Transformer fault sound detection and early warning method

Abstract

The invention discloses a transformer fault sound detection and early warning method, and belongs to the technical field of transformer detection, and the method comprises the steps: collecting the working sound data of a transformer in a fixed time period, including frequency, amplitude and duration; according to the working sound data of the transformer, analysis is carried out based on a pre-trained fault sound detection model, a fault type is output, and fault early warning is carried out; wherein a layered attention mechanism, residual connection and a Mish activation function are introduced to improve the time delay neural network, and the fault sound detection model is constructed based on the improved time delay neural network and a large language model. According to the invention, the problem that the fault type of the transformer cannot be accurately and timely identified and automatic early warning cannot be carried out in the prior art is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of transformer detection, and particularly to a method for detecting and warning transformer faults by sound. Background Art

[0002] In the current power system, as one of the key devices, the operating state of a transformer directly affects the stability and safety of the power grid. To monitor the operating state of the transformer, traditionally, manual inspections and regular detections are mainly adopted. Manual inspections usually rely on the experience and auditory judgment of operation and maintenance personnel, and initially judge whether there are abnormalities by listening to the sound of the transformer during operation. Regular detections include using professional detection instruments to test various performance indicators of the transformer, such as oil chromatographic analysis, insulation resistance measurement, etc. These methods can, to a certain extent, discover potential faults of the transformer and provide a basis for maintenance for operation and maintenance personnel. With the development of technology, some transformer fault detection methods based on signal processing technology have gradually been applied. These methods collect signals such as the sound and vibration of the transformer during operation, and use signal processing algorithms to analyze and process the signals, so as to extract fault features and then judge the fault type of the transformer. The emergence of these methods has improved the accuracy and efficiency of transformer fault detection and provided better protection for the safe operation of the transformer.

[0003] However, there are still some obvious defects in the existing technology for transformer fault detection. First, the methods of manual inspection and regular detection have subjectivity and lag. Manual inspection relies on the experience and auditory judgment of operation and maintenance personnel, is easily affected by human factors, and cannot real-time monitor the operating state of the transformer. Although regular detection can provide more accurate detection results, the detection period is long and sudden faults of the transformer cannot be discovered in time. Secondly, although the transformer fault detection methods based on signal processing technology have improved the accuracy and efficiency of detection to a certain extent, there are still some limitations. These methods usually require complex preprocessing and feature extraction of signals, and the recognition accuracy rate of fault types is limited by the performance of algorithms and models. At the same time, these methods can often only identify known fault types and may not be able to accurately identify unknown fault types. In addition, these methods also have deficiencies in automatic warning, and cannot give timely and effective warnings according to the severity and urgency of fault types, bringing potential safety hazards to the safe operation of the transformer. Summary of the Invention

[0004] The technical problem solved by the present invention is: the problem that the fault type of the transformer cannot be accurately and timely identified and automatically warned in the prior art.

[0005] To solve the above technical problem, the present invention provides the following technical solutions:

[0006] As a preferred solution of the transformer fault sound detection and early warning method described in the present invention, it includes:

[0007] Collect the transformer operating sound data within a fixed time period, including frequency, amplitude, and duration;

[0008] Based on the transformer operating sound data, analyze it using a pre-trained fault sound detection model, output the fault type, and conduct fault early warning;

[0009] Among them, a hierarchical attention mechanism, residual connection, and Mish activation function are introduced to improve the time-delay neural network. The fault sound detection model is constructed based on the improved time-delay neural network and large language model.

[0010] Furthermore, the data processing method of the fault sound detection model includes:

[0011] Separate the transformer operating sound data to obtain the transformer body noise and fault sound;

[0012] Extract the frequency, amplitude, and duration of the fault sound to obtain fault features;

[0013] Based on the fault features and the improved time-delay neural network, output the fault type;

[0014] According to the fault type, analyze it using a large language model and output the priority of the fault type;

[0015] Send a fault signal according to the priority of the fault type to conduct fault early warning.

[0016] Furthermore, the improved time-delay neural network includes:

[0017] An input layer for receiving fault features;

[0018] A hierarchical attention mechanism layer, including bottom-layer attention and top-layer attention. The bottom-layer attention is used to encode the fault features, capture local correlations, and obtain the bottom-layer encoded representation. The top-layer attention is used to further process the bottom-layer encoded representation to capture global semantic relationships;

[0019] A residual connection layer for adding a residual connection between the convolutional layer and the recurrent layer to alleviate the problem of gradient disappearance;

[0020] A Mish activation function layer for replacing the traditional activation function with the Mish activation function after each neuron to improve the training efficiency and generalization ability;

[0021] An output layer for outputting the fault type.

[0022] Further, according to the blind source separation algorithm, separate the transformer operating sound data to obtain the transformer body noise and fault sounds, including:

[0023] Based on the blind source separation algorithm, according to the statistical characteristics and independence principle of the sound signal, perform the separation process and output multiple independent sound components;

[0024] Perform spectral analysis on the multiple independent sound components, calculate the spectrogram, and extract the frequencies of the multiple independent sound components according to the frequency distribution and spectral characteristics;

[0025] Perform time-domain analysis on the multiple independent sound components, calculate the power spectral density, and extract the amplitudes of the multiple independent sound components according to the waveform changes;

[0026] According to the start and end points of the fault sound, extract the duration of the fault sound on the time axis

[0027] Perform on the multiple independent sound components to extract the durations of the multiple independent sound components;

[0028] According to the frequencies, amplitudes, and durations of the multiple independent sound components, obtain the transformer body noise and fault sounds.

[0029] Further, the method for extracting the frequencies, amplitudes, and durations of the fault sound to obtain fault characteristics includes:

[0030] Extract the fault sound with a frequency within a preset frequency range, an amplitude increment exceeding a preset first threshold, and a duration within a preset second threshold range as the fault characteristic.

[0031] Further, the training method for the improved time-delay neural network includes:

[0032] Obtain a large amount of sound data of transformers in normal and fault states within a historical time period, including frequencies, amplitudes, and durations;

[0033] The sound data in the fault state includes sound samples of different fault types;

[0034] Label the sound data in normal and fault states, and further label the fault types of the sound data in the fault state to obtain a labeled data set;

[0035] Separate the transformer body noise and fault sounds in the labeled data set according to the blind source separation algorithm;

[0036] Extract the frequencies, amplitudes, and durations of the fault sound to obtain fault characteristics;

[0037] Use the fault characteristics and fault types as the input labels and output labels of the training set, and train the improved time-delay neural network through the first loss function.

[0038] Further, the first loss function is expressed as:

[0039]

[0040] In the formula, represents the value of the first loss function, that is, the difference between the predicted value of the improved time-delay neural network and the true label. N represents the total number of samples, that is, the number of samples in the training set, and y represents the corresponding element of the true label of the i-th sample, i represents the corresponding element of the predicted value of the i-th sample, and log(·) represents the logarithmic function.

[0041] Further, the training method of the large language model includes:

[0042] Obtain transformer fault type data including fault descriptions, repair records, impact ranges, and durations;

[0043] After performing word segmentation on the transformer fault type data, extract the frequency, amplitude, and duration of the fault sound to obtain fault characteristics;

[0044] Label the fault types according to the severity, impact range, and repair difficulty of the fault characteristics to obtain a labeled data set;

[0045] According to the labeled data set, learn the correlation between fault descriptions, repair records, and fault types, and set the priority criteria for fault types. The priority criteria are used to reflect the urgency of fault types;

[0046] Quantify the priority criteria into priority values;

[0047] Use the fault type and the corresponding priority value as the input label and output label of the training set, and train the large language model through the second loss function to learn the mapping relationship between the fault type and the priority, and output the priority of the fault type.

[0048] Further, the second loss function is expressed as:

[0049]

[0050] In the formula, L 2 represents the value of the second loss function, that is, the difference between the predicted value of the large language model and the true label. N represents the total number of samples, that is, the number of samples in the training set, and w i ​$w_i$ represents the weight set according to the severity level of the $i$-th sample for different fault types, $C$ represents the total number of fault types, $y$ ij represents the priority value of the $j$-th fault type in the true label of the $i$-th sample, and $p_{ij}$ represents the probability that the $i$-th sample belongs to the $j$-th fault type with respect to the priority value.

[0051] Further, sending a fault signal according to the priority of the fault type for fault early warning includes:

[0052] Setting an early warning threshold according to the priority standard of the fault type, where 0 indicates no fault, 1 indicates a severe fault, 2 indicates a moderate fault, and 3 indicates a general fault;

[0053] When the priority of the fault type is greater than 0, triggering the fault signal corresponding to the early warning threshold and sending a fault early warning message including the fault type and the priority corresponding to the fault type.

[0054] Advantages of the present invention:

[0055] 1. By introducing a hierarchical attention mechanism, residual connections, and the Mish activation function to improve the time-delay neural network, and combining with a large language model to construct a fault sound detection model, the present invention can more accurately analyze transformer working sound data including frequency, amplitude, and duration, thereby effectively improving the recognition accuracy of fault types to achieve automatic early warning, providing stronger protection for the safe operation of transformers, and solving the problem in the prior art that the fault types of transformers cannot be accurately and timely identified and automatically warned;

[0056] 2. The present invention can also prioritize fault types using a large language model according to the severity, influence range, and repair difficulty of fault characteristics. This precise identification and prioritization provide clearer fault handling guidelines for maintenance personnel, helping to quickly respond to and prioritize critical faults.

[0057] 3. The present invention uses a blind source separation algorithm to separate the transformer body noise and fault sounds, extracts the frequency, amplitude, and duration of the fault sounds as fault characteristics, constructs the input labels of the training set, and combines with the first loss function to effectively train the improved time-delay neural network. At the same time, the Mish activation function is used to replace the traditional activation function, improving the training efficiency and generalization ability, so that the model can maintain a high recognition accuracy and stability when facing complex and variable transformer fault sounds;

[0058] 4. The present invention learns the correlation between fault descriptions, maintenance records, and fault types through a large language model and sets the priority criteria for fault types. During fault detection, corresponding fault signals are issued according to the priority of the fault type, and fault warnings are given. This not only provides immediate fault information but also gives warning levels according to the urgency of the fault, helping operation and maintenance personnel make quick decisions and reducing the impact of faults on the power system. Description of the Drawings

[0059] Figure 1 It is a schematic diagram of the basic process of a transformer fault sound detection and warning method provided by an embodiment of the present invention;

[0060] Figure 2 It is a schematic diagram of the improved method process of a time-delay neural network provided by an embodiment of the present invention. Detailed Embodiments

[0061] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following provides a detailed description of the specific embodiments of the present invention in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all embodiments.

[0062] Embodiment 1, referring to Figure 1 , which is an embodiment of the present invention, provides a transformer fault sound detection and warning method, including:

[0063] Step 1: Collect the working sound data of the transformer within a fixed time period, including frequency, amplitude, and duration.

[0064] In this embodiment, according to the operating characteristics and monitoring requirements of the transformer, considering the load changes, operating environment, and possible fault modes of the transformer, a representative time period is selected, such as the peak load period, the low load period, or the period under specific weather conditions, to collect the working sound data of the transformer.

[0065] In this embodiment, according to the requirements of the collected parameters, a suitable sound collection device is selected, such as a microphone, a sound sensor, or a professional sound data collection system, to ensure that the collection device has sufficient sensitivity and accuracy to accurately capture and record the sound information of the transformer during operation.

[0066] In this embodiment, the collection points are reasonably arranged around the transformer to ensure that the sound data of the transformer can be comprehensively and accurately collected. By considering the structure of the transformer, the sound propagation path, and possible interference factors, the best collection point positions are selected.

[0067] In this embodiment, according to the set acquisition parameters and acquisition point positions, the sound data acquisition device is started to collect the working sound data of the transformer within a fixed time period. At the same time, the acquisition process is monitored to ensure the smooth progress of data acquisition and to handle possible abnormal situations in a timely manner.

[0068] This embodiment also preprocesses the collected sound data, including denoising, filtering, etc., to improve the quality and accuracy of the data. The preprocessed sound data is sorted and stored according to parameters such as frequency, amplitude, and duration for subsequent analysis and processing.

[0069] Step 2: Analyze based on the pre-trained fault sound detection model according to the transformer working sound data, output the fault type, and issue a fault warning.

[0070] In this embodiment, the time-delay neural network is improved by introducing a hierarchical attention mechanism, residual connections, and the Mish activation function. The fault sound detection model is constructed based on the improved time-delay neural network and the large language model, as Figure 2 shown.

[0071] In this embodiment, a hierarchical attention mechanism is introduced into the time-delay neural network to better capture the key features and patterns in the sound signal. By setting up multiple layers of attention modules, the attention modules are integrated into each layer of the time-delay neural network to form a hierarchical attention structure. Each layer of the attention structure focuses on the frequency, amplitude, and duration of the sound signal respectively, and automatically learns the attention weights through the training process, enabling the model to dynamically adjust the attention to different features.

[0072] In this embodiment, the problem of gradient disappearance or gradient explosion in the deep neural network is alleviated through residual connections, improving the training stability and performance of the model. Residual connections are added between adjacent layers of the time-delay neural network, that is, the output of the previous layer is directly added to the input of the next layer. This structure allows information to flow directly between layers, helping to retain important features and information.

[0073] The Mish activation function is a smooth, non-linear, and almost everywhere differentiable activation function that can provide better gradient flow and expression ability. In the hidden layer of the time-delay neural network in this embodiment, the traditional ReLU or other activation functions are replaced with the Mish activation function.

[0074] Step 2.1: Construct and train the improved time-delay neural network and the large language model.

[0075] In this embodiment, the improved time-delay neural network includes:

[0076] An input layer for receiving fault features.

[0077] In this embodiment, the input layer is responsible for receiving the fault features extracted from the preprocessed working sound data of the transformer, including frequency, amplitude, and duration. Fault features can also be extracted by other methods such as Mel Frequency Cepstral Coefficients (MFCC).

[0078] The hierarchical attention mechanism layer includes bottom-layer attention and top-layer attention. The bottom-layer attention is used to encode the fault features, capture local correlations, and obtain the bottom-layer encoded representation. The top-layer attention is used to further process the bottom-layer encoded representation and capture global semantic relationships.

[0079] In this embodiment, the bottom-layer attention mechanism assigns a weight to each feature by calculating the similarity or correlation between features. The weight is used to reflect the importance of the fault feature in the local context. Then, these weights are used to perform weighted summation on the fault features to obtain the bottom-layer encoded representation.

[0080] In this embodiment, the top-layer attention mechanism further captures the global semantic relationships between features based on the bottom-layer encoded representation. This is achieved by calculating the similarity or correlation between the bottom-layer encoded representations, then assigning a weight to each bottom-layer encoded representation, and finally using these weights for weighted summation to obtain the global semantic representation.

[0081] The residual connection layer is used to add a residual connection between the convolutional layer and the recurrent layer to alleviate the problem of gradient vanishing.

[0082] The residual connection layer allows information to flow directly between layers, which helps to retain important features and information. In this embodiment, the output of the previous layer is directly added to the input of the next layer to form a skip connection, which helps to alleviate the problem of gradient vanishing in deep neural networks and improve the training stability and performance of the model.

[0083] The Mish activation function layer is used to replace the traditional activation function with the Mish activation function after each neuron to improve the training efficiency and generalization ability.

[0084] The Mish activation function is a smooth, non-linear, and almost everywhere differentiable activation function that can provide better gradient flow and expressive power. In this embodiment, using the Mish activation function after each neuron can replace the traditional ReLU or other activation functions, thereby improving the training efficiency and generalization ability of the model.

[0085] The output layer is used to output the fault type.

[0086] In this embodiment, the output layer is responsible for mapping the learned fault feature representation to specific fault types. This embodiment is implemented through a fully connected layer or a softmax layer. The fully connected layer is responsible for converting the fault feature representation into fault type probabilities, while the softmax layer is responsible for normalizing the fault type probabilities into a probability distribution. The improved time-delay neural network takes the fault type with the highest output probability as the prediction result.

[0087] In this embodiment, the training method of the improved time-delay neural network includes:

[0088] Step 2.1.1: Obtain a large amount of sound data of transformers in normal and fault states within a historical time period, including frequency, amplitude, and duration.

[0089] The sound data in the fault state includes sound samples of different fault types.

[0090] Step 2.1.2: Label the sound data in normal and fault states, and further label the fault types of the sound data in the fault state to obtain a labeled data set.

[0091] In this embodiment, each sound sample in the labeled data set has a corresponding label, that is, a normal state label or a specific fault type label.

[0092] Step 2.1.3: Separate the transformer body noise and fault sounds in the labeled data set according to the blind source separation algorithm.

[0093] Step 2.1.4: Extract the frequency, amplitude, and duration of the fault sounds to obtain fault features;

[0094] Use the fault features and fault types as the input labels and output labels of the training set, and train the improved time-delay neural network through the first loss function.

[0095] In this embodiment, the first loss function is expressed as:

[0096]

[0097] where represents the value of the first loss function, that is, the difference between the predicted value of the improved time-delay neural network and the true label y. N represents the total number of samples, that is, the number of samples in the training set. y i represents the corresponding element of the true label of the i-th sample, represents the corresponding element of the predicted value of the i-th sample, and log(·) represents the logarithmic function.

[0098] This embodiment aims to train a model that can identify different fault types, namely the improved time-delay neural network, by collecting, annotating, and analyzing the sound data of transformers. The use of blind source separation technology helps to more accurately extract fault features, while the improved time-delay neural network is used to effectively learn and identify the relationship between fault features and fault types. This embodiment has potential application value in the field of transformer fault diagnosis and can improve the accuracy and efficiency of fault diagnosis.

[0099] This embodiment also uses large language models to perform semantic understanding and context modeling on the sound data to improve the accuracy of fault sound detection. This embodiment takes the output of the improved time-delay neural network as the input of the large language model. The large language model adopted in this embodiment is a pre-trained model based on Transformer, including BERT, GPT, etc. Through the joint training or fine-tuning process, the time-delay neural network and the large language model work together to jointly improve the performance of fault sound detection.

[0100] This embodiment integrates the improved time-delay neural network and the large language model to form a fault sound detection model. The fault sound detection model is tested and evaluated by using an independent test data set to verify its performance and accuracy in practical applications. The fault sound detection model is optimized and adjusted according to the test results and finally deployed to the actual transformer fault sound detection system.

[0101] In this embodiment, the training method of the large language model includes:

[0102] Step 2.2.1: Obtain transformer fault type data including fault descriptions, repair records, impact ranges, and durations.

[0103] Step 2.2.2: After performing word segmentation on the transformer fault type data, extract the frequency, amplitude, and duration of the fault sound to obtain fault features.

[0104] Step 2.2.3: Label the fault types according to the severity, impact range, and repair difficulty of the fault features to obtain the labeled data set.

[0105] Step 2.2.4: According to the labeled data set, learn the association relationship between fault descriptions, repair records, and fault types, and set the priority criteria for fault types. The priority criteria are used to reflect the urgency of fault types.

[0106] Step 2.2.5: Quantify the priority criteria into priority values.

[0107] Step 2.2.6: Use the fault type and the corresponding priority value of the fault type as the input label and output label of the training set, and train the large language model through the second loss function to learn the mapping relationship between the fault type and the priority, and output the priority of the fault type.

[0108] In this embodiment, the second loss function is expressed as:

[0109]

[0110] In the formula, L 2 represents the value of the second loss function, that is, the difference between the predicted value of the large language model and the true label. N represents the total number of samples, that is, the number of samples in the training set. w i represents the weight set according to the urgency of the i-th sample's fault type. C represents the total number of fault types. y ij represents the priority value of the j-th fault type in the true label of the i-th sample. represents the probability that the i-th sample belongs to the priority value of the j-th fault type.

[0111] Step 2.3: Use the fault sound detection model for data processing.

[0112] Step 2.3.1: Separate the transformer working sound data to obtain the transformer body noise and the fault sound.

[0113] In some embodiments, according to the blind source separation algorithm, separating the transformer working sound data to obtain the transformer body noise and the fault sound includes:

[0114] Based on the blind source separation algorithm, according to the statistical characteristics and independence principle of the sound signal, perform the separation process and output multiple independent sound components;

[0115] Perform spectral analysis on the multiple independent sound components, calculate the spectrogram, and extract the frequencies of the multiple independent sound components according to the frequency distribution and spectral characteristics;

[0116] Perform time-domain analysis on the multiple independent sound components, calculate the power spectral density, and extract the amplitudes of the multiple independent sound components according to the waveform changes;

[0117] According to the start and end points of the fault sound, extract the duration of the fault sound on the time axis

[0118] Perform on the multiple independent sound components to extract the durations of the multiple independent sound components;

[0119] According to the frequencies, amplitudes and durations of the multiple independent sound components, obtain the transformer body noise and the fault sound.

[0120] Step 2.13.2: Extract the frequency, amplitude, and duration of the fault sound to obtain fault characteristics.

[0121] In some embodiments, the extracting the frequency, amplitude, and duration of the fault sound to obtain fault characteristics includes:

[0122] Extract the fault sound with a frequency within a preset frequency range, an amplitude increment exceeding a preset first threshold, and a duration within a preset second threshold range as fault characteristics.

[0123] In this embodiment, the preset frequency range is 500Hz - 1000Hz, and the preset second threshold is 2 seconds.

[0124] Step 2.3.3: Output the fault type according to the fault characteristics and the improved time-delay neural network.

[0125] Step 2.3.4: Analyze according to the fault type using a large language model and output the priority of the fault type;

[0126] Step 2.3.5: Send a fault signal according to the priority of the fault type for fault warning.

[0127] In some embodiments, sending a fault signal according to the priority of the fault type for fault warning includes:

[0128] Set a warning threshold according to the priority standard of the fault type, where 0 indicates no fault, 1 indicates a serious fault, 2 indicates a moderate fault, and 3 indicates a general fault.

[0129] In this embodiment, the warning threshold corresponds to different fault priorities, specifically as follows:

[0130] 0: Indicates a no-fault state and the system is running normally.

[0131] 1: Indicates a serious fault, which will cause the system to stop or a serious performance degradation.

[0132] 2: Indicates a moderate fault, which will have a certain impact on the system performance but usually will not cause the system to stop.

[0133] 3: Indicates a general fault, which will cause a slight system performance degradation or user experience problems.

[0134] According to the priority of the fault, the system will select the corresponding warning threshold and generate the corresponding fault signal

[0135] When the priority of the fault type is greater than 0, that is, when it is not in a no-fault state, trigger the fault signal corresponding to the warning threshold and send a fault warning message including the fault type and the priority corresponding to the fault type.

[0136] In this embodiment, the generated fault signal is encapsulated into a fault warning message and sent to the relevant system administrators or maintenance personnel. The fault warning message includes the fault type and the priority corresponding to the fault type, so that the receiver can quickly understand the nature and severity of the fault. Through this fault warning mechanism, a warning can be issued in a timely manner when a fault is detected, reminding the relevant personnel to pay attention and take corresponding maintenance measures. This not only helps to reduce the impact of system faults on business operations, but also improves the reliability and stability of the system.

[0137] Embodiment 2 is another embodiment of the present invention. The difference between this embodiment and the first embodiment is that it provides an experimental verification of the transformer fault sound detection and warning method. To verify and illustrate the technical effects adopted in this method, this embodiment uses the traditional technical solution and the method of the present invention for comparative testing, and scientifically demonstrates the test results to verify the true effects of this method.

[0138] I. Experimental purpose.

[0139] The purpose of this embodiment is to scientifically demonstrate the technical effects of the method of the present invention for detecting and warning transformer fault sounds by comparing and testing the traditional technical solution and the method proposed by the present invention, and to verify its accuracy and reliability in practical applications.

[0140] II. Experimental equipment and environment.

[0141] Experimental equipment:

[0142] Transformer fault simulation device: used to simulate different types of transformer fault sounds.

[0143] Sound acquisition equipment: such as high-sensitivity microphones, used to collect transformer sound signals.

[0144] Data processing and analysis system: including a computer, a data acquisition card and corresponding software, used to process and analyze the collected sound signals.

[0145] Traditional transformer fault sound detection equipment, as a comparative experiment.

[0146] Experimental environment:

[0147] It is carried out in a quiet laboratory environment to reduce external noise interference.

[0148] Ensure that environmental factors such as laboratory temperature and humidity are stable to avoid affecting the experimental results.

[0149] III. Experimental steps.

[0150] Preparation stage:

[0151] Connect the sound acquisition device to the data processing and analysis system.

[0152] Calibrate the sound acquisition device to ensure its accuracy and sensitivity.

[0153] Set up the transformer fault simulation device and prepare to simulate different types of fault sounds.

[0154] Testing of traditional technical solutions:

[0155] Use the traditional transformer fault sound detection device to detect the simulated fault sounds.

[0156] Record the detection results, including information such as the detected fault type, amplitude increment, duration, etc.

[0157] Testing of the method of the present invention:

[0158] Use the transformer fault sound detection and early warning method proposed by the present invention to detect the simulated fault sounds.

[0159] The specific steps include:

[0160] Collect sound signals and perform preprocessing such as filtering and noise reduction.

[0161] The frequency range for extracting sound features is 500Hz - 1000Hz, the amplitude increment is [value], and the duration is 2 seconds.

[0162] Judge whether there is a fault according to the preset threshold and record the detection results.

[0163] Data recording and analysis:

[0164] Compare the detection results of the traditional technical solution and the method of the present invention, and record the differences and similarities.

[0165] Analyze the performance of the two methods in terms of detection accuracy, sensitivity, false alarm rate, etc.

[0166] IV. Experimental results.

[0167] Comparison of detection accuracy:

[0168] The method of the present invention can accurately identify the simulated transformer fault sounds, and the detection accuracy is higher than that of the traditional technical solution.

[0169] Under various fault types, the accuracy rate of the method of the present invention reaches over 90%, while the accuracy rate of the traditional method is relatively low.

[0170] Comparison of sensitivity and false alarm rate:

[0171] The method of the present invention has high sensitivity and can detect weak fault sound signals. Note: There seems to be a value missing in the "amplitude increment" part of line in the original Chinese text. I've left it as [value] in the translation for now. You can correct it according to the actual situation.

[0172] Meanwhile, the false alarm rate of the method of the present invention is relatively low, effectively reducing the false alarm situations caused by noise or interference.

[0173] Duration and amplitude increment analysis:

[0174] The method of the present invention can accurately measure the duration and amplitude increment of the fault sound, providing a strong basis for fault diagnosis.

[0175] Traditional methods have relatively large errors and uncertainties in this regard.

[0176] V. Conclusion.

[0177] By comparing and testing the traditional technical solution with the transformer fault sound detection and early warning method proposed by the present invention, the experimental results show that the method of the present invention has higher detection accuracy, sensitivity and lower false alarm rate. The method of the present invention can effectively identify the transformer fault sound, providing a strong guarantee for the safe operation of the transformer. Therefore, the method of the present invention has broad application prospects and practical values.

[0178] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system or a computer program product. Therefore, the present invention can take the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. Among them, the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic memory, flash memory, a magnetic disk or an optical disc. These computer program instructions can also be stored in a computer-readable memory capable of guiding a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured article including an instruction device, and the instruction device realizes the functions specified in one process Figure 1 one process or multiple processes and / or boxes Figure 1 or multiple boxes.

[0179] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered within the scope of the claims of the present invention.

Claims

1. A transformer fault sound detection and early warning method, characterized in that: include: Collect transformer working sound data within a fixed time period, including frequency, amplitude and duration; According to the transformer working sound data, it is analyzed based on the pre-trained fault sound detection model, outputs the fault type and issues a fault warning; Among them, a hierarchical attention mechanism, residual connection and Mish activation function are introduced to improve the time-delay neural network, and the fault sound detection model is constructed based on the improved time-delay neural network and the large language model.

2. The transformer fault sound detection and early warning method according to claim 1, characterized in that: The data processing method of the fault sound detection model comprises: Separating the transformer working sound data to obtain transformer body noise and fault sound; Extracting the frequency, amplitude and duration of the fault sound to obtain fault characteristics; Outputting the fault type according to the fault characteristics and the improved time-delay neural network; According to the fault type, a large language model is used to analyze and output the priority of the fault type; A fault signal is issued according to the priority of the fault type to provide a fault warning.

3. The transformer fault sound detection and early warning method according to claim 1, characterized in that: The improved time-delay neural network comprises: Input layer, used to receive fault features; The hierarchical attention mechanism layer includes a bottom-level attention and a top-level attention. The bottom-level attention is used to encode fault features, capture local correlations, and obtain bottom-level encoding representations. The top-level attention is used to further process the bottom-level encoding representations and capture global semantic relationships. The residual connection layer is used to add residual connections between the convolutional layer and the loop layer to alleviate the gradient disappearance problem; The Mish activation function layer is used to replace the traditional activation function with the Mish activation function after each neuron to improve training efficiency and generalization ability; The output layer is used to output the fault type.

4. The transformer fault sound detection and early warning method according to claim 1, characterized in that: According to the blind source separation algorithm, the transformer working sound data is separated to obtain the transformer body noise and fault sound, including: Based on the blind source separation algorithm, according to the statistical characteristics and independence principle of the sound signal, the separation process is performed to output multiple independent sound components; Performing spectrum analysis on the multiple independent sound components, calculating a spectrum diagram, and extracting frequencies of the multiple independent sound components according to frequency distribution and spectrum characteristics; Performing time domain analysis on the multiple independent sound components, calculating power spectrum density, and extracting amplitudes of the multiple independent sound components according to waveform changes; According to the start and end points of the fault sound, extract the duration of the fault sound on the timeline Extracting durations of the plurality of independent sound components from the plurality of independent sound components; According to the frequency, amplitude and duration of multiple independent sound components, the transformer body noise and fault sound are obtained.

5. The transformer fault sound detection and early warning method according to claim 4, characterized in that: The extracting the frequency, amplitude and duration of the fault sound to obtain the fault characteristics includes: The fault sound with a frequency within a preset frequency range, an amplitude increment exceeding a preset first threshold, and a duration within a preset second threshold range is extracted as a fault feature.

6. The transformer fault sound detection and early warning method according to claim 1, characterized in that: The improved time-delay neural network training method includes: Obtain sound data of a large number of transformers in normal and fault conditions over a historical period of time, including frequency, amplitude and duration; The sound data under the fault state includes sound samples of different fault types; Labeling the sound data in the normal state and the fault state, and further labeling the fault type of the sound data in the fault state to obtain a labeled data set; Separating the transformer body noise and fault sound in the labeled data set according to a blind source separation algorithm; Extracting the frequency, amplitude and duration of the fault sound to obtain fault characteristics; The fault characteristics and fault types are used as input labels and output labels of a training set, and the improved time-delay neural network is trained through a first loss function.

7. The transformer fault sound detection and early warning method according to claim 6, characterized in that: The first loss function is expressed as: In the formula, Represents the value of the first loss function, that is, the predicted value of the improved time-delay neural network The difference between the actual label y, N represents the total number of samples, that is, the number of samples in the training set, y i Represents the corresponding element of the true label of the i-th sample, represents the corresponding element of the predicted value of the i-th sample, and log(·) represents the logarithmic function.

8. The transformer fault sound detection and early warning method according to claim 1, characterized in that: The training method of the large language model includes: Obtain transformer fault type data including fault description, maintenance records, impact area and duration; After the transformer fault type data is segmented, the frequency, amplitude and duration of the fault sound are extracted to obtain the fault characteristics; The fault types are labeled according to the severity, impact range, and repair difficulty of the fault characteristics to obtain a labeled data set; Based on the labeled data set, learn the association between fault description, maintenance record and fault type, and set the priority standard of the fault type, wherein the priority standard is used to reflect the urgency of the fault type; quantifying the priority criteria into a priority value; The fault type and the priority value corresponding to the fault type are used as the input label and output label of the training set. The large language model is trained through the second loss function to learn the mapping relationship between fault type and priority, and the priority of the fault type is output.

9. The transformer fault sound detection and early warning method according to claim 1, characterized in that: The second loss function is expressed as: Where L2 represents the value of the second loss function, that is, the difference between the predicted value of the large language model and the true label, N represents the total number of samples, that is, the number of samples in the training set, and w i represents the weight of the i-th sample according to the urgency of the fault type, C represents the total number of fault types, y ij Represents the priority value of the jth fault type in the true label of the i-th sample, Indicates the probability that the i-th sample belongs to the priority value of the j-th fault type.

10. The transformer fault sound detection and early warning method according to claim 1, characterized in that: The sending of a fault signal according to the priority of the fault type and the performing of a fault warning comprises: Set the warning threshold according to the priority standard of the fault type, where 0 means no fault, 1 means severe fault, 2 means moderate fault, and 3 means general fault; When the priority of the fault type is greater than 0, a fault signal corresponding to the warning threshold is triggered and fault warning information including the fault type and the priority corresponding to the fault type is sent.

Images