Transformer state monitoring method and system based on voiceprint perception, and medium

Abstract

The invention discloses a transformer state monitoring method and system based on voiceprint perception and a medium, and the method comprises the following steps: collecting a sound wave signal and a vibration signal of a transformer, carrying out the preprocessing of the collected signals, and carrying out the feature extraction; converting the obtained signal into frequency domain information, and detecting and marking an abnormal value in the signal; the method comprises the following steps: constructing a T-Manba model, and training the Mamba model by using known signal data of a transformer in a normal operation state; and inputting the obtained data into the trained Mamba model, and identifying the abnormal state of the transformer. According to the invention, real-time monitoring and diagnosis of the operation state of the transformer are realized through sound analysis, automation and intelligentization of state monitoring of the transformer are realized, and real-time and accurate monitoring and diagnosis of the operation state of the transformer are provided.

Description

Technical Field

[0001] The present application relates to the field of transformer operation and maintenance technology, and in particular to a transformer status monitoring method, system and medium based on voiceprint perception. Background Art

[0002] As an important part of the power system, the health status of power transformers is crucial to ensure the stability and reliability of power supply. During operation, transformers may experience various faults due to wear, aging or external factors. If these faults are not discovered and handled in time, they may lead to serious safety accidents, such as power outages or even equipment damage. Traditionally, the condition monitoring of power transformers mainly relies on regular physical inspections and analysis based on historical performance data. The state quantities used in common transformer condition monitoring and fault diagnosis methods include: oil chromatography, electromagnetics, temperature, etc. Usually, when the diagnosis is discovered, the equipment defects and faults have already formed, so there is a lag in the diagnosis. Although these methods can detect potential problems to a certain extent, these methods often rely on professional equipment and manual judgment, are inefficient, and may be misjudged due to human factors. They face key problems such as slow response speed, high maintenance cost and limited accuracy. In particular, for the early prediction and accuracy of faults, these methods often fail to effectively prevent serious equipment damage and power grid failures. Therefore, with the development of industrial automation and intelligent technology, more advanced monitoring methods are introduced to improve the accuracy and timeliness of fault diagnosis.

[0003] In recent years, fault diagnosis methods based on sound analysis have gradually attracted attention. Voiceprint perception, as a non-invasive diagnostic technology, provides new possibilities for transformer status monitoring. During the operation of the transformer, the core, winding and other structures will vibrate and transmit mechanical waves. The generated vibration and audible sound signals contain a lot of equipment status information. Audible noise and vibration are generated with the operation of the transformer. The amplitude, time domain waveform, and spectrum characteristics of sound and vibration are closely related to its operating voltage, current, mechanical state, excitation state, insulation state, etc., and can timely reflect the changes in the equipment operation status. According to DL / T573-2010 "Guidelines for Maintenance of Power Transformers", transformer overload, unbalanced load, harmonic load, severe overheating, DC bias, partial discharge, loose winding core, loose accessories and other defects are all related to transformer operation sound and vibration. By capturing and analyzing these sound signals, this technology can monitor and diagnose various potential mechanical and electrical problems, detect potential faults early and perform preventive maintenance. However, existing sound monitoring technologies have certain limitations. They are usually limited to using predefined, static sound signatures for diagnosis and lack the ability to adapt to various transformer models and operating conditions. In addition, due to the influence of external environmental noise, there are errors in sound data acquisition, which further limits the accuracy of early diagnosis and fault prediction. Secondly, most systems lack sufficient intelligence and cannot learn and adapt to new sound patterns in real time, which reduces the flexibility and accuracy of diagnosis. Finally, the data processing of existing systems is usually not real-time, which reduces the speed and efficiency of fault response.

[0004] Voiceprint sensing technology is based on the unique sound wave pattern generated when the transformer is in operation. When an abnormality occurs inside the transformer (such as winding deformation, insulation aging, core loosening, etc.), the frequency and intensity of the sound it emits will be different from that during normal operation. However, although this type of monitoring technology has great potential, its practical application faces many challenges.

[0005] First, the interference of environmental noise is a difficult problem to overcome. Transformers are usually located in industrial areas with high noise environments. Environmental noise often masks or blurs the tiny sound changes emitted by the transformer, resulting in inaccurate monitoring results. The second is the complexity of voiceprint data analysis. There are many types of transformer faults, and the same fault can present different voiceprints due to different transformer models, ages or operating conditions. This means that a large amount of data must be collected and processed to establish a reliable fault voiceprint database.

[0006] Then there is the issue of the installation of monitoring equipment. The performance of acoustic sensors is greatly affected by their installation position, angle and fastening method. Improper installation may lead to inaccurate data collection or even miss key fault signals. Finally, existing monitoring systems are lacking in data processing speed and intelligence level. The rapid identification and location of transformer faults requires real-time or near real-time data processing capabilities, but many current systems cannot meet this requirement. In addition, most existing systems lack the ability to learn and adapt to new fault modes, which requires constant manual intervention to update the system. Furthermore, current systems are unable to achieve real-time data analysis and instant feedback, which leads to delays in monitoring responses and may miss fault warnings. Summary of the invention

[0007] The present application provides a transformer status monitoring method, system and medium based on voiceprint perception, aiming to realize real-time monitoring and diagnosis of transformer operating conditions through sound analysis, realize automation and intelligence of transformer status monitoring, and aiming to provide real-time and accurate transformer operating status monitoring and diagnosis.

[0008] The technical solution of this application is as follows:

[0009] On the one hand, the present application provides a transformer state monitoring method based on voiceprint perception, characterized in that it includes the following steps:

[0010] Step 1: Collect the acoustic wave signal and vibration signal of the transformer, pre-process the collected signals and perform feature extraction;

[0011] Step 2: Convert the signal obtained in step 1 into frequency domain information, detect and mark the abnormal values ​​in the signal;

[0012] Step 3: Build the T-Manba model and use the known signal data of the transformer under normal operating conditions to train the Mamba model; the T-Manba model is a combination of the T-MambaBlock and the Temporal Convolutional Network framework;

[0013] Step 4: Input the data obtained in step 2 into the trained Mamba model to identify the abnormal status of the transformer.

[0014] Furthermore, in step 1, the preprocessing of the collected signal includes:

[0015] For acoustic wave signals, a Kalman filter algorithm is used to eliminate background noise, and peak detection is used to enhance the sound signal captured from the MEMS microphone;

[0016] For vibration signals, Fast Fourier Transform (FFT) is implemented to analyze the vibration characteristics at different frequencies, and a sliding window is used to smooth the time series data.

[0017] Furthermore, signal feature extraction includes:

[0018] For sound wave signals, advanced feature extraction algorithms are used to extract the features of sound signals;

[0019] For vibration signals, discrete wavelet transform (DWT) is used to analyze the multi-scale characteristics of vibration signals.

[0020] Furthermore, in step 1, the acoustic wave signal and vibration signal of the transformer are collected by multiple sensor devices; in step 2, the network time protocol NTP is used to determine the clock synchronization between all sensor devices, and the data from different sensor devices can be accurately matched by using timestamps and sequence numbers.

[0021] Furthermore, in step 2, the signal obtained in step 1 is converted into frequency domain information using fast Fourier transform (FFT).

[0022] Furthermore, in step 2, the isolation forest algorithm is used to detect outliers in the data and filter out outliers in the signal.

[0023] Furthermore, the T-Manba model in step 3 is implemented through a combination of linear layers, convolutional layers, and activation functions of T-MambaBlock based on the Temporal Convolutional Network framework; the linear layer is used to project the input data from a low-dimensional space to a high-dimensional space to enhance the feature expression capability; the convolutional layer is used to extract local features and capture the spatial pattern in the input data through convolution operations; the activation function introduces nonlinearity;

[0024] T-MambaBlock combines the state space model to describe the state of the system through a set of state variables, and describes the evolution of the state and the generation of observations through state transfer equations and observation equations to capture long-range dependencies. The state update equation of the state space model is:

[0025] h(t)=A·h(t-1)+B·x(t)+w(t)

[0026] The output equation is:

[0027] y(t)=C·h(t)+D·x(t)

[0028] Among them, h(t) is the hidden state vector at time t; A is the state transfer matrix, B is the input control matrix, x(t) is the input vector at time t, and w(t) is the process noise vector.

[0029] On the other hand, the present application provides a transformer state monitoring system based on voiceprint perception, characterized in that it includes:

[0030] A data collection layer, which is used to collect the acoustic wave signal and vibration signal of the transformer, and pre-process the collected signals and perform feature extraction;

[0031] The edge computing layer is used to build a Mamba model and train the Mamba model using known signal data of the transformer in normal operation; convert the signal obtained by the data collection layer into frequency domain information, detect and mark the abnormal values ​​in the signal, and then input them into the trained Mamba model to identify the abnormal state of the transformer;

[0032] And a software application layer, wherein the software application layer is used to output the operation results of the edge computing layer to the outside.

[0033] Furthermore, in the data collection layer, several MEMS microphones are used to collect sound wave signals of the transformer and the surrounding environment, and the several MEMS microphones are arranged in an equidistant polygon shape around the surface of the transformer; several acceleration sensors are used to collect transformer vibration data, and the several acceleration sensors are arranged in an equidistant polygon shape around the surface of the transformer.

[0034] Furthermore, the edge computing layer includes a memory, and the edge computing layer stores the collected data in the memory.

[0035] On the other hand, the present application provides a computer-readable medium, characterized in that the computer-readable medium stores a computer program, and when the computer program is called and executed by a computer, the transformer state monitoring method based on voiceprint perception as described above is implemented.

[0036] In summary, the beneficial effects of this application are:

[0037] 1. Improve the real-time and accuracy of monitoring: Use voiceprint sensing to monitor the real-time status of the transformer, making diagnosis more immediate and accurate, and timely discovering and preventing potential faults.

[0038] 2. Non-intrusive monitoring: The system uses non-intrusive monitoring technology, which does not require direct contact with the transformer, reducing the interference of the monitoring process on the operation of the transformer.

[0039] 3. Intelligent data processing: Integrate edge computing technology and machine learning algorithms to make data processing more intelligent and able to self-learn and adapt to new failure modes.

[0040] 4. Remote real-time diagnosis: Provides a remote monitoring front-end interface that can monitor the transformer status in real time, providing operators with intuitive data and trend analysis, reducing the need for on-site operations.

[0041] 5. Preventive maintenance: The system can predict the operating trends of the transformer based on the monitored data and analysis results, which helps to implement more effective preventive maintenance measures.

[0042] 6. Low maintenance cost and high efficiency: Since the system adopts automatic and intelligent monitoring, the frequency of manual inspection and maintenance workload are greatly reduced, thereby reducing maintenance costs.

[0043] 7. Data security and integrity: The system emphasizes data security and authorization management to ensure the security and accuracy of monitoring data.

[0044] 8. High scalability: The software design is highly compatible and can be expanded through API to adapt to future updates and hardware upgrade needs.

[0045] 9. Resistance to environmental noise: The designed voiceprint sensing system can effectively filter environmental noise, enhance the sensitivity to the sound of the transformer itself, and improve the detection accuracy.

[0046] In summary, the present invention provides a new solution for the health status monitoring of power transformers by integrating advanced voiceprint sensing technology, edge computing and intelligent data analysis, overcomes the limitations of traditional methods, and provides an efficient and reliable technical solution for the needs of smart grids. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is the overall architecture diagram of the transformer status monitoring system based on voiceprint perception proposed in this application;

[0048] Figure 2 This is the overall flow chart of the transformer status monitoring system based on voiceprint perception proposed in this application;

[0049] Figure 3 This is the structure diagram of the transformer voiceprint perception deep learning model based on T-Manba in the transformer state monitoring system based on voiceprint perception proposed in this application. DETAILED DESCRIPTION

[0050] The specific implementation of the present application is described in detail below with reference to the accompanying drawings.

[0051] In a specific embodiment of the present application, a transformer state monitoring method based on voiceprint perception is provided, comprising the following steps:

[0052] Step 1: Collect the acoustic wave signal and vibration signal of the transformer, pre-process the collected signals and perform feature extraction; the pre-processing of the collected signals includes:

[0053] For acoustic wave signals, a Kalman filter algorithm is used to eliminate background noise, and peak detection is used to enhance the sound signal captured from the MEMS microphone;

[0054] For vibration signals, Fast Fourier Transform (FFT) is implemented to analyze the vibration characteristics at different frequencies, and a sliding window is used to smooth the time series data.

[0055] Signal feature extraction includes:

[0056] For sound wave signals, advanced feature extraction algorithms are used to extract the features of sound signals;

[0057] For vibration signals, discrete wavelet transform (DWT) is used to analyze the multi-scale characteristics of vibration signals.

[0058] The acoustic wave signal and vibration signal of the transformer are collected through multiple sensor devices; Step 2: Use the network time protocol NTP to determine the clock synchronization between all sensor devices, and use timestamps and sequence numbers to accurately match data from different sensor devices.

[0059] Step 2: Convert the signal obtained in step 1 into frequency domain information, detect and mark the outliers in the signal; specifically, use fast Fourier transform (FFT) to convert the signal obtained in step 1 into frequency domain information, use the isolation forest algorithm to detect outliers in the data, and filter out outliers in the signal.

[0060] Step 3: Build the T-Manba model and use the known signal data of the transformer under normal operating conditions to train the Mamba model; the T-Manba model is a combination of the T-MambaBlock and the Temporal Convolutional Network framework;

[0061] The T-Manba model is based on the Temporal Convolutional Network framework and is implemented through a combination of the linear layer, convolutional layer, and activation function of T-MambaBlock. The linear layer is used to project the input data from a low-dimensional space to a high-dimensional space to enhance the feature expression capability. The convolutional layer is used to extract local features and capture the spatial pattern in the input data through convolution operations. The activation function introduces nonlinearity.

[0062] like Figure 3,T-MambaBlock combines the state space model, describes the state of the system through a set of state variables, and describes the evolution of the state and the generation of observations through state transfer equations and observation equations to capture long-range dependencies. The state update equation of the state space model is:

[0063] h(t)=A·h(t-1)+B·x(t)+w(t)

[0064] The output equation is:

[0065] y(t)=C·h(t)+D·x(t)

[0066] Among them, h(t) is the hidden state vector at time t; A is the state transfer matrix, B is the input control matrix, x(t) is the input vector at time t, and w(t) is the process noise vector.

[0067] Step 4: Input the data obtained in step 2 into the trained Mamba model to identify the abnormal status of the transformer.

[0068] Another specific implementation of the present application provides a transformer status monitoring system based on voiceprint perception, referring to Figure 1 , including a data collection layer, a data collection layer and a software application layer, the transformer status monitoring method based on voiceprint perception is applied to detect the status of the substation.

[0069] The data collection layer is used to collect the acoustic wave signal and vibration signal of the transformer, and preprocess the collected signals and perform feature extraction.

[0070] like Figure 1 As shown in the figure, the data collection layer is the first level of the transformer condition monitoring system based on voiceprint perception, and its main function is to capture the acoustic wave signals related to the transformer status. This layer contains a set of precision acoustic sensors and vibration sensors, which are used to capture the sound characteristics emitted by the transformer under different working conditions. The acoustic sensors are made of highly sensitive materials and can respond to the slight changes in the sound waves generated during the operation of the transformer. These sensors can identify the characteristics of the sound print within a certain frequency range, including infrasound and ultrasonic frequency bands that are inaudible to the human ear, ensuring that the system has a high detection capability for the changes in the sound print under mechanical and electrical faults. During the operation of the system, the sensor senses the acoustic wave activity of the transformer in real time and converts the acoustic wave signal into an electrical signal. Each sensor is precisely calibrated to ensure that accurate sound data can be captured in a noisy environment.

[0071] The captured signal is first preprocessed, including denoising, amplification and filtering. The purpose of preprocessing is to improve data quality and provide a cleaner and more representative signal for signal feature extraction and further analysis at the edge computing layer. Data preprocessing is a critical part of the transformer voiceprint condition monitoring system, which is mainly responsible for improving the quality of data for subsequent analysis. This stage usually involves several steps, including denoising, which is used to remove non-target background noise in the sound signal in order to clearly extract the sound characteristics of the transformer. The use of different types of filters and signal processing techniques can effectively reduce the impact of noise. Next is signal amplification, which is a step to enhance the voiceprint signal of interest to make it more prominent and facilitate the observation of the operating status of the transformer. After the signal is properly adjusted for gain, it is easier to process later. Then the signal is filtered, which will filter out unnecessary frequency components, such as components that are not in the analysis frequency band, thereby retaining the signal part that is useful for subsequent analysis. The specific type of filter used depends on the frequency range to be retained or removed. Signal normalization is also an important step, which adjusts the signal strength to a standard range to ensure that data from different time periods or different sensors can be compared fairly. Signal segmentation is the process of segmenting a continuous signal so that specific events or states can be analyzed separately. This step is performed according to specific characteristics of time, events or sound signals. Feature extraction is the simplification of complex sound signals into key data indicators. Such a transformation allows the data to be expressed in a lower dimension while retaining information that is critical to explaining the behavior of the transformer. Finally, encoding is the process of converting the preprocessed data into a format suitable for transmission, often through various encoding and decoding technologies to optimize data transmission efficiency and reliability. Together, these preprocessing steps ensure that the data is superior in terms of accuracy, consistency, and interpretability, which is critical for the system to accurately monitor the status of the transformer.

[0072] To ensure the consistency of the monitored acoustic wave data in time series, multiple sensors are formed into a sensor network to work in a time-synchronized manner. This data synchronization is critical for subsequent data fusion and analysis because it allows the system to accurately compare data from different sensors and identify abnormal patterns. The preprocessed signals are packaged in digital form and sent to the edge computing layer through a stable communication interface. Data packaging uses an efficient encoding method to minimize delays and data loss during transmission.

[0073] Through such a data collection layer design, this scheme can obtain high-quality transformer sound signals and provide reliable input data for the system, thereby greatly improving the accuracy and efficiency of transformer condition monitoring.

[0074] The edge computing layer is used to build a Mamba model, and the Mamba model is trained using known signal data of the transformer under normal operating conditions; the signal obtained by the data collection layer is converted into frequency domain information, the abnormal values ​​in the signal are detected and marked, and then input into the trained Mamba model to identify the abnormal state of the transformer.

[0075] like Figure 1 As shown in Figure 1, the edge computing layer is responsible for processing the sound and vibration data from the sensor module. This layer is usually located near the data source, such as close to the sound sensor. The main goal of the edge computing layer is to process the received data immediately and reduce the amount of data that must be sent to the cloud or central processor, thereby significantly improving the system response time and operating efficiency. The edge computing layer specifically performs the following functions:

[0076] First, it receives the voiceprint signal from the data preprocessing layer, that is, the signal that has been preliminarily processed through steps such as denoising, amplification, filtering, and standardization. These signals contain important acoustic features of the transformer under specific working conditions. This layer then performs a deeper analysis of the signal, such as using a fast Fourier transform (FFT) to convert the time domain signal to the frequency domain. The transformed frequency domain signal makes it easier to determine the unique frequencies and any abnormal frequencies generated during the operation of the transformer, which is critical for early fault detection and diagnosis. The edge computing layer also performs feature extraction, which extracts key statistical information from the processed signal or uses machine learning technology to identify and monitor voiceprint features, so that the working status and health of the transformer can be more accurately analyzed and understood.

[0077] Another important function of the edge computing layer is event detection and decision support. When the edge computing layer identifies state deviations or signs of failure, it can make local decisions in a timely manner, such as adjusting monitoring strategies or issuing alarms. This ability to process and respond locally greatly reduces dependence on central servers, optimizes bandwidth utilization, and improves the overall reliability of the system.

[0078] The edge computing layer also includes data compression and encryption functions to ensure that the data sent to the central processor or the cloud does not occupy too much network bandwidth and is not tampered with during transmission. Through efficient encoding and decoding methods, the edge computing layer can minimize the volume of transmitted data while retaining important information. In general, the edge computing layer provides fast local data processing capabilities in the transformer sound monitoring system, brings decentralized intelligent decision-making to the system and reduces the burden on the central server, which is crucial for building a real-time, efficient and reliable monitoring system.

[0079] The software application layer is used to output the operating results of the edge computing layer to the outside.

[0080] like Figure 1 As shown in the figure, the software application layer includes the remote monitoring front-end interface and the server backend. The front-end interface displays the equipment operation list, status information, status evaluation diagram, as well as database storage and characteristic parameter trend diagram, so that users can intuitively grasp the real-time operation status of the transformer. The back-end server is not only responsible for storage and further data processing, but also for historical data analysis, prediction of operation trends, and providing data support for maintenance decisions.

[0081] First is the user interface, through which users can observe and control the operating status of the transformer in real time. This interface intuitively displays information such as sound waveforms and spectra, and presents the status assessment results in the form of charts, so that the operator can understand the current situation at a glance. Then, data management and storage is another core component of the software application layer. It includes a database system for long-term storage of voiceprint data, characteristic parameters and monitoring history. This not only facilitates data retrieval and analysis, but also makes it possible to generate trend reports and charts. The back-end server also plays a key role in the system. It is responsible for further processing and analysis of data transmitted from the edge computing layer. Using data fusion technology and machine learning algorithms, the server can predict the operating trend of the transformer and identify possible failure modes early. When anomalies are detected, warnings and notifications can be quickly issued to ensure that technicians can respond in time. In addition, the system also provides support for maintenance decisions. It not only combines real-time monitoring data and historical analysis results, but also makes specific maintenance and repair suggestions to help operators remotely diagnose problems, thereby reducing on-site workload.

[0082] The system also attaches great importance to data security and authorization issues, ensuring that all sensitive information and management operations are strictly limited to authorized personnel through user authentication and permission management. Finally, the software is designed with interoperability with other systems in mind, allowing it to adapt to changing needs and can be expanded through APIs to adapt to new hardware or application requirements that may be added in the future. In general, the software application layer is the brain and control center of this monitoring system based on voiceprint sensing technology, which simplifies the transformer condition monitoring and maintenance process in an intelligent and user-friendly way.

[0083] The system has a clear hierarchical architecture and division of labor, modular functions, coordinated operation of each layer and module, and step-by-step call, which ensures the security, integrity and accuracy of data resources in the entire method. The beneficial effect of the present invention is that it uses a combination of mechanical vibration and sound wave data to achieve early detection and diagnosis of potential transformer faults, thereby improving the transformer's operation and maintenance efficiency and reliability. In addition, the system supports remote data transmission and intelligent analysis, reducing the workload of on-site maintenance and reducing operation and maintenance costs.

[0084] It is understandable that the matters not described in detail in this application are understood to be implemented by adopting existing technical solutions.

[0085] refer to Figure 2 The principles of a transformer state monitoring system based on voiceprint perception proposed in this embodiment include the following:

[0086] 1. In the voiceprint-aware transformer condition monitoring system proposed in the present invention, the data collection layer is the basis of the system architecture. The key lies in being able to capture high-quality sound wave signals and provide reliable input data for subsequent processing.

[0087] 1.1: Sensor deployment

[0088] In the data collection layer, several MEMS microphones are used to collect sound wave signals of the transformer and the surrounding environment. Several MEMS microphones are arranged in an equidistant polygon shape around the surface of the transformer; several acceleration sensors are used to collect transformer vibration data. Several acceleration sensors are arranged in an equidistant polygon shape around the surface of the transformer.

[0089] Specifically, high-performance MEMS microphones (typically Knowles SPU0410LR5H-QB) and precise accelerometers (typically Analog Devices ADXL345) are selected and arranged around the transformer surface in an equidistant triangle shape according to the geometric structure and radiation sound field characteristics of the transformer. The MEMS microphone is used to capture the sound wave signals of the surrounding environment, while the ADXL345 accelerometer is used to capture vibration data to ensure that the mechanical vibration state of the transformer can be accurately recorded.

[0090] 1.2: Signal preprocessing

[0091] Design a preprocessing algorithm flow, use the Kalman filter algorithm to eliminate background noise, and use peak detection to enhance the sound signal captured from the MEMS microphone. For acceleration data, implement fast Fourier transform (FFT) to analyze vibration characteristics at different frequencies, and use sliding window technology to smooth time series data. The above steps are implemented using the Signal Processing Toolbox provided by Analog Devices.

[0092] The Kalman filter algorithm is usually divided into two stages: prediction and update. The key to solving the algorithm is to recursively calculate the state estimate and covariance estimate, and update them according to the new measurement value at each time step.

[0093] This algorithm is efficient and suitable for real-time systems. Its strength lies in its ability to provide an optimal estimate of the system state in the presence of noise. In practical applications, the performance of the algorithm depends on the accuracy of the state transfer matrix, control matrix, measurement matrix, and noise covariance matrix.

[0094] 1.3: Feature extraction and data synchronization

[0095] After preprocessing the signal data, advanced feature extraction algorithms such as Mel Frequency Cepstral Coefficient (MFCC) are used to extract the characteristics of the sound signal, and Discrete Wavelet Transform (DWT) is used to analyze the multi-scale characteristics of the vibration signal. These features can reveal the key properties of the signal, which is crucial for accurate sound recognition and vibration pattern analysis.

[0096] After preprocessing the speech sequence, the time sequence (n) is obtained and then Fourier transformed into a frequency domain waveform: X(i,k) = FFT[X m (n)]

[0097] Wherein, X(i,k) represents the amplitude of the kth frequency component in the i-th frame.

[0098] Speech signal energy: E(i,k) = [X(i,k)] 2

[0099] Mel filter is used for H m (k) Filtering:

[0100]

[0101] Where m is the number of filters, f represents the boundary frequency of the filter, and f(m) is the center frequency of the mth Mel filter;

[0102] Energy after Mel filter (k):

[0103]

[0104] Calculate the MFCC feature parameters of the discrete cosine change cepstrum:

[0105]

[0106] Where S(i,k) is the energy of the kth Mel band in the i-th frame, E(i,k) is the amplitude spectrum of the kth frequency component in the i-th frame, and H m (k) is the frequency response of the mth Mel filter, and N is the total number of frequency components in the frequency domain.

[0107] For data synchronization, we use the Network Time Protocol (NTP) to ensure clock synchronization between all sensors. In addition, timestamps and sequence numbers are used to ensure that data from different sensors can be accurately matched. This synchronization mechanism ensures the correct fusion and analysis of multi-channel data.

[0108] 2: At the edge computing layer, the pre-processed acoustic signal is further converted to a frequency domain representation through a fast Fourier transform (FFT) to facilitate the identification of voiceprint features under normal working conditions and signal changes related to abnormal conditions. The extracted features are further processed by the feature analysis module, which uses machine learning techniques for event detection and pattern recognition. When identifying abnormal patterns or predicting potential failures, this layer can make quick decisions and trigger alarms or adjust data collection strategies in a timely manner. The data compression and encryption module is responsible for optimizing data transmission to ensure that data is both efficient and secure when transmitted to the upper layer.

[0109] 2.1: Hardware Selection In order to implement the edge computing layer, we need to select the right hardware devices. The edge computing hardware selected is Dell EMC PowerEdge XE2420, which provides powerful processing power, sufficient memory, and fast storage solutions, such as at least 16GB of RAM and 512GB of solid-state drives. At the same time, for sensors, Infineon IM69D120 microphones and Analog Devices' ADXL355 vibration sensors are selected to ensure that we can obtain high-quality monitoring data.

[0110] 2.2: Machine Learning Algorithms Deep learning algorithms are introduced to the edge computing layer to better process and analyze data. The first step is frequency domain analysis, where the collected signal is converted into frequency domain information using Fast Fourier Transform (FFT). This step is critical for capturing the periodic and non-periodic components within the signal. This process can use the SciPy library to implement Fast Fourier Transform to perform frequency domain analysis on the signal, while using the Isolation Forest algorithm to detect anomalies in the data.

[0111] 2.2.1: When we get the frequency domain data, the next step is to use the isolation forest algorithm to help us identify abnormal patterns. Isolation Forest (iForest) is a fast outlier detection method based on Ensemble, with linear time complexity and high accuracy. It is a state-of-the-art algorithm that meets the requirements of big data processing. This machine learning algorithm is particularly suitable for processing high-dimensional data and is suitable for anomaly detection of continuous numerical data. Unlike other anomaly detection algorithms that use quantitative indicators such as distance and density to characterize the degree of alienation between samples, the isolation forest algorithm detects outliers by isolating sample points.

[0112] Specifically, the algorithm uses a binary search tree structure called the isolation tree iTree to isolate samples. Due to the small number of outliers and their alienation from most samples, outliers will be isolated earlier, that is, outliers will be closer to the root node of iTree, while normal values ​​will be farther away from the root node. In addition, compared with traditional algorithms such as LOF and K-means, the isolation forest algorithm has better robustness to high-dimensional data and does not require us to predefine normal or abnormal data. By constructing multiple decision trees to isolate observed samples, we can identify which samples are more likely to be isolated.

[0113] Use the Ensemble method to get a convergence value (Monte Carlo method), that is, repeatedly cut from the beginning and then average the results of each cut. iForest consists of T iTrees, each iTree is a binary tree structure. The algorithm can be roughly divided into two stages. In the first stage, we need to train T isolated trees to form an isolated forest. Then we bring each sample point into each isolated tree in the forest, calculate the average height, and then calculate the outlier score of each sample point.

[0114] In the first stage, the steps are as follows:

[0115] 1. Randomly select Ψ sample points from the training data as the sample subset and put them into the root node of the tree.

[0116] 2. Randomly specify a dimension (feature) and randomly generate a cutting point p in the current node data (the cutting point is generated between the maximum and minimum values ​​of the specified dimension in the current node data).

[0117] 3. A hyperplane is generated with this cutting point, and then the data space of the current node is divided into two subspaces: the data less than p in the specified dimension is placed in the left child node of the current node, and the data greater than or equal to p is placed in the right child node of the current node.

[0118] 4. Recursively repeat steps (2) and (3) in the child node, and continuously construct new child nodes until there is only one data in the child node (no further cutting is possible) or the child node has reached the specified height.

[0119] 5. Loop (1) to (4) until T isolated trees iTree are generated.

[0120] Phase 2:

[0121] After obtaining T iTrees, the iForest training is completed, and then we can use the generated iForest to evaluate the test data. i , let it traverse each isolated tree iTree and calculate the point x i The average height in the forest h(x i ), normalize the average height of all points. The calculation formula of the outlier score is as follows:

[0122]

[0123] in

[0124]

[0125] in, is the anomaly score of the i-th data point, and its value range is [0,1]. The closer the score is to 1, the more likely it is an outlier, and the closer it is to 0, the more likely it is a normal value. h(i) represents the path length of data point i in each isolated tree iTree, which refers to the path from the root node to the sample point. is the expected path length of the isolated tree, which is used as the normalization benchmark. H(x) is the harmonic number, which can be estimated by ln(x)+0.5772156649 (Euler constant). The smaller the score, the more abnormal the data.

[0126] 2.2.2: Based on the detection results of the isolation forest algorithm, the data points marked as abnormal are screened out. These data points are considered potential abnormalities or fault signals and require further analysis. In order to further improve the performance of our system, using more advanced deep learning technology, we use the T-Manba model to process and analyze the sound and vibration signals of the transformer, and identify and classify the abnormal sound of the transformer.

[0127] In order to overcome the shortcomings of traditional methods in feature extraction and long-range dependency capture, we introduced the T-Manba model. The T-Manba model combines T-MambaBlock with the Temporal Convolutional Network (TCN) framework, and achieves enhanced feature extraction and long-range dependency capture through T-MambaBlock while maintaining the characteristics of time series data. The causal convolution and dilated convolution in TCN further improve the performance of the model when processing long sequence data, enabling the T-Manba model to efficiently capture long-range dependencies in time series data. Causal convolution ensures that the current time step of the output depends only on the input of the current time step and before. Dilated convolution expands the receptive field by inserting zero elements between the convolution kernels. The mathematical expression of dilated convolution is:

[0128]

[0129] Where y(t) is the output of the convolution operation, w(i) is the weight of the convolution kernel, x(nd·i) is the element of the input sequence, and d is the dilation rate.

[0130] Based on TCN, the T-Manba model achieves efficient feature extraction of input data through a combination of linear layers, convolutional layers, and activation functions of T-MambaBlock. The linear layer is used to project the input data from a low-dimensional space to a high-dimensional space to enhance the feature expression capability. The convolutional layer is used to extract local features and capture the spatial patterns in the input data through convolution operations. The activation function introduces nonlinearity, thereby enhancing the expressive power of the model.

[0131] T-MambaBlock combines the State Space Model (SSM), which is a mathematical model used to describe dynamic systems. It describes the state of the system through a set of state variables, and describes the evolution of the state and the generation of observations through state transition equations and observation equations. It is used to capture long-range dependencies. The state update equation of SSM is:

[0132] h(t)=A·h(t-1)+B·x(t)+w(t)

[0133] The output equation is:

[0134] y(t)=C·h(t)+D·x(t)

[0135] Among them, h(t) is the hidden state vector at time t. A is the state transfer matrix, B is the input control matrix, x(t) is the input vector at time t, w(t) is the process noise vector, matrix C: maps the state variable h(t) to the output y(t), and matrix D: directly maps the current input x(t) to the output y(t).

[0136] In this way, the T-Manba model can more effectively capture the complex patterns and dependencies in transformer sound and vibration signals, thereby improving the accuracy of identifying and classifying abnormal sounds. The screened abnormal data is input into the T-Manba model, which deeply analyzes the characteristics of abnormal signals by learning the time dependency and dynamic changes of the data to identify the specific fault type or state.

[0137] The T-Manba model performs well when processing long series of data, can identify subtle abnormal patterns in transformer sound and vibration signals, and is suitable for capturing long-term dependencies and dynamic characteristics of signals. The preprocessed sound and vibration signals are converted into a time series data format that the T-Manba model can process, and the model's built-in feature extraction capabilities are used to automatically identify key time series features from the data without the need for complex manual feature engineering. Next, the T-Manba model is trained using known signal data under normal operating conditions to learn the sound and vibration patterns of the transformer under normal conditions. By learning the inherent structure of this data, the model can accurately distinguish between normal and abnormal states in subsequent analysis. After the model training is completed, real-time or newly acquired signal data is input into the model for analysis. The T-Manba model evaluates the signal features and compares them with the learned normal patterns to identify deviations from abnormal patterns or states. In this way, the T-Manba model can not only identify the current abnormal state, but also continuously adapt to changes in the transformer's operating characteristics and emerging fault types through online learning. The T-Manba model supports online learning and can continuously adjust and update itself based on new data to adapt to changes in the transformer's operating characteristics and emerging fault types. This continuous learning and adaptation ensures that the model remains highly accurate and effective over time.

[0138] The T-Manba model significantly improves the efficiency of feature extraction by combining the advantages of TCN and T-MambaBlock. The multi-layer feature extraction method can capture the multi-scale features in the data. The ability to capture long-range dependencies makes the model perform better when processing long sequence data. The nonlinear feature expression method improves the model's expressiveness. The T-Manba model performs well when processing transformer sound and vibration signals. Through efficient feature extraction and capture of long-range dependencies, the T-Manba model can achieve real-time monitoring and rapid response to transformer status, and accurately identify and classify abnormal transformer sounds. The model's adaptive ability and online learning characteristics ensure that the model can adapt to changes in transformer operating characteristics and emerging fault types, maintaining a high degree of accuracy and effectiveness.

[0139] With this approach, the T-Manba model provides an efficient and adaptive solution for deep learning analysis of transformer sound and vibration signals, which helps to achieve real-time and accurate monitoring of transformer status and improve the stability and reliability of the power system.

[0140] The entire system will form a comprehensive transformer condition monitoring system including fast Fourier transform (FFT), isolation forest algorithm and T-Manba model to accurately monitor and classify the various conditions of the transformer. Combining these methods can achieve early diagnosis of potential problems of the transformer and optimize the efficiency of condition monitoring and maintenance work.

[0141] 2.3: The edge computing layer includes a memory, and the edge computing layer stores the collected data in the memory.

[0142] The data storage and analysis module manages and maintains the data in the database during equipment operation, including sensor data, transformer status data, etc. The data is recorded in a time series to ensure the integrity and availability of the data. The collected data is sorted and stored in an appropriate format, using the database format. Every three days, the module will sort and classify the data and evaluate the transformer status information to facilitate the subsequent optimization of the model.

[0143] 2.4: Network protocol selection To ensure efficient and secure data transmission, the selection of network protocols is crucial. The MQTT protocol is used to ensure that data can be transmitted in real time and at low cost. At the same time, the time-sensitive network standard is used to achieve accurate time synchronization to ensure the timing consistency of all sensor data. For the security of data transmission, we will enable the TLS protocol to encrypt all network communications. Through such a network configuration, the data transmission of the edge computing layer can be ensured to be both fast and secure.

[0144] 3: The software application layer provides users with an intuitive and easy-to-understand monitoring and control interface. Real-time data can be displayed intuitively through the interface, and combined with the status assessment results, it provides users with a basis for operating the transformer. In addition, the Web-based front-end interface allows users to remotely access the system to view equipment operation lists, status information, trend charts, etc. In addition to more in-depth data processing and long-term storage, the server backend can also perform more complex analysis based on historical trends, such as using machine learning algorithms to predict equipment operation trends and provide maintenance decision recommendations. All these functional modules work together to ensure the efficiency and reliability of the system and bring convenience to maintenance and monitoring.

[0145] When building this system, we also considered the security of data and user operation permissions. The system ensures the security of data and operations through multi-level user verification and permission control. At the same time, it maintains interoperability with other systems and provides API expansion to cope with the integration of new hardware or new business needs in the future.

[0146] Another embodiment of the present application provides a computer-readable medium, wherein the computer-readable medium stores a computer program, and when the computer program is called and executed by a computer, the transformer state monitoring method based on voiceprint perception as described above is implemented.

[0147] The above is only a preferred implementation of the present application. It should be pointed out that a person skilled in the art can make several modifications and improvements without departing from the inventive concept of the present application, and these all fall within the scope of protection of the present application.

Claims

1. A transformer state monitoring method based on voiceprint perception, characterized in that: The following steps are involved: Step 1: Collect the acoustic wave signal and vibration signal of the transformer, pre-process the collected signals and perform feature extraction; Step 2: Convert the signal obtained in step 1 into frequency domain information, detect and mark the abnormal values ​​in the signal; Step 3: Build the T-Manba model and use the known signal data of the transformer under normal operating conditions to train the Mamba model; the T-Manba model is a combination of the T-MambaBlock and the Temporal Convolutional Network framework; Step 4: Input the data obtained in step 2 into the trained Mamba model to identify the abnormal status of the transformer.

2. The transformer state monitoring method based on voiceprint perception according to claim 1 is characterized in that: In step 1, the preprocessing of the collected signal includes: For acoustic wave signals, a Kalman filter algorithm is used to eliminate background noise, and peak detection is used to enhance the sound signal captured from the MEMS microphone; For vibration signals, Fast Fourier Transform (FFT) is implemented to analyze the vibration characteristics at different frequencies, and a sliding window is used to smooth the time series data.

3. The transformer state monitoring method based on voiceprint perception according to claim 1 is characterized in that: Signal feature extraction includes: For sound wave signals, advanced feature extraction algorithms are used to extract the features of sound signals; For vibration signals, discrete wavelet transform (DWT) is used to analyze the multi-scale characteristics of vibration signals.

4. The transformer state monitoring method based on voiceprint perception according to claim 1 is characterized in that: In step 1, the acoustic wave signal and vibration signal of the transformer are collected through multiple sensor devices; the network time protocol NTP is used to determine the clock synchronization between all sensor devices, and the timestamp and sequence numbering method is used to enable the data from different sensor devices to be accurately matched.

5. The transformer state monitoring method based on voiceprint perception according to claim 1 is characterized in that: In step 2, the signal obtained in step 1 is converted into frequency domain information using fast Fourier transform (FFT); the isolation forest algorithm is used to detect outliers in the data and filter out outliers in the signal.

6. The transformer state monitoring method based on voiceprint perception according to claim 1 is characterized in that: The T-Manba model in step 3 is implemented based on the Temporal Convolutional Network framework through a combination of linear layers, convolutional layers, and activation functions of T-MambaBlock. The linear layer is used to project the input data from a low-dimensional space to a high-dimensional space to enhance the feature expression capability. The convolutional layer is used to extract local features and capture the spatial pattern in the input data through convolution operations. The activation function introduces nonlinearity. T-MambaBlock combines the state space model to describe the state of the system through a set of state variables, and describes the evolution of the state and the generation of observation values ​​through state transfer equations and observation equations to capture long-range dependencies. The state update equation of the state space model is: h(t)=A·h(t-1)+B·x(t)+w(t) The output equation is: y(t)=C·h(t)+D·x(t) Among them, h(t) is the hidden state vector at time t; A is the state transfer matrix, B is the input control matrix, x(t) is the input vector at time t, and w(t) is the process noise vector.

7. A transformer status monitoring system based on voiceprint perception, characterized in that: include: A data collection layer, which is used to collect the acoustic wave signal and vibration signal of the transformer, and pre-process the collected signals and perform feature extraction; The edge computing layer is used to build a Mamba model and train the Mamba model using known signal data of the transformer in normal operation; convert the signal obtained by the data collection layer into frequency domain information, detect and mark the abnormal values ​​in the signal, and then input them into the trained Mamba model to identify the abnormal state of the transformer; And a software application layer, wherein the software application layer is used to output the operation results of the edge computing layer to the outside.

8. The transformer state monitoring system based on voiceprint perception according to claim 7 is characterized in that: In the data collection layer, several MEMS microphones are used to collect sound wave signals of the transformer and the surrounding environment. Several MEMS microphones are arranged in an equidistant polygon shape around the surface of the transformer; several acceleration sensors are used to collect transformer vibration data. Several acceleration sensors are arranged in an equidistant polygon shape around the surface of the transformer.

9. The transformer state monitoring system based on voiceprint perception according to claim 7 is characterized in that: The edge computing layer includes a memory, and the edge computing layer stores the collected data in the memory.

10. A computer-readable medium, characterized in that The computer-readable medium stores a computer program, and when the computer program is called and executed by a computer, the transformer state monitoring method based on voiceprint perception according to any one of claims 1 to 6 is implemented.

Images