Multi-parameter anomaly feature extraction and ai fault diagnosis method and system

Abstract

The application discloses a multi-parameter abnormal feature extraction and AI fault diagnosis method and system, relates to the technical field of power equipment online monitoring and intelligent fault diagnosis, and comprises the following steps: acquiring multi-source sensing monitoring data, wherein the multi-source sensing monitoring data comprises partial discharge signals, overvoltage signals, leakage currents, temperatures and voltage currents; performing multi-domain fusion abnormal feature extraction on the multi-source sensing monitoring data to obtain multi-domain features, fusing and encoding the multi-domain features to obtain a comprehensive abnormal feature set, wherein the multi-domain features comprise time domain, frequency domain, time-frequency domain and working condition domain; inputting the comprehensive abnormal feature set into a pre-established lightweight AI model to output fault diagnosis results; and performing risk grading early warning output based on the fault diagnosis results, which can significantly improve the operation safety of power equipment.

Description

Technical Field

[0001] This invention relates to the field of online monitoring and intelligent fault diagnosis technology for power equipment, specifically to a method and system for fault diagnosis based on multi-parameter anomaly feature extraction and AI. Background Technology

[0002] High-voltage equipment in new energy power station systems (10kV or 35kV) includes switchgear or ring main units, prefabricated substations, and power collection cables, among others. The number of these devices is enormous and their distribution is widespread. In the categories of related electrical equipment failures, over 50% are caused by insulation defects, including the continued development of partial discharge, insulation dampness, overheating of conductor joints, overvoltage surges, and abnormally increased leakage current. These types of faults often progress from early signs to final breakdown within just a few days to a few weeks, and once they occur, they can directly lead to power outages, equipment burnout, and even personal injury accidents.

[0003] Existing on-site monitoring technologies have certain problems. They do not perform integrated synchronous acquisition of multi-source insulation parameters and adaptive extraction of multi-dimensional anomaly features, thus failing to achieve real-time diagnosis and graded early warning of power equipment faults, which means that the operational safety performance of the power system cannot be guaranteed. Summary of the Invention

[0004] To address the shortcomings mentioned in the background art, the present invention aims to provide a method and system for multi-parameter anomaly feature extraction and AI fault diagnosis.

[0005] Firstly, the objective of this invention can be achieved through the following technical solution: a method based on multi-parameter anomaly feature extraction and AI fault diagnosis, the method comprising the following steps: Acquire multi-source sensor monitoring data, including partial discharge signal, overvoltage signal, leakage current, temperature, and voltage and current. Multi-domain fusion anomaly feature extraction is performed on multi-source sensor monitoring data to obtain multi-domain features. The multi-domain features are then fused and encoded to obtain a comprehensive anomaly feature set. The multi-domain features include time domain, frequency domain, time-frequency domain, and operating condition domain. The comprehensive set of abnormal features is input into a pre-established lightweight AI model, and the output is the fault diagnosis result. Based on the fault diagnosis result, risk classification and early warning are output.

[0006] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the extraction process of the multi-domain features, comprising: Temporal feature extraction: Extract time-series statistical features such as mean, variance, peak value, kurtosis, waveform distortion rate, rate of change, and steady-state offset to reflect the steady-state deviation and slow deterioration trend of equipment operation; Frequency domain feature extraction: Through FFT spectrum analysis, harmonic component decomposition, and high-frequency component amplitude ratio analysis, abnormal frequency components caused by partial discharge, high-frequency vibration, and poor circuit contact are captured. Time-frequency domain feature extraction: Wavelet transform and Hilbert transform are used to perform joint time-frequency analysis on non-stationary transient fault signals to locate the fault occurrence time and characteristic frequency distribution, and to capture transient abrupt anomalies. Operating condition related domain characteristics: By associating external conditions such as unit load, ambient temperature and humidity, seasonal operating conditions, and operating time, the interference of operating condition fluctuations is removed, and the pure equipment degradation characteristics after eliminating the influence of operating conditions are extracted.

[0007] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the process of fusing and encoding multi-domain features, comprising: Multi-domain features are filtered, dimensionality reduced, and fused for encoding. Redundant and irrelevant features are eliminated to construct a comprehensive abnormal feature set with high discriminative power and low redundancy, laying the foundation for lightweight AI model input.

[0008] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the construction process of the pre-established lightweight AI model, as follows: Pruning, quantization, convolution kernel simplification, and structural distillation are performed on deep learning networks to compress model parameters and computational load, adapt to low-computing hardware deployment of local edge gateways and monitoring units, and balance diagnostic accuracy and inference real-time performance to obtain the processed deep learning network. The deep learning network is pre-trained based on an existing fault sample library of similar power equipment. It is then fine-tuned for transfer learning based on actual fault samples of equipment at field sites, and finally outputs a lightweight AI model.

[0009] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: an expression for the lightweight AI model, as follows: Let the original deep learning network model be... Its network parameters are The lightweight AI model The construction process satisfies the following function mapping relationship: Step 1, Model Compression: in: This represents the quantization operation function, which quantizes the original model parameters. The quantization model is obtained by mapping high-precision floating-point numbers to low-bit integers. , This refers to the quantization bit width parameter; This represents the pruning and convolution kernel simplification operation functions, and the quantization model is evaluated. The importance weights of each channel or convolutional kernel are used to remove those below a threshold. The redundant structure is used to obtain a simplified model. ; Represent the structural distillation operation function to simplify the model. As a student network, with the original model As the teacher network, a compressed model is obtained by minimizing the distillation loss of the student and teacher networks on the soft-label output and jointly training them with the hard-label loss. , This is the hyperparameter for balancing distillation temperature and losses. The second step, pre-training and transfer fine-tuning: This represents the pre-training operation function to compress the model. As the initial model, it is based on an existing fault sample library of similar power equipment. Pre-training is performed by updating the model parameters by minimizing the fault classification loss function to obtain the pre-trained model. ; Represents the few-shot transfer learning operation function, used in the pre-trained model. Based on this, and targeting the actual fault sample set of on-site electrical equipment Perform migration fine-tuning, among which, By employing a few-shot learning strategy, some low-level network parameters are frozen, and parameters are updated only for the high-level classification layer, ultimately resulting in a lightweight AI model adapted to the target site. ; The lightweight AI model The output is a fault diagnosis result vector y=[p1, p2, ..., pn], where pi represents the probability of the i-th type of fault and n is the total number of fault categories, which satisfies the real-time inference requirements under the computing power constraints of the edge gateway.

[0010] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the risk classification and early warning output based on the fault diagnosis results includes full-level control from early hidden danger warning, mid-term abnormal alarm to emergency fault trip warning.

[0011] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the process of acquiring the multi-source sensing monitoring data, including: For rapidly changing signals such as transient faults, partial discharges, and high-frequency abnormal noises, high-frequency synchronous sampling is used; for steady-state operating parameters, timed periodic sampling is used. All measurement points are packaged and sent up at the same time, frequency, and timestamp to eliminate subsequent analysis errors caused by timing misalignment.

[0012] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the preprocessing of the multi-source sensing monitoring data includes data denoising, outlier removal, and normalization.

[0013] Secondly, in order to achieve the above objectives, this invention discloses a multi-parameter anomaly feature extraction and AI fault diagnosis system, comprising: The data acquisition module is used to acquire multi-source sensor monitoring data, which includes partial discharge signal, overvoltage signal, leakage current, temperature, and voltage and current. The data processing module is used to extract multi-domain fusion anomaly features from multi-source sensor monitoring data to obtain multi-domain features, and to fuse and encode the multi-domain features to obtain a comprehensive anomaly feature set. The multi-domain features include time domain, frequency domain, time-frequency domain and operating condition domain. The diagnostic output module is used to input a comprehensive set of abnormal features into a pre-established lightweight AI model and output the fault diagnosis results. Based on the fault diagnosis results, risk classification and early warning output are performed.

[0014] This invention enables integrated synchronous acquisition of multi-source insulation parameters, high-precision signal purification under strong interference environments, adaptive extraction of multi-dimensional abnormal features, efficient deployment of lightweight AI models at the edge, and real-time fault diagnosis and graded early warning, significantly improving the operational safety, reliability, and intelligent operation and maintenance level of power equipment. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Example 1: like Figure 1 As shown, the method based on multi-parameter anomaly feature extraction and AI fault diagnosis includes the following steps: S101: Acquire multi-source sensor monitoring data, which includes partial discharge signal, overvoltage signal, leakage current, temperature, and voltage and current. Specifically, multi-source sensing monitoring data is obtained based on monitoring by multiple sensors, and the sensor design process is as follows: The integrated sensor features a one-piece metal shielded shell and a fully sealed potting structure. Its overall design employs a compact, modular layout, with internal sensing units arranged in isolation. Crosstalk between different types of sensing signals is suppressed through partitioned electromagnetic shielding partitions. The shell is made of high and low temperature resistant and corrosion-resistant engineering materials, combined with sealing rings and potting processes to achieve dustproof, waterproof, and condensation-proof protection levels. Internally, it integrates a multi-channel parallel signal acquisition link. Each sensing signal is independently configured with high-precision signal conditioning, impedance matching, programmable amplification, and adaptive filtering circuits, effectively suppressing power frequency interference, pulse interference, and high-frequency stray noise.

[0018] The process of acquiring multi-source sensor monitoring data includes: Employing a BeiDou / GPS dual-mode timing synchronization protocol, it provides a millisecond-level unified clock reference for on-site monitoring units, sensor terminals, and edge gateways, eliminating inherent clock deviations between devices from different manufacturers and achieving time axis alignment across all measurement points. It covers multi-dimensional heterogeneous parameters such as partial discharge signals, overvoltage signals, leakage current, temperature, and conventional electrical quantities (voltage, current). High-frequency synchronous sampling is used for rapidly changing signals such as transient faults, partial discharges, and high-frequency abnormal noises; timed periodic sampling is used for steady-state operating parameters. All measurement points are packaged and transmitted simultaneously, at the same frequency, and with the same timestamp, eliminating subsequent analysis errors caused by timing misalignment. Relying on the on-site monitoring unit, it realizes front-end signal conditioning, filtering, AD conversion, and synchronous packetization, possessing the ability to resume sampling after interruption and local buffering and retransmission capabilities, ensuring that acquired data is not lost or out of sync under extreme operating conditions.

[0019] The preprocessing of multi-source sensor monitoring data includes: Employing a terminal-edge preprocessing architecture, the raw collected data is cleaned, corrected, normalized, and dimensionality reduced to provide a high-quality standard dataset for feature extraction and AI modeling.

[0020] The algorithms used include: Denoising the raw data: The operating environment of local equipment is complex, with significant interference from power frequency, corona, switching pulses, and radio frequency interference, resulting in a signal-to-noise ratio as low as -12dB. Direct feature extraction would lead to complete diagnostic failure. This invention designs differentiated denoising strategies for different signal characteristics: 1) Partial discharge signal: Wavelet thresholding denoising + EMD empirical mode decomposition is adopted. First, the signal is decomposed into 6 levels of wavelet, and the high-frequency coefficients are shrunk by soft thresholding. Then, the 6th to 8th order intrinsic mode functions are obtained through EMD decomposition. The noise-dominant components are removed, which can improve the signal-to-noise ratio to more than 18dB.

[0021] 2) Leakage current signal: The sliding average filter + adaptive Kalman filter is adopted, and the sliding window length is set to 20 points to effectively suppress 50Hz power frequency interference and random pulse interference, so that the leakage current measurement error is controlled within ±1μA.

[0022] 3) Temperature signal: Median filtering is used with a window length of 5 points to eliminate instantaneous interference points caused by lightning strikes and switching actions.

[0023] Outlier removal: Double verification using the 3σ criterion and box method: 3σ criterion: Calculate the mean μ and standard deviation σ of the data, and remove extreme values ​​other than [μ-3σ, μ+3σ]. Box method: outliers are identified using the interquartile range (IQR). The dual strategy can improve data effectiveness to 99.2%, avoiding misdiagnosis due to single points of anomaly.

[0024] Normalization process: The maximum-minimum normalization method is adopted: X'=(X-Xmin) / (Xmax-Xmin), which maps data of different dimensions such as partial discharge (pC), leakage current (μA), temperature (°C), and voltage (kV) to the interval [0,1], thereby eliminating the impact of magnitude differences on AI model training and improving convergence speed and diagnostic accuracy.

[0025] S102: Extract multi-domain fusion anomaly features from multi-source sensor monitoring data to obtain multi-domain features, fuse and encode the multi-domain features to obtain a comprehensive anomaly feature set, wherein the multi-domain features include time domain, frequency domain, time-frequency domain and operating condition domain. The process of extracting multi-domain features includes: Temporal feature extraction: Extract time-series statistical features such as mean, variance, peak value, kurtosis, waveform distortion rate, rate of change, and steady-state offset to reflect the steady-state deviation and slow deterioration trend of equipment operation; Frequency domain feature extraction: Through FFT spectrum analysis, harmonic component decomposition, and high-frequency component amplitude ratio analysis, abnormal frequency components caused by partial discharge, high-frequency vibration, and poor circuit contact are captured. Time-frequency domain feature extraction: Wavelet transform and Hilbert transform are used to perform time-frequency joint analysis on non-stationary transient fault signals to locate the fault occurrence time and characteristic frequency distribution, and capture transient abrupt anomalies; Operating condition related domain characteristics: By associating external conditions such as unit load, ambient temperature and humidity, seasonal operating conditions, and operating time, the interference of operating condition fluctuations is removed, and the pure equipment degradation characteristics after eliminating the influence of operating conditions are extracted.

[0026] This process breaks through the limitations of single time-domain features and integrates multiple dimensions such as time domain, frequency domain, time-frequency domain, operating condition domain, and correlation domain to mine hidden anomalies and early fault characteristics of equipment, achieving a deep transformation from "data values" to "fault characteristics".

[0027] The process of fusing and encoding multi-domain features includes: Multi-domain features are filtered, dimensionality reduced, and fused for encoding. Redundant and irrelevant features are eliminated to construct a comprehensive abnormal feature set with high discriminative power and low redundancy, laying the foundation for lightweight AI model input.

[0028] S103: Input the comprehensive abnormal feature set into the pre-established lightweight AI model, output the fault diagnosis result, and output the risk classification warning based on the fault diagnosis result.

[0029] The process of building and training a pre-built lightweight AI model is as follows: Traditional deep learning networks are pruned, quantized, have their convolution kernels simplified, and their structure distilled to compress model parameters and computational load, making them suitable for low-computing hardware deployments in local edge gateways and monitoring units. This approach balances diagnostic accuracy and inference real-time performance, achieving a lightweight model design.

[0030] After completing model pruning, quantization, and structural simplification to obtain a lightweight model, in order to solve the accuracy recovery problem in small-sample scenarios, transfer fine-tuning training based on multi-parameter anomaly features is performed to conduct small-sample transfer learning. The specific configuration of the training process is as follows: Differentiated hyperparameter strategy: The learning rate of the backbone feature layer is set to 1e. -5 ~1e -4 The learning rate for the fault classification head layer is set to 1e. -3 ~1e -2The AdamW optimizer is selected, with a fixed weight decay of 0.01. The training batch size is limited to 8~16 to adapt to the limited hardware resources at the edge, and can be combined with gradient accumulation to simulate large batch training. A Dropout layer of 0.3~0.5 is set before the fully connected layer, and a label smoothing mechanism with a coefficient of 0.1 is introduced at the same time to suppress small sample overfitting.

[0031] Training round control logic: Before training, take 5%~10% of the total number of steps for linear warm-up of the learning rate to avoid gradient oscillation in the early stage; small sample fine-tuning limits the total number of training rounds to 20~50 rounds, with a forced early stopping mechanism: when the validation set loss does not decrease for 5~10 consecutive rounds, training is automatically terminated and the optimal weights are saved to avoid overfitting.

[0032] Loss function: A combined loss method is adopted, with Focal Loss used for the main classification and a fixed modulation factor γ of 2.0 to address the class imbalance problem between faulty and normal samples; Supporting... Divergence distillation loss, with a temperature coefficient T set to 3-5, allows the lightweight student model to fit the soft output of the pre-trained large model; the total loss is... The balance coefficients λ1 and λ2 are initially set to 1.0 and 0.5, respectively, and can be fine-tuned according to the accuracy of the validation set.

[0033] The model can automatically identify typical fault types such as insulation aging, partial discharge, and circuit overheating, and output quantitative assessments of fault confidence, fault location, and degree of deterioration.

[0034] It supports local offline diagnostics at the edge, enabling real-time assessment of device status without relying on the cloud; the cloud enables model iteration updates, version distribution, and parameter optimization, forming an AI diagnostic closed loop of "edge inference + cloud iteration".

[0035] The expression for a lightweight AI model is as follows: Let the original deep learning network model be... Its network parameters are The lightweight AI model The construction process satisfies the following function mapping relationship: Step 1, Model Compression: in: This represents the quantization operation function, which quantizes the original model parameters. The quantization model is obtained by mapping high-precision floating-point numbers to low-bit integers. , This refers to the quantization bit width parameter; This represents the pruning and convolution kernel simplification operation functions, and the quantization model is evaluated. The importance weights of each channel or convolutional kernel are used to remove those below a threshold. The redundant structure is used to obtain a simplified model. ; Represent the structural distillation operation function to simplify the model. As a student network, with the original model As the teacher network, a compressed model is obtained by minimizing the distillation loss of the student and teacher networks on the soft-label output and jointly training them with the hard-label loss. , This is the hyperparameter for balancing distillation temperature and losses.

[0036] The second step, pre-training and transfer fine-tuning: This represents the pre-training operation function to compress the model. As the initial model, it is based on an existing fault sample library of similar power equipment. Pre-training is performed by updating the model parameters by minimizing the fault classification loss function to obtain the pre-trained model. ; Represents the few-shot transfer learning operation function, used in the pre-trained model. Based on this, and targeting the actual fault sample set of on-site electrical equipment (in Transfer learning is performed to fine-tune the model, employing a few-shot learning strategy. Some low-level network parameters are frozen, and only the parameters of the high-level classification layers are updated, ultimately resulting in a lightweight AI model adapted to the target site. .

[0037] The lightweight AI model The output is a fault diagnosis result vector y=[p1, p2, ..., pn], where pi represents the probability of the i-th type of fault and n is the total number of fault categories. The model meets the real-time inference requirements under the computing power constraints of the edge gateway.

[0038] Risk classification and early warning output based on fault diagnosis results includes full-level control from early hidden danger warning, mid-term abnormal alarm to emergency fault trip warning.

[0039] Specifically, it includes: Level 1 (Minor Risk): Characteristic indicators deviate slightly from the baseline value, with no immediate operational risk. This is an early sign of slow equipment deterioration, triggering a trend warning, prompting inspection and monitoring recommendations, and inclusion in medium- to long-term status tracking.

[0040] Level 2 (General Anomaly): Key parameters or fusion characteristics significantly exceed the standard, showing a gradual deterioration and expansion trend, posing potential operational risks, triggering an anomaly alarm, and pushing an anomaly cause analysis, on-site verification points, and troubleshooting steps.

[0041] Level 3 (Severe Fault): Characteristic indicators exceed safety thresholds, typical fault characteristics appear, and there is a risk of equipment damage, unit shutdown, and safety accidents. This triggers an emergency warning, which is linked to audible and visual prompts, platform pop-ups, and SMS / message push notifications, providing emergency handling suggestions and prevention and control measures.

[0042] Adaptive threshold setting: Supports automatic setting of early warning thresholds based on historical operating data, benchmarking against similar equipment, and national / industry standard limits. It can be dynamically corrected according to season and load conditions to avoid false or missed early warnings.

[0043] This embodiment also includes a process of local rolling storage + cloud persistent management, as follows: Construct a two-tier data storage architecture consisting of edge local distributed storage and cloud centralized persistent storage to achieve full lifecycle security management of runtime data, feature data, alarm records, and diagnostic logs.

[0044] Local rolling storage: The local monitoring unit and edge gateway are configured with large-capacity local storage, which is divided into time slices for rolling storage, retaining recent high-frequency raw sampling data and real-time alarm data; it supports breakpoint caching and offline data caching, and automatically resumes transmission after network recovery, so as not to lose critical on-site data.

[0045] Cloud-based persistent management: The cloud platform permanently stores the standardized collected data, preprocessed data, feature data, AI diagnostic results, early warning events, and operation and maintenance records. It adopts a hybrid architecture of time-series database and relational database, which is suitable for high-speed reading and writing of massive time-series data and querying related business information.

[0046] Specifically, the present invention will be further illustrated below through embodiments: Hardware deployment scenarios Multidimensional sensors are installed in three key locations: the cable compartment, circuit breaker compartment, and busbar compartment of the 10kV (or 35kV) switchgear; the local monitoring unit is installed on the cabinet door panel; the edge AI module is mounted on the rail inside the cabinet and communicates with the local monitoring unit; the power supply is connected to AC220V.

[0047] Typical fault diagnosis demonstration: Case 1: Switchgear contact overheating fault One day, during a routine inspection using the online monitoring system, maintenance personnel discovered an abnormal fluctuation in the temperature of the B-phase contact of the 10kV switchgear. The initial monitored temperature was 38℃, which was basically the same as the contact temperatures of the A-phase and C-phase contacts in the same cabinet (both maintained at 35-37℃). This did not reach the warning threshold. After initially recording the data, the maintenance personnel continued to monitor its changing trend.

[0048] Subsequently, monitoring system data showed that the temperature of phase B contact continued to rise. After calculation, the temperature rise rate stabilized at 1.8℃ / hour. After about 10 hours of continuous heating, the temperature finally climbed to 56℃, far exceeding the normal operating temperature range of switchgear contacts (generally not exceeding 40℃). At this point, obvious signs of overheating had appeared.

[0049] The maintenance personnel immediately retrieved the associated monitoring data of the switchgear and found that, along with the temperature rise, the partial discharge level showed a slight abnormal increase, gradually increasing from the initial 45pC to 92pC. Although it did not reach the severe discharge threshold, it formed a clear abnormal correlation with the temperature change. Meanwhile, the leakage current remained stable at 5.2μA without any fluctuation, ruling out the possibility of overheating caused by insulation damage.

[0050] To further accurately determine the fault type, maintenance personnel extracted key monitoring features for that period and found that the temperature gradient (temperature difference with other phases in the same cabinet), temperature rise rate, and coupling degree between temperature and partial discharge of phase B contacts were significantly abnormal, with large deviations compared with historical normal operation data, exhibiting typical characteristics of overheating caused by poor contact.

[0051] After the above abnormal data is entered into the AI ​​fault diagnosis system, the system combines historical fault cases, equipment parameters and real-time monitoring data to quickly give a diagnosis result: contact overheating fault, with a diagnosis confidence of 92.3%. At the same time, based on the severity and development trend of the fault, a three-level warning is automatically triggered to remind maintenance personnel to arrive on site in time to avoid the fault from escalating further.

[0052] Upon receiving the warning notification, maintenance personnel immediately rushed to the site with specialized tools, temperature measuring instruments, and spare parts. Strictly following the switchgear maintenance operation procedures, they first disconnected the relevant power supply, implemented safety precautions, and then opened the switchgear door to conduct a comprehensive inspection of the B-phase contacts. The inspection revealed that the B-phase contact connections were loose, and the contact surfaces showed signs of oxidation and wear due to long-term operation, leading to increased contact resistance and overheating. Simultaneously, the loose contacts created tiny air gaps, resulting in minor partial discharge.

[0053] The maintenance personnel immediately began troubleshooting. First, they used specialized tools to tighten the loose contact connections one by one to ensure reliable connections. Then, they carefully sanded the contact surfaces with fine sandpaper to remove oxide layers and wear impurities, restoring the conductivity of the contact surfaces. After the troubleshooting was completed, they reconnected the power supply and started the monitoring system for real-time tracking.

[0054] After about 30 minutes of observation, the temperature of phase B contact gradually dropped and eventually stabilized at 36℃, consistent with the temperature of other phases in the same cabinet; the partial discharge also dropped to 48pC, returning to the normal range, and the leakage current remained stable at 5.2μA. All monitoring indicators met the normal operating standards, confirming that the overheating fault had been completely eliminated.

[0055] Table 1. Comparison of Implementation in Case 1 Case 2: Early Moisture Insulation During a routine inspection, maintenance personnel discovered an anomaly in the monitoring data of a 35kV switchgear through the online monitoring system. They paid particular attention to the significant fluctuation in the leakage current indicator—the leakage current, which was originally stable at 4.1μA, gradually increased in a short period of time, eventually reaching 26.3μA, exceeding the normal operating range. This anomaly drew the attention of the maintenance personnel.

[0056] To gain a comprehensive understanding of the equipment's status, maintenance personnel immediately retrieved other relevant monitoring data for the switchgear. They discovered that in addition to abnormal leakage current, the frequency of partial discharges had also increased significantly, rising from the normal 2 times / second to 18 times / second, indicating a marked increase in discharge activity. In contrast, the temperatures of each phase of the switchgear remained stable, without any significant temperature rise or overheating-related anomalies. Based on these data characteristics, the maintenance personnel initially determined that the equipment might have an insulation problem, most likely caused by early moisture absorption of the insulation layer leading to a decline in insulation performance.

[0057] To further confirm the type and severity of the fault, maintenance personnel entered all abnormal data into the AI ​​fault diagnosis system. The system, combining the equipment's operational history, historical maintenance records, on-site environmental parameters, and real-time monitoring data, conducted a comprehensive analysis and provided a clear diagnosis: insulation degradation / moisture absorption, with a diagnostic confidence level of 87.6%. Based on the fault's development stage and potential risks, the system automatically triggered a level-two warning, prompting maintenance personnel to take timely measures to prevent further insulation degradation and avoid developing into serious faults such as surface flashover, which could cause equipment damage and power outages.

[0058] Upon receiving the warning, maintenance personnel quickly formulated a response plan and rushed to the site with drying equipment and insulation testing instruments. Upon arrival, they first disconnected the equipment power supply, implemented safety precautions, and conducted a comprehensive inspection of the switchgear's interior, focusing on the appearance of the insulating components. No obvious damage or carbonization was found, but the insulation resistance data showed a decrease, further confirming the diagnosis of moisture in the insulation. Subsequently, the staff activated the specialized drying equipment to perform targeted drying treatment on the switchgear's interior. During the process, they monitored changes in insulation parameters in real time, adjusting the drying temperature and time based on the decrease in leakage current and discharge frequency to ensure effective drying.

[0059] After a period of drying, maintenance personnel retested the equipment's insulation parameters and found that the leakage current had dropped to 4.3μA, essentially returning to the normal range; the partial discharge frequency had also decreased to 3 times / second, and the discharge activity was stabilizing. All insulation indicators met the requirements for normal equipment operation. Subsequently, the equipment was put back into operation, and continuous monitoring was conducted for 24 hours, confirming stable insulation performance without any abnormalities. This successfully prevented the insulation from becoming damp and further developing into surface flashover, ensuring the safe and stable operation of the equipment.

[0060] Table 2 Comparison of Case 2 Implementation Example 2: To achieve the above objective, such as Figure 2 As shown, based on Embodiment 1, this invention discloses a multi-parameter anomaly feature extraction and AI fault diagnosis system, including: Data acquisition module 11 is used to acquire multi-source sensor monitoring data, including partial discharge signal, overvoltage signal, leakage current, temperature and voltage and current. Data processing module 12 is used to extract multi-domain fusion anomaly features from multi-source sensor monitoring data to obtain multi-domain features, and to fuse and encode the multi-domain features to obtain a comprehensive anomaly feature set. The multi-domain features include time domain, frequency domain, time-frequency domain and operating condition domain. The diagnostic output module 13 is used to input the comprehensive abnormal feature set into the pre-established lightweight AI model, output the fault diagnosis result, and output the risk classification warning based on the fault diagnosis result.

[0061] Based on the same inventive concept, this invention also provides a computer device, comprising: one or more processors, and a memory for storing one or more computer programs; the programs include program instructions, and the processor executes the program instructions stored in the memory. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, used to implement one or more instructions, specifically for loading and executing one or more instructions stored in a computer storage medium to implement the above-described method.

[0062] It should be further explained that, based on the same inventive concept, the present invention also provides a computer storage medium storing a computer program, which, when executed by a processor, performs the above-described method. This storage medium can be any combination of one or more computer-readable media. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In the present invention, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0063] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this disclosure. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0064] The foregoing has shown and described the basic principles, main features, and advantages of this disclosure. Those skilled in the art should understand that this disclosure is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of this disclosure. Various changes and modifications can be made to this disclosure without departing from its spirit and scope, and all such changes and modifications fall within the scope of this disclosure as claimed.

Claims

1. A method for multi-parameter anomaly feature extraction and AI-based fault diagnosis, characterized in that: The method includes the following steps: Acquire multi-source sensor monitoring data, including partial discharge signal, overvoltage signal, leakage current, temperature, and voltage and current. Multi-domain fusion anomaly feature extraction is performed on multi-source sensor monitoring data to obtain multi-domain features. The multi-domain features are then fused and encoded to obtain a comprehensive anomaly feature set. The multi-domain features include time domain, frequency domain, time-frequency domain, and operating condition domain. The comprehensive set of abnormal features is input into a pre-established lightweight AI model, and the output is the fault diagnosis result. Based on the fault diagnosis result, risk classification and early warning are output.

2. The method for multi-parameter anomaly feature extraction and AI fault diagnosis according to claim 1, characterized in that, The extraction process of the multi-domain features includes: Temporal feature extraction: Extract time-series statistical features such as mean, variance, peak value, kurtosis, waveform distortion rate, rate of change, and steady-state offset to reflect the steady-state deviation and slow deterioration trend of equipment operation; Frequency domain feature extraction: Through FFT spectrum analysis, harmonic component decomposition, and high-frequency component amplitude ratio analysis, abnormal frequency components caused by partial discharge, high-frequency vibration, and poor circuit contact are captured. Time-frequency domain feature extraction: Wavelet transform and Hilbert transform are used to perform time-frequency joint analysis on non-stationary transient fault signals to locate the fault occurrence time and characteristic frequency distribution, and capture transient abrupt anomalies; Operating condition related domain characteristics: By associating external conditions such as unit load, ambient temperature and humidity, seasonal operating conditions, and operating time, the interference of operating condition fluctuations is removed, and the pure equipment degradation characteristics after eliminating the influence of operating conditions are extracted.

3. The method for multi-parameter anomaly feature extraction and AI fault diagnosis according to claim 1, characterized in that, The process of fusing and encoding multi-domain features includes: Multi-domain features are filtered, dimensionality reduced, and fused for encoding. Redundant and irrelevant features are eliminated to construct a comprehensive abnormal feature set with high discriminative power and low redundancy, laying the foundation for lightweight AI model input.

4. The method for multi-parameter anomaly feature extraction and AI fault diagnosis according to claim 1, characterized in that, The construction process of the pre-built lightweight AI model is as follows: Pruning, quantization, convolution kernel simplification, and structural distillation are performed on deep learning networks to compress model parameters and computational load, adapt to low-computing hardware deployment of local edge gateways and monitoring units, and balance diagnostic accuracy and inference real-time performance to obtain the processed deep learning network. The deep learning network is pre-trained based on an existing fault sample library of similar power equipment. It is then fine-tuned for transfer learning based on actual fault samples of equipment at field sites, and finally outputs a lightweight AI model.

5. The method for multi-parameter anomaly feature extraction and AI fault diagnosis according to claim 4, characterized in that, The expression for the lightweight AI model is as follows: Let the original deep learning network model be... Its network parameters are The lightweight AI model The construction process satisfies the following function mapping relationship: Step 1, Model Compression: in: This represents the quantization operation function, which quantizes the original model parameters. The quantization model is obtained by mapping high-precision floating-point numbers to low-bit integers. , This refers to the quantization bit width parameter; This represents the pruning and convolution kernel simplification operation functions, and the quantization model is evaluated. The importance weights of each channel or convolutional kernel are used to remove those below a threshold. The redundant structure is used to obtain a simplified model. ; Represent the structural distillation operation function to simplify the model. As a student network, with the original model As the teacher network, a compressed model is obtained by minimizing the distillation loss of the student and teacher networks on the soft-label output and jointly training them with the hard-label loss. , This is the hyperparameter for balancing distillation temperature and losses. The second step, pre-training and transfer fine-tuning: This represents the pre-training operation function to compress the model. As the initial model, it is based on an existing fault sample library of similar power equipment. Pre-training is performed by updating the model parameters by minimizing the fault classification loss function to obtain the pre-trained model. ; Represents the few-shot transfer learning operation function, used in the pre-trained model. Based on this, and targeting the actual fault sample set of on-site electrical equipment Perform migration fine-tuning, among which, By employing a few-shot learning strategy, some low-level network parameters are frozen, and parameters are updated only for the high-level classification layer, ultimately resulting in a lightweight AI model adapted to the target site. ; The lightweight AI model The output is a fault diagnosis result vector y=[p1, p2, ..., pn], where pi represents the probability of the i-th type of fault and n is the total number of fault categories, which satisfies the real-time inference requirements under the computing power constraints of the edge gateway.

6. The method for multi-parameter anomaly feature extraction and AI fault diagnosis according to claim 1, characterized in that, The risk classification and early warning output based on fault diagnosis results includes full-level control from early hidden danger warning, mid-term abnormal alarm to emergency fault trip warning.

7. The method for multi-parameter anomaly feature extraction and AI fault diagnosis according to claim 1, characterized in that, The process of acquiring the multi-source sensor monitoring data includes: For rapidly changing signals such as transient faults, partial discharges, and high-frequency abnormal noises, high-frequency synchronous sampling is used; for steady-state operating parameters, timed periodic sampling is used. All measurement points are packaged and sent up at the same time, frequency, and timestamp to eliminate subsequent analysis errors caused by timing misalignment.

8. The method for multi-parameter anomaly feature extraction and AI fault diagnosis according to claim 7, characterized in that, The preprocessing of the multi-source sensor monitoring data includes data denoising, outlier removal, and normalization.

9. A multi-parameter anomaly feature extraction and AI fault diagnosis system, employing the multi-parameter anomaly feature extraction and AI fault diagnosis method as described in any one of claims 1 to 8, characterized in that, include: The data acquisition module is used to acquire multi-source sensor monitoring data, which includes partial discharge signal, overvoltage signal, leakage current, temperature, and voltage and current. The data processing module is used to extract multi-domain fusion anomaly features from multi-source sensor monitoring data to obtain multi-domain features, and to fuse and encode the multi-domain features to obtain a comprehensive anomaly feature set. The multi-domain features include time domain, frequency domain, time-frequency domain and operating condition domain. The diagnostic output module is used to input a comprehensive set of abnormal features into a pre-established lightweight AI model and output the fault diagnosis results. Based on the fault diagnosis results, risk classification and early warning output are performed.

10. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, The memory stores a computer program that can run on the processor. When the processor loads and executes the computer program, it employs the method based on multi-parameter abnormal feature extraction and AI fault diagnosis as described in any one of claims 1 to 8.

Images