A data fusion-based electrical equipment fault early warning system

Abstract

The application provides an electrical equipment fault early warning system based on data fusion, and relates to the technical field of intelligent monitoring and fault early warning of electrical equipment. The system adopts improved adaptive Kalman filtering to complete data layer noise reduction fusion, weighted cosine feature fusion to complete feature layer dimension fusion, improved D-S evidence theory to complete decision level fault confidence fusion, realizes early prediction of fault trend in combination with a long short-term memory network, and dynamically adjusts early warning judgment criteria through an adaptive threshold algorithm. The application breaks through the technical limitations of traditional single data and single layer fusion, realizes multi-source data redundancy verification, conflict resolution and deep coupling, has an early warning accuracy of greater than or equal to 98.5%, a false alarm rate of less than or equal to 0.3% / month, and an early warning time length of greater than or equal to 30 minutes, can be widely adapted to various electrical equipment such as high-voltage switch cabinets, transformers, distribution cabinets and motors, solves the core technical defects of poor data robustness, fault false alarm and low prediction accuracy of the prior art, and has significant technical novelty and engineering practicability.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring and fault early warning technology for electrical equipment, specifically to a fault early warning system for electrical equipment based on data fusion. Background Technology

[0002] Electrical equipment is the core equipment for power transmission and distribution and industrial power supply. Its operating status directly affects power supply safety and production continuity. Traditional electrical equipment monitoring often uses single-parameter threshold judgment, such as monitoring only temperature or only voltage. This has certain technical defects: the data source is single, the robustness is extremely poor, and relying solely on data from a single sensor makes it susceptible to noise, interference, and single-point faults. It cannot reflect the overall status of the equipment, resulting in high false alarm and false negative rates. Data fusion is simple and lacks deep coupling. Existing technologies mostly use simple weighted averaging without building a layered fusion architecture, which cannot handle data redundancy, conflicts, and uncertainties. Fault feature extraction is insufficient, lacking joint mining of multi-dimensional features such as electrical quantities, temperature, vibration, and partial discharge, making it difficult to identify early latent faults. The warning threshold is fixed and rigid, unable to adapt to equipment aging, load fluctuations, and environmental changes, resulting in low warning accuracy. It lacks forward-looking predictive capabilities, only providing alarms after the fact, unable to predict fault trends in advance, and failing to meet preventive maintenance needs. The system has poor scalability, unable to be compatible with multiple types of sensors and electrical equipment, resulting in weak engineering adaptability.

[0003] Meanwhile, existing technologies mostly use conventional Kalman filtering with fixed noise covariance, resulting in poor noise reduction performance in complex industrial environments. Furthermore, most solutions rely on fixed threshold warnings, lack adaptive adjustment mechanisms, LSTM time series prediction functions, and have not formed a complete three-level fusion system of data, features, and decisions, thus lacking fault characterization capabilities.

[0004] In summary, existing technologies cannot achieve deep fusion of multi-source data, accurate identification of early latent faults, dynamic adaptive early warning, and long-term trend prediction, severely restricting the level of safe operation and maintenance of electrical equipment. Therefore, those skilled in the art have provided an electrical equipment fault early warning system based on data fusion to solve the problems mentioned in the background. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an electrical equipment fault early warning system based on data fusion. This system solves the problems of existing electrical equipment fault early warning systems, such as single data source, rudimentary fusion algorithm, insufficient feature extraction, fixed thresholds, lack of predictive ability, and high false alarm / false negative rates. It can achieve the following objectives: 1. Construct a multi-source heterogeneous data acquisition system to simultaneously collect electrical quantities, status quantities, and environmental quantities, thereby improving data comprehensiveness; 2. Pioneering a three-tiered data fusion architecture: data layer (improved adaptive Kalman filtering), feature layer (weighted cosine fusion), and decision layer (improved DS evidence theory), achieving redundancy verification, conflict resolution, and uncertainty reduction; 3. Introduce LSTM time series prediction to enable early prediction and warning of fault trends, transforming "post-event alarm" into "pre-event prediction"; 4. Design an adaptive dynamic threshold algorithm to automatically update the warning standard based on the real-time status of the equipment, thereby improving adaptability to complex working conditions; 5. Improve the preprocessing workflow (noise reduction, interpolation, normalization, spatiotemporal alignment) to enhance data quality and fusion accuracy; 6. Enable edge-cloud collaboration, balancing real-time performance and computing power, and support large-scale deployment across multiple devices and scenarios; 7. Significantly improve early warning performance: Early warning accuracy rate ≥98.5%, false alarm rate ≤0.3% / month, and early warning lead time ≥30 minutes.

[0006] To achieve the above objectives, the present invention provides the following technical solution: An electrical equipment fault early warning system based on data fusion includes a multi-source heterogeneous data acquisition module, a data preprocessing module, a three-level hierarchical data fusion module, a fault feature extraction module, an LSTM fault prediction module, an adaptive threshold early warning module, and a human-computer interaction and linkage module. The multi-source heterogeneous data acquisition module simultaneously collects three types of data by deploying multiple types of sensors, specifically: Electrical operating data: RMS voltage, RMS current, active power, reactive power, total harmonic distortion (THD); Equipment status data: casing temperature, winding / contact temperature, vibration amplitude, vibration frequency, and local discharge level; Environmental data: ambient temperature, relative humidity, dust concentration, condensation status; all of the above data are accompanied by high-precision timestamps and device serial numbers to ensure spatiotemporal consistency; The data preprocessing module executes a standardized preprocessing workflow, specifically: Outlier removal: The 3σ criterion is used to remove outliers that significantly deviate from the normal range; Missing value imputation: Single missing values ​​are imputed by linear interpolation, and consecutive missing values ​​are imputed by the mean of adjacent periods; Wavelet packet denoising: db6 wavelet 8-level decomposition to remove industrial electromagnetic noise and mechanical interference noise; Normalization: Max-Min normalization to [0, 1] eliminates dimensional differences; Spatiotemporal alignment: Unify the sampling frequency to 10Hz, and complete data matching according to device ID and timestamp; The three-level hierarchical data fusion module sequentially performs three-level operations: data layer noise reduction fusion, feature layer weighted fusion, and decision layer evidence fusion. Through spatiotemporal alignment of multi-source data and resolution of uncertainties, it outputs high-confidence fault judgment results. The data layer noise reduction and fusion performs optimal estimation of multi-sensor observation data (such as multiple temperature and multiple voltage sources), adaptively adjusts the noise covariance, suppresses noise, and outputs smooth and consistent state estimates. The weighted fusion of the feature layer extracts time-domain features (such as mean, variance, and peak value), frequency-domain features (such as harmonics and spectral entropy), and time-frequency-domain features (wavelet coefficients). The feature contribution is calculated through cosine similarity and then weighted and coupled into a unified fused feature vector. The decision-level evidence fusion maps the data fusion results to fault propositions (such as insulation aging, overheating, loosening, partial discharge, etc.), generates a BPA function through fuzzy membership, and introduces a conflict coefficient K to correct the fusion rules, thereby solving the problem of distortion of high-conflict evidence. The fault feature extraction module is used to extract 18-dimensional feature data, including mean, variance, harmonics, spectral entropy, and wavelet coefficients. The LSTM fault prediction module takes the fused features as input, constructs an LSTM time series model, learns the equipment condition degradation law, predicts the fault probability change trend in the next 5 to 30 minutes, and outputs the predicted fault probability and remaining useful life (RUL) reference. The adaptive threshold early warning module uses a sliding window to statistically analyze historical health data, dynamically calculates the mean and variance, and automatically updates the early warning threshold to avoid misjudgment under fixed thresholds due to equipment aging or load fluctuations. The warning levels are as follows: Normal: Fusion features < No action was taken. Notice: ≤Fusion features<1.2 Please pay attention; Warning: 1.2 ≤Fusion feature<1.5 Audible and visual alarms and platform push notifications; Urgent: Fusion feature ≥ 1.5 Remote locking and maintenance dispatching; The multi-source heterogeneous data includes: electrical operation data (such as voltage, current, active power, harmonic distortion rate), equipment status data (such as shell temperature, winding temperature, vibration amplitude, partial discharge), and environmental monitoring data (such as ambient temperature and humidity, dust concentration, humidity).

[0007] Furthermore, the data layer denoising and fusion employs an improved adaptive Kalman filter algorithm to denoise and optimally estimate multi-sensor observation data from the same source. The state equation and observation equation are as follows: Equations of state:

[0008] Observation equation:

[0009] Adaptive noise adjustment formula:

[0010] in: Let k be the system state vector at time k. Here is the state transition matrix. This is process noise; For the observation vector, For the observation matrix, To observe noise; Let α be the adaptive process noise covariance, α be the adjustment coefficient, and N be the number of sensors.

[0011] Furthermore, the weighted fusion of the feature layer employs a weighted cosine similarity feature fusion algorithm to weight and couple the normalized multi-dimensional fault features, generating a unified fused feature vector. The calculation formula is:

[0012] Weight constraints:

[0013] Where M is the feature dimension. The adaptive weights for the i-th dimension feature are: Let i be the i-th eigenvector. This is a reference vector for standard health characteristics.

[0014] Furthermore, the decision-level evidence fusion employs an improved DS evidence theory fusion algorithm, which optimizes the basic probability assignment (BPA) through fuzzy membership degree optimization and introduces a conflict coefficient k to modify the fusion rule. The fusion formula is as follows: ,

[0015] Conflict coefficient:

[0016] in: Assign a basic probability value to the fused faulty proposition A, where K is the coefficient of evidence conflict. , A basic probability assignment function for different data sources.

[0017] Furthermore, the LSTM fault prediction module constructs a time series prediction model to extrapolate the fault trend over the next T steps based on the fused features. The hidden layer state update formula is as follows:

[0018]

[0019] in: Let the hidden state be at time t. As input fusion features, This is the output for fault probability prediction. Let W be the activation function, and b be the network weights and biases.

[0020] Furthermore, the adaptive threshold warning module employs a dynamic threshold iteration algorithm to automatically update the warning threshold based on historical health data and real-time fused features. The calculation formula is as follows:

[0021]

[0022]

[0023] in: Let t be the warning threshold. The mean of the sliding window. Let λ be the variance of the sliding window, λ be the confidence coefficient, and L be the length of the sliding window.

[0024] Furthermore, the data preprocessing module sequentially performs: outlier removal, linear interpolation of missing values, wavelet packet denoising, and max-min normalization. The normalization formula is:

[0025] Furthermore, timestamp alignment and spatial coordinate calibration are performed on multi-source data to ensure the spatiotemporal consistency of the fused data.

[0026] Furthermore, the fault types include: insulation aging, overheating fault, mechanical loosening, partial discharge, short circuit hazard, and overload operation; The system divides the warning levels into four levels: normal, alert, warning, and emergency, and links them to perform audible and visual alarms, data reporting, and remote locking protection actions.

[0027] Furthermore, the system adopts an edge-cloud collaborative architecture, where real-time data acquisition, preprocessing, and primary fusion are achieved at the edge, while deep learning training, decision fusion, and historical data management are achieved in the cloud. The system's early warning indicators are: fault early warning accuracy ≥ 98.5%, fault false alarm rate ≤ 0.5%, average early warning lead time ≥ 30 min, and continuous operation stability ≥ 99.9%.

[0028] This invention provides an electrical equipment fault early warning system based on data fusion. It has the following beneficial effects: 1. This invention provides an electrical equipment fault early warning system based on data fusion. It combines an improved adaptive Kalman filter, weighted cosine feature fusion, and an improved DS evidence theory into a three-level fusion system, which eliminates noise, redundancy, and conflict at each level. Compared with the traditional single-layer fusion, the fault characterization capability is improved by more than 60%. The improved Kalman filter realizes online adaptive adjustment of noise covariance, and improves the signal-to-noise ratio by more than 15dB in strong electromagnetic interference environment compared with the traditional Kalman filter. The improved DS evidence theory introduces a conflict coefficient K to correct the fusion rule, solves the decision distortion problem caused by multi-source data conflict, and improves the decision reliability by 40% in conflict environment.

[0029] 2. This invention provides an electrical equipment fault early warning system based on data fusion. It adapts to changes in equipment status by using adaptive thresholds, uses LSTM to predict trends in advance, and accurately captures early latent faults. The early warning time is ≥30 minutes. Furthermore, it simultaneously integrates five types of data: electrical quantity, temperature, vibration, partial discharge, and environmental quantity. Compared with a single data source, the early warning accuracy is improved by more than 30%.

[0030] 3. This invention provides an electrical equipment fault early warning system based on data fusion, which can take into account both edge real-time performance and cloud computing power, support large-scale deployment, adapt to all types of electrical equipment such as switchgear, transformers, and motors, reduce more than 70% of invalid inspections, extend equipment service life, and reduce the risk of power outages. Attached Figure Description

[0031] Figure 1 This is a block diagram of the overall architecture of the electrical equipment fault early warning system based on data fusion according to the present invention; Figure 2 This is a schematic diagram of the three-level hierarchical data fusion process of the present invention; Figure 3 The flowchart of the improved adaptive Kalman filter algorithm of this invention is shown below; Figure 4 This is the LSTM fault prediction and adaptive early warning logic diagram of the present invention; Figure 5 This is a schematic diagram of the edge-cloud collaborative deployment of the present invention. Detailed Implementation

[0032] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0033] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0034] like Figures 1-5 As shown, this embodiment of the invention provides an electrical equipment fault early warning system based on data fusion. The system adopts a seven-layer architecture, including a multi-source heterogeneous data acquisition module, a data preprocessing module, a three-level hierarchical data fusion module, a fault feature extraction module, an LSTM fault prediction module, an adaptive threshold early warning module, and a human-computer interaction and linkage module. Multi-source heterogeneous data includes: electrical operation data (such as voltage, current, active power, harmonic distortion rate), equipment status data (such as shell temperature, winding temperature, vibration amplitude, partial discharge), and environmental monitoring data (such as ambient temperature and humidity, dust concentration, humidity).

[0035] The multi-source heterogeneous data acquisition module simultaneously collects three types of data by deploying multiple types of sensors, specifically: Electrical operating data: RMS voltage, RMS current, active power, reactive power, total harmonic distortion (THD); Equipment status data: casing temperature, winding / contact temperature, vibration amplitude, vibration frequency, and local discharge level; Environmental data: ambient temperature, relative humidity, dust concentration, condensation status; all of the above data are accompanied by high-precision timestamps and device serial numbers to ensure spatiotemporal consistency.

[0036] The data preprocessing module executes a standardized preprocessing workflow, specifically: Outlier removal: The 3σ criterion is used to remove outliers that significantly deviate from the normal range; Missing value imputation: Single missing values ​​are imputed by linear interpolation, and consecutive missing values ​​are imputed by the mean of adjacent periods; Wavelet packet denoising: db6 wavelet 8-level decomposition to remove industrial electromagnetic noise and mechanical interference noise; Normalization: Max-Min normalization to [0, 1] eliminates dimensional differences; Spatiotemporal alignment: Unify the sampling frequency to 10Hz, and complete data matching according to device ID and timestamp; The normalization formula is:

[0037] Furthermore, timestamp alignment and spatial coordinate calibration are performed on multi-source data to ensure the spatiotemporal consistency of the fused data.

[0038] The three-level hierarchical data fusion module sequentially performs three-level operations: data layer noise reduction fusion, feature layer weighted fusion, and decision layer evidence fusion. Through spatiotemporal alignment of multi-source data and resolution of uncertainties, it outputs high-confidence fault judgment results. The data layer noise reduction and fusion performs optimal estimation of multi-sensor observation data (such as multiple temperature and multiple voltage sources), adaptively adjusts the noise covariance, suppresses noise, and outputs smooth and consistent state estimates. The data layer denoising and fusion employs an improved adaptive Kalman filter algorithm to denoise and optimally estimate multi-sensor observation data from the same source. The state equation and observation equation are as follows: Equations of state:

[0039] Observation equation:

[0040] Adaptive noise adjustment formula:

[0041] in: Let k be the system state vector at time k. Here is the state transition matrix. This is process noise; For the observation vector, For the observation matrix, To observe noise; Let α be the adaptive process noise covariance, α be the adjustment coefficient, and N be the number of sensors.

[0042] The feature layer uses weighted fusion to extract time-domain features (such as mean, variance, and peak value), frequency-domain features (such as harmonics and spectral entropy), and time-frequency-domain features (wavelet coefficients). The feature contribution is calculated through cosine similarity and then weighted and coupled into a unified fused feature vector. The weighted fusion of features employs a weighted cosine similarity feature fusion algorithm to weight and couple the normalized multi-dimensional fault features, generating a unified fused feature vector. The calculation formula is:

[0043] Weight constraints:

[0044] Where M is the feature dimension. The adaptive weights for the i-th dimension feature are: Let i be the i-th eigenvector. This is a reference vector for standard health characteristics.

[0045] The decision-level evidence fusion maps the data fusion results to fault propositions (such as insulation aging, overheating, loosening, partial discharge, etc.), generates a BPA function through fuzzy membership, and introduces a conflict coefficient K to correct the fusion rules, thereby solving the problem of distortion of high-conflict evidence. The decision-making level evidence fusion adopts an improved DS evidence theory fusion algorithm, which optimizes the basic probability assignment (BPA) through fuzzy membership degree optimization and introduces a conflict coefficient k to modify the fusion rule. The fusion formula is as follows: ,

[0046] Conflict coefficient:

[0047] in: Assign a basic probability value to the fused faulty proposition A, where K is the coefficient of evidence conflict. , A basic probability assignment function for different data sources.

[0048] The fault feature extraction module is used to extract 18-dimensional feature data, including mean, variance, harmonics, spectral entropy, and wavelet coefficients. Fault types include: insulation aging, overheating fault, mechanical loosening, partial discharge, short circuit hazard, and overload operation. The LSTM fault prediction module takes the fused features as input, constructs an LSTM time series model, learns the degradation law of equipment condition, predicts the trend of fault probability change in the next 5 to 30 minutes, and outputs the predicted fault probability and remaining useful life (RUL) reference. The LSTM fault prediction module constructs a time series prediction model and performs future T-factor analysis on the fused features. Based on the fault trend deduction, the hidden layer state update formula is:

[0049]

[0050] in: Let the hidden state be at time t. As input fusion features, This is the output for fault probability prediction. Let W be the activation function, and b be the network weights and biases.

[0051] The adaptive threshold warning module uses a sliding window to statistically analyze historical health data, dynamically calculates the mean and variance, and automatically updates the warning threshold to avoid misjudgment under fixed thresholds due to equipment aging or load fluctuations. The adaptive threshold warning module employs a dynamic threshold iteration algorithm, automatically updating the warning threshold based on historical health data and real-time fused features. The calculation formula is as follows:

[0052]

[0053]

[0054] in: Let t be the warning threshold. The mean of the sliding window. Let λ be the variance of the sliding window, λ be the confidence coefficient, and L be the length of the sliding window. The warning levels are as follows: Normal: Fusion features < No action was taken. Notice: ≤Fusion features<1.2 Please pay attention; Warning: 1.2 ≤Fusion feature<1.5 Audible and visual alarms and platform push notifications; Urgent: Fusion feature ≥ 1.5 Remote locking and maintenance dispatching.

[0055] The system adopts an edge-cloud collaborative architecture, specifically: Edge computing enables: data acquisition, preprocessing, data layer fusion, and real-time early warning. Cloud-based implementation: feature extraction, decision-level fusion, LSTM model training, historical data storage, report generation, and remote monitoring; The system's early warning indicators are: fault early warning accuracy ≥ 98.5%, fault false alarm rate ≤ 0.5%, average early warning lead time ≥ 30 minutes, and continuous operation stability ≥ 99.9%. Example

[0056] This embodiment uses a 10kV high-voltage switchgear as the early warning target. The system is deployed in an industrial power distribution room and includes 5 switchgear monitoring nodes. The implementation process and effects of the system are described in detail.

[0057] The system includes: data acquisition sensors Electrical quantities: Three-phase voltage transformers and current transformers; data acquisition includes U / I, power, and THD. State variables: contact temperature sensor, vibration acceleration sensor, partial discharge sensor; Environmental parameters: temperature and humidity sensor, dust sensor.

[0058] Edge acquisition terminal: ARM architecture industrial controller, sampling frequency 10Hz, supports RS485 / EtherCAT / 4G transmission; Cloud server: A cloud server for deploying LSTM models, databases, and web management platforms.

[0059] The system software implementation process is as follows: Step S1: Multi-source data acquisition Synchronously collect switchgear voltage, current, temperature, vibration, partial discharge, and ambient temperature and humidity data; Step S2: Data Preprocessing 3σ outlier removal → linear interpolation completion → wavelet packet denoising → normalization → spatiotemporal alignment; Step S3: Data Layer Fusion Improved adaptive Kalman filtering for noise reduction and fusion of multi-channel temperature and voltage data, outputting the optimal estimate; Step S4: Feature Extraction Extract 18 dimensions of features, including mean, variance, harmonics, spectral entropy, and wavelet coefficients; Step S5: Feature Layer Fusion Weighted cosine similarity is used to calculate feature weights and generate a fused feature vector. ; Step S6: LSTM prediction Input fusion features to predict the probability of failure in the next 15 minutes and 30 minutes; Step S7: Decision-making level integration Improve the DS evidence theory by integrating the confidence levels of various data sources to output the fault type and probability; Step S8: Adaptive Threshold Warning Sliding window calculation of dynamic threshold Determine the warning level; Step S9: Linkage Output It features four levels of output: normal, alert, warning, and emergency, along with audible and visual alarms, platform push notifications, and remote protection.

[0060] The algorithm parameters are configured as follows: Improved Kalman filter: adjustment coefficient α=0.85, number of sensors N=3; Feature fusion: Feature dimension M=18, weights optimized by stochastic gradient descent; Improved DS: Fault Proposition Set: Normal insulation aging, overheating, loosening, partial discharge; LSTM: 128 hidden layer nodes, time step T=30, prediction step size 15; Adaptive threshold: sliding window L=600, confidence coefficient λ=2.5.

[0061] The test results and comparisons are shown in the table below: Test conditions: Continuous operation for 30 days, including four operating conditions: normal operation, overload, contact overheating, and partial discharge.

[0062] Early warning accuracy 98.7% 72.3% 85.6% False alarm rate / month 0.28 times 12.7 times 5.4 times Average warning lead time 32min No (reported to the police afterward) 8min Anti-interference capability Strong (SNR≥25dB) Poor (easily affected by interference) Medium (SNR≈12dB) Threshold adaptability Adaptive dynamic adjustment Fixed threshold Fixed threshold Fault warning case: The switchgear has a potential risk of overheating contacts. The handling procedure is as follows: 1) The temperature sensor detected a rise in temperature from 65℃ to 82℃; 2) Data layer filtering and noise reduction to eliminate electromagnetic noise interference; 3) Feature fusion identification of the features "temperature rise + resistance increase + vibration stability"; 4) LSTM predicts that the temperature will exceed 95℃ after 15 minutes; 5) The confidence level of the DS fusion-based determination of overheating faults is 0.92; 5) Adaptive threshold triggering of early warning levels; 7) The system pushes alarms, and maintenance personnel handle them in a timely manner to avoid tripping accidents.

[0063] Although embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

Claims

1. A fault early warning system for electrical equipment based on data fusion, characterized in that, It includes a multi-source heterogeneous data acquisition module, a data preprocessing module, a three-level hierarchical data fusion module, a fault feature extraction module, an LSTM fault prediction module, an adaptive threshold early warning module, and a human-computer interaction and linkage module; The multi-source heterogeneous data acquisition module simultaneously collects three types of data by deploying multiple types of sensors, specifically: Electrical operating data: RMS voltage, RMS current, active power, reactive power, total harmonic distortion (THD); Equipment status data: casing temperature, winding / contact temperature, vibration amplitude, vibration frequency, and local discharge level; Environmental data: ambient temperature, relative humidity, dust concentration, condensation status; all of the above data are accompanied by high-precision timestamps and device serial numbers to ensure spatiotemporal consistency; The data preprocessing module executes a standardized preprocessing workflow, specifically: Outlier removal: The 3σ criterion is used to remove outliers that significantly deviate from the normal range; Missing value imputation: Single missing values ​​are imputed by linear interpolation, and consecutive missing values ​​are imputed by the mean of adjacent periods; Wavelet packet denoising: db6 wavelet 8-level decomposition to remove industrial electromagnetic noise and mechanical interference noise; Normalization: Max-Min normalization to [0, 1] eliminates dimensional differences; Spatiotemporal alignment: Unify the sampling frequency to 10Hz, and complete data matching according to device ID and timestamp; The three-level hierarchical data fusion module sequentially performs three-level operations: data layer noise reduction fusion, feature layer weighted fusion, and decision layer evidence fusion. Through spatiotemporal alignment of multi-source data and resolution of uncertainties, it outputs high-confidence fault judgment results. The data layer noise reduction and fusion performs optimal estimation of multi-sensor covariance observation data, adaptively adjusts the noise covariance, suppresses noise, and outputs a smooth and consistent state estimate. The feature layer weighted fusion extracts time-domain features, frequency-domain features, and time-frequency-domain features, calculates feature contribution through cosine similarity, and weighted couples them into a unified fused feature vector. The decision-level evidence fusion maps the data fusion results to fault propositions, generates a BPA function through fuzzy membership, introduces a conflict coefficient K to correct the fusion rules, and solves the problem of distortion of high-conflict evidence. The fault feature extraction module is used to extract 18-dimensional feature data, including mean, variance, harmonics, spectral entropy, and wavelet coefficients. The LSTM fault prediction module takes the fused features as input to construct an LSTM time series model, learns the degradation law of equipment status, predicts the trend of fault probability change in the next 5 to 30 minutes, and outputs the predicted fault probability and remaining service life reference. The adaptive threshold early warning module uses a sliding window to statistically analyze historical health data, dynamically calculates the mean and variance, and automatically updates the early warning threshold to avoid misjudgment under fixed thresholds due to equipment aging or load fluctuations. The warning levels are as follows: normal: Fusion features < No action was taken. Notice: ≤Fusion features<1.2 Please pay attention; Warning: 1.2 ≤Fusion feature<1.5 Audible and visual alarms and platform push notifications; urgent: Fusion feature ≥ 1.5 Remote locking and maintenance dispatching; The multi-source heterogeneous data includes: electrical operation data, equipment status data, and environmental monitoring data.

2. The electrical equipment fault early warning system based on data fusion according to claim 1, characterized in that, The data layer noise reduction and fusion uses an improved adaptive Kalman filter algorithm to perform noise reduction and optimal estimation on multi-sensor observation data from the same source. The state equation and observation equation are as follows: Equations of state: ; Observation equation: ; Adaptive noise adjustment formula: ; in: Let k be the system state vector at time k. Here is the state transition matrix. This is process noise; For the observation vector, For the observation matrix, To observe noise; Let α be the adaptive process noise covariance, α be the adjustment coefficient, and N be the number of sensors.

3. The electrical equipment fault early warning system based on data fusion according to claim 1, characterized in that, The weighted fusion of the feature layer employs a weighted cosine similarity feature fusion algorithm to weight and couple the normalized multi-dimensional fault features, generating a unified fused feature vector. The calculation formula is: ； Weight constraints: ; Where M is the feature dimension. The adaptive weights for the i-th dimension feature are: Let i be the i-th eigenvector. This is a reference vector for standard health characteristics.

4. The electrical equipment fault early warning system based on data fusion according to claim 1, characterized in that, The decision-level evidence fusion employs an improved DS evidence theory fusion algorithm, which optimizes the basic probability assignment (BPA) through fuzzy membership degree optimization and introduces a conflict coefficient k to modify the fusion rule. The fusion formula is as follows: ， ； Conflict coefficient: ; in: Assign a basic probability value to the fused faulty proposition A, where K is the coefficient of evidence conflict. , A basic probability assignment function for different data sources.

5. The electrical equipment fault early warning system based on data fusion according to claim 1, characterized in that, The LSTM fault prediction module constructs a time series prediction model and performs future T-factor analysis on the fused features. Based on the fault trend deduction, the hidden layer state update formula is: ； ； in: Let the hidden state be at time t. As input fusion features, This is the output for fault probability prediction. Let W be the activation function, and b be the network weights and biases.

6. The electrical equipment fault early warning system based on data fusion according to claim 1, characterized in that, The adaptive threshold warning module employs a dynamic threshold iteration algorithm to automatically update the warning threshold based on historical health data and real-time fused features. The calculation formula is as follows: ； ； ； in: Let t be the warning threshold. The mean of the sliding window. Let λ be the variance of the sliding window, λ be the confidence coefficient, and L be the length of the sliding window.

7. The electrical equipment fault early warning system based on data fusion according to claim 1, characterized in that, The data preprocessing module performs the following steps sequentially: outlier removal, linear interpolation of missing values, wavelet packet denoising, and max-min normalization. The normalization formula is as follows: ； Furthermore, timestamp alignment and spatial coordinate calibration are performed on multi-source data to ensure the spatiotemporal consistency of the fused data.

8. The electrical equipment fault early warning system based on data fusion according to claim 1, characterized in that, The fault types include: insulation aging, overheating, mechanical loosening, partial discharge, short circuit hazard, and overload operation; The system divides the warning levels into four levels: normal, alert, warning, and emergency, and links them to perform audible and visual alarms, data reporting, and remote locking protection actions.

9. The electrical equipment fault early warning system based on data fusion according to any one of claims 1-8, characterized in that, The system adopts an edge-cloud collaborative architecture, with the edge realizing real-time data acquisition, preprocessing and primary fusion, and the cloud realizing deep learning training, decision fusion and historical data management. The system's early warning indicators are: fault early warning accuracy ≥ 98.5%, fault false alarm rate ≤ 0.5%, average early warning lead time ≥ 30 min, and continuous operation stability ≥ 99.9%.

Images