Transformer health assessment method and system based on voiceprint recognition

Abstract

The invention discloses a transformer health assessment method and system based on voiceprint recognition, relates to the technical field of power equipment state monitoring, and constructs a non-intrusive assessment method by taking a voiceprint signal of a transformer as a data source of health assessment. A transformer voiceprint health recognition model is obtained through convolutional neural network-long short-term memory network combined modeling, the model is trained through voiceprint signal data generated in the operation process of transformers in different health states, the voiceprint signal data, collected in real time, in the operation process of the transformers are recognized, and real-time and accurate recognition and evaluation of the health states can be achieved. An adaptive noise cancellation algorithm removes environmental noise, and environmental noise interference factors are greatly reduced. And performing cross validation in combination with transformer operation parameters, and correcting a health state recognition result. And outputting a health assessment report containing a health state recognition result and a maintenance suggestion based on the recognition result, thereby providing effective support for operation and maintenance decision while realizing real-time, accurate and intelligent assessment of the health state of the transformer.

Description

Technical Field

[0001] This invention relates to the field of power equipment condition monitoring technology, and in particular to a transformer health assessment method and system based on voiceprint recognition. Background Technology

[0002] As a core piece of equipment in the power system, the health status of transformers directly affects the reliability and safety of the power grid operation. Currently, transformer health assessment mainly relies on traditional detection methods, including offline testing and periodic inspections. For example, dissolved gas analysis in oil requires on-site sampling and sending to a laboratory for analysis, with a testing cycle of several days or even weeks, making it impossible to achieve real-time fault detection and early warning. Manual inspections rely on the experience and judgment of maintenance personnel, which has problems such as strong subjectivity and limited coverage, especially in remote substations and high-altitude equipment, making it difficult to effectively cover these scenarios and easily leading to missed detection of early potential faults.

[0003] Although some online monitoring devices can collect operating parameters such as oil temperature and load current, these parameters are difficult to directly reflect the mechanical and electrical fault conditions inside the transformer. While a few acoustic fingerprint-based monitoring methods have been attempted, most only extract single types of acoustic fingerprint features, making them susceptible to environmental noise interference, resulting in a high false alarm rate. Furthermore, the models have poor generalization ability, making it difficult to distinguish between normal operation acoustic fingerprints and various fault acoustic fingerprints. In addition, existing methods lack effective linkage with operation and maintenance decision-making, and the assessment results often cannot directly guide on-site handling, leading to wasted operation and maintenance resources or delays in fault handling.

[0004] In summary, existing transformer health assessment methods suffer from technical problems such as poor real-time performance, low identification accuracy, susceptibility to environmental interference, and a disconnect between assessment results and operation and maintenance decisions. These methods are no longer sufficient to meet the power system's demand for "real-time, accurate, and intelligent" assessment of transformer health status. There is an urgent need for a transformer health assessment method that features non-intrusive monitoring, strong anti-interference capabilities, and the ability to integrate with the operation and maintenance system to solve the challenges of transformer health management. Summary of the Invention

[0005] To address the technical problems of existing transformer health assessment methods, such as poor real-time performance, low recognition accuracy, susceptibility to environmental interference, and disconnect between assessment results and operation and maintenance decisions, this invention proposes a transformer health assessment method and system based on voiceprint recognition. This method enables real-time, accurate, and intelligent assessment of transformer health status while providing effective support for operation and maintenance decisions.

[0006] The objective of this invention can be achieved through the following technical solutions: In a first aspect, this invention provides a transformer health assessment method based on voiceprint recognition, comprising: S1. Collect acoustic signature data generated during the operation of transformers in different health states. Perform preprocessing on the collected acoustic signature data, including noise reduction and normalization. Store and label the preprocessed acoustic signature data according to health state to obtain an acoustic signature sample library. S2. Based on the acoustic signature sample library, extract the time-domain and frequency-domain features of the acoustic signature to construct acoustic signature feature vectors. Divide the acoustic signature feature vectors into training and testing sets and input them into a convolutional neural network-long short-term memory network fusion model for training and testing, generating a transformer acoustic signature health recognition model. S3. Collect acoustic signature data of the transformer to be evaluated in real time. An adaptive noise cancellation algorithm is used to remove environmental noise, and effective voiceprint segments are extracted using endpoint detection technology. The feature extraction step in S2 is repeated to generate the voiceprint feature vector to be evaluated. In S4, the voiceprint feature vector to be evaluated is input into the voiceprint health recognition model that has been trained and tested. The model outputs the current health status category and confidence level of the transformer. If the confidence level meets the preset confidence level requirements, it is directly determined to be in a healthy state and an evaluation report is generated. If the confidence level does not meet the preset confidence level requirements, cross-validation is further performed in conjunction with the transformer operating parameters to correct the health status recognition result. Finally, a health evaluation report containing the health status recognition result and maintenance suggestions is output.

[0007] In a second aspect, this invention provides a transformer health assessment system based on voiceprint recognition, used to execute the method comprising: a transformer voiceprint sample library construction module, used to collect voiceprint signal data generated during the operation of transformers in different health states, perform preprocessing on the collected voiceprint signal data including noise reduction and normalization, and classify, store and label the preprocessed voiceprint signal data according to health state to obtain a voiceprint sample library; a voiceprint feature extraction and model building module, used to extract time-domain features, frequency-domain features, and time-frequency-domain features of the voiceprint based on the voiceprint sample library to construct a voiceprint feature vector, divide the voiceprint feature vector into training and testing sets and input them into a convolutional neural network-long short-term memory network fusion model for training and testing, generating a transformer voiceprint health recognition model; and a transformer to be assessed. The transformer acoustic signature acquisition and preprocessing module is used to acquire acoustic signature signal data of the transformer to be evaluated in real time. It uses an adaptive noise cancellation algorithm to remove environmental noise, extracts effective acoustic signature segments through endpoint detection technology, and repeats the feature extraction steps in S2 to generate the acoustic signature feature vector to be evaluated. The health status identification and evaluation module is used to input the acoustic signature feature vector to be evaluated into the acoustic signature health identification model that has been trained and tested, and outputs the current health status category and confidence level of the transformer. If the confidence level meets the preset confidence level requirements, it is directly determined to be in a healthy state and an evaluation report is generated. If the confidence level does not meet the preset confidence level requirements, it is further cross-validated with transformer operating parameters to correct the health status identification results. Finally, a health evaluation report containing the health status identification results and maintenance suggestions is output.

[0008] Compared with the prior art, the present invention can achieve the following beneficial effects: This invention uses the acoustic signature signal of a transformer as the data source for transformer health assessment, constructing a non-invasive health assessment method. Through joint modeling of a convolutional neural network and a long short-term memory network, a transformer acoustic signature health recognition model is established. Multi-dimensional features of the acoustic signature (time-domain features, frequency-domain features, and time-frequency-domain features) are extracted and fused to construct an acoustic signature feature vector. The model is trained using acoustic signature signal data generated during transformer operation at different health states, enabling real-time and accurate identification and assessment of different transformer health states. Furthermore, this invention employs an adaptive noise cancellation algorithm to remove environmental noise and uses endpoint detection technology to extract effective acoustic signature segments, significantly reducing environmental noise interference and further improving the accuracy of the assessment. The use of multi-dimensional feature fusion and adaptive noise cancellation technology significantly improves the signal-to-noise ratio of the acoustic signature signal, reduces the false positive rate, and effectively avoids missed fault detection and mishandling. Based on this, cross-validation is performed using transformer operating parameters to correct the health status identification results, further improving the accuracy of the identification results. Finally, based on the model recognition results, a health assessment report containing health status recognition results and maintenance suggestions is output. The assessment results are effectively linked with operation and maintenance decisions, ultimately achieving real-time, accurate, and intelligent assessment of transformer health status while providing effective support for operation and maintenance decisions. Attached Figure Description

[0009] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0010] Figure 1 A flowchart illustrating a transformer health assessment method based on voiceprint recognition provided in a specific embodiment of the present invention; Figure 2 This is an architecture diagram of a transformer health assessment system based on voiceprint recognition, provided for a specific embodiment of the present invention. Figure 3 A diagram illustrating the steps involved in constructing a transformer voiceprint sample library in a transformer health assessment method based on voiceprint recognition, as provided in another specific embodiment of the present invention. Figure 4 This diagram illustrates the voiceprint feature extraction steps in a transformer health assessment method based on voiceprint recognition, as provided in another specific embodiment of the present invention. Detailed Implementation

[0011] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0012] Please see Figure 1 In one optional implementation, a transformer health assessment method based on voiceprint recognition is provided, which is mainly achieved through the following steps: Construction of Transformer Acoustic Sample Library: S1. Acoustic signal data generated during the operation of transformers in different health states are collected. The collected acoustic signal data is preprocessed, including noise reduction and normalization. The preprocessed acoustic signal data is classified, stored and labeled according to health state to obtain the acoustic sample library. S2: Based on the voiceprint sample library, extract the time-domain features, frequency-domain features, and time-frequency-domain features of the voiceprint to construct the voiceprint feature vector. Divide the voiceprint feature vector into training set and test set and input them into the convolutional neural network-long short-term memory network fusion model for training and testing to generate the transformer voiceprint health recognition model. Acoustic fingerprint acquisition and preprocessing of the transformer to be evaluated: S3, real-time acquisition of acoustic fingerprint signal data of the transformer to be evaluated, removal of environmental noise by adaptive noise cancellation algorithm, extraction of effective acoustic fingerprint segments by endpoint detection technology, and repeating the feature extraction steps in S2 to generate the acoustic fingerprint feature vector to be evaluated. Health Status Identification and Assessment: S4. Input the voiceprint feature vector to be assessed into the voiceprint health recognition model that has been trained and tested, and output the current health status category and confidence level of the transformer. If the confidence level meets the preset confidence level requirements, it is directly determined to be in a healthy state and an assessment report is generated. If the confidence level does not meet the preset confidence level requirements, further cross-validation is performed in conjunction with the transformer operating parameters to correct the health status identification results. Finally, a health assessment report containing the health status identification results and maintenance suggestions is output.

[0013] Please see Figure 2 Corresponding to the above method, this embodiment also provides a transformer health assessment system based on voiceprint recognition for performing the above method, which mainly consists of the following parts: Transformer acoustic signature sample library construction module 1 is used to collect acoustic signature signal data generated during the operation of transformers in different health states, perform preprocessing including noise reduction and normalization on the collected acoustic signature signal data, and store and label the preprocessed acoustic signature signal data according to health state to obtain the acoustic signature sample library. The voiceprint feature extraction and model building module 2 is used to extract the time-domain features, frequency-domain features, and time-frequency-domain features of voiceprints based on the voiceprint sample library, construct voiceprint feature vectors, divide the voiceprint feature vectors into training sets and test sets, input them into the convolutional neural network-long short-term memory network fusion model for training and testing, and generate the transformer voiceprint health recognition model. The acoustic signature acquisition and preprocessing module 3 of the transformer to be evaluated is used to acquire the acoustic signature signal data of the transformer to be evaluated in real time, use an adaptive noise cancellation algorithm to remove environmental noise, use endpoint detection technology to extract effective acoustic signature segments, repeat the feature extraction steps in S2, and generate the acoustic signature feature vector to be evaluated. The health status identification and assessment module 4 is used to input the voiceprint feature vector to be evaluated into the voiceprint health identification model that has been trained and tested, and output the current health status category and confidence level of the transformer. If the confidence level meets the preset confidence level requirements, it is directly determined to be in a healthy state and an assessment report is generated. If the confidence level does not meet the preset confidence level requirements, it is further cross-validated in combination with the transformer operating parameters to correct the health status identification results. Finally, a health assessment report containing the health status identification results and maintenance suggestions is output.

[0014] This embodiment uses the acoustic signature signal of a transformer as the data source for transformer health assessment, constructing a non-invasive health assessment method. Through joint modeling of a convolutional neural network and a long short-term memory network, a transformer acoustic signature health recognition model is established. Multi-dimensional features of the acoustic signature (time-domain features, frequency-domain features, and time-frequency-domain features) are extracted and fused to construct an acoustic signature feature vector. The model is trained using acoustic signature signal data generated during transformer operation at different health states, enabling real-time and accurate identification and assessment of different transformer health states. Furthermore, this invention employs an adaptive noise cancellation algorithm to remove environmental noise and uses endpoint detection technology to extract effective acoustic signature segments, significantly reducing environmental noise interference and further improving the accuracy of the identification and assessment. The use of multi-dimensional feature fusion and adaptive noise cancellation technology significantly improves the signal-to-noise ratio of the acoustic signature signal, reduces the false positive rate, and effectively avoids missed fault detection and mishandling. Based on this, cross-validation is performed using transformer operating parameters to correct the health status identification results, further improving the accuracy of the identification results. Finally, based on the model recognition results, a health assessment report containing health status recognition results and maintenance suggestions is output. The assessment results are effectively linked with operation and maintenance decisions, ultimately achieving real-time, accurate, and intelligent assessment of transformer health status while providing effective support for operation and maintenance decisions.

[0015] Please see Figure 3 As shown, in a preferred embodiment, the process of constructing the S1 transformer acoustic signature sample library is as follows: S11. Determine the sample collection scenario, covering different voltage levels, different capacity transformers and different operating environments, including indoor substations and outdoor open-air substations; S12. A high-sensitivity microphone array is used to collect the acoustic signature of the transformer operation; S13. Preprocess the collected raw voiceprint signal data, including removing high-frequency noise using wavelet threshold denoising algorithm and mapping the voiceprint signal data amplitude to the required range (reasonable range) using linear normalization. S14. Use short-time Fourier transform to convert the preprocessed acoustic signature signal data into a time-frequency graph, and label the health status label and transformer basic parameters corresponding to each acoustic signature segment. The health status label includes normal, aging, partial discharge and core fault, and the transformer basic parameters include model, years of operation and load rate, thereby constructing a structured acoustic signature sample library.

[0016] This preferred implementation overcomes the problems of traditional sample libraries being "limited to a single scenario and of inconsistent data quality". By collecting samples in layers to cover multiple voltage levels and operating environments, it ensures the representativeness of the samples. At the same time, by combining wavelet denoising, normalization preprocessing and structured annotation, it not only improves the effectiveness of the data, but also provides rich label information for subsequent model training, avoiding insufficient generalization ability of the model due to poor sample quality.

[0017] Please see Figure 4 As shown, in a preferred embodiment, step S2, which involves extracting the time-domain features, frequency-domain features, and time-frequency-domain features of the voiceprint based on the voiceprint sample library to construct a voiceprint feature vector, is specifically implemented through the following steps, i.e., the voiceprint feature extraction process is as follows: S41. Temporal Feature Extraction: Calculate the short-time energy, zero-crossing rate, and peak factor of the voiceprint signal in the voiceprint sample library to obtain the temporal feature vector; where the short-time energy calculation sets the window length and step size, the zero-crossing rate counts the number of times the signal amplitude crosses zero points per unit time, and the peak factor is the ratio of the signal peak value to the effective value, reflecting the energy fluctuation and amplitude distribution of the voiceprint signal. S42. Frequency domain feature extraction: Perform fast Fourier transform on the voiceprint signals in the voiceprint sample library, calculate the power spectral density to obtain frequency distribution features, and extract Mel frequency cepstral coefficients and differential Mel frequency cepstral coefficients through Mel filter bank to characterize the frequency domain details of the voiceprint, thereby obtaining the frequency domain feature vector. S43. Time-frequency domain feature extraction: Multi-level wavelet packet decomposition of the voiceprint signal is performed using a specific wavelet basis. The wavelet packet entropy of each frequency band is calculated. The wavelet packet entropy reflects the signal complexity. Based on the wavelet packet entropy, the corresponding frequency band is selected to form the time-frequency domain feature vector. S44. Voiceprint Feature Vector Generation: The time-domain feature vector, frequency-domain feature vector, and time-frequency-domain feature vector are weighted and fused to generate a fixed-dimensional voiceprint feature vector for subsequent model training and health recognition.

[0018] This preferred implementation method abandons the limitations of traditional single feature extraction and adopts multi-dimensional feature fusion of "time domain + frequency domain + time-frequency domain" to fully capture the differences in acoustic signatures between normal and fault states of transformers. By optimizing feature effectiveness through weighted fusion, it avoids insufficient recognition accuracy caused by incomplete information from a single feature and provides more comprehensive feature support for model training.

[0019] In a preferred embodiment, step S2, which involves dividing the voiceprint feature vector into training and testing sets and inputting them into a convolutional neural network-long short-term memory network fusion model for training and testing, to generate the transformer voiceprint health recognition model, is specifically implemented through the following steps: The voiceprint health recognition model establishment process is as follows: A convolutional neural network-long short-term memory network fusion model is constructed. The convolutional neural network part includes multiple convolutional layers and max pooling layers. The convolutional layers are used to extract local spatial features of the voiceprint feature vector, and the max pooling layers are used to compress the feature dimension. The long short-term memory network part includes multiple bidirectional long short-term memory structures to capture the temporal dependencies of voiceprint features. The model output layer uses the Softmax activation function to output the probability distribution of various health states. The voiceprint feature vector is divided into a training set and a test set. The model training parameters are set, including configuring the learning rate and decay coefficient, establishing the cross-entropy loss function, and setting the training epochs and batch size. The training set is then input into the constructed model for training. The model performance is verified using a test set. If the performance parameters, including the model accuracy and recall, are all up to standard, the model training is complete. If the model accuracy or recall is not up to standard, the number of samples is increased, the feature weights are adjusted, or the model structure is optimized. Model structure optimization includes adding a Dropout layer to suppress overfitting, retraining until the performance parameters are up to standard, saving the optimal model parameters, and obtaining the transformer voiceprint health recognition model.

[0020] This preferred implementation overcomes the shortcomings of traditional single network models. The CNN-LSTM fusion model can simultaneously extract the spatial and temporal features of the voiceprint, improving the model's adaptability to complex working conditions. At the same time, through optimization techniques such as the Adam optimizer and Dropout layer, the model overfitting problem is solved, ensuring the recognition accuracy and stability of the model in different scenarios.

[0021] In a preferred embodiment, step S3, which involves real-time acquisition of the acoustic signature signal data of the transformer to be evaluated, removal of environmental noise using an adaptive noise cancellation algorithm, and extraction of valid acoustic signature segments using endpoint detection technology, is specifically implemented as follows: The acoustic signature acquisition and preprocessing process for the transformer to be evaluated is as follows: Real-time voiceprint acquisition: A microphone array is deployed at the transformer body to synchronously and in real time acquire multiple voiceprint signals, avoiding the microphone being directly facing the fan outlet or other noise sources; Environmental noise cancellation: An adaptive noise cancellation algorithm is adopted, using the ambient sound far away from the transformer as the reference noise, constructing an adaptive filter, setting the filter order and convergence factor, removing environmental interference from the collected real-time soundprint, and retaining the transformer's own operating soundprint. Effective voiceprint extraction: Using a dual-threshold endpoint detection algorithm, a first energy threshold and a second energy threshold (i.e., low and high energy thresholds) are set. The second energy threshold is higher than the first energy threshold. Segments with energy higher than the second energy threshold are marked as effective voiceprints. Segments with energy between the first and second energy thresholds and whose continuous duration meets the requirements are added as effective voiceprints. Silent segments with energy lower than the first energy threshold are removed. Finally, effective voiceprint segments of fixed duration are obtained for feature extraction.

[0022] This preferred embodiment addresses the problem of significant environmental noise interference by employing an adaptive noise cancellation algorithm to track and remove environmental noise in real time, resulting in improved noise suppression compared to traditional fixed filtering methods. Furthermore, it accurately extracts valid voiceprints through dual-threshold endpoint detection, avoiding the impact of silent or invalid segments on feature extraction, thus ensuring the quality of voiceprint data input to the model and further reducing the subsequent misjudgment rate.

[0023] In a preferred embodiment, for step S4, the voiceprint feature vector to be evaluated is input into the voiceprint health recognition model that has completed training and testing, and the current health status category and confidence level of the transformer are output. If the confidence level meets the preset confidence level requirements, it is directly determined to be in a healthy state and an evaluation report is generated. If the confidence level does not meet the preset confidence level requirements, cross-validation is further performed in conjunction with the transformer operating parameters to correct the health status recognition result. The specific steps are as follows, that is, the health status recognition process is as follows: The acoustic signature feature vector to be evaluated is input into the transformer acoustic signature health recognition model, which outputs four health states: normal operation, component aging, partial discharge, and core fault, along with their corresponding confidence levels. If the confidence level of a certain health state meets the preset confidence level requirement, the transformer is directly determined to be in that health state. If the confidence levels of all health states do not meet the preset confidence level requirements, retrieve the real-time operating parameters of the transformer, calculate the operating parameter matching confidence level, and re-determine whether the operating parameters are normal based on the operating parameter matching confidence level; if the operating parameters are normal, correct to normal operating state and adjust the confidence level; operating parameters include oil temperature, load current, and dissolved gas in oil; A parameter-based voiceprint cross-validation mechanism is constructed. If abnormal oil temperature or load current occurs and the voiceprint is biased towards component aging, it is corrected to component aging state and the confidence level is adjusted. If abnormal dissolved gas in oil occurs and the voiceprint is biased towards partial discharge or iron core failure state, it is corrected to partial discharge state or iron core failure state and the confidence level is adjusted. The risk level is marked according to the corresponding health status: normal operation is marked as low risk, component aging is marked as medium risk, and partial discharge and core failure are both marked as high risk.

[0024] In the above steps, the formula for calculating the confidence level of the running parameters is: In the formula, To match confidence levels to runtime parameters, M For the number of running parameters, For the first i One actual operating parameter value, For the first i Standard values ​​for each parameter For the first i The allowable deviation of each parameter.

[0025] This preferred implementation overcomes the limitation that it cannot make a judgment when the confidence of a single voiceprint recognition is insufficient. It introduces a cross-validation mechanism based on transformer operating parameters, and corrects the voiceprint recognition results by quantitatively calculating the confidence of the operating parameters and adjusting the confidence to a high confidence range, thus solving the judgment problem in ambiguous scenarios. At the same time, it divides risk levels according to health status, providing a clear priority basis for subsequent operation and maintenance, and avoiding delays in fault handling or waste of resources.

[0026] The final output of a health assessment report, which includes health status identification results and maintenance recommendations, is achieved through the following steps, i.e., the health assessment report generation process is as follows: The output includes a health assessment report containing basic information, health status identification results, feature comparison analysis, and maintenance recommendations. The basic information includes the transformer model, assessment time, and data collection location. The health status identification results include the health status type, confidence level, and risk level. The feature comparison analysis includes the feature differences between the voiceprint being assessed and the standard voiceprints in the sample library. Maintenance recommendations are given for different risk levels: for low-risk conditions, regular voiceprint reviews are recommended; for medium-risk conditions, short-term inspections of aging components are recommended, including winding insulation and the cooling system; and for high-risk conditions, immediate shutdown and maintenance are recommended, with a focus on detecting partial discharge areas or core grounding. The assessment report will be output in a visual format, including charts and text descriptions, and can be uploaded to the transformer operation and maintenance management platform and pushed to operation and maintenance personnel.

[0027] This preferred implementation overcomes the shortcomings of traditional assessment reports that only output results and lack decision support. The report includes maintenance recommendations based on feature comparison analysis and risk level matching, enabling a direct link between assessment results and maintenance actions. Through visual output, platform uploads, and SMS reminders, it ensures that maintenance personnel can quickly obtain key information, improve response efficiency, and solve the problems of slow delivery and weak guidance of traditional reports.

[0028] In a preferred embodiment, the transformer health assessment method based on voiceprint recognition further includes model updating and optimization: S5, periodically collecting new transformer voiceprint samples, including samples of newly added fault types or special operating conditions, supplementing them to the voiceprint sample library, and updating the voiceprint sample library; using an incremental learning algorithm, based on the updated voiceprint sample library, extracting new voiceprint feature vectors and inputting them into the trained model, updating only the parameters of the last layer of the model (avoiding full retraining to reduce computational costs); periodically verifying the model performance, if the model accuracy drops beyond a preset range, repeating steps S2-S4 to perform full optimization of the model. Model optimization ensures the model's ability to identify new operating conditions and fault types, maintaining assessment accuracy.

[0029] This preferred implementation breaks through the limitations of traditional models that remain unchanged after training and cannot adapt to new scenarios. It achieves lightweight model updates through incremental learning, avoiding the high computational cost of full training. Combined with regular performance verification and full optimization, it ensures that the model continuously iterates with new fault types and new working conditions, maintains high recognition accuracy in the long term, solves the problem of evaluation failure caused by model aging, and extends the life cycle and application value of the method.

[0030] Based on the above embodiments, the present invention can achieve the following beneficial technical effects: Abandoning the traditional approach of extracting only single features in the time or frequency domain, this method integrates three types of features: time domain, frequency domain, and time-frequency domain. It comprehensively covers the differences in acoustic signatures between normal and fault states of transformers, resulting in richer feature dimensions and laying a data foundation for accurate identification.

[0031] To avoid the limitation of a single model that can only capture spatial or temporal features, a fusion model is constructed that uses CNN to extract spatial frequency features and bidirectional LSTM to capture temporal dynamic features. This model can not only identify the specific frequency distribution of fault voiceprints, but also track the voiceprint change trend of fault development. The model's recognition accuracy is significantly improved compared to a single structure.

[0032] To address the issue of multiple environmental noise interferences in substations, a combined approach of adaptive noise cancellation algorithm and dual-threshold endpoint detection is adopted. Using ambient noise far from the transformer as reference noise, the filter weights are dynamically adjusted to suppress interference in real time. Then, effective acoustic text segments are extracted through energy thresholding to avoid invalid segments affecting feature quality and significantly reduce the false positive rate. Breaking away from the limitations of traditional sample libraries that only cover a single voltage level and a fixed environment, the sample library covers different voltage levels, different operating environments, and multiple health states, with sufficient sample size for each state. This ensures that the model still has high adaptability in complex power grid scenarios, and the recognition accuracy remains stable at a high level under different operating conditions.

[0033] To address scenarios where the confidence level of a single voiceprint recognition is insufficient, real-time operating parameters of the transformer are introduced for cross-validation. The confidence level of the operating parameters is quantitatively calculated, and a weighted fusion model of voiceprint and parameter confidence is constructed to correct low-confidence evaluation results to a high level, thus avoiding missed or false judgments.

[0034] Unlike traditional methods that only output binary results, this approach categorizes risks into three levels based on the identification results, provides targeted maintenance recommendations, and generates a visual assessment report. This report can be uploaded to the operations and maintenance platform and receive push notifications, enabling direct linkage between assessment results and operations and maintenance actions, thus significantly improving operations and maintenance response efficiency. To address the issue that traditional models cannot adapt to new fault types or special working conditions, incremental learning is adopted. When new samples are added, only the output layer parameters of the model are updated, without the need for full retraining, which greatly reduces the computational cost and enables the model to quickly adapt to new scenarios. Establish a model iteration mechanism with quarterly incremental updates and semi-annual full verification. If the verification finds that the accuracy drops beyond the preset range, the entire process optimization is re-executed to ensure that the model tracks the changes in the health status of the transformer in the long term, maintains the assessment accuracy, extends the model life cycle, and reduces long-term operation and maintenance costs. The entire process uses non-invasive voiceprint acquisition, with the microphone array positioned at a reasonable distance outside the transformer body. This eliminates the need to disassemble the equipment, extract oil samples, or conduct power outages for testing, avoiding interference with equipment operation caused by traditional offline testing. It enables 24-hour real-time monitoring, reducing the early fault detection cycle from days to minutes and minimizing power outage losses caused by fault escalation.

[0035] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for transformer health assessment based on voiceprint recognition, characterized in that, The method comprises the following steps: S1, collecting the voiceprint signal data generated in the operation process of transformers in different health states, preprocessing the collected voiceprint signal data, including noise reduction and normalization, storing and labeling the preprocessed voiceprint signal data according to health states, and obtaining a voiceprint sample library; S2, based on the voiceprint sample library, extracting time domain features, frequency domain features and time-frequency domain features of the voiceprint to construct a voiceprint feature vector, dividing the voiceprint feature vector into a training set and a test set, inputting the convolutional neural network-long short-term memory network fusion model for training and testing, and generating a transformer voiceprint health recognition model; S3, real-time collection of voiceprint signal data of the transformer to be evaluated, removal of environmental noise by using an adaptive noise cancellation algorithm, interception of effective voiceprint segments by using an endpoint detection technology, and repeated feature extraction steps in S2 to generate a voiceprint feature vector to be evaluated; S4, inputting the voiceprint feature vector to be evaluated into the voiceprint health recognition model, outputting the current health state category and confidence of the transformer, if the confidence meets the pre-set confidence requirement, directly determining the health state and generating an evaluation report, if the confidence does not meet the pre-set confidence requirement, further cross-verification is performed in combination with the transformer operation parameters, the health state recognition result is corrected, and finally a health evaluation report containing the health state recognition result and maintenance suggestions is output.

2. The voiceprint recognition based transformer health assessment method according to claim 1, wherein, S1 specifically comprises: determining a sample collection scene, covering different voltage levels, different capacity transformers and different operating environments, different operating environments including indoor substation and outdoor open-air substation; collecting transformer operation voiceprints by using a high-sensitivity microphone array; preprocessing the collected original voiceprint signal data, including removing high-frequency noise by using a wavelet threshold denoising algorithm, and mapping the voiceprint signal data amplitude to a required interval by using linear normalization; converting the preprocessed voiceprint signal data into a time-frequency graph by using a short-time Fourier transform, labeling the health state label and transformer basic parameters corresponding to each voiceprint, wherein the health state label includes normal, aging, partial discharge and core fault, and the transformer basic parameters include model, operation life and load rate, thereby constructing a structured voiceprint sample library.

3. The voiceprint recognition based transformer health assessment method according to claim 2, wherein, In S2, the step of extracting time domain features, frequency domain features and time-frequency domain features of the voiceprint to construct a voiceprint feature vector based on the voiceprint sample library comprises: calculating the short-time energy, zero-crossing rate and peak factor of the voiceprint signal in the voiceprint sample library to obtain a time domain feature vector; wherein the short-time energy calculation sets the window length and step, the zero-crossing rate counts the number of signal amplitude zero-crossing points per unit time, and the peak factor is the ratio of signal peak value to effective value, reflecting the energy fluctuation and amplitude distribution of the voiceprint signal; performing a fast Fourier transform on the voiceprint signal in the voiceprint sample library, calculating the power spectral density to obtain frequency distribution features, extracting mel frequency cepstral coefficients and differential mel frequency cepstral coefficients by using a mel filter bank to characterize the frequency domain detail information of the voiceprint, and obtaining a frequency domain feature vector; The voiceprint signal is decomposed by a specific wavelet base to obtain a wavelet packet entropy of each frequency band, and the wavelet packet entropy reflects the complexity of the signal, and a time-frequency domain feature vector is constructed based on the wavelet packet entropy; The time domain feature vector, the frequency domain feature vector, and the time-frequency domain feature vector are fused by weighting to generate a fixed-dimensional voiceprint feature vector, which is used for subsequent model training and health identification.

4. The voiceprint recognition based transformer health assessment method of claim 1, wherein, In S2, the step of dividing the voiceprint feature vector into a training set and a test set and inputting the voiceprint feature vector into a convolutional neural network-long short-term memory network fusion model for training and testing to generate a transformer voiceprint health identification model includes: A convolutional neural network-long short-term memory network fusion model is constructed, the convolutional neural network part includes multiple convolutional layers and max-pooling layers, wherein the convolutional layers are used to extract local spatial features of the voiceprint feature vector, and the max-pooling layers are used to compress the feature dimension; the long short-term memory network part includes multiple bidirectional long short-term memory structures, which are used to capture the time sequence dependence of the voiceprint feature; the model output layer adopts a Softmax activation function to output the probability distribution of each health state; The voiceprint feature vector is divided into a training set and a test set, and the model training parameters are set, including configuring the learning rate and the attenuation coefficient, establishing the cross-entropy loss function, setting the training rounds and the batch size; the training set is input into the constructed model for training; The performance of the model is verified using the test set, and if the performance parameters including the model accuracy and the recall rate meet the standards, the model training is completed; if the model accuracy or the recall rate does not meet the standards, the sample size is increased, the feature weight is adjusted, or the model structure is optimized; the model structure optimization includes adding a Dropout layer to suppress overfitting, retraining until the performance parameters meet the standards, saving the optimal model parameters, and obtaining the transformer voiceprint health identification model.

5. The voiceprint recognition based transformer health assessment method of claim 1, wherein, In S3, the step of collecting the voiceprint signal data of the transformer to be evaluated in real time, removing the environmental noise by using an adaptive noise cancellation algorithm, and intercepting the effective voiceprint segment by using an endpoint detection technology includes: A microphone array is arranged at the transformer body to synchronously collect multiple voiceprint signals in real time, avoiding the microphone being directly opposite to the fan outlet or other noise sources; An adaptive noise cancellation algorithm is used to remove the environmental interference in the collected real-time voiceprint and retain the transformer body operation voiceprint by using the environmental sound far from the transformer as the reference noise, constructing an adaptive filter, and setting the filter order and the convergence factor. A double-threshold endpoint detection algorithm is used to set a first energy threshold and a second energy threshold, the second energy threshold is higher than the first energy threshold, a segment with energy higher than the second energy threshold is marked as an effective voiceprint, a segment with energy between the first energy threshold and the second energy threshold and a continuous time length meeting the requirements is supplemented as an effective voiceprint, and a mute segment with energy lower than the first energy threshold is removed, so that a fixed-length effective voiceprint segment is obtained for feature extraction.

6. The voiceprint recognition based transformer health assessment method of claim 1, wherein, In S4, the voiceprint feature vector to be evaluated is input into the trained and tested voiceprint health recognition model, and the current health state category and confidence of the transformer are output. If the confidence meets the pre-set confidence requirement, the health state is directly determined and an evaluation report is generated. If the confidence does not meet the pre-set confidence requirement, further cross-validation is performed in combination with the transformer operating parameters to correct the health state recognition result, which includes: The voiceprint feature vector to be evaluated is input into the transformer voiceprint health recognition model, and the four health states of normal operation, component aging, partial discharge and core fault and the corresponding confidence are output. If the confidence of a certain health state meets the pre-set confidence requirement, it is directly determined that the transformer is currently in that health state. If the confidence of all health states does not meet the pre-set confidence requirement, the real-time operating parameters of the transformer are retrieved, the operating parameter matching confidence is calculated, and whether the operating parameters are normal is re-determined according to the operating parameter matching confidence. If the operating parameters are normal, the normal operation state is corrected and the confidence is corrected. The operating parameters include oil temperature, load current and dissolved gas in oil. A parameter voiceprint cross-validation mechanism is constructed. If the oil temperature or load current is abnormal and the voiceprint is biased towards component aging, the component aging state is corrected and the confidence is corrected. If the dissolved gas in oil is abnormal and the voiceprint is biased towards partial discharge or core fault, the partial discharge state or core fault state is corrected and the confidence is corrected. According to the corresponding health state, the risk level is marked. The normal operation state is marked as low risk, the component aging state is marked as medium risk, and the partial discharge and core fault states are both marked as high risk.

7. The voiceprint recognition based transformer health assessment method according to claim 6, characterized in that, The calculation formula of the operating parameter matching confidence is: In the formula, To match confidence levels to runtime parameters, M For the number of running parameters, For the first i One actual operating parameter value, For the first i Standard values ​​for each parameter For the first i The allowable deviation of each parameter.

8. The voiceprint recognition based transformer health assessment method of claim 6, wherein, In S4, the health evaluation report finally output includes the health state recognition result and maintenance suggestions, which includes: The health evaluation report including basic information, health state recognition result, feature comparison analysis and maintenance suggestions is output. The basic information includes transformer model, evaluation time and collection position. The health state recognition result includes health state type, confidence and risk level. The feature comparison analysis includes the feature difference between the voiceprint to be evaluated and the standard voiceprint in the sample library. Maintenance suggestions are given for different risk levels. Low-risk state suggests regular review of voiceprint, medium-risk state suggests checking aging components in the short term, aging components include winding insulation and cooling system, and high-risk state suggests immediate shutdown for maintenance, focusing on detecting partial discharge parts or core grounding conditions. The evaluation report is output in a visual form, including charts and text explanations, and can be uploaded to the transformer operation and maintenance management platform and pushed to the operation and maintenance personnel.

9. The voice-print recognition based transformer health assessment method as claimed in claim 1, wherein, It also includes: S5、 New transformer voiceprint samples are collected regularly, including new fault types or special working condition samples, which are supplemented to the voiceprint sample library and the voiceprint sample library is updated. An incremental learning algorithm is used to extract new voiceprint feature vectors based on the updated voiceprint sample library and input them into the trained model to update only the last layer parameters of the model. The model performance is verified regularly. If the model accuracy decreases by more than a pre-set amplitude, repeat steps S2-S4 to optimize the model.

10. A voiceprint recognition based transformer health assessment system for performing the method of any one of the preceding claims 1 to 9, characterized in that, It includes: The transformer voiceprint sample library construction module is configured to collect voiceprint signal data generated in the operation of transformers in different health states, pre-process the collected voiceprint signal data, including noise reduction and normalization, store and label the pre-processed voiceprint signal data according to health states, and obtain a voiceprint sample library; The voiceprint feature extraction and model establishment module is configured to extract time-domain features, frequency-domain features and time-frequency domain features of voiceprints based on the voiceprint sample library to construct a voiceprint feature vector, divide the voiceprint feature vector into a training set and a test set, input the voiceprint feature vector into a convolutional neural network-long short-term memory network fusion model for training and testing, and generate a transformer voiceprint health recognition model; The voiceprint feature extraction and model establishment module is configured to extract time-domain features, frequency-domain features and time-frequency domain features of voiceprints based on the voiceprint sample library to construct a voiceprint feature vector, divide the voiceprint feature vector into a training set and a test set, input the voiceprint feature vector into a convolutional neural network-long short-term memory network fusion model for training and testing, and generate a transformer voiceprint health recognition model; The health state recognition and evaluation module is configured to input the voiceprint feature vector to be evaluated into the voiceprint health recognition model trained and tested, output the current health state category and confidence of the transformer, and if the confidence meets the pre-set confidence requirement, directly determine the health state and generate an evaluation report, if the confidence does not meet the pre-set confidence requirement, further cross-verify in combination with the transformer operation parameters, correct the health state recognition result, and finally output a health evaluation report containing the health state recognition result and maintenance suggestions.

Images