Electrical equipment multi-sensor fault feature fusion diagnosis method

Abstract

The invention relates to a multi-sensor fault feature fusion diagnosis method for electrical equipment, which comprises the following steps: synchronously acquiring operation data of the electrical equipment through a vibration sensor, a temperature sensor, a current sensor and an ultrasonic sensor, dynamically adjusting the sampling frequency according to the physical characteristics of each sensor, and the sampling rate of the temperature signal is not lower than 1Hz. Through a multi-source sensor data synchronous acquisition and time sequence alignment technology and a signal alignment method combining a dynamic time warping (DTW) algorithm and Hilbert-Huang transformation, the problem of time asynchronization of heterogeneous sensor data such as vibration and temperature is solved, so that the time alignment precision of multi-source data is improved, the feature extraction accuracy is improved, and the accuracy of feature extraction is improved. Through hierarchical feature extraction and graph convolutional network fusion, a feature incidence matrix based on mutual information is constructed, deep correlation between vibration signal TKEO features and cross-modal features such as current harmonics is mined by using GCN, the feature dimension is reduced, and the fault feature separability index is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electrical equipment fault diagnosis, in particular to a multi-sensor fault feature fusion diagnosis method for electrical equipment. BACKGROUND

[0002] The fault diagnosis of electrical equipment (such as transformers, circuit breakers, motors, etc.) mainly relies on sensor monitoring technology. Common sensors include: vibration sensors (detecting mechanical looseness, bearing wear), temperature sensors (monitoring overheating, poor contact), partial discharge sensors (detecting insulation deterioration), current / voltage sensors (analyzing electrical abnormalities), acoustic / ultrasonic sensors (identifying partial discharge or mechanical impact). Traditional methods usually use single sensor analysis, such as vibration signal spectrum analysis (FFT, wavelet transform), infrared thermal imaging temperature monitoring, and partial discharge pattern recognition (PRPD spectrum).

[0003] For example, the rotating machinery equipment fault diagnosis method based on multi-sensor related feature fusion disclosed in Chinese patent 202311080133.X fuses fault information in time and frequency domains to avoid the defects of existing technologies and further improve the accuracy of fault diagnosis.

[0004] However, the existing multi-sensor fault feature fusion diagnosis method for electrical equipment still has the following defects:

[0005] 1. The traditional method does not consider the physical correlation between features (such as the coupling relationship between vibration and temperature), and direct fusion of high-dimensional features leads to "dimension disaster", reducing the efficiency of the model.

[0006] 2. Fixed weights cannot adapt to different operating states of the equipment (such as the vibration weight should be higher than the steady state during the starting stage), and do not consider the dynamic changes in sensor reliability (such as the accuracy of temperature sensors decreases in high temperature environment).

[0007] 3. The calculation overhead of complex fusion algorithms is large, making it difficult to deploy on edge devices.

[0008] Therefore, a multi-sensor fault feature fusion diagnosis method combining signal processing, feature engineering, and machine learning is proposed to solve the above problems. SUMMARY

[0009] In view of the deficiencies of the prior art, the application provides a multi-sensor fault feature fusion diagnosis method for electrical equipment, which has the advantages of improving diagnosis accuracy, strong anti-interference ability, real-time diagnosis and the like, and solves the problems of the prior art that the physical correlation between features is not considered, direct fusion of high-dimensional features leads to "dimension disaster", reduces model efficiency, fixed weights cannot adapt to different operating states of equipment, sensor reliability dynamic changes are not considered, complex fusion algorithm has large calculation overhead and is difficult to deploy on edge devices.

[0010] To achieve the above object, the application provides the following technical scheme: a multi-sensor fault feature fusion diagnosis method for electrical equipment, comprising the following steps:

[0011] S1, multi-source sensor data synchronous acquisition

[0012] The vibration sensor, temperature sensor, current sensor and ultrasonic sensor are used to synchronously acquire the electrical equipment operation data, and the sampling frequency is dynamically adjusted according to the physical characteristics of each sensor, wherein the vibration signal sampling rate is not less than 100 kHz, and the temperature signal sampling rate is not less than 1 Hz;

[0013] S2, time sequence alignment and preprocessing

[0014] An asynchronous signal alignment algorithm based on dynamic time warping (DTW) is used to eliminate the time sequence deviation caused by sensor response delay, and wavelet denoising and normalization processing are performed on the sensor data;

[0015] S3, hierarchical feature extraction:

[0016] S3.1, primary feature layer: Teager-Kaiser energy operator (TKEO) time-frequency features of the vibration signal, gradient change rate of the temperature signal, harmonic distortion rate of the current signal and PRP D spectrum feature of the ultrasonic signal are extracted;

[0017] S3.2, advanced feature layer: a cross-sensor feature correlation matrix is constructed through a graph convolution network (GCN), nodes represent each feature parameter, and edge weights are calculated through a mutual information algorithm;

[0018] S4, dynamic weight fusion:

[0019] S4.1, real-time weight coefficients are generated based on a double-channel LSTM network, a first channel inputs current sensor confidence (calculated through residual analysis), and a second channel inputs device historical fault mode library data;

[0020] S4.2, a Softmax function is used to constrain the weight output, so that the sum of the weights of each sensor is 1;

[0021] S5, fault diagnosis decision

[0022] The fused features are input into a pre-trained deep residual network (ResNet-50), and the fault type and probability distribution are output, and a fault traceability report is generated.

[0023] Further, the time sequence alignment algorithm in S2 specifically includes:

[0024] S2.1, taking the current signal as a reference signal, calculating the cross-correlation coefficient of other sensor signals and the reference signal;

[0025] S2.2, when the correlation coefficient is lower than the threshold value 0.85, a dynamic time warping algorithm is used for nonlinear stretching compensation, and the compensation accuracy is controlled within ±5 sampling points;

[0026] S2.3, performing Hilbert-Huang transform (HHT) on the vibration signal, and extracting the instantaneous frequency as an auxiliary alignment feature.

[0027] Further, the graph convolution network construction method in S3.2 includes:

[0028] S3.2a, define the feature map G=(V,E), where the vertex set V={v1,v2,...,vn} represents n feature parameters, and the edge set E is generated by the following rules:

[0029] If the Pearson correlation coefficient of two feature parameters |ρ|>0.7, a connection edge is established, and the edge weight w=1-|ρ|;

[0030] S3.2b, use two-layer GCN for feature aggregation, the first layer activation function is ReLU, and the second layer is Sigmoid;

[0031] S3.2c, output the topological feature vector F_G∈R 128 , and the final fused features are formed after splicing with the original features.

[0032] Further, the dynamic weight generation algorithm in S4 specifically includes:

[0033] S4a, design a two-channel LSTM network structure, each LSTM unit contains 256 hidden nodes;

[0034] S4b, the input dimension of the first channel is a 4-dimensional tensor, including: the signal-to-noise ratio (SNR) of each sensor signal, the number of recent fault alarms, data integrity, and environmental interference intensity;

[0035] S4c, the second channel input device historical fault mode code adopts One-hot encoding method, and the dimension is 5 times the number of fault categories;

[0036] S4d, the weight adjustment period is configurable, and the default setting is 10 seconds.

[0037] Further, the depth residual network improvement in S5 includes:

[0038] S5.1, add attention mechanism module (CBAM) on the basis of the original ResNet-50, and insert it after each residual block;

[0039] S5.2, the output layer adopts an adaptive loss function:

[0040] l = a · l CE +(1-a)·l Focal

[0041] Wherein, α is dynamically adjusted according to the sample balance degree, ranging from 0.3 to 0.7;

[0042] S5.3, deploy model distillation technology to migrate the knowledge of the teacher model (100 million parameters) to the student model (10 million parameters).

[0043] Further, it also includes a diagnostic system for implementing a multi-sensor fault feature fusion diagnosis method of an electrical device, and the diagnostic system is specifically as follows:

[0044] Edge computing module: adopt FPGA+ARM heterogeneous architecture, FPGA is responsible for signal preprocessing and feature extraction, and ARM runs lightweight diagnosis model;

[0045] Dynamic weight controller: integrate special ASIC chip to realize hardware acceleration of LSTM network, with delay less than 2ms;

[0046] Cloud collaborative platform:

[0047] 1) Data center: store at least 5 years of device full life cycle data, and use time series database (InfluxDB) to optimize query efficiency;

[0048] 2) Knowledge graph engine: build a fault causal reasoning network based on Neo4j, supporting multi-dimensional association query;

[0049] Human-computer interaction terminal: provides AR visualization interface, supports fault part three-dimensional positioning and maintenance guidance.

[0050] Further, the specific implementation of the edge computing module is as follows:

[0051] T1, FPGA part design special pipeline:

[0052] Vibration signal processing channel: contains FIR filter (cutoff frequency 10kHz), Hilbert transformer, TKEO calculation unit;

[0053] Temperature signal processing channel: integrated moving average window (window length adjustable), first-order difference calculation module;

[0054] T2, ARM part of the pruned TinyML model, model size is not more than 3MB, inference time ≤50ms;

[0055] The diagnostic system deployment optimization method is as follows:

[0056] U1, network topology optimization:

[0057] The edge computing node is arranged in the transformer substation, and the hop number with any sensor is ensured to be ≤2;

[0058] TSN (Time Sensitive Network) is adopted to ensure that the critical data transmission delay is <10ms;

[0059] U2, energy consumption management:

[0060] According to the device load rate, the sampling frequency is dynamically adjusted, and the vibration sampling rate is reduced to 50kHz when the load is light;

[0061] A solar power supply unit is used to supply power for the wireless sensor node.

[0062] Further, the sensor networking device of the diagnostic system is as follows:

[0063] V1, industrial Ethernet (EtherCAT) and wireless (LoRaWAN) dual-mode communication architecture is adopted, wired mode is used for high real-time data (vibration, current), and wireless mode is used for low-frequency data (temperature);

[0064] V2, built-in self-diagnosis unit, continuously monitors the sensor health status, and automatically switches to redundant sensors when sensor failure is detected;

[0065] V3, provide IP67 protection level shell, built-in temperature compensation circuit, working temperature range-40℃~85℃.

[0066] Further, it further includes an electrical equipment fault knowledge graph construction method, and the diagnostic system is applied:

[0067] Entity definition: including three types of entities including device components (winding, insulator), fault modes (mechanical wear, partial discharge), and characteristic parameters (vibration amplitude, temperature rise rate);

[0068] Relationship modeling:

[0069] Causal chain: "vibration amplitude exceeds threshold" → "causes" → "bearing loosening" → "triggers" → "temperature anomaly";

[0070] Probability correlation: P(insulation aging | THD>5%) = 0.82 is calculated through the Bayesian network;

[0071] Dynamic updating mechanism: automatically trigger atlas reconstruction every 100 new diagnostic records.

[0072] Further, it further includes an electrical equipment fault prediction method, which realizes:

[0073] 1) Construct an LSTM-Attention prediction model, and the input window length is 30 sampling periods;

[0074] 2) Define the device health index (HI):

[0075]

[0076] Where w i is a dynamic weight, f i is the i-th normalized feature;

[0077] 3) Trigger an early warning when HI decreases continuously for 5 times and the slope is greater than 0.1.

[0078] Compared with the prior art, the technical scheme of the present application has the following beneficial effects:

[0079] 1) The present application solves the time asynchronous problem of heterogeneous sensor data such as vibration and temperature through multi-source sensor data synchronous acquisition and time alignment technology, adopts a signal alignment method combining dynamic time warping (DTW) algorithm and Hilbert-Huang transform, improves the time alignment accuracy of multi-source data, and improves the feature extraction accuracy. Through hierarchical feature extraction and graph convolution network fusion, a feature correlation matrix based on mutual information is constructed, the deep correlation between vibration signal TKEO features and current harmonics and other cross-modal features is mined using GCN, the feature dimension is reduced, and the fault feature separability index is improved. Through dynamic weight generation technology, a double-channel LSTM network is used to calculate the sensor weight in real time, combined with residual analysis and historical fault mode library, the current sensor weight is automatically adjusted from 0.3 to 0.72 under the overload working condition of the transformer, so that the robustness of the diagnostic system is improved when the sensor is partially failed.

[0080] 2、The application improves the deep residual network, embeds the CBAM attention module in the ResNet-50, and adopts the adaptive loss function, so that the identification accuracy of the rare fault with insufficient sample quantity is improved, and the reasoning speed is accelerated through the model distillation technology. Through the heterogeneous architecture of the edge computing module, the TKEO feature extraction pipeline accelerated by FPGA reduces power consumption and delay compared with pure CPU implementation, and meets the real-time monitoring demand of the circuit breaker opening and closing process. Through the dual-mode communication sensor networking, the EtherCAT+LoRaWAN hybrid networking scheme ensures the real-time transmission of vibration signals, prolongs the battery life of wireless temperature nodes, and reduces the wiring cost.

[0081] 3、The application shortens the fault root cause analysis time from manual diagnosis through the fault knowledge graph construction and the causal reasoning engine based on the Bayesian network, and improves the maintenance scheme recommendation accuracy. Through the health index prediction model, the LSTM-Attention model predicts the motor bearing failure trend 3-5 weeks in advance, compared with the traditional vibration threshold alarm, the number of unplanned shutdowns is reduced, and the annual maintenance cost is reduced. Through TSN network deployment, the time sensitive network guarantees the end-to-end transmission delay of key data <10ms, and in the strong electromagnetic interference environment of the substation, the data packet loss rate is reduced. BRIEF DESCRIPTION OF DRAWINGS

[0082] Figure 1 A flowchart of the electrical equipment multi-sensor fault feature fusion diagnosis method is shown in the figure.

[0083] Figure 2 A flowchart of the electrical equipment multi-sensor fault feature fusion diagnosis method is shown in the figure. DETAILED DESCRIPTION

[0084] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0085] Please refer to Figures 1-2 The electrical equipment multi-sensor fault feature fusion diagnosis method in the embodiment includes the following steps:

[0086] S1, multi-source sensor data synchronous acquisition

[0087] The electrical equipment operation data is synchronously collected by vibration sensors, temperature sensors, current sensors, and ultrasonic sensors, and the sampling frequency is dynamically adjusted according to the physical characteristics of each sensor, wherein the vibration signal sampling rate is not less than 100 kHz, and the temperature signal sampling rate is not less than 1 Hz;

[0088] S2, time alignment and pretreatment

[0089] An asynchronous signal alignment algorithm based on dynamic time warping (DTW) is adopted to eliminate the time deviation caused by sensor response delay, and wavelet denoising and normalization processing are performed on the data of each sensor. The time alignment algorithm specifically includes:

[0090] S2.1, taking the current signal as the reference signal, calculating the cross-correlation coefficient of other sensor signals and the reference signal;

[0091] S2.2, when the correlation coefficient is lower than the threshold value 0.85, the dynamic time warping algorithm is used for nonlinear stretching compensation, and the compensation accuracy is controlled within ±5 sampling points;

[0092] S2.3, Hilbert-Huang transform (HHT) is performed on the vibration signal to extract the instantaneous frequency as an auxiliary alignment feature;

[0093] S3, hierarchical feature extraction:

[0094] S3.1, primary feature layer: extracting Teager-Kaiser energy operator (TKEO) time-frequency features of vibration signals, gradient change rate of temperature signals, harmonic distortion rate of current signals, and PRP D spectrum feature of ultrasonic signals;

[0095] S3.2, high-level feature layer: a cross-sensor feature association matrix is constructed through a graph convolution network (GCN), nodes represent each feature parameter, and edge weights are calculated through a mutual information algorithm. The graph convolution network construction method includes:

[0096] S3.2a, define a feature graph G=(V,E), where the vertex set V={v1,v2,...,vn} represents n feature parameters, and the edge set E is generated by the following rules:

[0097] If the Pearson correlation coefficient |ρ| of two feature parameters is greater than 0.7, a connection edge is established, and the edge weight w=1-|ρ|;

[0098] S3.2b, two-layer GCN is used for feature aggregation, the first layer activation function is ReLU, and the second layer is Sigmoid;

[0099] S3.2c, output topological feature vector F_G∈R 128 , and the final fusion feature is formed after splicing with the original feature;

[0100] S4、Dynamic weight fusion:

[0101] S4.1, Real-time weight coefficient is generated based on a dual-channel LSTM network, the first channel inputs the current sensor confidence (calculated by residual analysis), and the second channel inputs the device historical fault mode library data;

[0102] S4.2, Softmax function is used to constrain the weight output, ensuring that the sum of the weights of each sensor is 1

[0103] The dynamic weight generation algorithm specifically includes:

[0104] S4a, Design a dual-channel LSTM network structure, each LSTM unit contains 256 hidden nodes;

[0105] S4b, The first channel inputs a 4-dimensional tensor, including: signal-to-noise ratio (SNR) of each sensor signal, recent fault alarm times, data integrity, and environmental interference intensity;

[0106] S4c, The second channel inputs the device historical fault mode encoding, using One-hot encoding method, with a dimension of 5 times the number of fault categories;

[0107] S4d, The weight adjustment period is configurable, with a default setting of 10 seconds;

[0108] S5, Fault diagnosis decision

[0109] The fused features are input into a pre-trained deep residual network (ResNet-50), which outputs the fault type and probability distribution, and generates a fault traceability report. The improvements of the deep residual network include:

[0110] S5.1, Add attention mechanism module (CBAM) to each residual block after the original ResNet-50;

[0111] S5.2, The output layer uses an adaptive loss function:

[0112] l = a • l CE +(1-a)•l Focal

[0113] Where α is dynamically adjusted according to the sample balance, ranging from 0.3 to 0.7;

[0114] S5.3, Deploy model distillation technology to transfer the knowledge of the teacher model (100 million parameters) to the student model (10 million parameters).

[0115] Specifically, it also includes a diagnostic system, which is as follows:

[0116] Edge computing module: adopts FPGA+ARM heterogeneous architecture, FPGA is responsible for signal preprocessing and feature extraction, ARM runs lightweight diagnostic model, the specific implementation of the edge computing module is as follows:

[0117] T1, FPGA part design special pipeline:

[0118] Vibration signal processing channel: contains FIR filter (cutoff frequency 10 kHz), Hilbert transformer, TKEO calculation unit;

[0119] Temperature signal processing channel: integrated moving average window (window length adjustable), first-order difference calculation module;

[0120] T2, ARM part deploys pruned TinyML model, model size is not more than 3MB, inference time ≤50ms;

[0121] Dynamic weight controller: integrated with special ASIC chip, realizes hardware acceleration of LSTM network, delay is less than 2ms;

[0122] Cloud collaborative platform:

[0123] 1) Data center: store at least 5 years of equipment full life cycle data, use time series database (InfluxDB) to optimize query efficiency;

[0124] 2) Knowledge graph engine: build fault causal reasoning network based on Neo4j, support multi-dimensional association query;

[0125] Human-computer interaction terminal: provides AR visualization interface, supports three-dimensional positioning and maintenance guidance of fault parts.

[0126] The specific implementation of the diagnostic system deployment optimization method is as follows:

[0127] U1, network topology optimization:

[0128] Arrange edge computing nodes in the substation, ensure that the hop number with any sensor is ≤2;

[0129] Use TSN (time sensitive network) to ensure that the critical data transmission delay is <10ms;

[0130] U2, energy consumption management:

[0131] Adjust the sampling frequency dynamically according to the device load rate, reduce the vibration sampling rate to 50kHz when the load is light;

[0132] Use solar power supply unit to supply power for wireless sensor nodes.

[0133] Specifically, the sensor networking device of the diagnostic system is as follows:

[0134] V1, dual-mode communication architecture with industrial Ethernet (EtherCAT) and wireless (LoRaWAN), wired mode for high real-time data (vibration, current), wireless mode for low-frequency data (temperature);

[0135] V2, built-in self-diagnosis unit, continuously monitors sensor health status, automatically switches to redundant sensor when sensor failure is detected;

[0136] V3, provide IP67 protection level shell, built-in temperature compensation circuit, working temperature range-40℃~85℃.

[0137] In this embodiment, it also includes an electrical equipment fault knowledge graph construction method:

[0138] Entity definition: including three types of entities including device components (winding, insulator), failure modes (mechanical wear, partial discharge), and characteristic parameters (vibration amplitude, temperature rise rate);

[0139] Relationship modeling:

[0140] Causal chain: "vibration amplitude exceeds threshold" → "causes" → "loose bearing" → "triggers" → "temperature anomaly";

[0141] Probability association: through Bayesian network calculation P(insulation aging | THD>5%) = 0.82;

[0142] Dynamic updating mechanism: automatically trigger graph reconstruction every 100 new diagnosis records.

[0143] Specifically, it also includes an electrical equipment fault prediction method:

[0144] 1) Construct LSTM-Attention prediction model, input window length is 30 sampling periods;

[0145] 2) Define device health index (HI):

[0146]

[0147] Where w i is the dynamic weight, f i is the i-th normalized feature;

[0148] 3) Trigger warning when HI decreases continuously for 5 times and slope >0.1.

[0149] Example 1: High-voltage circuit breaker mechanical fault diagnosis

[0150] 1, system configuration

[0151] 1.1, Device under test: ZF12-252 GIS circuit breaker.

[0152] 1.2, Sensor deployment:

[0153] Vibration sensor (PCB 352C03, 200 kHz sampling rate).

[0154] Temperature sensor (PT100, 1 Hz sampling, LoRaWAN transmission).

[0155] Current sensor (Rogowski coil, 50 kHz sampling, EtherCAT transmission).

[0156] Ultrasonic sensor (UE Systems, bandwidth 40 kHz).

[0157] 2, Fault simulation

[0158] In the opening operation, the following faults are artificially created (right 11 health index monitoring target).

[0159] Fault A: Stuck operating mechanism (vibration signal dominant frequency from 120 Hz to 85 Hz).

[0160] Fault B: Increased contact resistance (temperature increased by 28°C compared to normal).

[0161] 3, Implementation steps

[0162] (1) Data synchronization and preprocessing

[0163] Align the vibration signal with the current signal zero-crossing point by DTW algorithm, maximum time delay compensation 12 ms

[0164] 6-layer db4 wavelet packet decomposition is performed on the vibration signal, and the 3rd layer detail coefficient entropy value is extracted

[0165] (2) Feature fusion

[0166] Primary features:

[0167] \text{TKEO feature} = \sum_{k=1}^N[x(k)^2-x(k-1)x(k+1)]\quad(N=2000).

[0168] GCN topology construction:

[0169] Vertex set: {vibration entropy, temperature rise rate, current THD, ultrasonic pulse count}.

[0170] Edge weight: Mutual information of vibration entropy and current THD = 0.82 (threshold > 0.7).

[0171] (3) Dynamic weight calculation

[0172] Dual-channel LSTM input:

[0173]

[0174] Output weights: vibration 0.68, temperature 0.21, current 0.11.

[0175] (4) Fault diagnosis

[0176] ResNet-50+CBAM model output:

[0177]

[0178] 4、Effect comparison

[0179] Indicator Traditional D-S evidence theory Method of the invention Diagnostic accuracy 76.2％ 94.8％ Fault location time 3.2s 0.8s Weight adjustment response speed Fixed weight <10 ms

[0180] Example 2: Transformer insulation deterioration online monitoring system

[0181] 1、Hardware deployment

[0182] Edge computing node:

[0183] FPGA: Xilinx Zynq UltraScale+;

[0184] Vibration processing pipeline: 3-stage FIR filtering (125 MHz clock);

[0185] Resource occupation: LUT 58%, DSP 42%;

[0186] ARM: Cortex-A72 running pruned model (3.1MB),

[0187] 2、Knowledge graph construction

[0188] Dynamic update: When 50 new records of oil chromatogram data are added, the edge weight of "discharge -> CO2" is automatically strengthened.

[0189] 3、Storage optimization

[0190] NVMe partitioning strategy:

[0191]

[0192] 4、Actual performance

[0193] Edge computing delay:

[0194] Feature extraction: 0.95ms (FPGA acceleration vs. software implementation 8.3ms);

[0195] Model inference: 43ms (meets requirement of < 50ms).

[0196] Cloud query efficiency: Knowledge graph traces "similar faults in the past 3 years": average response time 1.4s.

[0197] In summary, the electrical equipment multi-sensor fault feature fusion diagnosis method solves the time asynchronous problem of heterogeneous sensor data such as vibration and temperature through multi-source sensor data synchronous acquisition and time alignment technology, adopts a signal alignment method combining dynamic time warping (DTW) algorithm and Hilbert-Huang transform, improves the time alignment accuracy of multi-source data, and improves the feature extraction accuracy. Through hierarchical feature extraction and graph convolution network fusion, a feature correlation matrix based on mutual information is constructed, and GCN is used to mine the deep correlation between vibration signal TKEO features and cross-modal features such as current harmonics, reduce feature dimension, and improve fault feature separability index. Through dynamic weight generation technology, a dual-channel LSTM network is used to calculate sensor weights in real time, combined with residual analysis and historical fault mode library, the current sensor weight is automatically adjusted from 0.3 to 0.72 under the overload working condition of the transformer, and the robustness of the diagnosis system is improved when the sensor is partially failed. Through the improved deep residual network, the CBAM attention module is embedded in ResNet-50 and the adaptive loss function is used, the recognition accuracy of rare faults with insufficient sample size is improved, and the inference speed is accelerated through model distillation technology. Through the heterogeneous architecture of the edge computing module, the TKEO feature extraction pipeline accelerated by FPGA reduces power consumption and delay compared to pure CPU implementation, meeting the real-time monitoring needs of the circuit breaker opening and closing process. Through dual-mode communication sensor networking, the EtherCAT+LoRaWAN hybrid networking scheme ensures real-time transmission of vibration signals while extending the battery life of wireless temperature nodes and reducing wiring costs. Through the construction of the fault knowledge graph, the causal reasoning engine based on Bayesian networks can shorten the fault root cause analysis time from manual diagnosis and improve the accuracy of maintenance scheme recommendations. Through the health index prediction model, the LSTM-Attention model predicts the motor bearing failure trend 3-5 weeks in advance, compared with the traditional vibration threshold alarm, the number of unplanned shutdowns is reduced, and the annual maintenance cost is reduced. Through TSN network deployment, the time-sensitive network ensures that the end-to-end transmission delay of key data is < 10ms, and the data packet loss rate is reduced in the strong electromagnetic interference environment of the substation.

[0198] It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting; it is not intended to exclude myriad other embodiments of the present application that other inventors can develop based on the same general inventive concepts embodied by the described embodiments. That is, although the present application is described in terms of particular embodiments and implementations, it is to be understood that the terminology used is for the purpose of descriptive clarity and that it should be taken in a descriptive sense and not a limiting sense.

[0199] While the embodiments of the application have been shown and described herein, it is to be understood that the application is not limited to these embodiments. Rather, many modifications, changes, substitutions, and alterations can be made to the embodiments of the application without departing from the spirit and scope of the application as defined by the appended claims and their equivalents.

Claims

1. A multi-sensor fault feature fusion diagnosis method for electrical equipment, characterized in that: The following steps are involved: S1. Synchronous acquisition of multi-source sensor data: Synchronous acquisition of electrical equipment operating data through vibration sensors, temperature sensors, current sensors, and ultrasonic sensors. The sampling frequency is dynamically adjusted based on the physical characteristics of each sensor. The sampling rate of vibration signals is not less than 100kHz, and the sampling rate of temperature signals is not less than 1Hz. S2. Timing alignment and preprocessing: An asynchronous signal alignment algorithm based on dynamic time warping (DTW) is used to eliminate timing deviations caused by sensor response delays, and wavelet denoising and normalization are performed on the sensor data. S3, hierarchical feature extraction: S3.1, primary feature layer: extract the Teager-Kaiser energy operator (TKEO) time-frequency features of the vibration signal, the gradient change rate of the temperature signal, the harmonic distortion rate of the current signal, and the PRP D spectrum features of the ultrasonic signal; S3.2, High-level feature layer: Build a cross-sensor feature correlation matrix through the graph convolutional network (GCN), where nodes represent feature parameters and edge weights are calculated using the mutual information algorithm; S4. Dynamic weight fusion: S4.

1. Generate real-time weight coefficients based on a dual-channel LSTM network. The first channel inputs the current sensor confidence (calculated through residual analysis), and the second channel inputs the device's historical failure mode library data. S4.

2. Use the Softmax function to constrain the weight output to ensure that the sum of the weights of each sensor is 1; S5. Fault diagnosis decision: The fused features are input into the pre-trained deep residual network (ResNet-50), the fault type and probability distribution are output, and a fault traceability report is generated.

2. The method for fusion diagnosis of multi-sensor fault characteristics of electrical equipment according to claim 1, characterized in that: The timing alignment algorithm in S2 specifically includes: S2.

1. Using the current signal as a reference signal, calculate the cross-correlation coefficients of other sensor signals and the reference signal. S2.

2. When the correlation coefficient is lower than the threshold of 0.85, the dynamic time warping algorithm is used to perform nonlinear stretch compensation, and the compensation accuracy is controlled within ±5 sampling points; S2.

3. Perform Hilbert-Huang transform (HHT) on the vibration signal and extract the instantaneous frequency as an auxiliary alignment feature.

3. The method for fusion diagnosis of multi-sensor fault characteristics of electrical equipment according to claim 1, characterized in that: The graph convolutional network construction method in S3.2 includes: S3.2a. Define a feature graph G = (V, E), where the vertex set V = {v1, v2, ..., vn} represents n feature parameters, and the edge set E is generated by the following rules: If the Pearson correlation coefficient of two feature parameters |ρ|>0.7, a connecting edge is established with edge weight w=1-|ρ|; S3.2b, use two layers of GCN for feature aggregation, the activation function of the first layer is ReLU, and the second layer is Sigmoid; S3.2c. Output topological feature vector F_G∈R 128 , and is concatenated with the original features to form the final fusion features.

4. The method for fusion diagnosis of multi-sensor fault characteristics of electrical equipment according to claim 1, characterized in that: The dynamic weight generation algorithm in S4 specifically includes: S4a, design a dual-channel LSTM network structure, each LSTM unit contains 256 hidden nodes; S4b, the first channel input dimension is 4 tensors, including: the signal-to-noise ratio (SNR) of each sensor signal, the number of recent fault alarms, data integrity, and environmental interference intensity; S4c, the second channel inputs the historical fault mode encoding of the device, using one-hot encoding with a dimension of 5 times the number of fault categories; S4d, weight adjustment period is configurable and is set to 10 seconds by default.

5. The method for fusion diagnosis of multi-sensor fault characteristics of electrical equipment according to claim 1, characterized in that: The improvements to the deep residual network in S5 include: S5.

1. Add an attention mechanism module (CBAM) to the original ResNet-50 and insert it after each residual block. S5.

2. The output layer uses an adaptive loss function: l=a·l CE +(1-a)·l Focal Among them, α is dynamically adjusted according to the sample balance, ranging from 0.3-0.7; S5.

3. Deploy model distillation technology to transfer the knowledge of the teacher model (with 100 million parameters) to the student model (with 10 million parameters).

6. The multi-sensor fault feature fusion diagnosis method for electrical equipment according to claim 1, further comprising a diagnostic system for implementing the method according to claims 1-5, characterized in that: The diagnostic system is specifically as follows: Edge computing module: Adopting FPGA+ARM heterogeneous architecture, FPGA is responsible for signal preprocessing and feature extraction, and ARM runs lightweight diagnostic models; Dynamic weight controller: Integrates a dedicated ASIC chip to achieve hardware acceleration of the LSTM network with a latency of less than 2ms; Cloud collaboration platform: 1) Data center: Stores at least five years of equipment lifecycle data and uses a time series database (InfluxDB) to optimize query efficiency. 2) Knowledge Graph Engine: Builds a fault causal reasoning network based on Neo4j, supporting multi-dimensional correlation queries; Human-computer interaction terminal: Provides an AR visualization interface, supports three-dimensional positioning of fault parts and maintenance guidance.

7. According to the multi-sensor fault feature fusion diagnosis method for electrical equipment according to claim 6, the edge computing module is specifically implemented as follows: T1, FPGA part design dedicated pipeline: Vibration signal processing channel: includes FIR filter (cutoff frequency 10kHz), Hilbert transformer, and TKEO calculation unit; Temperature signal processing channel: integrated sliding average window (window length adjustable) and first-order difference calculation module; T2, ARM partially deploys the pruned TinyML model, with a model size of no more than 3MB and an inference time of ≤50ms; The diagnostic system deployment optimization method is specifically as follows: U1. Network topology optimization: Deploy edge computing nodes within the substation to ensure that the number of hops to any sensor is ≤ 2; Adopt TSN (Time Sensitive Network) to ensure that the critical data transmission delay is less than 10ms; U2, Energy Consumption Management: Dynamically adjust the sampling frequency according to the equipment load rate, and reduce the vibration sampling rate to 50kHz when lightly loaded; Solar power supply units are used to power wireless sensor nodes.

8. The method for fusion diagnosis of multi-sensor fault characteristics of electrical equipment according to claim 6 is characterized in that ,The sensor networking device of the diagnostic system is as follows: V1, adopts industrial Ethernet (EtherCAT) and wireless (LoRaWAN) dual-mode communication architecture, wired mode is used for high real-time data (vibration, current), and wireless mode is used for low-frequency data (temperature); V2, built-in self-diagnostic unit, continuously monitors the health status of the sensor, and automatically switches to the redundant sensor when a sensor failure is detected; V3, provides IP67 protection grade housing, built-in temperature compensation circuit, operating temperature range -40℃~85℃.

9. The method for multi-sensor fault feature fusion diagnosis of electrical equipment according to claim 1, further comprising a method for constructing a knowledge graph of electrical equipment faults, applied to the system according to claim 6, characterized in that: Entity definition: includes three types of entities: equipment components (windings, insulators), failure modes (mechanical wear, partial discharge), and characteristic parameters (vibration amplitude, temperature rise rate); Relational Modeling: Causal chain: "Vibration amplitude exceeds threshold" → "Causes" → "Bearing looseness" → "Trigger" → "Temperature abnormality"; Probabilistic association: calculated by Bayesian network: P(insulation aging|THD>5%) = 0.82; Dynamic update mechanism: Every 100 new diagnostic records automatically trigger graph reconstruction.

10. The method for fusion diagnosis of multi-sensor fault features of electrical equipment according to claim 1, further comprising an electrical equipment fault prediction method, implemented based on the fusion features of claim 1, characterized in that: 1) Build an LSTM-Attention prediction model with an input window length of 30 sampling periods; 2) Define the device health index (HI): where w i is the dynamic weight, f i is the i-th normalized feature; 3) When HI drops for 5 consecutive times and the slope is greater than 0.1, an alarm is triggered.

Images