as temperature field features.

[0131] Based on the extracted modal features, Z-score standardization is adopted for standardization processing, and mutual information of each feature f i and the fault label y is calculated.

[0132] ;

[0133] Wherein, is the joint probability density; is the independent probability distribution of feature f i ; and p(y) is the prior probability of the fault label y.

[0134] Features F MI with Top-K (such as K=50) high mutual information are retained, and redundant features satisfying ∣p(fi,fj)∣>0.7 are deleted; p(f i ,fj ) is f i and f j ;

[0135] The multi-modal deep network learns the internal association between high mutual information features to obtain a high expression representation of the fusion features, and constructs a joint multi-modal feature h fusion .

[0136] In the embodiment, the multi-modal deep network includes a feature embedding layer, a feature interaction layer, an attention mechanism and a feature fusion layer, and specifically as follows:

[0137] The feature embedding layer maps the features of different modalities to a unified feature space:

[0138] h pd =W pd ·f pd +b pd ;

[0139] h sound =W sound ·f sound +b sound ;

[0140] h temp =W temp ·f temp +b temp ;

[0141] Wherein: f pd , f sound , f temp are feature vectors of partial discharge, acoustic signal and temperature field respectively; W pd , W sound , W temp are learnable weight matrices; b pd , b sound , b temp are bias vectors; h pd , h sound , h temp are feature vectors embedded by partial discharge, acoustic signal and temperature field respectively;

[0142] The feature interaction layer models the complex association between different modalities, and uses a multi-modal interaction network to learn the complex association Z between features:

[0143] Z=σ(W v ·[h pd ;h sound ;h temp ]+b v );

[0144] where [;] denotes the concatenation operation of features; W v and b v are the weights and bias of the feature interaction layer; and σ is the activation function.

[0145] The self-attention mechanism MultiHead is used to capture long-distance dependencies between features:

[0146] Q = W Q · Z, K = W K · Z, and V = W V · Z;

[0147] ;

[0148] ;

[0149] where Q, K, and V represent the query, key, and value matrices, respectively, and W Q , W K , and W V are the parameter matrices corresponding to Q, K, and V, respectively. , , are the learned projection matrices corresponding to Q, K, and V of the hth attention head; head h is the hth attention head; H is the number of attention heads; Attention represents the attention mechanism calculation function; W O is the output weight matrix; and Concat represents the concatenation function.

[0150] The importance of different modalities is adaptively weighted:

[0151] ;

[0152] where is the attention weight of each modality; W a1 , W a2 , W a3 and b a1 , b a2 , b a3 are learned parameters.

[0153] The feature fusion layer combines the results of multi-head attention and modality attention to generate the final joint multi-modal feature representation h fusion :

[0154] ;

[0155] ;

[0156] ;

[0157] where h attn is the output of the self-attention mechanism; h modal is the output of the modal attention mechanism; W f and b f are the weights and bias of the feature fusion layer.

[0158] In this embodiment, a fault type recognition model is constructed based on Bayesian inference and Monte Carlo sampling, and a fault type is discriminated according to the joint multi-modal features to obtain a fault type classification result, as follows:

[0159] First, the joint multi-modal features are subjected to nonlinear transformation and feature enhancement:

[0160] F enhanced = σ (W1·h fusion + b1) + h fusion ;

[0161] F attention = Softmax (W2·F enhanced + b2) 0 F enhanced ;

[0162] where 0 represents element-wise multiplication; F enhanced represents the output after nonlinear transformation; F attention represents the output after feature enhancement; W1 and b1 are the weight and bias of the nonlinear transformation, respectively; and W2 and b2 are the weight and bias of the feature enhancement, respectively.

[0163] A multi-scale feature extraction module is constructed and processed by maximum pooling MaxPool and average pooling AvgPool:

[0164] F local = MaxPool (W3·F attention + b3) ;

[0165] F global = AvgPool (W4·F attention + b4) ;

[0166] F multi-scale = Concat (F local , F global ) ;

[0167] where F local represents the output after maximum pooling; F global represents the output after average pooling; W3 and b3 are the weight and bias of the maximum pooling MaxPool, respectively; and W4 and b4 are the weight and bias of the average pooling AvgPool, respectively; and F multi-scale represents the extracted multi-scale features.

[0168] Constructing a classifier based on Bayesian inference, outputting a probability distribution :

[0169] ;

[0170] ;

[0171] ;

[0172] where, is random noise, is the probability density function of noise; and are the mean vector and standard deviation vector of class y, respectively; W σ and b σ are the weight and bias of the standard deviation prediction, respectively; and are the weight and bias of the mean prediction, respectively;

[0173] ;

[0174] ;

[0175] where, is the average probability expectation; T is the number of samples; is the random noise of the t-th sample; Uncertainty is the information entropy of the model prediction distribution; c represents the class c;

[0176] Simultaneously optimize classification and uncertainty estimation, the objective function is:

[0177] ;

[0178] where, is the cross-entropy loss; is the regularization loss; is the uncertainty constraint loss, using entropy regularization; , are weight coefficients;

[0179] Final classification decision, predict the fault type:

[0180] ;

[0181] where, is the probability of predicting the fault type as c.

[0182] In this embodiment, according to the joint multi-modal feature, the deep learning and the physical model are combined to obtain the interval and positioning information of fault occurrence, specifically as follows:

[0183] The spatial-temporal feature is extracted from the joint multi-modal feature h fusion :

[0184] F temp =TCN(h fusion );

[0185] F space =GCN([F temp ,A]);

[0186] Wherein, TCN is a time convolution network, capturing the time sequence feature F temp ; GCN is a graph convolution network, modeling the spatial relationship F space in combination with the adjacency matrix A;

[0187] The cable physical parameters are fused with the spatial-temporal feature to embed the positioning process:

[0188] F phys =Embed(R,L,C,G);

[0189] F joint =a'·F space +(1-a')·F phys ;

[0190] Wherein, R, L, C, G are the physical parameters of the cable, respectively, resistance, inductance, capacitance and conductance; Embed is a physical parameter embedding function; a' is a learnable fusion weight; F phys is a cable physical feature vector; F joint is a fusion feature;

[0191] The probability mapping from the feature to the position is constructed:

[0192] ;

[0193] ;

[0194] Wherein, p(l|F joint ) is the probability distribution of the position l; l i is the pre-defined i-th candidate position; is the predicted fault position; W loc and b loc are the weight and bias of the probability mapping respectively; N is the number of candidate positions.

[0195] In this embodiment, based on the fault type classification result and the interval and positioning information of fault occurrence, the fault level is obtained, and a diagnosis report is generated for early warning push, as follows:

[0196] In combination with historical data and expert experience, a fault level evaluation model is constructed, and based on the fault type classification result and the interval and positioning information of fault occurrence, the fault level is obtained.

[0197] For different types of faults, corresponding weights are set, for example, partial discharge fault is given a higher weight due to its potential severity of damaging cable insulation; while the weight of slight temperature anomaly is relatively low. At the same time, considering the location of fault occurrence, if the fault is located at the key node of the cable or the power supply line in the densely populated area, its influence range and potential harm are greater, which will also increase the evaluation score of fault level. Through weighted calculation of the comprehensive score of fault type and location, the fault is divided into different levels, such as emergency (need to be handled immediately), serious (need to be handled within 24 hours), and general (can be arranged for regular maintenance).

[0198] Specifically, the fault type is scored according to the severity (such as 1-10 points): important fault 10 points, general fault 5 points, and slight fault 2 points; the fault location is scored according to the importance (such as 1-10 points): core component 10 points, key component 6 points, and secondary component 3 points, then the comprehensive score = (fault type score × fault type weight) + (fault location score × fault location weight).

[0199] For example, when a serious partial discharge fault is detected at the cable joint, combined with its location on the power supply line in the urban core area, the evaluation model calculates and determines it as an emergency level fault.

[0200] According to the fault type classification result, the interval and positioning information of fault occurrence, and the fault level, a detailed diagnosis report is generated; the diagnosis report content includes multi-source evidence chain, showing the original information and processing process of multi-source data such as partial discharge signal, acoustic signal, and temperature field used for diagnosis; the diagnosis conclusion is clearly given, such as “after analysis, the cable at XX location has XX type fault, and the fault level is urgent”; at the same time, according to the fault situation and historical maintenance experience, disposal suggestions are provided, such as “immediately cut off the power supply in this area, dispatch professional maintenance team to replace the fault cable joint, and conduct comprehensive detection on the surrounding cable”. In addition, the report can also include information such as fault occurrence probability and impact range prediction, presented in the form of intuitive charts, to facilitate the operation and maintenance personnel to quickly understand the overall situation of the fault.

[0201] Different pre-warning mechanisms are set according to the fault levels. For the emergency level fault, strong reminding notifications such as pop-up windows, vibration, voice broadcast, etc. are immediately sent to the operation and maintenance managers through the Web system and mobile APP to ensure that the information is timely conveyed. For the serious level fault, detailed fault information and disposal suggestions are pushed within the preset time (10 minutes). For the general level fault, messages can be pushed within the preset time (1 hour), and they are included in the daily inspection task reminders.

[0202] For example: a sudden fault occurs in the operation of a 10kV high-voltage cable line in a certain city. The operation and maintenance center collects multi-source data, including: partial discharge signals (time series electrical data), temperature distribution data (infrared thermal imaging), geographic location information (GIS coordinates), historical operation and maintenance records, and current / voltage waveforms. Based on the multi-information fusion-based high-voltage cable insulation fault intelligent diagnosis method of the present scheme, the specific process is as follows:

[0203] Obtain multi-source data of high-voltage cables (as shown in Table 1), and construct a multi-source original data matrix.

[0204] Table 1 Original data:

[0205]

[0206] Output joint multi-modal features, h fusion =[0.72, 0.10,..., 0.94], which represents the multi-dimensional comprehensive health state feature of the space-time point.

[0207] And through Bayesian inference Monte Carlo sampling, the fault type recognition network output (fault classification probability) is obtained, as shown in Table 2.

[0208] Table 2 Fault classification probability

[0209]

[0210] Infer that the high probability is an insulation breakdown event. According to the features, the deep neural network infers the high probability fault interval, which is “1.4km to 1.8km from substation A”. Finally, the fault level is automatically judged as an emergency level fault, which needs to be urgently repaired, and a diagnosis report is automatically generated, which is pushed to the operation and maintenance APP and the large screen, and suggestions are given:

[0211] Diagnosis report abstract:

[0212] Fault type: insulation breakdown;

[0213] Fault interval: 1.4km~1.8km;

[0214] Fault level: emergency level fault;

[0215] Speculation: Local discharge and high temperature superposition, insulation continues to deteriorate;

[0216] Recommended measures: Isolate the fault section immediately, send staff to the site for detection and confirmation, and prepare a standby cable.

[0217] The embodiment also provides a high-voltage cable insulation fault intelligent diagnosis system based on multi-information fusion, including a processor, a memory, and a computer program stored in the memory, and when the processor executes the computer program, the steps in the high-voltage cable insulation fault intelligent diagnosis method based on multi-information fusion are specifically executed.

[0218] The embodiment also provides a computer storage medium, which stores a plurality of instructions, and the instructions are suitable for being loaded and executed by a processor to execute the steps of the high-voltage cable insulation fault intelligent diagnosis method based on multi-information fusion.

[0219] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can adopt a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer usable program codes.

[0220] The present application is described with reference to flowcharts and / or block diagrams of the method, device (system), and computer program product according to the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of the flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to produce a machine, so that the instructions executed by the computer or other programmable data processing devices produce a device implemented in the flowcharts and / or block diagrams. Figure 1 The function specified in one flow or multiple flows and / or blocks Figure 1 The device for implementing the function specified in one block or multiple blocks.

[0221] These computer program instructions can also be stored in a computer readable storage medium, which can guide the computer or other programmable data processing devices to work in a specific way, so that the instructions stored in the computer readable storage medium produce a product including instruction devices, which implement the flowcharts and / or block diagrams. Figure 1 The function specified in one flow or multiple flows and / or blocks Figure 1 The device for implementing the function specified in one block or multiple blocks.

[0222] These computer program instructions can also be loaded into a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flowcharts and / or blocks Figure 1 one or more flowcharts and / or blocks

[0223] The above descriptions are only the preferred embodiments of the present application, not intended to limit the present application to other forms described. Any person skilled in the art may make changes or modifications to the above-described technical contents as equivalent embodiments without departing from the technical solutions of the present application. However, any simple modification, equivalent change and modification of the above embodiments according to the technical essence of the present application without departing from the technical solutions of the present application still belongs to the protection scope of the technical solutions of the present application.

Claims

1. A multi-information fusion-based intelligent diagnosis method for high-voltage cable insulation faults, characterized in that, The method comprises the following steps: S1: Obtain the operation multi-source data of the high-voltage cable, and construct a multi-source original data matrix; S2: Preprocess the obtained operation multi-source data of the high-voltage cable to obtain a spatio-temporally aligned standardized data matrix; S3: According to the spatio-temporally aligned standardized data matrix, feature extraction is performed to obtain a multi-dimensional feature vector, and a multi-modal deep network is used to learn the inherent correlation between the features to obtain joint multi-modal features; The multi-modal deep network comprises a feature embedding layer, a feature interaction layer, an attention mechanism and a feature fusion layer, and specifically as follows: The feature embedding layer maps the features of different modalities to a unified feature space: h pd =W pd ·f pd +b pd ; h sound =W sound ·f sound +b sound ; h temp =W temp ·f temp +b temp ; Where: f pd , f sound , f temp are the feature vectors of partial discharge, acoustic signal and temperature field, respectively; W pd , W sound , W temp are learnable weight matrices; b pd , b sound , b temp are bias vectors; h pd , h sound , h temp are the feature vectors after embedding of partial discharge, acoustic signal and temperature field respectively; The feature interaction layer models the complex correlation between different modal features, and uses a multi-modal interaction network to learn the complex correlation between the features Z: Z = σ(W v · [h pd ; h sound ; h temp ] + b v ); where [; ] denotes the concatenation operation of features; W v and b v are the weights and bias of the feature interaction layer; σ is the activation function. The self-attention mechanism MultiHead is used to capture the long-distance dependence relationship between the features: Q = W Q • Z, K = W K • Z, V = W V • Z; ； ； where Q, K, V represent table query, key and value matrix respectively, W Q , K , V is the parameter matrix corresponding to Q, K, V; , , is the learned projection matrix corresponding to Q, K, V of the hth attention head; head h is the hth attention head; H is the number of attention heads; Attention represents the attention mechanism calculation function; W O is the output weight matrix; Concat represents the concatenation function; The importance of different modalities is adaptively weighted: ； wherein, are attention weights for each modality; W a1 , W a2 , W a3 and b a1 , b a2 , b a3 are learnable parameters; The feature fusion layer combines the results of the multi-head attention and the modality attention to generate the final joint multi-modal feature representation h fusion : ； ； ； wherein h attn is the output of the self-attention mechanism; h modal is the output of the modal attention mechanism; W f and b f are the weights and bias of the feature fusion layer; S4: A fault type recognition model is constructed based on Bayesian inference and Monte Carlo sampling, and according to the joint multi-modal features, fault type discrimination is performed to obtain fault type classification results; S5: According to the joint multi-modal features, deep learning and physical models are combined to obtain fault occurrence interval and positioning information, specifically as follows: spatio-temporal features from the joint multi-modal features h fusion extracted from the joint multi-modal features h F temp = TCN(h fusion ); F space = GCN([F temp , A]); where TCN is a time convolutional network that captures temporal features F temp ; GCN is a graph convolutional network that models spatial relationships F space ; Fuse the cable physical parameters with the spatio-temporal features to embed the positioning process: F phys = Embed(R, L, C, G); F joint = a' · F space + (1 - a') · F phys ; Wherein, R, L, C, G are physical parameters of the cable, respectively resistance, inductance, capacitance, conductance; Embed is the physical parameter embedding function; a' is the learnable fusion weight; F phys is the cable physical feature vector; F joint is the fusion feature; Construct a probability mapping from features to positions: ； ； where p(l|F) is the probability distribution of the location l; l joint is the predefined ith candidate location; i is the predicted failure location; W loc and b loc are the weight and bias of the probability mapping, respectively; N is the number of candidate locations;​ S6: Based on the fault type classification results and the fault occurrence interval and positioning information, the fault level is obtained, and a diagnosis report is generated for early warning push.

2. The method according to claim 1, wherein, The operation multi-source data of the high-voltage cable comprises partial discharge signals, dielectric loss factors, cable temperature and humidity, environmental parameters, operation and maintenance historical data and fault labels, specifically as follows: A high-frequency current transformer is deployed at the cable joint to detect electric pulse signals, piezoelectric ceramic sensors are installed along the cable body to capture ultrasonic signals, and an optical fiber Bragg grating array is embedded in the cable shielding layer to monitor optical pulse signals in real time; A digital dielectric loss tester is installed on the cable grounding wire, a test signal is injected through a low-frequency superposition method, and dielectric loss factor and capacitance change rate time series data are obtained; A distributed optical fiber temperature measurement system is spirally wound along the cable outer sheath to obtain an axial temperature distribution matrix, a three-axis MEMS vibration sensor is installed at the cable joint to monitor X / Y / Z three-axis vibration acceleration time series signals, and environmental parameter time series data and gas concentration spectrum are obtained through an environmental sensor; The cable load curve, operation and maintenance historical data and fault labels are extracted through the OPC UA protocol to access the SCADA system; The above data is collected through the main control cabinet node, the main control node is equipped with GPS and Beidou dual-mode clock sources, all sensor clocks are synchronized through the PTP protocol, and when any partial discharge channel detects a >50pC pulse, all sensors are triggered to record 1ms window data synchronously; A cable three-dimensional coordinate system is established, and the installation position of each sensor is labeled, and the temperature data is automatically mapped to the corresponding coordinate point.

3. The method according to claim 2, characterized in that, The preprocessing comprises filtering and denoising processing, multi-channel synchronization processing distortion correction and abnormal value screening, specifically as follows: According to S1, an original multi-source data matrix X is acquired raw : X raw = {x1(t), x2(t),..., x a (t),..., x n (t)}; wherein x a (t) denotes a sensor signal of type a, including electrical signals, acoustic signals, temperature fields, t is a time series; Corresponding filtering and denoising processing is adopted for different sensor signals, elliptical filtering is adopted for electric signals, EMD denoising is adopted for sound signals, and wavelet threshold is adopted for light signals to obtain denoised signal data X denoised ; The signal data X after noise reduction denoised , using multi-channel synchronous processing distortion correction, through the transmission line model correction of electric signal amplitude, through the heat conduction compensation temperature field: obtaining physical quantity true value signal data X corrected ; Finally, combined with local outlier factor and Mahalanobis distance, the abnormal data caused by hardware failure or environmental disturbance is removed to obtain the standardized space-time alignment matrix X proc .

4. The method according to claim 1, wherein, The normalized data matrix according to the spatio-temporal alignment is subjected to feature extraction, multi-dimensional feature vectors are obtained and fused to obtain a joint multi-modal feature representation, specifically as follows: The standardized spatio-temporal alignment matrix X obtained according to S2 proc Features are extracted from each modality data, including partial discharge signal features, acoustic signal features and temperature field features; the partial discharge signal adopts ResNet-18 to extract spatial features as partial discharge signal features; the ultrasonic signal adopts bidirectional LSTM to capture the evolution trend as acoustic signal features; and the temperature field adopts a Transformer encoder to model spatial dependence as temperature field features; Based on the extracted modal features, Z-score standardization is used for standardization processing, and each feature f i Mutual information with fault label y : ； wherein is the joint probability density; is the independent probability distribution of the feature f i p(y) is the prior probability of the failure label y; Features F reserved Top-K high mutual information MI , delete redundant features satisfying ∣p(f i ,f j )∣>0.7, where p(f i ,f j ) is the joint probability distribution of f i and f j ; The multi-modal deep network is used to learn the internal correlation between high mutual information features, to obtain a high expression representation of the fusion features, and to construct a joint multi-modal feature h fusion .

5. The method according to claim 1, wherein The fault type recognition model is constructed based on Bayesian inference and Monte Carlo sampling, and the joint multi-modal feature is used for fault type discrimination to obtain a fault type classification result, specifically as follows: First, the joint multi-modal feature is subjected to nonlinear transformation and feature enhancement: F enhanced =σ(W1·h fusion +b1)+h fusion ; F attention = Softmax(W2 · F enhanced + b2) 0 F enhanced ; wherein, ⊙ represents element-wise multiplication; F enhanced represents the output after nonlinear transformation; F attention represents the output after feature enhancement; W1 and b1 are the weight and bias of the nonlinear transformation, respectively; W2 and b2 are the weight and bias of the feature enhancement, respectively. A multi-scale feature extraction module is constructed, and MaxPool and AvgPool are used for processing: F local =MaxPool(W3·F attention +b3); F global =AvgPool(W4·F attention +b4); F multi-scale = Concat(F local , F global ) ; wherein F local represents the output after max-pooling; F global represents the output after average-pooling; W3, b3 are the weight and bias of the max-pooling MaxPool, respectively; W4, b4 are the weight and bias of the average-pooling AvgPool, respectively; F multi-scale represents the extracted multi-scale feature; constructing a classifier based on bayesian inference, outputting a probability distribution : ； ； ； wherein, is random noise, is the probability density function of the noise; and are the mean vector and standard deviation vector for class y, respectively; W σ and b σ are the weight and bias for the standard deviation prediction, respectively; and are the weight and bias for the mean prediction, respectively; The classification uncertainty is estimated by Monte Carlo sampling: ； ； wherein, is the average probability expectation; T is the number of samples; is the random noise of the t-th sample; Uncertainty is the information entropy of the model prediction distribution; c denotes the class c; Simultaneously optimize classification and uncertainty estimation, objective function is: ； wherein, is a cross-entropy loss; is a regularization loss; is an uncertainty constraint loss, employing entropy regularization; , is a weight coefficient; Finally, the classification decision is made to predict the fault type: ； wherein, is the probability of predicting the failure type as c.

6. The method according to claim 1, wherein Based on the fault type classification result, the interval and positioning information of the fault occurrence, the fault level is obtained, and a diagnosis report is generated for early warning push, specifically as follows: A fault level evaluation model is constructed based on historical data and expert experience, and the fault level is obtained based on the fault type classification result and the interval and positioning information of the fault occurrence; A detailed diagnosis report is generated based on the fault type classification result, the interval and positioning information of the fault occurrence, and the fault level; the diagnosis report content includes a multi-source evidence chain, showing the original information and processing process of the partial discharge signal, acoustic signal, and temperature field multi-source data used for diagnosis; at the same time, disposal suggestions are provided based on the fault condition and historical maintenance experience; different early warning mechanisms are set according to the fault level, and for urgent level faults, strong reminder notifications are immediately sent to the operation and maintenance personnel through the Web system and mobile APP; For serious level faults, detailed fault information and disposal suggestions are pushed within a preset time; for general level faults, messages can be pushed within a preset time, and they are included in the daily inspection task reminders.

7. A multi-information fusion based intelligent diagnosis system for high voltage cable insulation faults, characterized in that, The computer program is stored in the memory, and the processor executes the computer program to specifically execute the steps of the method.

8. A computer storage medium, characterized in that, The computer storage medium stores a plurality of instructions, which are suitable for being loaded and executed by the processor to execute the steps of the method.

Images