A transformer defect intelligent identification and early warning method based on multi-source sensing data

Abstract

The present application relates to the technical field of power intelligent substation monitoring, and particularly relates to a transformer defect intelligent identification and early warning method based on multi-source sensing data, which relies on six types of multi-source sensing terminals of substation infrared thermal imaging, CT passive wireless temperature measurement, acoustic voiceprint imaging, ultrahigh frequency / ultrasonic partial discharge, AI visual meter identification, and environmental perception to synchronously collect transformer global operation data; layered denoising and time sequence alignment preprocessing are implemented on the heterogeneous multi-source data, and a temperature, voiceprint, partial discharge, visual, and environmental multi-modal feature set is constructed; a multi-channel attention fusion neural network is introduced to realize cross-modal feature deep fusion, and combined with a transformer defect fault knowledge graph, intelligent discrimination of defect types is completed; a hierarchical early warning signal is output based on a device degradation time sequence trend prediction model, and a substation centralized control system, access control video, fire extinguishing, and dehumidification auxiliary control equipment are linked to execute a linkage disposal strategy.

Description

Technical Field

[0002] This invention relates to the field of intelligent power substation monitoring technology, specifically to a method for intelligent identification and early warning of transformer defects based on multi-source sensing data. Background Technology

[0003] The 110kV substation is a combined 110kV to 10kV / 6kV substation. Previously, all equipment operation, inspection, and defect monitoring relied on manual on-site procedures. Before the upgrade, there were no automated intelligent monitoring methods, resulting in the following existing technical challenges: 1. Single data source for monitoring, resulting in low accuracy in defect identification. Traditional transformer monitoring often uses only a single infrared thermometry or a single-channel partial discharge acquisition. Electromagnetic interference, changes in ambient temperature and humidity, and light intensity can easily cause false alarms. Relying on a single sensor cannot distinguish between "ambient temperature rise" and "equipment contact overheating defects." The single channel of partial discharge is susceptible to interference from power grid harmonics, leading to missed detection of latent insulation defects such as floating potential and tip discharge.

[0004] 2. Multiple types of sensing data operate independently, lacking a cross-modal fusion and analysis mechanism. Infrared, acoustic, partial discharge, and video monitoring systems are deployed independently, resulting in severe data silos. It is impossible to synchronously correlate multiple features such as abnormal temperature, unusual noises, discharge, and oil level drops to comprehensively determine faults. The only solution is for maintenance personnel to manually compare data from multiple systems, which is inefficient and has a strong lag.

[0005] 3. The existing monitoring system only provides over-limit alarms, lacking intelligent identification and trend prediction. It only sets fixed thresholds; alarms are triggered when parameters exceed the limits. It cannot automatically distinguish defect types, locate defect points, or predict equipment degradation trends based on time-series data. Therefore, it cannot provide early warnings, often only issuing alarms when the fault has progressed to a severe stage, which can easily lead to transformer burnout and power outages.

[0006] 4. Lack of integrated response capability for all equipment after alarm. Traditional monitoring systems only push alarm information and cannot automatically link switchgear electric operating mechanisms, fire extinguishing, dehumidification, access control, and video equipment to implement protective measures. After a fault occurs, manual on-site handling is still required, which increases the risk of the fault escalating.

[0007] 5. Incompatible with the centralized control system upgrade requirements for unattended substations. The existing monitoring algorithm lacks a multi-source data access and model self-updating mechanism adapted to the remote centralized control backend, making it impossible to achieve one-click visualization and remote intelligent analysis in the backend, which does not meet the intelligent control requirements of GB / T37546-2019 "Technical Specification for Monitoring Systems of Unattended Substations".

[0008] In existing publicly available technologies, some solutions use a single image or a single partial discharge data for fault identification, and a few multi-source monitoring solutions only perform simple data stitching. They do not design a complete technical link such as layered noise reduction and timing alignment, multi-channel attention weighted fusion, defect knowledge graph correction, and hierarchical linkage handling. As a result, the false detection and false detection rates of defects are high, which cannot meet the engineering requirements of unmanned operation and fully automatic intelligent prevention and control of 110kV substations. Summary of the Invention

[0009] The purpose of this invention is to provide a method for intelligent identification and early warning of transformer defects based on multi-source sensing data, thereby solving the problems mentioned in the background art.

[0010] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: A method for intelligent identification and early warning of transformer defects based on multi-source sensing data, characterized by the following steps: S1. Synchronous acquisition of multi-source sensing data: Deploy multiple types of sensing terminals to collect multi-source heterogeneous sensing data corresponding to the transformer in real time; S2. Hierarchical preprocessing and time-series alignment of multi-source heterogeneous sensing data: Denoising, outlier removal, and normalization preprocessing are performed on the multi-source heterogeneous sensing data respectively. Time-series interpolation alignment of data with different sampling frequencies is completed based on a unified system clock to obtain a synchronous time-series multi-source dataset. S3. Multimodal feature extraction: Extract multi-dimensional defect-related features from the synchronous time-series multi-source dataset and construct a multi-dimensional joint feature library for transformers; S4. Multi-channel attention fusion neural network for defect identification: Construct a multi-channel attention fusion neural network, input the multi-dimensional joint feature library into the multi-channel attention mechanism fusion neural network, complete cross-modal feature weighted fusion, and output the initial judgment result of transformer defects - that is, the initial judgment probability vector of various defects; S5. Precise judgment by matching defect knowledge graph: The preliminary judgment result of the defect is associated and matched with the precise defect knowledge graph of the transformer to accurately identify the subdivided defect types and output the corresponding defect confidence, defect location and deterioration degree parameters. S6. Degradation Trend Prediction and Graded Early Warning: Based on the synchronous time-series multi-source dataset, a time-series prediction model is constructed to deduce the development trend of defect degradation. Combined with defect confidence and degradation rate, multiple early warning levels are divided, and standardized early warning messages are generated. S7. Centralized control system linkage: The standardized early warning messages are pushed to the substation remote centralized control backend, and corresponding automated handling strategies are executed according to different early warning levels, while simultaneously pushing diversified alarm information.

[0011] A further technical solution is that the multi-type sensing terminals mentioned in step S1 include an infrared thermal imager, a CT power-generating passive wireless temperature sensor, an acoustic voiceprint imaging sensing terminal, a dual-channel partial discharge monitoring sensor, an AI visual recognition camera, and an environmental monitoring sensor, with each sensing terminal collecting multi-source sensing data of a corresponding matching type. The multi-source heterogeneous sensing data includes infrared thermal imaging temperature measurement data, CT power supply wireless contact temperature data, acoustic fingerprint imaging data, UHF and ultrasonic partial discharge monitoring data, AI visual image data, and power distribution room environmental monitoring data.

[0012] A further technical solution is that the layered preprocessing in step S2 adopts a differentiated processing method for the multi-source sensing data, including: the infrared thermal imaging temperature measurement data and the CT power-taking wireless contact temperature data are filtered by sliding mean to remove noise, an effective temperature measurement range is set to remove outliers and complete the standardization process. The acoustic signature imaging data is filtered out by an adaptive noise cancellation algorithm to extract effective frequency band features. The UHF and ultrasonic partial discharge monitoring data are filtered to remove power frequency interference and generate a partial discharge statistical spectrum. The AI ​​visual image data is used to enhance contrast, correct distortion, and crop targets to eliminate environmental interference. The environmental monitoring data of the power distribution room is processed by smoothing and filtering and then used as interference correction features in subsequent calculations.

[0013] A further technical solution is that the multi-dimensional defect-related features mentioned in step S3 include temperature features, acoustic signature features, partial discharge PRPD / PRPS statistical features, visual image defect features, and temperature and humidity environmental interference features; the temperature features are obtained by merging and extracting the infrared thermal imaging temperature measurement data and the temperature data of the CT power supply wireless contact.

[0014] A further technical solution is that the multi-channel attention fusion neural network is constructed in step S4 as follows: An independent feature extraction subnetwork corresponding to each type of perceived data is constructed, and a cross-modal attention weight allocation module is added. The weights are dynamically allocated based on the correlation between channel features and transformer defects, and environmental interference features are suppressed. The weighted channel features are then concatenated and input into a fully connected classification layer, and a multi-defect category preliminary judgment probability matrix is ​​output.

[0015] A further technical solution is that the preset transformer defect fault knowledge graph mentioned in step S5 is an entity and relation triplet constructed by collecting historical transformer fault repair cases, power defect judgment standards, and equipment operation and maintenance procedures. The entities mentioned include defect types, defect causes, failure consequences, and remedial measures; the relationship attributes between the entities include perception feature thresholds and corresponding defect probability intervals. The judgment process uses the initial judgment probability of various defects and the partial discharge feature quantity to search and match the knowledge graph, correcting the single-mode recognition bias, and can identify subdivided defects such as metal particle discharge, contact overheating, abnormal oil level, breather failure, mechanical vibration and abnormal noise, porcelain bottle cracks, and insulation gap discharge.

[0016] A further technical solution is that the multi-level early warning system in step S6 is divided into four levels: normal, early warning, alarm, and emergency fault. The rules for each level of early warning are as follows: Normal: All features are within the standard threshold range, with no deterioration trend and no alarm output; Warning: A single modal feature slightly exceeds the standard, the defect confidence level is <60%, the degradation rate is slow, and an inspection reminder is pushed out and the monitoring point is automatically located via a pop-up window; Alarm: Two or more types of multi-source features exceed the standard simultaneously, the defect confidence level is 60%~90%, the degradation rate continues to rise, triggering the background audio and visual alarm and locking the area for continuous recording; Emergency Fault: Multimodal characteristics are severely out of control, with a defect confidence level of ≥90%, posing a risk of fire and insulation breakdown, triggering emergency response.

[0017] A further technical solution is that the automated processing strategy described in step S7 includes: The electric operating mechanisms of the transformer-related switchgear are locked, and remote closing is prohibited. Activate the automatic fire extinguishing device for heptafluoropropane in the power distribution room and increase the operating power of the dehumidification equipment to suppress condensation faults; Remotely unlock substation access control and push fault location videos to the operation and maintenance mobile terminal; Improve the sensitivity of electronic fence intrusion alarms and record fault data in the entire process; The corresponding handling strategy for the warning level is to only push alarm information and not to perform equipment operations; The corresponding handling strategy for the alarm level is to activate the audible and visual alarm and extend the video recording storage period; The emergency fault level response strategy is to simultaneously execute all of the following linkage measures: lock operation, activate fire extinguishing device, unlock access control, increase fence alarm sensitivity, and record fault data on video throughout the process.

[0018] A further technical solution is that the diversified alarm information includes alarm level, defect type, defect location, fault location video and handling suggestions, and alarm forms include background pop-up window, audible and visual alarm signal and mobile terminal push text.

[0019] A further technical solution also includes: S8, incremental model update: periodically expand the labeled samples of defect types and location information confirmed by on-site manual review to the training dataset, carry out incremental transfer learning on the multi-channel attention fusion neural network, dynamically iterate and update the defect identification model to adapt to the long-term operating conditions of the substation.

[0020] Compared with the prior art, the beneficial effects of the present invention are: Heterogeneous data fusion significantly reduces false positive and false negative rates: This invention simultaneously integrates six types of perception data: infrared, passive wireless temperature measurement, soundprint, partial discharge, AI vision, and environment. It automatically suppresses the weight of environmental interference features through a multi-channel attention mechanism and corrects the single-modal recognition bias by combining fault knowledge graph. It can accurately distinguish between environmental temperature rise and equipment heat generation, external noise and transformer mechanical noise, and power grid harmonics and real partial discharge.

[0021] Automatic defect classification and location, and prediction of degradation trends: Traditional solutions can only alarm for exceeding limits. This invention can automatically identify more than ten subdivided defects such as overheated contacts, low oil level, breather failure, porcelain insulator cracks, metal particle discharge, floating potential discharge, and abnormal mechanical vibration and noise, and accurately output the defect location. The time-series prediction model predicts the rate of equipment degradation in advance, realizes early warning of faults, and avoids small defects from escalating into major power outages.

[0022] The four-level early warning system, combined with fully automated linkage of all equipment in the station, divides the system into four levels based on the confidence level of defects and the rate of deterioration. Different levels are matched with differentiated control strategies. In the event of an emergency fault, the system automatically locks the electric operating mechanism of the switchgear, activates heptafluoropropane fire suppression, enhances dehumidification inside the cabinet, remotely opens the access control, and starts full video recording. No maintenance personnel are required to operate the system on-site, which greatly shortens the fault handling window and is suitable for unattended substation operation. Attached Figure Description

[0023] Figure 1 This is a flowchart of the steps of the intelligent identification and early warning method for transformer defects based on multi-source sensing data according to the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0025] All electrical components involved in this application are prior art. Those skilled in the art understand their connection methods. With the help of those skilled in the art, all electrical components in this application and their compatible power supplies can be connected by wires. According to the actual situation, a suitable controller can be selected to meet the control requirements. For specific connections and control sequences, please refer to the description below. The electrical connection between each electrical component is completed in the order of operation. The detailed connection methods are well known in the art. The following mainly introduces the working principle and process, and will not describe the electrical control.

[0026] See Figure 1 The illustrations shown are for illustrative purposes and are intended to be embodiments of the present invention.

[0027] Example 1 This embodiment provides a method for intelligent identification and early warning of transformer defects based on multi-source sensing data, characterized by the following steps: S1. Synchronous acquisition of multi-source sensing data: Deploy multiple types of sensing terminals to collect multi-source heterogeneous sensing data corresponding to the transformer in real time; S2. Hierarchical preprocessing and time-series alignment of multi-source heterogeneous sensing data: Denoising, outlier removal, and normalization preprocessing are performed on the multi-source heterogeneous sensing data respectively. Time-series interpolation alignment of data with different sampling frequencies is completed based on a unified system clock to obtain a synchronous time-series multi-source dataset. S3. Multimodal feature extraction: Infrared thermal imaging temperature measurement data and CT power supply wireless contact temperature data are uniformly classified into temperature feature categories. Multi-dimensional defect-related features are extracted from the synchronous time-series multi-source dataset to construct a transformer multi-dimensional joint feature library. The multi-dimensional defect-related features include temperature features, acoustic signature features, partial discharge PRPD / PRPS statistical features, visual image defect features, and temperature and humidity environmental interference features. S4. Multi-channel attention fusion neural network for defect identification: Construct a multi-channel attention fusion neural network, input the multi-dimensional joint feature library into the multi-channel attention mechanism fusion neural network, complete cross-modal feature weighted fusion, and output the initial judgment result of transformer defects - that is, the initial judgment probability vector of various defects; S5. Precise judgment by matching defect knowledge graph: The preliminary judgment result of the defect is associated and matched with the precise defect knowledge graph of the transformer to accurately identify the subdivided defect types and output the corresponding defect confidence, defect location and deterioration degree parameters. S6. Deterioration Trend Prediction and Graded Early Warning: Based on the synchronous time-series multi-source dataset, a time-series prediction model is constructed. The time-series prediction model adopts a long short-term memory network (LSTM), takes the multi-source feature sequence within a continuous time window as input, iteratively extrapolates the feature evolution trajectory at future moments, outputs the defect deterioration degree prediction curve and deterioration rate index, infers the defect deterioration development trend, and divides multiple early warning levels by combining defect confidence and deterioration rate to generate standardized early warning messages. S7. Centralized control system linkage: The standardized early warning messages are pushed to the substation remote centralized control backend, and corresponding automated handling strategies are executed according to different early warning levels, while simultaneously pushing diversified alarm information.

[0028] As a preferred implementation, the multi-type sensing terminals in step S1 include an infrared thermal imager, a CT power-driven passive wireless temperature sensor, an acoustic voiceprint imaging sensing terminal, a dual-channel partial discharge monitoring sensor, an AI visual recognition camera, and an environmental monitoring sensor. Each sensing terminal collects multi-source sensing data of a corresponding matching type. The multi-source heterogeneous sensing data includes infrared thermal imaging temperature measurement data, CT power supply wireless contact temperature data, acoustic fingerprint imaging data, UHF and ultrasonic partial discharge monitoring data, AI visual image data, and power distribution room environmental monitoring data.

[0029] As a preferred implementation, the layered preprocessing in step S2 adopts a differentiated processing method for the multi-source sensing data, including: the infrared thermal imaging temperature measurement data and the CT power-on wireless contact temperature data are subjected to sliding mean filtering for noise reduction, outliers are removed by setting an effective temperature measurement range, and standardization processing is completed. The acoustic signature imaging data is filtered out by an adaptive noise cancellation algorithm to extract effective frequency band features. The UHF and ultrasonic partial discharge monitoring data are filtered to remove power frequency interference and generate a partial discharge statistical spectrum. The AI ​​visual image data is used to enhance contrast, correct distortion, and crop targets to eliminate environmental interference. The environmental monitoring data of the power distribution room is processed by smoothing and filtering and then used as interference correction features in subsequent calculations.

[0030] As a preferred embodiment, the multi-dimensional defect-related features in step S3 include temperature features, acoustic signature features, partial discharge PRPD / PRPS statistical features, visual image defect features, and temperature and humidity environmental interference features; the temperature features are obtained by merging and extracting the infrared thermal imaging temperature measurement data and the temperature data of the CT power supply wireless contact.

[0031] As a preferred embodiment, the multi-channel attention fusion neural network is constructed in step S4 as follows: An independent feature extraction subnetwork corresponding to each type of perceived data is constructed, and a cross-modal attention weight allocation module is added. The weights are dynamically allocated based on the correlation between channel features and transformer defects, and environmental interference features are suppressed. The weighted channel features are then concatenated and input into a fully connected classification layer, and a multi-defect category preliminary judgment probability matrix is ​​output.

[0032] As a preferred implementation, the preset transformer defect fault knowledge graph mentioned in step S5 is an entity and relation triplet constructed by collecting historical transformer fault repair cases, power defect judgment standards, and equipment operation and maintenance procedures. The entities mentioned include defect types, defect causes, failure consequences, and remedial measures; the relationship attributes between the entities include perception feature thresholds and corresponding defect probability intervals. The judgment process uses the initial judgment probability of various defects and the partial discharge feature quantity to search and match the knowledge graph, correcting the single-mode recognition bias, and can identify subdivided defects such as metal particle discharge, contact overheating, abnormal oil level, breather failure, mechanical vibration and abnormal noise, porcelain bottle cracks, and insulation gap discharge.

[0033] The defect location is obtained by pixel-level fusion registration of the target detection box coordinates of the AI ​​visual recognition camera and the hotspot coordinates of the infrared thermal imager.

[0034] As a preferred implementation, the multi-level early warning system in step S6 is divided into four levels: normal, early warning, alarm, and emergency fault. The rules for each level of early warning are as follows: Normal: All features are within the standard threshold range, with no deterioration trend and no alarm output; Warning: A single modal feature slightly exceeds the standard, the defect confidence level is <60%, the degradation rate is slow, and an inspection reminder is pushed out and the monitoring point is automatically located via a pop-up window; Alarm: Two or more types of multi-source features exceed the standard simultaneously, the defect confidence level is 60%~90%, the degradation rate continues to rise, triggering the background audio and visual alarm and locking the area for continuous recording; Emergency Fault: Multimodal characteristics are severely out of control, with a defect confidence level of ≥90%, posing a risk of fire and insulation breakdown, triggering emergency response.

[0035] As a preferred implementation, the automated processing strategy in step S7 includes: The electric operating mechanisms of the transformer-related switchgear are locked, and remote closing is prohibited. Activate the automatic fire extinguishing device for heptafluoropropane in the power distribution room and increase the operating power of the dehumidification equipment to suppress condensation faults; Remotely unlock substation access control and push fault location videos to the operation and maintenance mobile terminal; Improve the sensitivity of electronic fence intrusion alarms and record fault data in the entire process; The corresponding handling strategy for the warning level is to only push alarm information and not to perform equipment operations; The corresponding handling strategy for the alarm level is to activate the audible and visual alarm and extend the video recording storage period; The emergency fault level response strategy is to simultaneously execute all of the following linkage measures: lock operation, activate fire extinguishing device, unlock access control, increase fence alarm sensitivity, and record fault data on video throughout the process.

[0036] Specifically, the system is divided into four levels of early warning based on defect confidence and degradation rate, with different levels matched with differentiated control strategies. In emergency fault conditions, the system automatically locks the electric operating mechanism of the switchgear, activates heptafluoropropane fire suppression, enhances dehumidification inside the cabinet, remotely opens the access control, and starts full video recording. No maintenance personnel are required to operate on-site, which greatly shortens the fault handling window and is suitable for unattended substation operation mode.

[0037] As a preferred implementation, the diversified alarm information includes alarm level, defect type, defect location, fault location video and handling suggestions, and alarm forms include background pop-up window, audible and visual alarm signal and mobile terminal push text.

[0038] As a preferred implementation, it also includes: S8, incremental model update: pre-constructing a training dataset for training the multi-channel attention fusion neural network, periodically expanding the training dataset with labeled samples of defect types and location information confirmed after on-site manual review, performing incremental transfer learning on the multi-channel attention fusion neural network, dynamically iteratively updating the defect identification model, and adapting to the long-term operating conditions changes of the substation.

[0039] This method simultaneously integrates six types of perception data: infrared, passive wireless temperature measurement, acoustic signature, partial discharge, AI vision, and environmental data. It automatically suppresses the weight of environmental interference features through a multi-channel attention mechanism and corrects the single-modal recognition bias by combining fault knowledge graphs. It can accurately distinguish between ambient temperature rise and equipment heat generation, external noise and transformer mechanical noise, and power grid harmonics and real partial discharge, thereby achieving multi-source heterogeneous data fusion and significantly reducing the false detection and false negative rates.

[0040] This method can automatically identify more than ten subcategories of defects, such as overheated contacts, low oil level, breather failure, porcelain insulator cracks, metal particle discharge, floating potential discharge, and abnormal mechanical vibration and noise, and accurately output the defect location; the time-series prediction model can predict the rate of equipment deterioration in advance, realize early warning of faults, and prevent small defects from escalating into major power outages; and realize automatic defect classification and location and deterioration trend prediction.

[0041] Example 2 Based on Example 1, a method for intelligent identification and early warning of transformer defects based on multi-source sensing data includes the following steps: Step S1: Synchronous acquisition of multi-source sensing data; Deploy a complete set of sensing terminals according to the project's bill of quantities, and collect six types of multi-source data: Seven infrared thermal imagers are deployed at the highest points around the 110kV isolator and main transformer, with a temperature measurement range of -20 to 350℃ and an output of infrared temperature field images at 25 frames per second. The CT-powered passive wireless temperature sensor has 300 points. It is installed on the main transformer terminal block, conductive arm, and switch cabinet busbar connector. It uses 433MHz wireless transmission and the acquisition interval is adaptively adjusted from 2 to 10 seconds. One set of acoustic fingerprint imaging sensing terminal is installed on the side of the main transformer body, with a detection frequency band of 2kHz-65kHz. It automatically shields fan and air conditioner noise and outputs sound pressure contour maps. One set of dual-channel partial discharge monitoring sensor, a dual-channel sensor of ultra-high frequency + ultrasonic, with a sensitivity of 10PC, which collects time-domain and frequency-domain data of discharge pulses and generates PRPD / PRPS spectrum; 50 sets of AI visual recognition cameras, with no less than 5 cameras configured in the main transformer area, each with 8 million visible light pixels, to identify oil level, breather, porcelain insulator, and meter readings; One set of environmental monitoring sensors is installed to collect data on power distribution room temperature and humidity, smoke, water immersion, and harmful gas concentration. All terminals synchronously upload raw data to the centralized control redundant server via Modbus TCP, 433MHz wireless, and IEC61850 protocol. Each sensing terminal synchronizes its clock with the centralized control server via NTP protocol, with an accuracy of no less than 10ms.

[0042] Step S2: Hierarchical preprocessing and temporal alignment of multi-source heterogeneous data; Preprocessing of infrared thermal imaging temperature measurement data and CT power supply wireless contact temperature data: sliding window mean filtering to eliminate electromagnetic jumps, limiting the effective temperature measurement range to -25-125℃, removing outliers outside the range, and standardizing to the [0,1] range. Acoustic acoustic fingerprinting data preprocessing: Adaptive noise cancellation algorithm filters out low-frequency environmental noise below 2kHz, extracts the effective frequency band of 2kHz-65kHz, and extracts spectral energy, peak value, and kurtosis features; Preprocessing of UHF and ultrasonic partial discharge monitoring data: power frequency harmonic threshold filtering of interference pulses, extraction of single discharge amplitude, phase and pulse frequency, and generation of two-dimensional PRPD statistical spectrum; AI visual image data preprocessing: image dehazing, contrast enhancement, target box cropping, and separation of target areas such as transformer body, oil level gauge, breather, and porcelain insulator; Preprocessing of power distribution room environmental monitoring data: Temperature and humidity time series sliding smoothing filter, which is stored in the feature library as interference correction feature; Timing alignment: Based on the 1s system clock, linear interpolation is performed on multi-source data with sampling periods of 2s, 5s, 10s, and 25fps to generate a synchronized multi-source time-series dataset per second.

[0043] Step S3: Multimodal feature extraction; Five categories of features were extracted to construct a joint feature library: Temperature characteristics: maximum temperature, temperature difference, rate of temperature rise, coordinates of hot spot areas; Acoustic signature features: peak frequency, sound pressure level, and anomalous vibration frequency components; Statistical characteristics of partial discharge PRPD / PRPS: discharge frequency, maximum discharge quantity, discharge phase distribution, and spectral texture features; Visual image defect features: oil level scale, breather color change grayscale value, porcelain bottle crack edge features, and gauge count value; Environmental interference features: ambient temperature, humidity, and smoke concentration, used for subsequent feature weight correction.

[0044] Step S4: Defect identification using a multi-channel attention fusion neural network; Network structure construction: Five independent feature extraction subnetworks are set up for temperature, voiceprint, partial discharge, vision, and environment. The subnetworks use lightweight convolution + fully connected layers to extract basic features. Cross-modal attention module: Calculates the correlation similarity between features of each channel and historical fault samples, dynamically assigns feature weights in the 0-1 interval, and automatically reduces the weight of environmental interference features; Feature fusion and classification: The weighted multi-channel features are spliced ​​into a three-layer fully connected classification layer, and the output is a preliminary probability matrix for transformer defects, including categories such as normal, overheated contacts, low oil level, breather failure, tip discharge, floating potential discharge, mechanical noise, and porcelain insulator cracks.

[0045] In step 4, the lightweight convolution can be implemented using grouped convolution or deformable convolution structures.

[0046] As a preferred implementation, lightweight convolution is achieved using grouped convolution: The input features are divided into four groups based on channels, with each group undergoing independent convolution operations using a 3×3 convolution kernel. A total of three convolutional layers are used, all with a 3×3 kernel size and strides of 1, 2, and 1 respectively, resulting in 32, 64, and 128 output channels. ReLU activation is employed, and 2×2 max pooling is used. Global average pooling and two fully connected layers are then used to complete the basic feature output. This grouped convolutional structure reduces the number of parameters and computational cost of convolution operations by grouping channels, while still achieving lightweight basic feature extraction.

[0047] Step S5: Precise assessment using defect knowledge graph matching; Construct a transformer defect and fault knowledge graph triple in advance, with each entity containing the defect type, defect cause, fault consequence, and handling measures.

[0048] Step S6: Deterioration trend prediction and four-level early warning; A time-series prediction model is constructed based on historical time-series characteristic data to infer the future deterioration trend of defects, and a four-level early warning system is established based on the comprehensive confidence level. Normal: All features are within the threshold, with no deterioration trend and no alarms; Warning: A single modal feature slightly exceeds the standard, with a confidence level of <60%. A text message inspection reminder will be sent, and the corresponding monitoring video will be automatically displayed in the background. Alarm: When two or more types of multi-source features exceed the standard simultaneously, with a confidence level of 60%-90%, the background will issue a continuous audio-visual alarm and the video will automatically lock the defect location and record the video in a loop. Emergency Fault: Multimodal characteristics are severely out of control with a confidence level of ≥90%, posing a risk of fire and insulation breakdown, triggering a station-wide emergency response.

[0049] Step S7: Centralized control system coordinated response; Warning level: Only maintenance reminders are pushed, no equipment operation is performed; Alarm level: Activate on-site audible and visual alarms; extend video recording storage period to 30 days. Emergency fault linkage execution: ① Issue remote control commands to lock the electric chassis vehicle and electric grounding switch of the main transformer's related 10kV / 6kV switchgear, prohibiting remote closing operations; ② Activate the cabinet-type heptafluoropropane fire extinguishing device in the distribution room and start the automatic fire extinguishing system inside the cabinet; ③ Increase the full power operation of 3 top-mounted dehumidifiers and 4 industrial dehumidifiers to reduce humidity inside the cabinet and inhibit insulation degradation; ④ Remotely send commands to open the substation's facial recognition access control, allowing maintenance personnel to quickly enter the site by facial recognition; ⑤ Electronic fences improve the sensitivity of intrusion alarms, and cameras in the defective area record continuously for 24 hours; ⑥ Push alarm messages with defect location, temperature, and partial discharge data to the maintenance personnel's mobile phones.

[0050] Step S8: Incremental model update; A training dataset for training the multi-channel attention fusion neural network is pre-constructed. Labeled samples with defect types and locations confirmed by on-site manual review are periodically added to the training dataset. The bottom feature extraction layer of the network is frozen, and transfer learning is performed only on the attention module and classification layer. The model weights are updated to continuously optimize the defect recognition accuracy under different seasons and aging conditions.

[0051] Steps S1-S7 are executed continuously in a second-level cycle, while step S8 is triggered independently in a monthly cycle.

[0052] System deployment and adaptation instructions: The entire algorithm is deployed on two operator stations with uninterrupted power supply from a 6kVA UPS. The system database is authorized to collect 10,000 data points and supports power protocols such as IEC104, IEC61850, and Modbus. It can directly connect to the existing protection devices, 110kV electric pressure plates, electric transfer switches, and one-button sequential control logic in the substation. All the upgrades were completed during the substation's shutdown and maintenance window without affecting the existing power supply operation.

[0053] Example 3 Based on Example 1, a method for intelligent identification and early warning of transformer defects based on multi-source sensing data is applied to a remote centralized control system for 110kV substations. The system adopts a dual-machine hot standby redundancy architecture, supports IEC61850, 104, and Modbus TCP power protocols, and realizes unified access, calculation, and remote control output of multi-source data.

[0054] Example 4 This embodiment provides a computer-readable storage medium storing a computer program. When executed by a processor, the computer program is used to complete the corresponding steps of the intelligent identification and early warning method for transformer defects based on multi-source sensing data in Embodiment 1.

[0055] Although the invention has been described herein with reference to several illustrative embodiments, it should be understood that many other modifications and implementations can be devised by those skilled in the art, which will fall within the scope and spirit of the principles disclosed herein. More specifically, various variations and modifications can be made to the components and / or layout of the subject matter combination within the scope of the disclosure, drawings, and claims. Besides variations and modifications to the components and / or layout, other uses will be apparent to those skilled in the art.

Claims

1. A method for intelligent identification and early warning of transformer defects based on multi-source sensing data, characterized in that, Includes the following steps: S1. Synchronous acquisition of multi-source sensing data: Deploy multiple types of sensing terminals to collect multi-source heterogeneous sensing data corresponding to the transformer in real time; S2. Hierarchical preprocessing and time-series alignment of multi-source heterogeneous sensing data: Denoising, outlier removal, and normalization preprocessing are performed on the multi-source heterogeneous sensing data respectively. Time-series interpolation alignment of data with different sampling frequencies is completed based on a unified system clock to obtain a synchronous time-series multi-source dataset. S3. Multimodal feature extraction: Extract multi-dimensional defect-related features from the synchronous time-series multi-source dataset and construct a multi-dimensional joint feature library for transformers; S4. Multi-channel attention fusion neural network for defect identification: Construct a multi-channel attention fusion neural network, input the multi-dimensional joint feature library into the multi-channel attention mechanism fusion neural network, complete cross-modal feature weighted fusion, and output the initial judgment result of transformer defects - that is, the initial judgment probability vector of various defects; S5. Precise judgment by matching defect knowledge graph: The preliminary judgment result of the defect is associated and matched with the precise defect knowledge graph of the transformer to accurately identify the subdivided defect types and output the corresponding defect confidence, defect location and deterioration degree parameters. S6. Degradation Trend Prediction and Graded Early Warning: Based on the synchronous time-series multi-source dataset, a time-series prediction model is constructed to deduce the development trend of defect degradation. Combined with defect confidence and degradation rate, multiple early warning levels are divided, and standardized early warning messages are generated. S7. Centralized control system linkage: The standardized early warning messages are pushed to the substation remote centralized control backend, and corresponding automated handling strategies are executed according to different early warning levels, while simultaneously pushing diversified alarm information.

2. The intelligent identification and early warning method for transformer defects based on multi-source sensing data according to claim 1, characterized in that, The multiple types of sensing terminals mentioned in step S1 include infrared thermal imagers, CT power-driven passive wireless temperature sensors, acoustic voiceprint imaging sensing terminals, dual-channel partial discharge monitoring sensors, AI visual recognition cameras, and environmental monitoring sensors. Each sensing terminal collects multi-source sensing data of a corresponding matching type. The multi-source heterogeneous sensing data includes infrared thermal imaging temperature measurement data, CT power supply wireless contact temperature data, acoustic fingerprint imaging data, UHF and ultrasonic partial discharge monitoring data, AI visual image data, and power distribution room environmental monitoring data.

3. The intelligent identification and early warning method for transformer defects based on multi-source sensing data according to claim 2, characterized in that, The layered preprocessing in step S2 adopts a differentiated processing method for the multi-source sensing data, including: the infrared thermal imaging temperature measurement data and the CT power-taking wireless contact temperature data are subjected to sliding mean filtering for noise reduction, outliers are removed by setting an effective temperature measurement range, and standardization processing is completed. The acoustic signature imaging data is filtered out by an adaptive noise cancellation algorithm to extract effective frequency band features. The UHF and ultrasonic partial discharge monitoring data are filtered to remove power frequency interference and generate a partial discharge statistical spectrum. The AI ​​visual image data is used to enhance contrast, correct distortion, and crop targets to eliminate environmental interference. The environmental monitoring data of the power distribution room is processed by smoothing and filtering and then used as interference correction features in subsequent calculations.

4. The intelligent identification and early warning method for transformer defects based on multi-source sensing data according to claim 1, characterized in that, The multi-dimensional defect-related features mentioned in step S3 include temperature features, acoustic signature features, partial discharge PRPD / PRPS statistical features, visual image defect features, and temperature and humidity environmental interference features.

5. The intelligent identification and early warning method for transformer defects based on multi-source sensing data according to claim 1, characterized in that, The method for constructing the multi-channel attention fusion neural network in step S4 is as follows: An independent feature extraction subnetwork corresponding to each type of perceived data is constructed, and a cross-modal attention weight allocation module is added. The weights are dynamically allocated based on the correlation between channel features and transformer defects, and environmental interference features are suppressed. The weighted channel features are then concatenated and input into a fully connected classification layer, and a multi-defect category preliminary judgment probability matrix is ​​output.

6. The intelligent identification and early warning method for transformer defects based on multi-source sensing data according to claim 1, characterized in that, The preset transformer defect and fault knowledge graph mentioned in step S5 is an entity and relation triplet constructed by collecting historical transformer fault repair cases, power defect judgment standards, and equipment operation and maintenance procedures. The entities mentioned include defect types, defect causes, failure consequences, and remedial measures; the relationship attributes between the entities include perception feature thresholds and corresponding defect probability intervals. The judgment process uses the initial judgment probability of various defects and the partial discharge feature quantity to search and match the knowledge graph, correcting the single-mode recognition bias, and can identify subdivided defects such as metal particle discharge, contact overheating, abnormal oil level, breather failure, mechanical vibration and abnormal noise, porcelain bottle cracks, and insulation gap discharge.

7. The intelligent identification and early warning method for transformer defects based on multi-source sensing data according to claim 1, characterized in that, The multi-level early warning system described in step S6 is divided into four levels: normal, early warning, alarm, and emergency fault. The rules for each level of early warning are as follows: normal: All features are within the standard threshold range, with no deterioration trend and no alarm output; Warning: A single modal feature slightly exceeds the standard, the defect confidence level is <60%, the degradation rate is slow, and an inspection reminder is pushed out and the monitoring point is automatically located via a pop-up window; Alarm: Two or more types of multi-source features exceed the standard simultaneously, the defect confidence level is 60%~90%, the degradation rate continues to rise, triggering the background audio and visual alarm and locking the area for continuous recording; Emergency Fault: Multimodal characteristics are severely out of control, with a defect confidence level of ≥90%, posing a risk of fire and insulation breakdown, triggering emergency response.

8. The intelligent identification and early warning method for transformer defects based on multi-source sensing data according to claim 1, characterized in that, The automated processing strategy described in step S7 includes: The electric operating mechanisms of the transformer-related switchgear are locked, and remote closing is prohibited. Activate the automatic fire extinguishing device for heptafluoropropane in the power distribution room and increase the operating power of the dehumidification equipment to suppress condensation faults; Remotely unlock substation access control and push fault location videos to the operation and maintenance mobile terminal; Improve the sensitivity of electronic fence intrusion alarms and record fault data in the entire process; The corresponding handling strategy for the warning level is to only push alarm information and not to perform equipment operations; The corresponding handling strategy for the alarm level is to activate the audible and visual alarm and extend the video recording storage period; The emergency fault level response strategy is to simultaneously execute all of the following linkage measures: lock operation, activate fire extinguishing device, unlock access control, increase fence alarm sensitivity, and record fault data on video throughout the process.

9. The intelligent identification and early warning method for transformer defects based on multi-source sensing data according to claim 1, characterized in that, The diversified alarm information includes alarm level, defect type, defect location, fault location video and handling suggestions. Alarm formats include background pop-ups, audible and visual alarm signals and mobile push text.

10. The intelligent identification and early warning method for transformer defects based on multi-source sensing data according to claim 1, characterized in that, Also includes: S8. Incremental Model Update: A training dataset for training the multi-channel attention fusion neural network is pre-constructed. Labeled samples of defect types and locations confirmed by on-site manual review are periodically added to the training dataset. Incremental transfer learning is carried out on the multi-channel attention fusion neural network to dynamically iterate and update the defect identification model to adapt to the long-term operating conditions of the substation.

Images