A method and system for identifying typical fault defects of a large oil filling device

Abstract

The present application relates to oil filling equipment technical field, provide a kind of large oil filling equipment typical fault defect identification method and system, it includes the following steps: step 1, acquisition large oil filling equipment in the ultrasonic signal of running process, and the ultrasonic signal is preprocessed, obtain the voiceprint data to be analyzed;Step 2, extract the multi-dimensional acoustic feature in the voiceprint data to be analyzed;Step 3, construct multi-dimensional feature fusion coding mechanism, the multi-dimensional acoustic feature is weighted and fused, obtain fusion feature vector;Step 4, the fusion feature vector is input into the deep learning model that is pre-constructed and trained, and the fault type and confidence corresponding to the current running state of equipment are output;Step 5, based on the fault type and confidence, calculate fault risk index, and combine preset multi-level early warning threshold, generate equipment state evaluation result and operation and maintenance suggestion.The present application can preferably identify fault defect.

Description

Technical Field

[0001] This invention relates to the field of oil filling equipment technology, and more specifically, to a method and system for identifying typical faults and defects in large oil filling equipment. Background Technology

[0002] Large oil-filled equipment, especially converter transformers and power transformers, are core components of ultra-high voltage direct current transmission systems and AC power grids. Their operational reliability directly affects the safety and stability of the entire power grid. Partial discharge is an early sign of insulation degradation in equipment. If it is not detected and addressed in time, it may develop into insulation breakdown, leading to serious accidents. Therefore, accurate and efficient partial discharge fault diagnosis of large oil-filled equipment is of great significance.

[0003] Typical faults and defects that may occur during the operation of large oil-filled equipment mainly include four types of partial discharge: corona discharge, which usually occurs in areas with extremely uneven electric fields (such as electrode tips and conductor surface burrs), manifesting as a hissing sound with broadband high-frequency noise in the spectrum; suspension discharge, generated when internal metal parts of the equipment are at a floating potential, producing intermittent "crackling" pulses with discrete energy peaks in the spectrum; surface discharge, occurring on the surface of the insulating medium, producing a continuous "humming" sound accompanied by intermittent discharge sounds, forming a continuous energy band in the spectrum; and internal discharge, occurring inside solid insulation or between oil-paper insulation layers, often accompanied by the continuous development of insulation degradation. These four types of defects have different acoustic fingerprint characteristics, and accurate identification of them is crucial for equipment condition assessment and operation and maintenance decisions.

[0004] Currently, partial discharge detection technologies mainly include pulsed current method, dissolved gas analysis method, and ultrasonic detection method. Among them, ultrasonic detection method is widely used in the field of online monitoring of power equipment due to its advantages such as non-contact, strong resistance to electromagnetic interference, and accurate positioning. However, existing fault diagnosis methods based on ultrasonic detection still have the following technical shortcomings: First, traditional ultrasound diagnostic methods mostly use amplitude threshold judgment or single feature analysis, which can only determine the presence or absence of partial discharge, making it difficult to accurately identify the type of discharge. They cannot effectively distinguish between different defects such as corona discharge, suspension discharge, surface discharge, and internal discharge, nor can they distinguish between normal operating noise of the equipment and partial discharge signals, resulting in a high false alarm rate and false negative rate.

[0005] Second, existing fault diagnosis methods based on voiceprint recognition typically extract only single-type features in the time domain, frequency domain, or time-frequency domain, lacking the fusion and utilization of multi-dimensional acoustic features, making it difficult to comprehensively characterize complex fault modes. Although some methods incorporate deep learning, they mostly employ a single network structure, failing to fully combine the complementary advantages of convolutional neural networks in local feature extraction and recurrent neural networks in temporal dependency modeling, thus limiting recognition accuracy.

[0006] Third, the on-site operating environment of power equipment is complex. UHV converter stations are subject to multiple sources of interference, including strong electromagnetic interference, mechanical vibration, cooling fan noise, and environmental noise. Existing noise reduction methods are insufficient to effectively separate weak partial discharge signals from strong background noise. Furthermore, ultrasonic waves undergo multiple reflections and reverberation when propagating in the complex transformer oil-solid medium, leading to signal distortion and further increasing the difficulty of feature extraction and fault identification.

[0007] Fourth, partial discharge fault samples are scarce, especially data on different types and severity levels, leading to imbalanced training data for deep learning models and insufficient generalization ability. Existing methods lack effective sample augmentation and model pre-training strategies, making it difficult to achieve high-precision identification under small sample conditions.

[0008] Fifth, existing technologies only output fault type judgment results, lacking a quantitative assessment of the reliability of the identification results, and failing to establish a complete closed loop from fault identification to operation and maintenance decisions. They cannot perform secondary verification based on confidence levels, nor can they comprehensively calculate risk indices based on fault type, probability, and confidence levels. This results in a simplistic early warning mechanism, making it difficult to achieve hierarchical management of equipment status and the automatic generation of precise operation and maintenance recommendations.

[0009] In summary, existing fault diagnosis technologies for large-scale oil-filled equipment have significant shortcomings in terms of accurate identification of typical faults (corona discharge, suspension discharge, surface discharge, and internal discharge), multi-dimensional feature fusion, adaptability to complex environments, small-sample learning capabilities, and intelligent operation and maintenance decision-making. Therefore, there is an urgent need to develop an intelligent fault identification method and system that can integrate multi-dimensional acoustic features, employ an advanced deep learning architecture, possess strong anti-interference capabilities, support small-sample learning, and automatically generate operation and maintenance suggestions to improve the accuracy and intelligence level of condition monitoring for large-scale oil-filled equipment. Summary of the Invention

[0010] The present invention provides a method and system for identifying typical faults and defects in large oil-filled equipment, which can overcome some or more defects of the prior art.

[0011] A method for identifying typical faults and defects in a large oil-filled equipment according to the present invention includes the following steps: Step 1: Collect ultrasonic signals during the operation of large oil filling equipment, and preprocess the ultrasonic signals to obtain acoustic signature data to be analyzed; Step 2: Extract multi-dimensional acoustic features from the voiceprint data to be analyzed. The multi-dimensional acoustic features include at least one of time-domain features, frequency-domain features, time-frequency-domain features, and deep learning features. Step 3: Construct a multi-dimensional feature fusion coding mechanism to weight and fuse the multi-dimensional acoustic features to obtain a fused feature vector; Step 4: Input the fused feature vector into the pre-built and trained deep learning model, and output the fault type and confidence level corresponding to the current operating state of the device; Step 5: Based on the fault type and confidence level, calculate the fault risk index, and combine it with the preset multi-level early warning thresholds to generate equipment status assessment results and operation and maintenance suggestions.

[0012] Preferably, the preprocessing includes dereverberation processing using a weighted prediction error algorithm; The weighted prediction error algorithm estimates the reverberation component using a linear prediction model and subtracts the reverberation component from the original signal. Its estimation model is as follows: ; in The signal after de-reverberation. The original signal, Predictive coefficient To predict delays, For the prediction order; For the current moment, This is the delay step number; The prediction coefficients are optimized by minimizing the power estimate of the output signal, which effectively suppresses the reflection and reverberation interference generated during the propagation of the ultrasonic signal inside the device.

[0013] As a preferred embodiment, when extracting deep learning features from the voiceprint data to be analyzed, a memory-enhanced autoencoder is used for feature extraction and anomaly detection. The memory-enhanced autoencoder includes an encoding network, a memory module, and a decoding network; The extraction process includes: converting voiceprint data into Mel frequency spectrograms; extracting original encoded features through an encoding network; calculating the similarity between the original encoded features and the normal state feature codes stored in the memory module to obtain reconstructed encoded features; decoding the reconstructed encoded features through a decoding network to obtain a reconstructed spectrogram; calculating the generation error between the reconstructed spectrogram and the original spectrogram, and using the generation error as an anomaly score. The generation error The calculation formula is: ; in To reconstruct the spectrogram, This is the original input spectrogram. The total number of samples; When the generation error exceeds a preset threshold, the device is determined to be in an abnormal state.

[0014] Preferably, the multidimensional feature fusion coding mechanism includes: Normalize various features: ; in, For normalized features, The characteristic mean, The characteristic standard deviation; This is the original feature vector; Constructing a multidimensional feature fusion vector : ; in, These are the weighting coefficients. Each weight coefficient is calculated using feature mutual information or feature importance. , , , These are time-domain feature vectors, frequency-domain feature vectors, time-frequency-domain feature vectors, and deep learning feature vectors, respectively.

[0015] Preferably, the deep learning model adopts a dual-branch structure, including a convolutional neural network branch for extracting local spatial features and a recurrent neural network branch for capturing temporal dependencies, and uses an attention mechanism to weight and enhance the features.

[0016] Preferably, the deep learning model employs a dual-path feature decoupling and recombination mechanism, including: The extracted multi-dimensional acoustic features are decoupled into two orthogonal subspaces: inherent state features of the equipment and transient fault features. The decoupling constraint is as follows: ; in, The original feature vector, These are inherent state characteristics of the equipment. It is a transient fault characteristic. As a measure of correlation between features, The regularization coefficient is used. The transient fault characteristics are enhanced, while the inherent state characteristics of the equipment are suppressed, thereby achieving accurate extraction of partial discharge pulse signals.

[0017] Preferably, the fault identification confidence level Defined as: ; in The maximum class probability is output in the probability vector of the model. Let i be the probability value of the i-th type of fault. Total number of fault categories; It also includes a secondary verification based on feature similarity of the confidence level of the fault type output by the deep learning model; The secondary verification includes: calculating the similarity between the feature vector of the voiceprint data to be identified and the feature vectors of various fault templates in the voiceprint database to obtain a similarity score; and constructing a fusion confidence score by combining the softmax probability output by the model with the similarity score. : ; in The softmax probability output by the model. To score feature similarity, For fusion weighting coefficients; When the fusion confidence exceeds a preset threshold, the final recognition result is output.

[0018] Preferably, the fault risk index The calculation method is as follows: ; And introduce fault weight coefficient Calculate the final risk index : ; in Risk weights corresponding to different fault types are used to characterize the degree of impact of different fault types on equipment operation risk; The multi-level early warning thresholds include: The equipment operating status is divided into normal status, attention status, warning status, and severe warning status; the corresponding threshold relationships are as follows: Normal state: ; Attention status: ; Warning status: ; Severe alert status: ; in The preset threshold is determined based on historical operational data statistics, expert experience rules, or equipment failure sample analysis.

[0019] Preferably, the deep learning model is pre-trained using a contrastive learning strategy to enhance its ability to distinguish features of different types of partial discharge. Objective function of contrastive learning for: ; in The feature representation of positive sample pairs. The cosine similarity function is used. For temperature coefficient, For indicator functions; Through the comparative learning, voiceprint features of similar faults are brought closer to each other in the feature space, while voiceprint features of dissimilar faults are moved away from each other.

[0020] This invention provides a typical fault and defect identification system for large oil-filled equipment, which adopts the aforementioned method for identifying typical faults and defects in large oil-filled equipment and includes: The acoustic signature acquisition module is used to acquire ultrasonic signals during the operation of large oil-filled equipment. The signal preprocessing module is used to perform noise reduction, filtering, and feature extraction on the ultrasound signal; The feature fusion encoding module is used to execute a multi-dimensional feature fusion encoding mechanism to generate a fused feature vector. The deep learning diagnostic module is equipped with a pre-trained deep learning model, which is used to output the fault type and confidence level. The early warning and decision-making module is used to generate equipment status assessment results and operation and maintenance suggestions based on the model output results and preset multi-level early warning thresholds. The voiceprint database module is used to store voiceprint feature data under normal device operation and various fault conditions, and supports model training and updates.

[0021] The beneficial effects of this invention are as follows: This invention introduces a weighted prediction error algorithm for dereverberation processing during the preprocessing stage. By using a linear prediction model to estimate and eliminate multiple reflections and reverberation components generated when ultrasound propagates in complex oil-solid media, it effectively suppresses signal distortion and significantly improves the signal-to-noise ratio. This algorithm solves the problems of severe signal attenuation and waveform distortion when ultrasound propagates through complex structures inside transformers in existing technologies, providing high-quality input data for subsequent feature extraction and fault identification.

[0022] This invention extracts multi-dimensional acoustic features encompassing time-domain, frequency-domain, time-frequency-domain, and deep learning features, and innovatively constructs a multi-dimensional feature fusion coding mechanism. Through normalization and weighted fusion, the four types of features are organically combined, with weight coefficients dynamically determined through feature mutual information or feature importance calculations, achieving information synergy among features. Compared to existing methods using only a single type of feature, this invention can more comprehensively and accurately characterize complex fault modes, significantly improving the accuracy of fault identification.

[0023] This invention employs a dual-branch deep learning model structure. A convolutional neural network branch extracts local spatial features of the voiceprint signal, while a bidirectional long short-term memory network branch captures the temporal dependencies of the signal. An attention mechanism is used to assign weights to different feature dimensions, reinforcing key information related to fault characteristics. This architecture combines the complementary advantages of CNNs in image / spectral feature extraction and BiLSTMs in temporal modeling. Compared to a single network structure, it can more effectively identify complex patterns in partial discharge voiceprints, significantly improving recognition accuracy.

[0024] This invention introduces a dual-path feature decoupling and recombination mechanism, decoupling multi-dimensional acoustic features into two orthogonal subspaces: inherent equipment state features and transient fault features. Regularization constraints enhance the orthogonality of these two subspaces. This mechanism effectively distinguishes between steady-state noise during normal equipment operation (such as magnetostrictive vibration of the iron core and cooling fan noise) and transient pulse signals generated by partial discharge, achieving accurate extraction of fault features and solving the technical challenge of separating weak partial discharge signals from strong background noise in existing technologies.

[0025] This invention employs a contrastive learning strategy to pre-train a deep learning model. By constructing positive and negative sample pairs, it ensures that voiceprint features of similar faults are close to each other in the feature space, while voiceprint features of dissimilar faults are far apart, forming a highly discriminative feature space. This strategy can effectively learn feature representations under small sample conditions, significantly improving the model's ability to identify rare fault types and overcoming the problem of insufficient model generalization ability caused by scarce fault samples and imbalanced data in practical engineering.

[0026] This invention defines a formula for calculating the confidence level of fault identification based on probability distribution, and further introduces a secondary verification mechanism based on feature similarity to construct a fused confidence level. This mechanism combines the softmax probability output by the model with the similarity score of the feature vector to perform dual verification of the identification results, effectively solving the problem of misjudgment that is easily generated by a single probability threshold. When the confidence level is lower than a preset threshold, the system can automatically trigger manual review to ensure the reliability and accuracy of the diagnostic results.

[0027] This invention constructs a fault risk index calculation model, comprehensively considering fault probability, identification confidence, and fault type weights to obtain a comprehensive risk index. Based on this index, a multi-level early warning threshold system is established, dividing equipment operating status into four levels: normal, attention, early warning, and severe early warning. Differentiated operation and maintenance suggestions (such as increasing monitoring frequency, arranging special inspections, and immediate power outage for maintenance) are automatically generated according to different levels. This mechanism achieves a complete closed loop from fault identification to operation and maintenance decision-making, solving the problem of disconnect between diagnostic results and operation and maintenance strategies in existing technologies, and significantly improving the intelligence level of equipment status monitoring and operation and maintenance efficiency. Attached Figure Description

[0028] Figure 1 This is a flowchart of a typical fault and defect identification method for a large oil-filled equipment in an embodiment. Figure 2 This is a schematic diagram of a typical fault and defect identification system for a large oil-filled equipment in an embodiment. Detailed Implementation

[0029] To further understand the content of this invention, a detailed description of the invention will be provided in conjunction with the accompanying drawings and embodiments. It should be understood that the embodiments are merely illustrative and not limiting of the invention. Example

[0030] like Figure 1 As shown in the figure, this embodiment provides a method for identifying typical faults and defects in large oil-filled equipment, which includes the following steps: Step 1: Collect ultrasonic signals during the operation of large oil-filled equipment, and preprocess the ultrasonic signals to obtain acoustic data to be analyzed.

[0031] A multi-sensor array, including four surface-mounted piezoelectric ultrasonic sensors and two non-contact microphones, was arranged on the outer wall and surrounding space of the converter transformer tank. The surface-mounted sensors were installed on different sides at 2 / 5 to 3 / 5 of the tank height, while the non-contact microphones were fixed to the transformer body at a distance of 1.5 meters using brackets. All sensors adopted a synchronous acquisition method, with a sampling frequency set to 200kHz to meet the Nyquist sampling requirements for partial discharge ultrasonic signals (30kHz~160kHz).

[0032] The acquired raw ultrasound signals are first preprocessed. This preprocessing includes dreverberation processing using a weighted prediction error algorithm. This algorithm estimates the reverberation component using a linear prediction model and subtracts the reverberation component from the original signal. The estimation model is as follows: ; in The signal after de-reverberation. The original signal, Predictive coefficient For complex conjugate, To predict delays, For the prediction order, this embodiment , ; For the current moment, The delay step is used to optimize the prediction coefficients by minimizing the power estimate of the output signal, thereby effectively suppressing the reflection and reverberation interference generated during the propagation of the ultrasonic signal inside the device.

[0033] Preprocessing also includes wavelet thresholding and bandpass filtering (passband 30kHz~160kHz), as well as signal framing (frame length 1024 points, frame shift 512 points) and normalization to obtain the voiceprint data to be analyzed.

[0034] Step 2: Extract multi-dimensional acoustic features from the voiceprint data to be analyzed. These multi-dimensional acoustic features include at least one of time-domain features, frequency-domain features, time-frequency-domain features, and deep learning features. Time-domain features include peak value, root mean square (RMS), kurtosis, and impulse repetition rate; frequency-domain features include dominant frequency, bandwidth, spectral centroid, and spectral entropy; time-frequency-domain features include wavelet packet energy, Hilbert spectral features, and Mel frequency cepstral coefficients.

[0035] When extracting deep learning features from the voiceprint data to be analyzed, a memory-enhanced autoencoder is used for feature extraction and anomaly detection. The memory-enhanced autoencoder includes an encoding network, a memory module, and a decoding network; The extraction process includes: converting voiceprint data into Mel frequency spectrograms; extracting original encoded features through an encoding network; calculating the similarity between the original encoded features and the normal state feature codes stored in the memory module to obtain reconstructed encoded features; decoding the reconstructed encoded features through a decoding network to obtain a reconstructed spectrogram; calculating the generation error between the reconstructed spectrogram and the original spectrogram, and using the generation error as an anomaly score.

[0036] Specifically, the voiceprint data is converted into a Mel frequency spectrogram, and the original encoded features are extracted through an encoding network (3 convolutional layers, each with a kernel size of 3×3 and 32, 64, and 128 channels respectively). The cosine similarity between the original encoded features and the normal-state feature codes stored in the memory module is calculated, and the reconstructed encoded features are obtained through soft addressing. Then, the spectrogram is reconstructed by a decoding network (3 deconvolutional layers).

[0037] The generation error The calculation formula is: ; in To reconstruct the spectrogram, This is the original input spectrogram. The total number of samples; When the generation error exceeds a preset threshold (e.g., 0.15, determined based on normal sample statistical distribution, ROC optimization, safety-first false alarm rate constraints, expert experience, and on-site iterative feedback), the equipment is determined to be in an abnormal state.

[0038] Step 3: Construct a multi-dimensional feature fusion coding mechanism to weight and fuse the multi-dimensional acoustic features to obtain a fused feature vector.

[0039] The multidimensional feature fusion coding mechanism includes: Normalize various features: ; in, For normalized features, The characteristic mean, The characteristic standard deviation, This is the original feature vector.

[0040] Constructing a multidimensional feature fusion vector : ; in, In this embodiment, the weighting coefficient is used. , Each weight coefficient is calculated using feature mutual information or feature importance. , , , These are time-domain feature vectors, frequency-domain feature vectors, time-frequency-domain feature vectors, and deep learning feature vectors, respectively.

[0041] Step 4: Input the fused feature vector into the pre-built and trained deep learning model, and output the fault type and confidence level corresponding to the current operating state of the device.

[0042] The deep learning model adopts a dual-branch structure, including a convolutional neural network branch for extracting local spatial features and a recurrent neural network branch for capturing temporal dependencies, and uses an attention mechanism to weight and enhance the features.

[0043] The deep learning model includes: (1) The convolutional feature extraction module uses three convolutional layers, each with a kernel size of 3×3 and channels of 32, 64, and 128 respectively. The activation function is ReLU, and each layer is followed by a max pooling layer (pooling window 2×2). The calculation process is as follows: ; in, For the first i Feature maps output by each convolutional layer For convolution kernel parameters, For the input feature matrix, For bias terms; For activation functions; (2) Temporal feature learning module, which uses a bidirectional long short-term memory network for temporal modeling:

[0044] in, The timing features are those obtained after concatenation using bidirectional LSTM. In the forward-hidden state, This is a backward hidden state; the hidden layer dimension is 128, and the bidirectional state concatenation is 256 dimensions.

[0045] (3) Attention enhancement module: A multi-head attention mechanism is introduced after the feature fusion layer, with the number of heads set to 8. The calculation formula is as follows: ; in: Q For querying the matrix, K The key matrix, V For value matrices, As a dimension, Indicates transpose. It is a normalized exponential function.

[0046] The attention mechanism assigns weights to different feature dimensions, thereby strengthening key information related to fault characteristics.

[0047] (4) Model training: cross-entropy loss function, Adam optimizer, initial learning rate 0.001, batch size 32, training epochs 50. Before training, a contrastive learning strategy was used for pre-training to enhance the ability to distinguish features of different partial discharge types.

[0048] Objective function of contrastive learning for: ; in The feature representation of positive sample pairs. The cosine similarity function is used. The temperature coefficient is set to 0.07. The function is an indicator function; through the comparative learning, the voiceprint features of similar faults are brought closer to each other in the feature space, while the voiceprint features of dissimilar faults are moved away from each other.

[0049] The fused feature vectors are input into the trained deep learning model, and the model outputs probability vectors for various types of faults. , where k=5, corresponding to corona discharge, levitation discharge, surface discharge, internal discharge and normal operation, respectively.

[0050] The confidence level of fault identification Defined as: ; in The maximum class probability is output in the probability vector of the model. Let i be the probability value of the i-th type of fault. This represents the total number of fault categories; in the example, when the model output probability vector is [0.02, 0.85, 0.05, 0.03, 0.05], The confidence level C = 0.85 / 1.0 = 0.85.

[0051] To further improve reliability, a secondary verification based on feature similarity is performed.

[0052] The secondary verification includes: calculating the cosine similarity between the feature vector of the voiceprint data to be identified and the feature vectors of various fault templates in the voiceprint database, and obtaining a similarity score. Combined with the softmax probability output by the model Construct a fusion confidence score using the similarity score. : ; in The softmax probability output by the model. To score feature similarity, The fusion weighting coefficient is 0.6; calculated as follows: =0.6×0.85+0.4×0.82=0.838, which exceeds the preset threshold of 0.70 (determined based on validation set performance, safety priority principle, confidence statistical distribution and on-site iterative feedback). The output identification result is "suspended discharge".

[0053] The deep learning model employs a dual-path feature decoupling and recombination mechanism, including: The extracted multi-dimensional acoustic features are decoupled into two orthogonal subspaces: inherent state features of the equipment and transient fault features. The decoupling constraint is as follows: ; in, The original feature vector, These are inherent state characteristics of the equipment. It is a transient fault characteristic. As a measure of correlation between features, The regularization coefficient is used to enhance the transient fault characteristics and suppress the inherent state characteristics of the equipment, thereby achieving accurate extraction of the partial discharge pulse signal.

[0054] Step 5: Based on the fault type and confidence level, calculate the fault risk index, and combine it with the preset multi-level early warning thresholds to generate equipment status assessment results and operation and maintenance suggestions.

[0055] The failure risk index The calculation method is as follows: ;

[0056] It also introduces a fault weighting coefficient and a risk weighting for levitation discharge. Calculate the final risk index : ; The multi-level early warning thresholds include: The equipment operating status is divided into normal status, attention status, warning status, and severe warning status; the corresponding threshold relationships are as follows: Normal state: ; Attention status: ; Warning status: ; Severe alert status: ; in The preset threshold is determined based on historical operational data statistics, expert experience rules, or equipment failure sample analysis. It is based on preset multi-level early warning thresholds (…). The current state is determined to be a warning state. ).

[0057] Based on the warning status and fault type "floating discharge", the system automatically generates maintenance suggestions: "The equipment shows obvious abnormality trend. It is recommended to arrange on-site special inspection, further locate the partial discharge area, focus on checking the fixation of metal parts, and formulate a maintenance plan." This embodiment provides a typical fault and defect identification system for large oil-filled equipment, which adopts the above-mentioned method for identifying typical faults and defects in large oil-filled equipment and includes: The acoustic signature acquisition module is used to acquire ultrasonic signals during the operation of large oil-filled equipment. It includes 8 surface-mount piezoelectric ultrasonic sensors and 4 non-contact microphones, and adopts a multi-sensor array arrangement. The sensors are fixed to the outer wall of the transformer tank by magnetic attraction. The sampling frequency is 200kHz, the quantization bit is 24 bits, and it supports multi-channel synchronous acquisition and sound source localization.

[0058] The signal preprocessing module is used to perform noise reduction, filtering, and feature extraction on the ultrasonic signal. Deployed on the edge computing device, it uses a weighted prediction error algorithm for dereverberation processing, combined with wavelet threshold noise reduction and bandpass filtering, to output high-quality voiceprint data.

[0059] The feature fusion encoding module is used to execute a multi-dimensional feature fusion encoding mechanism to generate a fused feature vector; it extracts time-domain, frequency-domain, time-frequency-domain, and deep learning features, and generates a fused feature vector through the multi-dimensional feature fusion encoding mechanism, with weight coefficients dynamically calculated through feature mutual information.

[0060] The deep learning diagnostic module, deployed on a cloud server, employs a dual-branch CNN-BiLSTM-attention model. Model parameters are obtained through contrastive learning pre-training and transfer learning fine-tuning, supporting online learning and incremental update mechanisms. The model outputs the fault type and confidence level, and its credibility is evaluated through a secondary verification mechanism.

[0061] The early warning and decision-making module generates equipment status assessment results and maintenance recommendations based on fault type, confidence level, and risk index, combined with multi-level early warning thresholds. Early warning information is pushed to maintenance personnel via SMS, email, and mobile app.

[0062] The voiceprint database module adopts a distributed storage architecture to store voiceprint feature data under normal operation and various fault conditions of the equipment. Currently, it has collected voiceprint data from 40 transformers / converter transformers, covering 25 models, with operating years ranging from 9 to 33 years, and includes more than 6,000 typical fault samples such as corona discharge, floating discharge, surface discharge, and internal discharge.

[0063] This embodiment significantly improves the accuracy, robustness, and intelligence of partial discharge fault identification in large oil-filled equipment through the synergy of signal processing, feature fusion, deep learning model architecture, feature decoupling, few-shot learning, confidence assessment, risk warning, and system integration. It provides reliable technical support for condition monitoring and predictive maintenance of power equipment, and has important engineering application value and significant economic and social benefits.

[0064] The identification results for different fault types are as follows: This embodiment demonstrates the effectiveness of the method in identifying four typical fault defects. The identification results on an independent test set containing 600 test samples are shown in Table 1:

[0065] The overall accuracy reached 98.2%, which is significantly better than the traditional amplitude threshold method (about 65%) and the single feature deep learning method (about 85%).

[0066] To verify the adaptability of this embodiment in complex environments, tests were conducted under simulated mixed interference conditions including strong electromagnetic interference, mechanical vibration, and wind and rain noise. Experimental results show that after weighted prediction error déreverberation and dual-path feature decoupling processing, the signal-to-noise ratio (SNR) is improved by 15.3 dB compared to the original signal. Under strong interference conditions with an SNR of -10 dB, the fault identification accuracy remains above 92%, which is better than the 54% of traditional methods and the 76% of conventional deep noise reduction methods.

[0067] The present invention and its embodiments have been described above illustratively. This description is not restrictive, and the figures shown are only one embodiment of the present invention; the actual structure is not limited thereto. Therefore, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the present invention, such designs should fall within the protection scope of the present invention.

Claims

1. A method for identifying typical faults and defects in large oil-filled equipment, characterized in that, Includes the following steps: Step 1: Collect ultrasonic signals during the operation of large oil filling equipment, and preprocess the ultrasonic signals to obtain acoustic signature data to be analyzed; Step 2: Extract multi-dimensional acoustic features from the voiceprint data to be analyzed. The multi-dimensional acoustic features include at least one of time-domain features, frequency-domain features, time-frequency-domain features, and deep learning features. Step 3: Construct a multi-dimensional feature fusion coding mechanism to weight and fuse the multi-dimensional acoustic features to obtain a fused feature vector; Step 4: Input the fused feature vector into the pre-built and trained deep learning model, and output the fault type and confidence level corresponding to the current operating state of the device; Step 5: Based on the fault type and confidence level, calculate the fault risk index, and combine it with the preset multi-level early warning thresholds to generate equipment status assessment results and operation and maintenance suggestions.

2. The method for identifying typical faults and defects in a large oil-filled equipment according to claim 1, characterized in that, The preprocessing includes de-reverberation processing using a weighted prediction error algorithm; The weighted prediction error algorithm estimates the reverberation component using a linear prediction model and subtracts the reverberation component from the original signal. Its estimation model is as follows: ； in The signal after de-reverberation. The original signal, Predictive coefficient For complex conjugate, To predict delays, For the prediction order; For the current moment, This is the delay step number; The prediction coefficients are optimized by minimizing the power estimate of the output signal, which effectively suppresses the reflection and reverberation interference generated during the propagation of the ultrasonic signal inside the device.

3. The method for identifying typical faults and defects in large oil-filled equipment according to claim 2, characterized in that, When extracting deep learning features from the voiceprint data to be analyzed, a memory-enhanced autoencoder is used for feature extraction and anomaly detection. The memory-enhanced autoencoder includes an encoding network, a memory module, and a decoding network; The extraction process includes: converting the voiceprint data into a Mel frequency spectrogram, and extracting the original encoded features through a coding network; Calculate the similarity between the original encoded features and the normal-state feature codes stored in the memory module to obtain reconstructed encoded features; decode the reconstructed encoded features through a decoding network to obtain a reconstructed spectrogram; calculate the generation error between the reconstructed spectrogram and the original spectrogram, and use the generation error as an anomaly score; The generation error The calculation formula is: ； in To reconstruct the spectrogram, The original input spectrogram, The total number of samples; When the generation error exceeds a preset threshold, the device is determined to be in an abnormal state.

4. The method for identifying typical faults and defects in a large oil-filled equipment according to claim 3, characterized in that, The multidimensional feature fusion coding mechanism includes: Normalize various features: ； in, For normalized features, The characteristic mean, The characteristic standard deviation; This is the original feature vector; Constructing a multidimensional feature fusion vector : ； in, These are the weighting coefficients. Each weight coefficient is calculated using feature mutual information or feature importance. , , , These are time-domain feature vectors, frequency-domain feature vectors, time-frequency-domain feature vectors, and deep learning feature vectors, respectively.

5. The method for identifying typical faults and defects in a large oil-filled equipment according to claim 4, characterized in that, The deep learning model adopts a dual-branch structure, including a convolutional neural network branch for extracting local spatial features and a recurrent neural network branch for capturing temporal dependencies, and uses an attention mechanism to weight and enhance the features.

6. The method for identifying typical faults and defects in a large oil-filled equipment according to claim 5, characterized in that, The deep learning model employs a dual-path feature decoupling and recombination mechanism, including: The extracted multi-dimensional acoustic features are decoupled into two orthogonal subspaces: inherent state features of the equipment and transient fault features. The decoupling constraint is as follows: ； in, The original feature vector, These are inherent state characteristics of the equipment. It is a transient fault characteristic. As a measure of correlation between features, The regularization coefficient is used. The transient fault characteristics are enhanced, while the inherent state characteristics of the equipment are suppressed, thereby achieving accurate extraction of partial discharge pulse signals.

7. The method for identifying typical faults and defects in a large oil-filled equipment according to claim 6, characterized in that, The confidence level of fault identification Defined as: ； in The maximum class probability is output in the probability vector of the model. Let i be the probability value of the i-th type of fault. Total number of fault categories; It also includes a secondary verification based on feature similarity of the confidence level of the fault type output by the deep learning model; The secondary verification includes: calculating the similarity between the feature vector of the voiceprint data to be identified and the feature vectors of various fault templates in the voiceprint database to obtain a similarity score; and constructing a fusion confidence score by combining the softmax probability output by the model with the similarity score. : ； in The softmax probability output by the model. To score feature similarity, For fusion weighting coefficients; When the fusion confidence exceeds a preset threshold, the final recognition result is output.

8. The method for identifying typical faults and defects in a large oil-filled equipment according to claim 7, characterized in that, The failure risk index The calculation method is as follows: ； And introduce fault weight coefficient Calculate the final risk index : ； in Risk weights corresponding to different fault types are used to characterize the degree of impact of different fault types on equipment operation risk; The multi-level early warning thresholds include: The equipment operating status is divided into normal status, attention status, warning status, and severe warning status; the corresponding threshold relationships are as follows: Normal state: ; Attention status: ; Warning status: ; Severe alert status: ; in The preset threshold is determined based on historical operational data statistics, expert experience rules, or equipment failure sample analysis.

9. The method for identifying typical faults and defects in a large oil-filled equipment according to claim 8, characterized in that, The deep learning model is pre-trained using a contrastive learning strategy to enhance its ability to distinguish features of different types of partial discharge. Objective function of contrastive learning for: ； in The feature representation of positive sample pairs. The cosine similarity function is used. For temperature coefficient, For indicator functions; Through the comparative learning, voiceprint features of similar faults are brought closer to each other in the feature space, while voiceprint features of dissimilar faults are moved away from each other.

10. A typical fault and defect identification system for large-scale oil-filled equipment, characterized in that, It employs a method for identifying typical faults and defects in large oil-filled equipment as described in any one of claims 1-9, and includes: The acoustic signature acquisition module is used to acquire ultrasonic signals during the operation of large oil-filled equipment. The signal preprocessing module is used to perform noise reduction, filtering, and feature extraction on the ultrasound signal; The feature fusion encoding module is used to execute a multi-dimensional feature fusion encoding mechanism to generate a fused feature vector. The deep learning diagnostic module is equipped with a pre-trained deep learning model, which is used to output the fault type and confidence level. The early warning and decision-making module is used to generate equipment status assessment results and operation and maintenance suggestions based on the model output results and preset multi-level early warning thresholds. The voiceprint database module is used to store voiceprint feature data under normal device operation and various fault conditions, and supports model training and updates.

Images