A method and system for early warning of faults in large oil-filled equipment

Abstract

This invention relates to the field of fault early warning technology, and provides a method and system for early warning of faults in large oil-filled equipment, comprising: Step 1, collecting acoustic fingerprint data throughout the entire lifecycle of the large oil-filled equipment and constructing an acoustic fingerprint sample set; Step 2, preprocessing the acoustic fingerprint sample set, extracting acoustic fingerprint feature parameters, and establishing a standardized acoustic fingerprint database; Step 3, constructing a multi-channel deep learning acoustic fingerprint recognition model and a discharge severity assessment model based on deep learning; Step 4, based on the output of the trained acoustic fingerprint recognition model and combined with a preset multi-level early warning threshold system, performing real-time assessment of the operating status of the large oil-filled equipment, and generating corresponding early warning information when the assessment result meets the early warning triggering conditions; Step 5, pushing the early warning information to the operation and maintenance terminal, and automatically generating operation and maintenance suggestions based on the defect type and severity. This invention can effectively provide early warning of faults in large oil-filled equipment.

Description

Technical Field

[0001] This invention relates to the field of fault early warning technology, and more specifically, to a method and system for early warning of faults in large oil-filled equipment. Background Technology

[0002] Ultra-high voltage direct current (UHVDC) transmission systems are critical infrastructure for optimizing energy resource allocation and supporting large-scale renewable energy integration. Converter transformers, as core equipment in UHVDC transmission systems, directly impact the safety and stability of the entire power grid. However, due to prolonged exposure to high voltage, high current, complex electromagnetic environments, and fluctuating loads, the internal insulation of converter transformers is prone to aging and deterioration, leading to early insulation defects such as partial discharge (PD). Partial discharge is a significant indicator of insulation degradation in power equipment; failure to detect and address it promptly can result in insulation breakdown, equipment damage, and even power grid accidents, causing severe economic losses and safety risks.

[0003] Currently, the main methods for detecting partial discharge include electrical measurement, chemical analysis, and ultrasonic testing. While electrical measurement methods (such as pulse current methods) offer high sensitivity, they are susceptible to electromagnetic interference and struggle to pinpoint the exact location of the fault. Chemical analysis methods (such as dissolved gas analysis in oil, DGA) can reflect the degradation trend of insulating oil, but they suffer from long detection cycles, poor real-time performance, and insufficient ability to specifically identify fault types. Ultrasonic testing, due to its non-contact nature, resistance to electromagnetic interference, and high positioning accuracy, is widely used in partial discharge detection. However, traditional ultrasonic diagnostic methods often rely on amplitude threshold judgment or manual experience, making it difficult to effectively distinguish between normal equipment noise and partial discharge signals, and even more difficult to accurately identify discharge types (such as corona discharge, surface discharge, suspended discharge, and internal discharge).

[0004] In recent years, deep learning technology has made breakthroughs in fields such as image recognition and speech recognition, providing new technical pathways for fault diagnosis of power equipment. Combining deep learning with ultrasonic partial discharge acoustic signature recognition can automatically extract discriminative features from complex acoustic signals, enabling accurate identification of partial discharge types and early warning of faults. However, existing technologies still have the following shortcomings: Data acquisition difficulties: In actual operation, partial discharge samples are scarce, especially labeled voiceprint data covering multiple working conditions and multiple defect types, which leads to insufficient model training and limited generalization ability; Severe environmental interference: The UHV substation site is subject to multiple sources of background noise, including strong electromagnetic interference, mechanical vibration, and cooling system noise, which affect the signal-to-noise ratio and recognition accuracy of ultrasonic signals. Insufficient feature extraction: Traditional voiceprint features rely heavily on manual design, making it difficult to fully characterize the nonlinear and non-stationary characteristics of partial discharge; Disconnect between diagnosis and early warning: Existing systems mostly remain at the level of fault identification, lacking quantitative assessment of the severity of discharge, prediction of status trends, and intelligent operation and maintenance decision support.

[0005] Therefore, there is an urgent need for a comprehensive method that can automatically identify the acoustic signature of partial discharge in large oil-filled equipment, assess its severity, provide early warnings, and offer intelligent operation and maintenance suggestions, in order to improve the reliability of equipment operation and the level of intelligent operation and maintenance. Summary of the Invention

[0006] The present invention provides a method and system for early warning of faults in large oil-filled equipment, which can overcome some or more defects of the prior art.

[0007] A method for early warning of faults in large oil-filled equipment according to the present invention includes the following steps: Step 1: Collect full lifecycle acoustic fingerprint data of large oil-filled equipment under various operating conditions, and construct an acoustic fingerprint sample set including normal and abnormal states; wherein, the abnormal states include at least partial discharge types such as corona discharge, suspension discharge, surface discharge and internal discharge. Step 2: Preprocess the voiceprint sample set, extract voiceprint feature parameters, and establish a standardized acoustic fingerprint database; the voiceprint feature parameters include time-domain features, frequency-domain features, time-frequency-domain features, and deep learning features; The establishment of an acoustic fingerprint database includes the following steps: frame processing of acoustic fingerprint signals, multi-domain acoustic fingerprint feature extraction, adaptive fusion of acoustic fingerprint features, acoustic fingerprint encoding, and construction of a standardized acoustic fingerprint database. Step 3: Construct a multi-channel deep learning voiceprint recognition model and a discharge severity assessment model based on deep learning. Use samples from the acoustic fingerprint database to train and optimize the model to obtain a trained voiceprint recognition model. The voiceprint recognition model is used to identify partial discharge types and assess the severity of discharge. Step 4: Based on the output of the trained voiceprint recognition model and combined with the preset multi-level early warning threshold system, the operating status of the large oil filling equipment is evaluated in real time. When the evaluation result meets the early warning triggering conditions, the corresponding early warning information is generated. Step 5: Push the warning information to the operation and maintenance terminal, and automatically generate operation and maintenance suggestions including inspection suggestions, maintenance priorities and risk levels based on the defect type and severity.

[0008] Preferably, step 1, which involves collecting full-lifecycle voiceprint data, specifically includes: Under uninterrupted power supply, a high-sensitivity ultrasonic sensor array deployed in key parts of the equipment is used to continuously collect acoustic data under normal operation, load change, DC bias, and partial discharge conditions. By constructing a physical simulation model of transformer partial discharge, supplementary voiceprint data with different discharge types and locations are generated to expand the voiceprint sample set.

[0009] Preferably, step 2 involves the following specific steps: 2.1) Frame-by-frame processing of voiceprint signals; Short-time analysis method is used to analyze the preprocessed voiceprint signal Perform frame segmentation processing. t Representing time; assuming a frame length of 20ms to 40ms, a frame shift of 10ms to 20ms, and using a Hamming window function to window each frame signal, the result is obtained. n The frame signal is represented as: in, The amplitude of the kth sampling point of the original voiceprint signal. For window functions, R The frame shift length, Indicates the first n The first frame k Each sampling point signal value; through the above processing, a series of short-time frame sequences are obtained, providing basic data for subsequent voiceprint feature extraction; 2.2) Temporal feature extraction; For each frame of the audioprint signal, time-domain statistical features reflecting signal energy changes and impact characteristics are extracted, specifically including: 2.21) Root mean square (RMS) value, used to reflect the energy change of the acoustic signal; 2.22) Kurtosis K, used to characterize the impulse characteristics of a signal; 2.23) Peak factor CF, used to characterize the transient impulse component in a signal; 2.24) Short-time energy It is used to reflect the change of acoustic signal energy over time; The time-domain feature vector is constructed using the above parameters. : 2.3) Frequency domain feature extraction; Perform a Fast Fourier Transform (FFT) on each frame of the signal to convert the time-domain signal to the frequency domain, thus obtaining the spectrum. ; The following frequency domain feature parameters are extracted based on the spectral information: 2.31) Spectral centroid , used to indicate the concentrated location of spectral energy; 2.32) Spectral Entropy This is used to measure the complexity of the spectral distribution; 2.33) Spectral flatness This is used to reflect the uniformity of the spectrum; 2.34) Spectral roll-off frequency , is used to indicate the frequency position corresponding to when the accumulated energy reaches a certain proportion; The frequency domain feature vector is formed using the above parameters. : 2.4) Time-frequency domain feature extraction; Feature extraction is performed using a multi-resolution time-frequency analysis method. First, continuous wavelet transform (CWT) is used to perform time-frequency analysis on the acoustic signal. The energy distribution of the acoustic signal at different times and frequencies can be obtained through wavelet transform, forming a time-frequency energy matrix. Based on this, the S-transform is further used to enhance and analyze the signal. The following time-frequency feature parameters are extracted based on the time-frequency energy distribution: 2.41) Energy distribution in different frequency bands; 2.42) Bandwidth energy ratio; 2.43) Time-frequency energy texture features; Finally, time-frequency domain feature vectors are formed. : This represents the frequency band energy, frequency band energy ratio, and time-frequency energy texture features. 2.5) Deep learning-based voiceprint feature extraction; To further explore the potential features in voiceprint signals, deep learning methods are introduced for automatic feature extraction. First, the voiceprint signal is converted into a Mel spectrogram. The generated Mel spectrogram and the original voiceprint signal are used as inputs to construct a multi-channel deep feature extraction network. The network structure includes a convolutional neural network (CNN) feature extraction layer, an attention fusion module, and a Transformer encoding module. Among them, CNN is used to extract local time-frequency features, attention mechanism is used to highlight important acoustic information, and Transformer encoder is used to capture long-term dependencies in acoustic signals. The deep voiceprint feature vector is obtained through the aforementioned deep learning network. : in, This represents the m-dimensional features extracted by the deep learning network; 2.6) Voiceprint feature fusion method; By fusing time-domain features, frequency-domain features, time-frequency-domain features, and deep learning features, a comprehensive voiceprint feature vector is formed. : 2.7) Acoustic fingerprint coding method; To achieve a standardized representation of voiceprint features, the fused features are normalized using the following formula: in, For normalized features, The characteristic mean, The characteristic standard deviation; Subsequently, the normalized features are encoded using a feature hashing mapping method, mapping them into a fixed-length acoustic fingerprint vector. : in, Let W be the sign function and W be the random mapping matrix; Finally, generate the device acoustic fingerprint. : in, This is the nth element of the acoustic fingerprint; 2.8) Establish a standardized acoustic fingerprint database; An acoustic fingerprint database is established based on equipment type, operating conditions, and time information. Each record in the database includes equipment number, operating status label, acoustic fingerprint characteristics, and sampling time information. When the system acquires new voiceprint data, it extracts voiceprint features and generates an acoustic fingerprint, which is then matched with fingerprints in the database. The similarity calculation formula is as follows: For the first i , j Acoustic fingerprint of a device; Cosine similarity; When the similarity is lower than the set threshold, it is determined that the equipment may be in an abnormal operating state, thereby realizing early warning of faults in large oil-filled equipment.

[0010] As a preferred option, in step 3, a multi-channel deep learning voiceprint recognition model is constructed to identify the type of partial discharge. The overall structure of the model includes a voiceprint feature input layer, a convolutional feature extraction module, a temporal feature modeling module, a global feature attention module, and a discharge type recognition layer. Specifically: 3.1) Voiceprint feature input layer; The acoustic fingerprint data is converted into multi-dimensional input features, including: 3.11) Mel spectrum; Mel spectrum can better reflect the spectral structure of acoustic signals; 3.12) Time-frequency energy diagram; The time-frequency distribution matrix of the acoustic signature signal is obtained using continuous wavelet transform; 3.13) Voiceprint feature vector; This includes time-domain, frequency-domain, and time-frequency domain feature parameters; ultimately, a three-channel input is constructed: Original voiceprint signal, Mel spectrum and time-frequency energy graph; 3.2) Convolutional feature extraction module; The convolutional feature extraction module is used to extract local feature information from the voiceprint signal; this module consists of multiple convolutional layers and pooling layers, and its calculation process is as follows: in: For the first i Feature maps output by each convolutional layer For activation function, For convolution kernel parameters, X As input features, As a bias term, convolution operations can extract local texture features and spectral features from the voiceprint signal; 3.3) Temporal Feature Modeling Module; After convolutional feature extraction, a bidirectional long short-term memory network (BiLSTM) is introduced for temporal feature modeling. The calculation process is as follows: This represents the LSTM hidden state at the current time t. The input features are at the current time t. For the previous moment The hidden state is used to remember temporal information; The two-way structure is: The time-series characteristics are obtained after concatenation using bidirectional LSTM. For the output of the forward LSTM, This is the output of an inverted LSTM; 3.4) Global Feature Attention Module; A Transformer attention mechanism is introduced for global feature modeling; the attention calculation formula is as follows: in: For normalized exponential functions, Q For querying the matrix, K The key matrix, For matrix transpose, V For value matrices, For dimensions; 3.5) Discharge type identification layer; After feature extraction, partial discharge type classification is performed using a fully connected layer; the classification output function uses the Softmax function. in: Indicates the first i The probability of a discharge type, For the fully connected layer to the first i、j The original output score of the class. This represents the total number of discharge types.

[0011] Preferably, in step 3, based on the completion of discharge type identification, a discharge severity assessment model is constructed; and a discharge severity index is defined. : in: A The amplitude of the sound signal. N The number of discharge events. E Voiceprint energy characteristics, , , These are the weighting coefficients.

[0012] Preferably, in step 3, the model is trained using samples from the acoustic fingerprint database; the training steps are as follows: Step a: Construct the training dataset; The voiceprint data was divided into a training set, a validation set, and a test set, with a ratio of 70% : 15% : 15%. Step b: Data augmentation; Data augmentation can improve the generalization ability of a model by augmenting training samples through random noise, time shifting, and frequency perturbation. Step c: Model training; Using the cross-entropy loss function: This is the loss value. For real labels, The model predicts the first i Class probability; The Adam optimization algorithm is used for parameter updates; Step d: Iterative training of the model; The model parameters are continuously optimized through multiple rounds of training, and training is stopped when the accuracy on the validation set reaches a stable level. The model optimization method is as follows: A) Adaptive feature weight fusion; The input feature weights are dynamically adjusted based on the importance of different voiceprint features. B) Multi-scale training strategies; Simultaneously, voiceprint samples at different time scales are used for training to improve the model's ability to recognize different discharge cycles. C) Transfer learning optimization; The model is pre-trained using existing device voiceprint data and then fine-tuned using target device data to improve recognition accuracy under small sample conditions. Once trained, the voiceprint recognition model can provide early warning of abnormal equipment conditions. When the severity of discharge exceeds a preset threshold, the system issues a fault warning signal, thus enabling early warning of faults in large oil-filled equipment.

[0013] Preferably, in step 4, after obtaining the trained voiceprint recognition model, a multi-level early warning threshold system is constructed, and the device operating status is evaluated in real time based on the model output results; when the evaluation results meet the early warning triggering conditions, early warning information of the corresponding level is generated; specifically: 4.1) Construction of early warning indicators; Based on a trained voiceprint recognition model, the real-time acquired voiceprint signals are identified and analyzed to obtain multiple evaluation indicators of the device's current operating status; these evaluation indicators include: 4.11) Discharge type identification probability; The voiceprint recognition model outputs probability values ​​for different discharge types through the Softmax function. When the probability of a certain type of discharge is higher than a set threshold, it is considered that the device has the corresponding type of partial discharge characteristics. 4.12) Voiceprint anomaly index; The anomaly index of voiceprint is obtained by calculating the degree of difference between real-time voiceprint features and normal features in the acoustic fingerprint database. : in: This is the current voiceprint feature vector. This represents the average voiceprint characteristics under normal conditions. is the standard deviation of features; this index is used to measure the deviation degree between the current voiceprint features of the device and the normal state; 4.13) Discharge energy index; The discharge energy index is obtained by statistically analyzing the energy of the voiceprint signal : This index is used to reflect the energy intensity of partial discharge; is the amplitude of the i th sampling point of the voiceprint signal, is the total number of sampling points; 4.14) Discharge event frequency index; Statistically analyze the number of partial discharge events detected within a unit time : Among them: is the number of detected discharge events, T is the statistical time window; Construct a comprehensive operation status evaluation parameter through the above indexes : Among them: a、b、c is the weight coefficient; this comprehensive index is used to describe the overall risk level of the current operation status of the device; 4.2) Construction of a multi-level early warning threshold system; According to the historical operation data, acoustic fingerprint database and equipment operation experience, divide the equipment operation status into four levels: normal operation status, slight abnormal status, developing fault status, serious fault status, and set corresponding multi-level early warning thresholds: normal threshold T1, attention threshold T2, warning threshold T3, alarm threshold T4; satisfying the relationship: T1 < T2 < T3 < T4; Determine the equipment status according to the relationship between the comprehensive status evaluation parameter S and each threshold: When S < T2, the equipment is in normal operation status; When T2 ≤ S < T3, the equipment is in a slight abnormal status and needs operation attention; When T3 ≤ S < T4, the equipment is in a developing fault status and the system triggers a warning message; When S ≥ T4, the equipment is in a serious fault status and the system triggers an emergency alarm message; 4.3) Dynamic threshold adaptive adjustment; In order to improve the adaptability of the early warning system to different devices and different operating environments, introduce a dynamic threshold adjustment mechanism on the basis of the basic threshold system; by statistically analyzing the historical operation data of the equipment, calculate the mean and standard deviation of the voiceprint feature parameters: in: The threshold is dynamically adjusted. m Historical data average, s For standard deviation, k To adjust the coefficients, this method enables adaptive updating of the early warning threshold. 4.4) Real-time operational status assessment; During operation, the system continuously collects acoustic signals from large oil-filled equipment and analyzes and processes them using a trained acoustic recognition model to obtain real-time acoustic feature parameters and discharge recognition results. Subsequently, the system compares the comprehensive state evaluation parameter S with a multi-level early warning threshold system to determine the current operating status of the equipment. The evaluation process includes: voiceprint data acquisition, voiceprint feature extraction, voiceprint recognition model inference, state evaluation parameter calculation, threshold comparison and state determination; 4.5) Early warning triggering mechanism; When the device status meets the warning triggering conditions, the system automatically generates a warning message of the corresponding level. The conditions for triggering an early warning include: 4.51) The comprehensive status evaluation parameters exceed the warning threshold; When S≥T3, the system triggers a device fault warning; 4.52) Continuous anomaly detection; When an abnormal state is detected within multiple consecutive time windows, the system triggers a trend warning. 4.53) The discharge type persists; When the probability of identifying a certain type of discharge is higher than the set threshold for an extended period, the system triggers a fault warning for the corresponding type. The warning information includes: device number, detection time, discharge type, fault level, and status assessment parameters. The system will send the warning information to the monitoring platform or operation and maintenance system to provide decision-making basis for operation and maintenance personnel. 4.6) Early warning information output; Depending on the warning level, the system generates different levels of warning information, including: alert information, attention information, warning information, and alarm information; By providing alerts to maintenance personnel through visual interfaces, alarm systems, or remote monitoring platforms, early warnings and proactive maintenance management of large oil-filled equipment failures can be achieved.

[0014] Preferably, in step 5, after determining the equipment's operating status level and fault type, the system automatically matches the corresponding operation and maintenance strategy; the specific strategy is as follows: 5.1) Normal operation and maintenance strategy; When the equipment is in normal operation, the system records monitoring data and maintains a regular inspection cycle, while continuing online voiceprint monitoring in order to detect potential anomalies in a timely manner. 5.2) Focus on operational status and maintenance strategies; When the equipment is in a slightly abnormal state, the system recommends that maintenance personnel strengthen equipment operation monitoring and appropriately shorten the inspection cycle, while performing the following checks on the equipment: Check the equipment load operation, check the oil temperature and oil level, and check the cooling system operation status; The above methods are used to further confirm whether there are any potential abnormalities in the equipment; 5.3) Early warning operation status maintenance strategy; When equipment is in a state of progressive failure, the system recommends that maintenance personnel arrange a special inspection and take preventative maintenance measures, including: Conduct partial discharge detection or ultrasonic testing, check the quality and gas content of insulating oil, and check whether the winding and core structure is loose. If necessary, arrange for equipment to be shut down for maintenance to prevent the fault from worsening. 5.4) Operation and maintenance strategy for critical alarm status; When the equipment is in a critical malfunction state, the system recommends immediate emergency response measures, including: Reduce equipment operating load, start backup equipment or backup lines, and arrange emergency power outage maintenance; At the same time, a comprehensive fault diagnosis is performed on the equipment to prevent equipment damage or accidents from occurring.

[0015] This invention provides an early warning system for faults in large oil-filled equipment, which employs the aforementioned method for early warning of faults in large oil-filled equipment.

[0016] The beneficial effects of this invention are as follows: This invention constructs a comprehensive and well-labeled acoustic fingerprint database by collecting acoustic fingerprint data of large oil-filled equipment throughout its entire lifecycle under normal operation, load changes, DC bias, and various partial discharge conditions without power interruption. This data is then combined with supplementary data generated from a physical simulation model. This effectively solves the problems of scarce partial discharge samples and imbalanced data in actual operation, providing a high-quality data foundation for training deep learning models and significantly improving the model's generalization ability and accuracy in identifying rare faults.

[0017] This invention extracts features from the time domain, frequency domain, time-frequency domain, and deep learning domain to construct a multi-dimensional acoustic fingerprint feature vector. An adaptive feature weight fusion method is used to dynamically adjust the contribution of each type of feature. Compared to traditional single-feature or simple concatenation methods, this invention can more comprehensively and accurately characterize the acoustic fingerprint characteristics of different types of partial discharges, significantly improving the accuracy of discharge type identification and severity assessment.

[0018] This invention constructs a multi-channel deep learning voiceprint recognition model that integrates CNN, BiLSTM and Transformer attention mechanisms. It can simultaneously process multi-dimensional inputs such as original voiceprint signals, Mel spectrograms and time-frequency energy maps, automatically extract local texture features, temporal dependencies and global context information, achieve end-to-end accurate identification of partial discharge types, and construct a discharge severity assessment model to provide quantitative basis for device status judgment.

[0019] This invention constructs a comprehensive operational status evaluation parameter system that includes indicators such as discharge type identification probability, voiceprint anomaly degree, discharge energy, and discharge event frequency. It also establishes a multi-level early warning threshold system (normal, watchful, early warning, alarm) and introduces a dynamic threshold adaptive adjustment mechanism. Compared to traditional single-threshold judgment methods, this invention can more sensitively capture early, weak abnormal signals, enabling early warning and trend prediction of equipment failures, and effectively preventing accidents.

[0020] Based on the identification of fault type and severity, this invention automatically matches the corresponding operation and maintenance strategy, generates operation and maintenance suggestions including inspection recommendations, maintenance priorities and risk levels, and pushes them to operation and maintenance personnel through a visual interface or remote monitoring platform, which significantly improves the intelligent operation and maintenance level of equipment and reduces the intensity of manual inspection and operation and maintenance costs. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating a method for early warning of faults in large oil-filled equipment, as described in the embodiment. Figure 2 This is a schematic diagram of the sensor arrangement scheme in the embodiment. Detailed Implementation

[0022] 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.

[0023] Example like Figure 1 As shown, this embodiment provides a method for early warning of faults in large oil-filled equipment, which includes the following steps: Step 1: Collect full lifecycle acoustic fingerprint data of large oil-filled equipment under various operating conditions, and construct an acoustic fingerprint sample set including normal and abnormal states; wherein, the abnormal states include at least partial discharge types such as corona discharge, suspension discharge, surface discharge and internal discharge.

[0024] To obtain comprehensive and representative acoustic signature data, this step involves continuously collecting acoustic signature data under normal operation, load variation, DC bias, and partial discharge conditions using a high-sensitivity ultrasonic sensor array deployed at key parts of the equipment, without power interruption. Simultaneously, a physical simulation model of transformer partial discharge is constructed to generate supplementary acoustic signature data for different discharge types and locations, thereby expanding the acoustic signature sample set and addressing the problem of scarce partial discharge samples in actual operation.

[0025] In this embodiment, a converter transformer is used as an example of a large oil-filled device. Figure 2 As shown, a high-sensitivity ultrasonic sensor array is arranged on the converter transformer body. The sensors are piezoelectric accelerometers with a frequency response range of 20kHz to 160kHz, sensitivity ≥100 mV / g, and electromagnetic interference resistance ≥100 V / m. One contact sensor is installed on each of the A, B, and C phase sidewalls of the transformer tank (avoiding the cooler), and fixed using magnetic attraction. Three non-contact acoustic sensors are installed in a triangular arrangement at a distance of 1.5 meters from the transformer body, forming an acoustic array for sound source localization.

[0026] The acquisition system operates continuously, acquiring voiceprint data without power interruption, with a sampling rate of 256 kHz and quantization precision of 24 bits. The acquisition conditions include: Normal operating conditions (load rate 30%~80%); Overload operation status (load rate > 100%) DC bias state (simulated by DC injection at neutral point); Partial discharge state: By constructing a physical simulation model of transformer partial discharge (such as a finite element multiphysics coupling model), four typical discharge types, namely corona discharge, floating discharge, surface discharge, and internal discharge, are simulated, and the simulated acoustic fingerprint data is recorded to expand the sample set.

[0027] The collected raw voiceprint data is segmented into 10-second segments to form voiceprint samples. Simultaneously, the device number, operating condition, load rate, and discharge type (if applicable) corresponding to each sample segment are recorded. A voiceprint sample set containing normal samples and various abnormal samples is constructed, with no fewer than 500 samples in each category, to meet the training requirements of the deep learning model.

[0028] Step 2: Preprocess the voiceprint sample set, extract voiceprint feature parameters, and establish a standardized acoustic fingerprint database; the voiceprint feature parameters include time-domain features, frequency-domain features, time-frequency-domain features, and deep learning features.

[0029] The acquired raw voiceprint signal is preprocessed, including: Noise reduction: Wavelet thresholding (wavelet basis selected as db4, decomposition layer of 5, soft thresholding function) and a noise reduction network based on deep autoencoder are used for joint noise reduction.

[0030] Framed windowing: The signal is divided into frames with a frame length of 30 ms and a frame shift of 15 ms, and a Hamming window is used for windowing.

[0031] Endpoint detection: Based on short-time energy and short-time zero-crossing rate, silent segments are eliminated, while valid voiceprint segments are retained.

[0032] The establishment of an acoustic fingerprint database includes the following steps: frame processing of voiceprint signals, extraction of multi-domain voiceprint features, adaptive fusion of voiceprint features, acoustic fingerprint encoding, and construction of a standardized acoustic fingerprint database.

[0033] The specific steps are as follows: 2.1) Frame-by-frame processing of voiceprint signals; Short-time analysis method is used to analyze the preprocessed voiceprint signal Perform frame segmentation processing. t Representing time; assuming a frame length of 20ms to 40ms, a frame shift of 10ms to 20ms, and using a Hamming window function to window each frame signal, the result is obtained. n The frame signal is represented as: in, The amplitude of the kth sampling point of the original voiceprint signal. For window functions, R The frame shift length, Indicates the first n The first frame k Each sampling point signal value; through the above processing, a series of short-time frame sequences are obtained, providing basic data for subsequent voiceprint feature extraction; 2.2) Temporal feature extraction; For each frame of the audioprint signal, time-domain statistical features reflecting signal energy changes and impact characteristics are extracted, specifically including: 2.21) Root mean square (RMS) value, used to reflect the energy change of the acoustic signal, is calculated using the following formula: This represents the number of sampling points for a single frame of signal. For the first iThe signal amplitude at each sampling point; 2.22) Kurtosis K, used to characterize the impulse characteristics of a signal, is calculated using the following formula: in, For mathematical expectation, It is a sequence of signal amplitudes within a single frame. The mean of the signal. Standard deviation; 2.23) Peak factor CF, used to characterize the transient impulse component in a signal, is calculated using the following formula: The maximum amplitude of a single frame signal; 2.24) Short-time energy This is used to reflect the change of acoustic signal energy over time, and its calculation formula is: For the first n Frame number i The signal amplitude at each sampling point; The time-domain feature vector is constructed using the above parameters. : 2.3) Frequency domain feature extraction; Perform a Fast Fourier Transform (FFT) on each frame of the signal to convert the time-domain signal to the frequency domain, thus obtaining the spectrum. expression: For frequency, The imaginary unit, The amplitude of the nth sampling point within a single frame; The following frequency domain feature parameters are extracted based on the spectral information: 2.31) Spectral centroid , is used to represent the concentrated location of spectral energy, and its calculation formula is: For the first i Frequency values ​​at each frequency point For frequency Spectral amplitude at that location; 2.32) Spectral Entropy This is used to measure the complexity of the spectral distribution, and its calculation formula is: in, Represents the normalized spectral probability; Indicates Logarithm to base 0; 2.33) Spectral flatness This is used to reflect the uniformity of the spectrum, and its calculation formula is: Power spectral density; 2.34) Spectral roll-off frequency , is used to indicate the frequency position corresponding to when the accumulated energy reaches a certain proportion; The frequency domain feature vector is formed using the above parameters. : 2.4) Time-frequency domain feature extraction; Because the acoustic signature of partial discharge exhibits non-stationary characteristics, it is difficult to fully describe it using only time-domain or frequency-domain features. Therefore, a multi-resolution time-frequency analysis method is employed for feature extraction. First, continuous wavelet transform (CWT) is used to perform time-frequency analysis on the acoustic signal; the calculation formula is as follows: in, These are wavelet coefficients, representing the signal at different scales. Translation The time-frequency components below, For scale parameters, For translation parameters, For the mother wavelet function; Wavelet transform can be used to obtain the energy distribution of the acoustic signal at different times and frequencies, forming a time-frequency energy matrix. ; Based on this, the S-transform is further used to enhance and analyze the signal; The following time-frequency feature parameters are extracted based on the time-frequency energy distribution: 2.41) Energy distribution in different frequency bands; 2.42) Bandwidth energy ratio; 2.43) Time-frequency energy texture features; Finally, time-frequency domain feature vectors are formed. : This represents the frequency band energy, frequency band energy ratio, and time-frequency energy texture features. 2.5) Deep learning-based voiceprint feature extraction; To further explore the potential features in voiceprint signals, deep learning methods are introduced for automatic feature extraction. First, the voiceprint signal is converted into a Mel frequency spectrum; the conversion formula for Mel frequency scales is: Indicates linear frequency; This frequency scale is more in line with the laws of sound signal perception; The generated Mel spectrogram and the original voiceprint signal are used as inputs to construct a multi-channel deep feature extraction network. The network structure includes a convolutional neural network (CNN) feature extraction layer, an attention fusion module, and a Transformer encoding module. Among them, CNN is used to extract local time-frequency features, attention mechanism is used to highlight important acoustic information, and Transformer encoder is used to capture long-term dependencies in acoustic signals. The deep voiceprint feature vector is obtained through the aforementioned deep learning network. : in, This represents the m-dimensional features extracted by the deep learning network; 2.6) Voiceprint feature fusion method; By fusing time-domain features, frequency-domain features, time-frequency-domain features, and deep learning features, a comprehensive voiceprint feature vector is formed. : To avoid excessively high feature dimensionality and improve feature expressive power, an adaptive feature weight fusion method is proposed, the expression of which is: in: For adaptive fusion features, These are the feature weight coefficients; Each weight coefficient is automatically calculated using the feature mutual information evaluation method, enabling different types of features to be adaptively adjusted according to their contribution during the fusion process; 2.7) Acoustic fingerprint coding method; To achieve a standardized representation of voiceprint features, the fused features are normalized using the following formula: in, For normalized features, The characteristic mean, The characteristic standard deviation; The normalized features are then encoded using a feature hashing mapping method, mapping them into a fixed-length acoustic fingerprint vector: in, Let W be the sign function and W be the random mapping matrix; Finally, generate the device acoustic fingerprint. : Let n be the nth element of the acoustic fingerprint.

[0034] 2.8) The method for establishing a standardized acoustic fingerprint database is as follows; An acoustic fingerprint database is established based on equipment type, operating conditions, and time information. Each record in the database includes equipment number, operating status label, acoustic fingerprint characteristics, and sampling time information. When the system acquires new voiceprint data, it extracts voiceprint features and generates an acoustic fingerprint, which is then matched with fingerprints in the database. The similarity calculation formula is as follows: For the first i , j Acoustic fingerprint of a device; Cosine similarity; When the similarity is lower than the set threshold, it is determined that the equipment may be in an abnormal operating state, thereby realizing early warning of faults in large oil-filled equipment.

[0035] Step 3: Construct a multi-channel deep learning voiceprint recognition model and a discharge severity assessment model based on deep learning. Use samples from the acoustic fingerprint database to train and optimize the model to obtain a trained voiceprint recognition model. The voiceprint recognition model is used to identify the type of partial discharge and assess the severity of the discharge.

[0036] A multi-channel deep learning voiceprint recognition model is constructed to identify partial discharge types. The overall structure of the model includes a voiceprint feature input layer, a convolutional feature extraction module, a temporal feature modeling module, a global feature attention module, and a discharge type recognition layer.

[0037] Specifically: 3.1) Voiceprint feature input layer; To provide rich and complementary voiceprint information for subsequent deep networks, this layer converts the raw data in the acoustic fingerprint database into three-channel input features, including: 3.11) Mel spectrum; Mel spectrum can better reflect the spectral structure of acoustic signals; 3.12) Time-frequency energy diagram; The time-frequency distribution matrix of the voiceprint signal is obtained by using continuous wavelet transform, preserving the transient details of the non-stationary signal; 3.13) Voiceprint feature vector; This includes the time-domain, frequency-domain, and time-frequency-domain feature parameters extracted in step 2; ultimately, a three-channel input is constructed: Original voiceprint signal, Mel spectrum and time-frequency energy graph; 3.2) Convolutional feature extraction module (local feature extraction); The convolutional feature extraction module receives the above three-channel input and extracts local feature information from the voiceprint signal; this module consists of multiple convolutional layers and pooling layers, and its calculation process is as follows: in: For the first i Feature maps output by each convolutional layer For activation function, For convolution kernel parameters, X As input features, As a bias term, convolution operations can extract local texture features and spectral features from the voiceprint signal; 3.3) Temporal Feature Modeling Module; To capture the continuity and temporal evolution of partial discharge events, a bidirectional long short-term memory network (BiLSTM) is introduced after convolutional feature extraction to model temporal features. The calculation process is as follows: This represents the LSTM hidden state at the current time t. The input features are at the current time t. For the previous moment The hidden state is used to remember temporal information; The two-way structure is: The time-series characteristics are obtained after concatenation using bidirectional LSTM. For the output of the forward LSTM, This is the output of an inverted LSTM; 3.4) Global Feature Attention Module; While BiLSTM can capture temporal dependencies within a limited range, more robust global modeling capabilities are still needed to address intermittent discharges, periodic disturbances, or long-term evolution trends that may occur in large oil-filled equipment. This module introduces a Transformer multi-head attention mechanism to globally interact with features across all time steps, highlighting the key moments most relevant to the current discharge type determination. The attention calculation formula is as follows: in: For normalized exponential functions, Q For querying the matrix, K The key matrix, For matrix transpose, V For value matrices, For dimensions.

[0038] BiLSTM output As input to the attention module, global attention weights are generated by calculating the correlation between any two time steps. These weights are then weighted and summed over V to output a global context feature vector G. This vector integrates all voiceprint information within the entire time window and significantly enhances features related to partial discharge types (such as the high-frequency periodicity of corona discharge and the broadband bursts of surface discharge), while suppressing background noise. G is finally fed into the discharge type recognition layer.

[0039] 3.5) Discharge type identification layer; After obtaining the global context feature vector G, a fully connected layer is used for partial discharge type classification, mapping G to the scores of each discharge type; the classification output function uses the Softmax function. in: Indicates the first i The probability of a discharge type, For the fully connected layer to the first i、j The original output score of the class. This represents the total number of discharge types.

[0040] In step 3, based on the completion of discharge type identification, a discharge severity assessment model is constructed; and discharge severity indices are defined. : in: A The amplitude of the sound signal. N The number of discharge events. E Voiceprint energy characteristics, , , These are the weighting coefficients.

[0041] In step 3, the model is trained using samples from the acoustic fingerprint database; the training steps are as follows: Step a: Construct the training dataset; The voiceprint data was divided into a training set, a validation set, and a test set, with a ratio of 70% : 15% : 15%. Step b: Data augmentation; Data augmentation can improve the generalization ability of a model by augmenting training samples through random noise, time shifting, and frequency perturbation. Step c: Model training; Using the cross-entropy loss function: This is the loss value. For real labels, The model predicts the first i Class probability; The Adam optimization algorithm is used for parameter updates; Step d: Iterative training of the model; The model parameters are continuously optimized through multiple rounds of training, and training is stopped when the accuracy on the validation set reaches a stable level. The model optimization method is as follows: A) Adaptive feature weight fusion; The input feature weights are dynamically adjusted based on the importance of different voiceprint features. B) Multi-scale training strategies; Simultaneously, voiceprint samples at different time scales are used for training to improve the model's ability to recognize different discharge cycles. C) Transfer learning optimization; The model is pre-trained using existing device voiceprint data and then fine-tuned using target device data to improve recognition accuracy under small sample conditions. Once trained, the voiceprint recognition model can provide early warning of abnormal equipment conditions. When the severity of discharge exceeds a preset threshold, the system issues a fault warning signal, thus enabling early warning of faults in large oil-filled equipment.

[0042] Step 4: Based on the output of the trained voiceprint recognition model and combined with the preset multi-level early warning threshold system, the operating status of the large oil filling equipment is evaluated in real time. When the evaluation result meets the early warning triggering conditions, the corresponding early warning information is generated.

[0043] In step 4, after obtaining the trained voiceprint recognition model, a multi-level early warning threshold system is constructed, and the device operating status is evaluated in real time based on the model output results; when the evaluation results meet the early warning triggering conditions, the corresponding level of early warning information is generated; specifically: 4.1) Construction of early warning indicators; Based on a trained voiceprint recognition model, the real-time acquired voiceprint signals are identified and analyzed to obtain multiple evaluation indicators of the device's current operating status; these evaluation indicators include: 4.11) Discharge type identification probability; The voiceprint recognition model outputs probability values ​​for different discharge types through the Softmax function. When the probability of a certain type of discharge is higher than a set threshold, it is considered that the device has the corresponding type of partial discharge characteristics. 4.12) Voiceprint anomaly index; The anomaly index of voiceprint is obtained by calculating the degree of difference between real-time voiceprint features and normal features in the acoustic fingerprint database. : in: This is the current voiceprint feature vector. This represents the average voiceprint characteristics under normal conditions. The standard deviation of the characteristic is used to measure the degree of deviation between the current voiceprint characteristics of the device and its normal state. 4.13) Discharge energy index; Discharge energy index is obtained by statistically analyzing the voiceprint signal energy. : This indicator is used to reflect the energy intensity of partial discharge; For the voiceprint signal i The amplitude of each sampling point This represents the total number of sampling points; 4.14) Discharge event frequency index; Count the number of partial discharge events detected per unit time. : in: The number of discharge events detected. T For statistical time windows; The above indicators are used to construct comprehensive operational status evaluation parameters. : in: a、b、c This is a weighting coefficient; this comprehensive indicator is used to describe the overall risk level of the equipment's current operating status. 4.2) Construction of a multi-level early warning threshold system; According to historical operation data, the acoustic fingerprint database, and equipment operation experience, the equipment operation status is divided into four levels: normal operation status, slight anomaly status, developing fault status, and severe fault status. Corresponding multi-level early warning thresholds are set: normal threshold T1, attention threshold T2, warning threshold T3, and alarm threshold T4; the relationship is satisfied: T1 < T2 < T3 < T4; Determine the equipment status according to the relationship between the comprehensive status evaluation parameter S and each threshold: When S < T2, the equipment is in a normal operation status; When T2 ≤ S < T3, the equipment is in a slight anomaly status and requires operation attention; When T3 ≤ S < T4, the equipment is in a developing fault status, and the system triggers a warning message; When S ≥ T4, the equipment is in a severe fault status, and the system triggers an emergency alarm message; 4.3) Dynamic threshold adaptive adjustment; To improve the adaptability of the early warning system to different equipment and different operating environments, a dynamic threshold adjustment mechanism is introduced to the basic threshold system; by statistically analyzing the historical operation data of the equipment, calculate the mean and standard deviation of the voiceprint characteristic parameters: Where: is the threshold after dynamic adjustment, m is the mean of historical data, s is the standard deviation, k is the adjustment coefficient; the adaptive update of the early warning threshold is realized through this method; 4.4) Real-time operation status evaluation; During the operation of the system, continuously collect the voiceprint signals of large oil-filled equipment, and analyze and process them through a trained voiceprint recognition model to obtain real-time voiceprint characteristic parameters and discharge recognition results; then compare the comprehensive status evaluation parameter S with the multi-level early warning threshold system to judge the current operation status of the equipment; The evaluation process includes: voiceprint data collection, voiceprint feature extraction, voiceprint recognition model inference, status evaluation parameter calculation, threshold comparison, and status determination; 4.5) Early warning trigger mechanism; When the equipment status meets the early warning trigger condition, the system automatically generates an early warning message of the corresponding level; The early warning trigger conditions include: 4.51) The comprehensive status evaluation parameter exceeds the warning threshold; When S ≥ T3, the system triggers an equipment fault warning; 4.52) Continuous anomaly detection; When an abnormal state is detected within multiple consecutive time windows, the system triggers a trend warning. 4.53) The discharge type persists; When the probability of identifying a certain type of discharge is higher than the set threshold for an extended period, the system triggers a fault warning for the corresponding type. The warning information includes: device number, detection time, discharge type, fault level, and status assessment parameters. The system will send the warning information to the monitoring platform or operation and maintenance system to provide decision-making basis for operation and maintenance personnel. 4.6) Early warning information output; Depending on the warning level, the system generates different levels of warning information, including: alert information, attention information, warning information, and alarm information; By providing alerts to maintenance personnel through visual interfaces, alarm systems, or remote monitoring platforms, early warnings and proactive maintenance management of large oil-filled equipment failures can be achieved.

[0044] Step 5: Push the warning information to the operation and maintenance terminal, and automatically generate operation and maintenance suggestions including inspection suggestions, maintenance priorities and risk levels based on the defect type and severity.

[0045] In step 5, after determining the equipment's operating status level and fault type, the system automatically matches the corresponding operation and maintenance strategy; the specific strategies are as follows: 5.1) Normal operation and maintenance strategy; When the equipment is in normal operation, the system records monitoring data and maintains a regular inspection cycle, while continuing online voiceprint monitoring in order to detect potential anomalies in a timely manner. 5.2) Focus on operational status and maintenance strategies; When the equipment is in a slightly abnormal state, the system recommends that maintenance personnel strengthen equipment operation monitoring and appropriately shorten the inspection cycle, while performing the following checks on the equipment: Check the equipment load operation, check the oil temperature and oil level, and check the cooling system operation status; The above methods are used to further confirm whether there are any potential abnormalities in the equipment; 5.3) Early warning operation status maintenance strategy; When equipment is in a state of progressive failure, the system recommends that maintenance personnel arrange a special inspection and take preventative maintenance measures, including: Conduct partial discharge detection or ultrasonic testing, check the quality and gas content of insulating oil, and check whether the winding and core structure is loose. If necessary, arrange for equipment to be shut down for maintenance to prevent the fault from worsening. 5.4) Operation and maintenance strategy for critical alarm status; When the equipment is in a critical malfunction state, the system recommends immediate emergency response measures, including: Reduce equipment operating load, start backup equipment or backup lines, and arrange emergency power outage maintenance; At the same time, a comprehensive fault diagnosis is performed on the equipment to prevent equipment damage or accidents from occurring.

[0046] This embodiment provides an early warning system for faults in large oil-filled equipment, which adopts the above-described method for early warning of faults in large oil-filled equipment.

[0047] This embodiment provides a comprehensive early fault warning method and system for large oil-filled equipment, integrating acoustic signature acquisition, feature extraction, intelligent identification, status assessment, early warning, and operation and maintenance decision-making. It has significant engineering application value and promotion prospects for improving the reliability of power equipment operation and promoting intelligent operation and maintenance of power grids.

[0048] 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 early warning of faults in large oil-filled equipment, characterized in that, Includes the following steps: Step 1: Collect full lifecycle acoustic fingerprint data of large oil-filled equipment under various operating conditions, and construct an acoustic fingerprint sample set including normal and abnormal states; wherein, the abnormal states include at least partial discharge types such as corona discharge, suspension discharge, surface discharge and internal discharge. Step 2: Preprocess the voiceprint sample set, extract voiceprint feature parameters, and establish a standardized acoustic fingerprint database; the voiceprint feature parameters include time-domain features, frequency-domain features, time-frequency-domain features, and deep learning features; The establishment of an acoustic fingerprint database includes the following steps: frame processing of acoustic fingerprint signals, multi-domain acoustic fingerprint feature extraction, adaptive fusion of acoustic fingerprint features, acoustic fingerprint encoding, and construction of a standardized acoustic fingerprint database. Step 3: Construct a multi-channel deep learning voiceprint recognition model and a discharge severity assessment model based on deep learning. Use samples from the acoustic fingerprint database to train and optimize the model to obtain a trained voiceprint recognition model. The voiceprint recognition model is used to identify partial discharge types and assess the severity of discharge. Step 4: Based on the output of the trained voiceprint recognition model and combined with the preset multi-level early warning threshold system, the operating status of the large oil filling equipment is evaluated in real time. When the evaluation result meets the early warning triggering conditions, the corresponding early warning information is generated. Step 5: Push the warning information to the operation and maintenance terminal, and automatically generate operation and maintenance suggestions including inspection suggestions, maintenance priorities and risk levels based on the defect type and severity.

2. The method for early warning of faults in large oil-filled equipment according to claim 1, characterized in that, Step 1, which involves collecting full-lifecycle voiceprint data, specifically includes: Under uninterrupted power supply, a high-sensitivity ultrasonic sensor array deployed in key parts of the equipment is used to continuously collect acoustic data under normal operation, load change, DC bias, and partial discharge conditions. By constructing a physical simulation model of transformer partial discharge, supplementary voiceprint data with different discharge types and locations are generated to expand the voiceprint sample set.

3. The method for early warning of faults in large oil-filled equipment according to claim 2, characterized in that, In step 2, the specific steps are as follows: 2.1) Frame-by-frame processing of voiceprint signals; Short-time analysis method is used to analyze the preprocessed voiceprint signal Perform frame segmentation processing. t Indicates time; Assuming a frame length of 20ms to 40ms and a frame shift of 10ms to 20ms, and applying a Hamming window function to window each frame signal, we obtain the... n The frame signal is represented as: in, The amplitude of the kth sampling point of the original voiceprint signal. For window functions, R The frame shift length, Indicates the first n The first frame k Each sampling point signal value; through the above processing, a series of short-time frame sequences are obtained, providing basic data for subsequent voiceprint feature extraction; 2.2) Temporal feature extraction; For each frame of the audioprint signal, time-domain statistical features reflecting signal energy changes and impact characteristics are extracted, specifically including: 2.21) Root mean square (RMS) value, used to reflect the energy change of the acoustic signal; 2.22) Kurtosis K, used to characterize the impulse characteristics of a signal; 2.23) Peak factor CF, used to characterize the transient impulse component in a signal; 2.24) Short-time energy It is used to reflect the change of acoustic signal energy over time; The time-domain feature vector is constructed using the above parameters. : 2.3) Frequency domain feature extraction; Perform a Fast Fourier Transform (FFT) on each frame of the signal to convert the time-domain signal to the frequency domain, thus obtaining the spectrum. ; The following frequency domain feature parameters are extracted based on the spectral information: 2.31) Spectral centroid , used to indicate the concentrated location of spectral energy; 2.32) Spectral Entropy This is used to measure the complexity of the spectral distribution; 2.33) Spectral flatness This is used to reflect the uniformity of the spectrum; 2.34) Spectral roll-off frequency , is used to indicate the frequency position corresponding to when the accumulated energy reaches a certain proportion; The frequency domain feature vector is formed using the above parameters. : 2.4) Time-frequency domain feature extraction; Feature extraction is performed using a multi-resolution time-frequency analysis method. First, continuous wavelet transform (CWT) is used to perform time-frequency analysis on the acoustic signal. The energy distribution of the acoustic signal at different times and frequencies can be obtained through wavelet transform, forming a time-frequency energy matrix. Based on this, the S-transform is further used to enhance and analyze the signal. The following time-frequency feature parameters are extracted based on the time-frequency energy distribution: 2.41) Energy distribution in different frequency bands; 2.42) Bandwidth energy ratio; 2.43) Time-frequency energy texture features; Finally, time-frequency domain feature vectors are formed. : This represents the frequency band energy, frequency band energy ratio, and time-frequency energy texture features. 2.5) Deep learning-based voiceprint feature extraction; To further explore the potential features in voiceprint signals, deep learning methods are introduced for automatic feature extraction. First, the voiceprint signal is converted into a Mel spectrogram. The generated Mel spectrogram and the original voiceprint signal are used as inputs to construct a multi-channel deep feature extraction network. The network structure includes a convolutional neural network (CNN) feature extraction layer, an attention fusion module, and a Transformer encoding module. Among them, CNN is used to extract local time-frequency features, attention mechanism is used to highlight important acoustic information, and Transformer encoder is used to capture long-term dependencies in acoustic signals. The deep voiceprint feature vector is obtained through the aforementioned deep learning network. : in, This represents the m-dimensional features extracted by the deep learning network; 2.6) Voiceprint feature fusion method; By fusing time-domain features, frequency-domain features, time-frequency-domain features, and deep learning features, a comprehensive voiceprint feature vector is formed. : 2.7) Acoustic fingerprint coding method; To achieve a standardized representation of voiceprint features, the fused features are normalized using the following formula: in, For normalized features, The characteristic mean, The characteristic standard deviation; Subsequently, the normalized features are encoded using a feature hashing mapping method, mapping them into a fixed-length acoustic fingerprint vector. : in, Let W be the sign function and W be the random mapping matrix; Finally, generate the device acoustic fingerprint. : in, The nth element of the acoustic fingerprint; 2.8) Establish a standardized acoustic fingerprint database; An acoustic fingerprint database is established based on equipment type, operating conditions, and time information. Each record in the database includes equipment number, operating status label, acoustic fingerprint characteristics, and sampling time information. When the system acquires new voiceprint data, it extracts voiceprint features and generates an acoustic fingerprint, which is then matched with fingerprints in the database. The similarity calculation formula is as follows: For the first i , j Acoustic fingerprint of a device; Cosine similarity; When the similarity is lower than the set threshold, it is determined that the equipment may be in an abnormal operating state, thereby realizing early warning of failure of large oil filling equipment.

4. The method for early warning of faults in large oil-filled equipment according to claim 3, characterized in that, In step 3, a multi-channel deep learning voiceprint recognition model is constructed to identify partial discharge types. The overall structure of the model includes a voiceprint feature input layer, a convolutional feature extraction module, a temporal feature modeling module, a global feature attention module, and a discharge type recognition layer. Specifically: 3.1) Voiceprint feature input layer; The acoustic fingerprint data is converted into multi-dimensional input features, including: 3.11) Mel spectrum; Mel spectrum can better reflect the spectral structure of acoustic signals; 3.12) Time-frequency energy diagram; The time-frequency distribution matrix of the acoustic signature signal is obtained using continuous wavelet transform; 3.13) Voiceprint feature vector; This includes time-domain, frequency-domain, and time-frequency domain feature parameters; ultimately, a three-channel input is constructed: Original voiceprint signal, Mel spectrum and time-frequency energy graph; 3.2) Convolutional feature extraction module; The convolutional feature extraction module is used to extract local feature information from the voiceprint signal; this module consists of multiple convolutional layers and pooling layers, and its calculation process is as follows: in: For the first i Feature maps output by each convolutional layer For activation function, For convolution kernel parameters, X For input features, As a bias term, convolution operations can extract local texture features and spectral features from the voiceprint signal; 3.3) Temporal Feature Modeling Module; After convolutional feature extraction, a bidirectional long short-term memory network (BiLSTM) is introduced for temporal feature modeling. The calculation process is as follows: This represents the LSTM hidden state at the current time t. The input features are at the current time t. For the previous moment The hidden state is used to remember temporal information; The two-way structure is: The timing features are those obtained after concatenation using bidirectional LSTM. For the output of the forward LSTM, This is the output of an inverted LSTM; 3.4) Global Feature Attention Module; A Transformer attention mechanism is introduced for global feature modeling; the attention calculation formula is as follows: in: For normalized exponential functions, Q For querying the matrix, K The key matrix, For matrix transpose, V For value matrices, For dimensions; 3.5) Discharge type identification layer; After feature extraction, partial discharge type classification is performed using a fully connected layer; the classification output function uses the Softmax function. in: Indicates the first i The probability of a discharge type, For the fully connected layer to the first i, j The original output score of the class. This represents the total number of discharge types.

5. A method for early warning of faults in large oil-filled equipment according to claim 4, characterized in that, In step 3, based on the completion of discharge type identification, a discharge severity assessment model is constructed; and discharge severity indices are defined. : in: A The amplitude of the sound signal. N The number of discharge events. E Voiceprint energy characteristics, , , These are the weighting coefficients.

6. A method for early warning of faults in large oil-filled equipment according to claim 5, characterized in that, In step 3, the model is trained using samples from the acoustic fingerprint database; the training steps are as follows: Step a: Construct the training dataset; The voiceprint data was divided into a training set, a validation set, and a test set, with a ratio of 70% : 15% : 15%. Step b: Data augmentation; Augment the training samples by the following methods: adding random noise, time translation, and frequency perturbation. Data augmentation can improve the generalization ability of the model; Step c: Model training; Adopt the cross-entropy loss function: This is the loss value. For real labels, The model predicts the first i Class probability; And use the Adam optimization algorithm for parameter update; Step d: Model iterative training; Continuously optimize the model parameters through multiple rounds of training, and stop training when the accuracy of the validation set reaches stability; The model optimization method is: A) Adaptive feature weight fusion; Dynamically adjust the input feature weights according to the importance of different voiceprint features; B) Multi-scale training strategy; Use voiceprint samples with different time scales for training simultaneously to improve the model's recognition ability for different discharge cycles; C) Transfer learning optimization; Pre-train the model using the existing device voiceprint data, and then fine-tune it using the target device data to improve the recognition accuracy under the condition of small samples; The voiceprint recognition model after training can achieve early warning of the abnormal state of the device. When it is recognized that the discharge severity exceeds the preset threshold, the system issues a device fault warning signal, thereby realizing early warning of large oil-filled equipment faults.

7. A method for early warning of faults in large oil-filled equipment according to claim 6, characterized in that, In step 4, after obtaining the trained voiceprint recognition model, construct a multi-level warning threshold system, and combine the model output results to evaluate the real-time operation state of the device; when the evaluation result meets the warning trigger condition, generate a warning message of the corresponding level; specifically: 4.1) Construction of warning indicators; Based on the trained voiceprint recognition model, identify and analyze the real-time collected voiceprint signals to obtain multiple evaluation indicators of the current operation state of the device; the evaluation indicators include: 4.11) Recognition probability of discharge type; The voiceprint recognition model outputs the probability values of different discharge types through the Softmax function. When the probability of a certain type of discharge is higher than the set threshold, it is considered that the device has local discharge characteristics of the corresponding type; 4.12) Voiceprint abnormality index; The anomaly index of voiceprint is obtained by calculating the degree of difference between real-time voiceprint features and normal features in the acoustic fingerprint database. : in: This is the current voiceprint feature vector. This represents the average voiceprint characteristics under normal conditions. The standard deviation of the characteristic is used to measure the degree of deviation between the current voiceprint characteristics of the device and its normal state. 4.13) Discharge energy index; Discharge energy index is obtained by statistically analyzing the voiceprint signal energy. : This indicator is used to reflect the energy intensity of partial discharge; For the voiceprint signal i The amplitude of each sampling point This represents the total number of sampling points; 4.14) Discharge event frequency index; Count the number of partial discharge events detected per unit time. : in: The number of discharge events detected. T For statistical time windows; The above indicators are used to construct comprehensive operational status evaluation parameters. : in: α, β, γ This is a weighting coefficient; this comprehensive indicator is used to describe the overall risk level of the equipment's current operating status. 4.2) Construction of multi-level warning threshold system; According to the historical operation data, acoustic fingerprint library, and device operation experience, divide the device operation state into four levels: normal operation state, slight abnormal state, developing fault state, and serious fault state, and set corresponding multi-level warning thresholds: normal threshold T1, attention threshold T2, warning threshold T3, alarm threshold T4; satisfy the relationship: T1 < T2 < T3 < T4; Determine the device state according to the relationship between the comprehensive state evaluation parameter S and each threshold: When S < T2, the device is in a normal operation state; When T2 ≤ S < T3, the device is in a slight abnormal state and needs operation attention; When T3 ≤ S < T4, the device is in a developing fault state, and the system triggers a warning message; When S ≥ T4, the device is in a serious fault state, and the system triggers an emergency alarm message; 4.3) Dynamic threshold adaptive adjustment; In order to improve the adaptability of the warning system to different devices and different operating environments, introduce a dynamic threshold adjustment mechanism on the basis of the basic threshold system; by statistically analyzing the historical operation data of the device, calculate the mean and standard deviation of the voiceprint feature parameters: in: The threshold is dynamically adjusted. μ For historical data average, σ For standard deviation, k To adjust the coefficients, this method enables adaptive updating of the early warning threshold. 4.4) Real-time operation state evaluation; During operation, the system continuously collects acoustic signals from large oil-filled equipment and analyzes and processes them using a trained acoustic recognition model to obtain real-time acoustic feature parameters and discharge recognition results. Subsequently, the system compares the comprehensive state evaluation parameter S with a multi-level early warning threshold system to determine the current operating status of the equipment. The evaluation process includes: voiceprint data acquisition, voiceprint feature extraction, voiceprint recognition model inference, state evaluation parameter calculation, threshold comparison and state determination; 4.5) Early warning triggering mechanism; When the device status meets the warning triggering conditions, the system automatically generates a warning message of the corresponding level. The conditions for triggering an early warning include: 4.51) The comprehensive status evaluation parameters exceed the warning threshold; When S≥T3, the system triggers a device fault warning; 4.52) Continuous anomaly detection; When an abnormal state is detected within multiple consecutive time windows, the system triggers a trend warning. 4.53) The discharge type persists; When the probability of identifying a certain type of discharge is higher than the set threshold for an extended period, the system triggers a fault warning for the corresponding type. The warning information includes: device number, detection time, discharge type, fault level, and status assessment parameters. The system will send the warning information to the monitoring platform or operation and maintenance system to provide decision-making basis for operation and maintenance personnel. 4.6) Early warning information output; Depending on the warning level, the system generates different levels of warning information, including: alert information, attention information, warning information, and alarm information; By providing alerts to maintenance personnel through visual interfaces, alarm systems, or remote monitoring platforms, early warnings and proactive maintenance management of large oil-filled equipment failures can be achieved.

8. A method for early warning of faults in large oil-filled equipment according to claim 7, characterized in that, In step 5, after determining the equipment's operating status level and fault type, the system automatically matches the corresponding operation and maintenance strategy; the specific strategies are as follows: 5.1) Normal operation and maintenance strategy; When the equipment is in normal operation, the system records monitoring data and maintains a regular inspection cycle, while continuing online voiceprint monitoring in order to detect potential anomalies in a timely manner. 5.2) Focus on operational status and maintenance strategies; When the equipment is in a slightly abnormal state, the system recommends that maintenance personnel strengthen equipment operation monitoring and appropriately shorten the inspection cycle, while performing the following checks on the equipment: Check the equipment load operation, check the oil temperature and oil level, and check the cooling system operation status; The above methods are used to further confirm whether there are any potential abnormalities in the equipment; 5.3) Early warning operation status maintenance strategy; When equipment is in a state of progressive failure, the system recommends that maintenance personnel arrange a special inspection and take preventative maintenance measures, including: Conduct partial discharge detection or ultrasonic testing, check the quality and gas content of insulating oil, and check whether the winding and core structure is loose. If necessary, arrange for equipment to be shut down for maintenance to prevent the fault from worsening. 5.4) Operation and maintenance strategy for critical alarm status; When the equipment is in a critical malfunction state, the system recommends immediate emergency response measures, including: Reduce equipment operating load, start backup equipment or backup lines, and arrange emergency power outage maintenance; At the same time, a comprehensive fault diagnosis is performed on the equipment to prevent equipment damage or accidents from occurring.

9. An early warning system for faults in large oil-filled equipment, characterized in that, It employs a method for early warning of faults in large oil-filled equipment as described in any one of claims 1-8.

Images