Cable vibration event identification method and system based on deep neural network

Abstract

The invention discloses a cable vibration event identification method and system based on a deep neural network, and belongs to the field of cable vibration identification, and the method comprises the steps: obtaining an original vibration signal sequence, carrying out the preprocessing of the original vibration signal sequence, and obtaining a preprocessed vibration signal sequence, and the preprocessing comprises the denoising and filtering; the feature set and the multi-scale feature set are fused to form a comprehensive feature set, and the comprehensive feature set is used for enhancing the representation ability and discrimination of the features; vibration signals are input into the vibration event recognition model in a flow mode, parameters of the vibration event recognition model are dynamically updated through a sliding window and an incremental learning mechanism, an online recognition framework is constructed, and real-time recognition and early warning of vibration events are achieved.

Description

Technical Field

[0001] The present invention belongs to the field of cable vibration identification, and in particular relates to a cable vibration event identification method and system based on a deep neural network. Background Art

[0002] In the power system, the vibration problem of cables has always been a difficult technical problem. During operation, cables are affected by various internal and external factors, generating vibrations of different frequencies and amplitudes. These vibrations may cause mechanical damage to the cables, insulation aging, and even cause serious safety accidents. Traditional cable vibration detection methods, such as manual inspections and regular inspections, have problems such as low efficiency and poor accuracy, and are difficult to meet the needs of real-time monitoring and rapid response.

[0003] Moreover, cable vibration signals have complex characteristics such as non-stationary, nonlinear, and multi-scale. How to extract effective features from these complex signals is a key technical problem that needs to be solved urgently. In addition, there are various types of cable vibration events, and different types of vibration events have different degrees of impact on cables. How to accurately classify them while identifying them is also a difficult problem that needs to be overcome. Summary of the invention

[0004] In order to solve the above technical problems, the present invention provides a cable vibration event recognition method based on a deep neural network, comprising:

[0005] Acquire an original vibration signal sequence of the cable, and preprocess the original vibration signal sequence to obtain a preprocessed vibration signal sequence;

[0006] Performing wavelet transform processing on the preprocessed vibration signal sequence, extracting time domain features, frequency domain features and time-frequency domain features of the preprocessed vibration signal sequence, and constructing a feature set;

[0007] Adaptively decomposing the preprocessed vibration signal sequence by empirical mode decomposition to obtain a plurality of intrinsic mode function components, extracting local features at different scales from the intrinsic mode function components to obtain a multi-scale feature set; wherein the intrinsic mode function components represent local features at different scales;

[0008] The feature set is integrated with the multi-scale feature set to form a comprehensive feature set, a training data set is constructed based on the comprehensive feature set, and the training data set is input into a convolutional neural network model for training to obtain a vibration event recognition model;

[0009] Cable vibration events are identified based on the vibration event identification model.

[0010] Preferably, the process of obtaining the preprocessed vibration signal sequence includes:

[0011] Obtain the original vibration signal sequence of the cable. For the problem of noise interference existing in the original vibration signal sequence, use a denoising algorithm to perform noise reduction processing on the original vibration signal sequence, remove the high-frequency noise components therein, and obtain the denoised vibration signal sequence;

[0012] According to the actual working environment characteristics of the cable, for the denoised vibration signal sequence, use a digital filtering algorithm to perform filtering processing, design filter parameters, perform frequency-domain filtering on the denoised vibration signal sequence, remove the useless frequency components therein, improve the signal-to-noise ratio of the vibration signal, and obtain the preprocessed vibration signal sequence.

[0013] Preferably, the process of constructing the feature set includes:

[0014] Obtain the preprocessed vibration signal sequence, perform wavelet transform processing on this sequence, and obtain a wavelet coefficient matrix;

[0015] According to the wavelet coefficient matrix, extract the time-domain features of the vibration signal, including mean, variance, and peak statistical features;

[0016] According to the wavelet coefficient matrix, extract the frequency-domain features of the vibration signal, including spectral center and spectral width morphological features;

[0017] According to the wavelet coefficient matrix, extract the time-frequency domain features of the vibration signal, including wavelet energy distribution energy features;

[0018] Fuse the extracted time-domain features, frequency-domain features, and time-frequency domain features to construct a feature set including statistical features, morphological features, and energy features.

[0019] Preferably, the process of obtaining the multi-scale feature set includes:

[0020] Use the empirical mode decomposition method to perform adaptive signal decomposition on the preprocessed vibration signal sequence, and decompose it into multiple intrinsic mode function components;

[0021] Extract local features at different scales based on the multiple intrinsic mode function components, and construct a multi-scale feature matrix based on the local features;

[0022] Use the principal component analysis method to perform dimensionality reduction on the multi-scale feature matrix to obtain the multi-scale feature set.

[0023] Preferably, the process of fusing the feature set and the multi-scale feature set to form a comprehensive feature set includes:

[0024] Obtain the given feature set and the multi-scale feature set, and analyze the feature attributes for each feature therein;

[0025] Using feature selection algorithms, we can quantitatively score each feature by evaluating its contribution and relevance to the target task.

[0026] According to the scoring results of the feature selection algorithm, the features are sorted according to their importance;

[0027] The scale threshold of the optimal feature subset is determined, and the sorted features are included in the optimal feature subset according to the importance from high to low, and are outputted after reaching the scale threshold to obtain the comprehensive feature set.

[0028] Preferably, the process of obtaining the vibration event recognition model includes:

[0029] A convolutional neural network model is adopted, and the comprehensive feature set is used as training data. The convolutional neural network model is trained by setting the number of network layers, convolution kernel size and pooling method to obtain a trained vibration event recognition model.

[0030] The trained vibration event recognition model is tested, and the performance of the trained vibration event recognition model is evaluated by calculating the accuracy, recall rate and F1 value of the model on the test set. If the performance of the vibration event recognition model reaches a preset performance threshold, it is saved as the final vibration event recognition model. Otherwise, the training step is returned to retrain the model by adjusting the network parameters and adding training data.

[0031] Preferably, the process of identifying the cable vibration event based on the vibration event identification model includes:

[0032] Acquire the real-time vibration signal data collected by the vibration sensor and convert it into a data format and feature representation suitable for input into the vibration event recognition model;

[0033] The sliding window mechanism is used to divide the continuous vibration signal into data segments of fixed length, and each data segment corresponds to one input and recognition of the model;

[0034] According to the characteristics of each data segment, reasoning is performed based on the vibration event recognition model to identify the type and occurrence probability of the vibration event corresponding to the segment;

[0035] If the identified vibration event type is a warning event and the probability of occurrence exceeds the preset threshold, a warning signal is triggered to notify relevant personnel to handle the situation;

[0036] During the recognition process, the incremental learning mechanism is used to dynamically update the model parameters, and the recognition results of each data segment are used as new training samples to fine-tune the model parameters.

[0037] On the other hand, the present invention also provides a cable vibration event recognition system based on a deep neural network, comprising:

[0038] A data acquisition module is used to acquire an original vibration signal sequence of the cable, and preprocess the original vibration signal sequence to obtain a preprocessed vibration signal sequence;

[0039] A feature set module, used to perform wavelet transform processing on the preprocessed vibration signal sequence, extract the time domain features, frequency domain features and time-frequency domain features of the preprocessed vibration signal sequence, and construct a feature set;

[0040] A fusion module is used to perform adaptive signal decomposition on the pre-processed vibration signal sequence by using empirical mode decomposition to obtain a plurality of intrinsic mode function components, extract local features at different scales from the intrinsic mode function components, and obtain a multi-scale feature set; wherein the intrinsic mode function components represent local features at different scales;

[0041] A model building module, used for fusing the feature set with the multi-scale feature set to form a comprehensive feature set, building a training data set based on the comprehensive feature set, and inputting the training data set into a convolutional neural network model for training to obtain a vibration event recognition model;

[0042] The identification module is used to identify the cable vibration event based on the vibration event identification model.

[0043] On the other hand, the present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the cable vibration event identification method based on a deep neural network when executing the computer program.

[0044] On the other hand, the present invention also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements a cable vibration event identification method based on a deep neural network.

[0045] Compared with the prior art, the present invention has the following advantages and technical effects:

[0046] The present invention discloses a method for real-time identification and early warning of cable vibration events. The method proposes a complete set of solutions to the problems of complex signals, difficult feature extraction, and high real-time requirements in cable vibration monitoring. First, the original vibration signal is preprocessed and feature extracted, and a comprehensive feature set is constructed using methods such as wavelet transform and empirical mode decomposition. Then, a convolutional neural network is trained using the comprehensive feature set to obtain a vibration event recognition model. Finally, online real-time identification and early warning of vibration signals are achieved through a sliding window and incremental learning mechanism. The present invention integrates technologies such as signal processing, feature engineering, and deep learning, and can effectively identify vibration events such as normal, loose, fatigue, and resonance of cables, thereby improving the accuracy and real-time performance of cable vibration monitoring and providing important guarantees for the safe operation of power systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0048] Figure 1 A schematic diagram of a method flow of an embodiment of the present invention;

[0049] Figure 2 Schematic diagram of the system structure of an embodiment of the present invention. DETAILED DESCRIPTION

[0050] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0051] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0052] Embodiment 1

[0053] like Figure 1-2 As shown, this embodiment provides a cable vibration event recognition method based on a deep neural network, including:

[0054] S101, acquiring a cable vibration signal to form an original vibration signal sequence. Preprocessing the original vibration signal sequence to obtain a preprocessed vibration signal sequence, wherein the preprocessing includes denoising and filtering.

[0055] Obtain the original vibration signal sequence of the cable. For the problem of noise interference existing in the original vibration signal sequence, use a denoising algorithm to perform noise reduction processing on the original vibration signal sequence, remove the high-frequency noise components therein, and obtain the denoised vibration signal sequence. According to the actual working environment characteristics of the cable, for the denoised vibration signal sequence, use a digital filtering algorithm to perform filtering processing, design appropriate filter parameters, perform frequency-domain filtering on the denoised vibration signal sequence, remove the useless frequency components therein, improve the signal-to-noise ratio of the vibration signal, and obtain the filtered vibration signal sequence. For the filtered vibration signal sequence, use the short-time Fourier transform algorithm to convert the vibration signal sequence in the time domain to the time-frequency domain, extract the time-frequency domain characteristics of the vibration signal, including characteristic parameters such as the frequency, amplitude, and phase of the vibration signal, and obtain the time-frequency domain characteristic set of the vibration signal. According to the empirical knowledge of cable fault diagnosis, establish the mapping relationship between the time-frequency domain characteristics of the vibration signal and the cable fault types, construct a cable fault diagnosis model, and use the support vector machine algorithm to train this model to obtain the trained cable fault diagnosis model. Input the time-frequency domain characteristic set of the vibration signal into the trained cable fault diagnosis model, use the classification function of the model to judge whether the current cable has a fault. If there is a fault, output the type and degree of the fault to obtain the cable fault diagnosis result. According to the cable fault diagnosis result, if the cable has a fault risk, automatically generate a cable fault warning message, and send the warning message to the relevant operation and maintenance personnel in a timely manner through the wireless communication network to prompt them to repair and maintain the cable to avoid the further deterioration of the fault. Record and store the whole process information of cable fault diagnosis and warning to form a historical database of cable fault diagnosis and warning. Through the mining and analysis of historical data, summarize the regular characteristics of cable faults to provide data support for the subsequent improvement of the cable fault diagnosis and warning system.

[0056] Specifically, obtaining the original vibration signal sequence of the cable is the first step in cable fault diagnosis. Assuming that an acceleration sensor is installed on the surface of the cable, the collected original vibration signal sequence may contain multiple frequency components, including low-frequency vibration during normal operation of the cable and high-frequency interference caused by environmental noise. In view of the noise interference problem existing in the original vibration signal sequence, a wavelet denoising algorithm is used for noise reduction. Wavelet denoising decomposes the signal into wavelet coefficients of different frequencies, and then thresholds the high-frequency coefficients to remove noise. For example, there may be high-frequency noise with a frequency of 1000Hz in the original signal. After wavelet denoising, this part of the noise is effectively suppressed, and the denoised vibration signal sequence is obtained, and the signal becomes smoother. According to the actual working environment characteristics of the cable, the denoised vibration signal sequence needs to be further filtered. Assuming that there is 50Hz power frequency interference in the cable working environment, a bandpass filter can be designed with parameters set to 20Hz to 200Hz to retain the main frequency components of the cable vibration and remove useless frequencies. Through this frequency domain filtering, the signal-to-noise ratio of the vibration signal is significantly improved. For the filtered vibration signal sequence, the short-time Fourier transform (STFT) algorithm is used to convert it to the time-frequency domain. Assume that the signal is divided into several short time periods, and Fourier transform is performed on each time period to obtain the spectrum of the time period. Through STFT, characteristic parameters such as frequency, amplitude, and phase of the vibration signal can be extracted. For example, the spectrum of a certain time period shows that the component with a frequency of 60Hz has a large amplitude and a phase of 30 degrees. These characteristic parameters constitute the time-frequency domain feature set of the vibration signal.

[0057] S102, using wavelet transform to process the preprocessed vibration signal sequence, extracting time domain features, frequency domain features and time-frequency domain features of the preprocessed vibration signal sequence, and constructing a feature set, wherein the feature set includes statistical features, morphological features and energy features.

[0058] Obtain the preprocessed vibration signal sequence, perform wavelet transform on the sequence, and obtain the wavelet coefficient matrix. According to the wavelet coefficient matrix, extract the time domain features of the vibration signal, including statistical features such as mean, variance, and peak. According to the wavelet coefficient matrix, extract the frequency domain features of the vibration signal, including morphological features such as spectrum center and spectrum width. According to the wavelet coefficient matrix, extract the time and frequency domain features of the vibration signal, including energy features such as wavelet energy distribution. Fuse the extracted time domain features, frequency domain features, and time and frequency domain features to construct a feature set containing statistical features, morphological features, and energy features. Use principal component analysis to reduce the dimension of the feature set and obtain a feature subset after dimension reduction. Input the feature subset after dimension reduction into the support vector machine model for training and classification, determine the fault type to which the vibration signal belongs, and output the fault diagnosis result.

[0059] Specifically, after obtaining the preprocessed vibration signal sequence, wavelet transform processing is first performed. Wavelet transform is a time-frequency analysis method that can decompose the signal into components of different frequencies and is suitable for the analysis of non-stationary signals. Suppose there is a preprocessed vibration signal sequence with a length of 1024 data points. Select a suitable wavelet basis function, such as Daubechies wavelet, and perform multi-scale decomposition to obtain a wavelet coefficient matrix. This matrix contains detail coefficients and approximate coefficients at different scales, reflecting the characteristics of the signal in different frequency bands. According to the wavelet coefficient matrix, the time domain characteristics of the vibration signal are extracted.

[0060] S103, using empirical mode decomposition to perform adaptive signal decomposition on the preprocessed vibration signal sequence to obtain multiple intrinsic mode function components, wherein the intrinsic mode function components represent local features at different scales. Local features at different scales are extracted from the intrinsic mode function components to obtain a multi-scale feature set.

[0061] The vibration signal sequence is obtained, preprocessed, and noise interference is removed to obtain the preprocessed vibration signal sequence. The empirical mode decomposition method is used to perform adaptive signal decomposition on the preprocessed vibration signal sequence, and decompose it into multiple intrinsic mode function components, each of which represents local features at different scales. Local features at different scales are extracted from the decomposed intrinsic mode function components, including time domain statistical features, frequency domain features, and time-frequency domain features, to form a multi-scale local feature set. According to the local features extracted at each scale, a multi-scale feature matrix is ​​constructed, each row represents a feature vector at a scale, and each column represents a feature dimension. The principal component analysis method is used to reduce the dimension of the multi-scale feature matrix, and a low-dimensional feature subset that best represents the original signal features is extracted to reduce feature redundancy. The extracted multi-scale low-dimensional features are input into the support vector machine classifier for training to establish a fault diagnosis model for vibration signals. Using the trained fault diagnosis model, the newly collected vibration signal is preprocessed and feature extracted, and then input into the model for fault type identification to achieve automatic fault diagnosis of the equipment.

[0062] Specifically, obtaining the vibration signal sequence is the starting point of fault diagnosis, which is usually obtained through sensors installed on the equipment. Assuming that a rotating machine is monitored, the original vibration signal collected by the sensor contains the operating status information of the equipment, but is also mixed with environmental noise and the noise of the sensor itself. In order to ensure the accuracy of subsequent analysis, the signal needs to be preprocessed first to remove these noise interferences. Preprocessing methods can include filtering, demeaning, normalization, etc. For example, a low-pass filter is used to filter out high-frequency noise to ensure that the main frequency components of the signal are within the range of interest. After the preprocessed vibration signal sequence, the empirical mode decomposition (EMD) method is then used for adaptive signal decomposition. EMD is a method that decomposes a complex signal into several intrinsic mode functions (IMFs), each of which represents the local characteristics of the signal at different scales. Assume that the preprocessed signal is decomposed into five IMF components, each of which reflects the vibration characteristics in a different frequency range. For example, the first IMF component may mainly contain high-frequency impact components, while the last IMF component may reflect low-frequency periodic fluctuations. From these IMF components, local features at different scales can be extracted. Time domain statistical features include mean, variance, peak, etc., which reflect the overall statistical characteristics of the signal. For example, a large mean of an IMF component may indicate that there is continuous vibration energy at this scale. Frequency domain features such as spectrum center and spectrum width describe the distribution characteristics of the signal in the frequency domain. Assuming that the spectrum center of an IMF component is offset, it may indicate that the frequency component at this scale has changed. Time-frequency domain features such as wavelet energy distribution provide information about the signal in the two-dimensional space of time and frequency, which helps to capture transient events. These features are integrated to form a multi-scale local feature set. Assuming that 10 features are extracted from each IMF component, the five IMF components constitute a 50-dimensional feature vector. These feature vectors are arranged by scale to construct a multi-scale feature matrix, where each row represents a feature vector at a scale and each column represents a feature dimension. Since the feature matrix has a high dimension and may contain redundant information, principal component analysis (PCA) is used for dimensionality reduction. PCA maps the original features to a new coordinate system through linear transformation and selects several principal components with the largest variance as low-dimensional feature subsets. Assuming that through PCA analysis, it is found that the cumulative variance contribution rate of the first 15 principal components reaches 95%, these 15 principal components are selected as the feature subset after dimensionality reduction, which effectively reduces feature redundancy.

[0063] S104: Fusing the feature set with the multi-scale feature set to form a comprehensive feature set, where the comprehensive feature set is used to enhance the representation capability and discrimination of features.

[0064] Based on the feature set and the multi-scale feature set, a feature fusion algorithm is used to obtain a comprehensive feature set. By scaling the feature set, a multi-scale feature set is obtained. Based on the feature attributes of the feature set, a feature selection algorithm is used to determine the optimal feature subset. If the dimension of the comprehensive feature set is too high, a dimensionality reduction algorithm is used to obtain a comprehensive feature set after dimensionality reduction. Based on the comprehensive feature set, a support vector machine algorithm is used to train a classification model. The classification model is used to predict the test sample and determine its category. Based on the classification results, the discrimination and recognition accuracy of the test sample are obtained to evaluate the model performance.

[0065] Specifically, in the process of vibration signal fault diagnosis, feature fusion, multi-scale feature extraction, feature selection, dimensionality reduction and classification model training are key steps, which are analyzed and explained in detail through specific examples below. First, consider the fusion of feature sets and multi-scale feature sets. Assume that there is a set of vibration signal data. After preprocessing, time domain statistical features (such as mean, variance), frequency domain features (such as spectrum center, spectrum entropy) and time-frequency domain features (such as wavelet transform coefficients) are extracted. These features describe the characteristics of the signal from different angles. Using feature fusion algorithms, such as weighted average method or principal component analysis method, these three types of features are fused to obtain a comprehensive feature set. For example, the mean of the time domain feature is 50 and the variance is 20; the spectrum center of the frequency domain feature is 100Hz and the spectrum entropy is 0.8; the wavelet transform coefficient of the time-frequency domain feature is 0.5. By weighted average method, assuming that the weights of time domain, frequency domain and time-frequency domain features are 0.3, 0.4 and 0.3 respectively, a feature value in the comprehensive feature set may be 0.3×50+0.4×100+0.3×0.5=40.15. Next, the multi-scale feature set is obtained by scaling the feature set. Assuming that the original signal exhibits different characteristics at different time scales, the wavelet transform is used to decompose the signal into sub-signals at multiple scales, and each sub-signal corresponds to a set of features. For example, the mean of the signal at scale 1 is 60 and the variance is 25; the mean at scale 2 is 55 and the variance is 30. Combining these features to form a multi-scale feature set can more comprehensively describe the characteristics of the signal at different time scales. Then, according to the feature attributes of the feature set, the feature selection algorithm is used to determine the optimal feature subset. Assuming there are 10 features, the importance of each feature is evaluated by the ReliefF algorithm, and it is found that feature 1, feature 3 and feature 5 have the highest importance scores. Selecting these three features to form the optimal feature subset can effectively reduce feature redundancy and improve model training efficiency. If the dimension of the comprehensive feature set is too high, a dimensionality reduction algorithm such as principal component analysis (PCA) is used. Assuming that the comprehensive feature set has 20 features, PCA analysis shows that the first five principal components can explain 95% of the variance. Selecting these five principal components to form the comprehensive feature set after dimensionality reduction not only retains most of the information, but also reduces the computational complexity.

[0066] According to the feature attributes of the feature set, a feature selection algorithm is adopted to determine the optimal feature subset.

[0067] Obtain the given feature set. For each feature therein, analyze its feature attributes. Adopt a feature selection algorithm to quantitatively score the features by evaluating the contribution degree and correlation of each feature to the target task. According to the scoring results of the feature selection algorithm, sort the features according to their importance. Determine the scale threshold of the optimal feature subset, and incorporate the sorted features into the optimal feature subset from high to low importance until the scale threshold is reached. Input the optimal feature subset into a machine learning algorithm for modeling and prediction tasks. Evaluate the performance of the machine learning model trained based on the optimal feature subset through methods such as cross-validation. If the model performance meets the expectations, determine the final optimal feature subset; if not, return to step 2, adjust the feature selection algorithm or threshold, and re-select the optimal feature subset until the model performance meets the requirements.

[0068] Specifically, in the fields of machine learning and data analysis, feature selection is one of the key steps to improve model performance. First, obtain the given feature set, and this step usually involves data preprocessing, including data cleaning, missing value filling, etc. Suppose there is a set of operation data of mechanical equipment, and the feature set includes multiple features such as temperature, humidity, vibration frequency, and current intensity. For each feature, analyze its feature attributes, which include the type of the feature (such as numerical type, categorical type), distribution (such as normal distribution, skewed distribution), and correlation with the target variable. For example, temperature may be a continuous numerical feature, vibration frequency may be a discrete numerical feature, and equipment model may be a categorical feature. Through analysis, it is found that temperature and vibration frequency have a strong correlation with the occurrence of equipment failures. Adopt a feature selection algorithm, such as ReliefF, information gain, chi-square test, etc., to quantitatively score the features by evaluating the contribution degree and correlation of each feature to the target task. Suppose the ReliefF algorithm is used to score each feature, and it is found that the score of temperature is the highest, followed by vibration frequency, while the scores of humidity and other features are lower. The ReliefF algorithm evaluates the importance of each feature by calculating its ability to distinguish neighboring samples. According to the scoring results of the feature selection algorithm, sort the features according to their importance. Suppose the sorting result is: temperature, vibration frequency, current intensity, humidity, equipment model. This sorting helps to clarify which features are more critical for fault diagnosis. Determine the scale threshold of the optimal feature subset, which is usually based on the complexity of the model and the limitation of computing resources. Suppose the scale threshold of the optimal feature subset is set to 3, that is, the three most important features are selected. Incorporate the sorted features into the optimal feature subset from high to low importance until the scale threshold is reached. Therefore, the optimal feature subset is: temperature, vibration frequency, current intensity.

[0069] S105, constructing a labeled data set, wherein the labeled data set includes normal vibration events, loose vibration events, fatigue vibration events, and resonant vibration events. Using a convolutional neural network, the labeled data set is used for training to obtain a vibration event recognition model.

[0070] The original vibration data including normal vibration events, loose vibration events, fatigue vibration events and resonant vibration events are obtained, and the original vibration data are preprocessed to remove the noise interference components, so as to obtain the preprocessed vibration data. The preprocessed vibration data are feature extracted to extract the time domain features, frequency domain features and time-frequency domain features that can characterize the characteristics of different vibration events, and construct the vibration event feature vector. According to the constructed vibration event feature vector, the vibration data are labeled, and the categories of each vibration event are labeled to obtain the labeled data set. The convolutional neural network model is adopted, and the labeled data set is used as the training data. By setting appropriate parameters such as the number of network layers, convolution kernel size, and pooling method, the convolutional neural network model is trained to obtain the trained vibration event recognition model. A certain amount of labeled data is used as the test set to test the trained vibration event recognition model, and the performance of the vibration event recognition model is evaluated by calculating the accuracy, recall rate, F1 value and other evaluation indicators of the model on the test set. If the performance of the vibration event recognition model reaches the preset performance threshold, it will be saved as the final vibration event recognition model; otherwise, return to the training step and retrain the model by adjusting network parameters, adding training data, etc. The trained vibration event recognition model is applied to the actual vibration data analysis. According to the input vibration data, the trained model is used to predict it, identify the vibration event category to which the vibration data belongs, and realize the automatic recognition of vibration events.

[0071] Specifically, obtaining the original vibration data that includes normal vibration events, loose vibration events, fatigue vibration events, and resonant vibration events is the basis for vibration event recognition. For example, in the operation monitoring of wind turbines, vibration signals under different working conditions can be collected through acceleration sensors installed on key components. These signals include the stable vibration during normal operation, the irregular vibration when bolts are loose, the weak vibration when blades are fatigued, and the intense vibration during system resonance. Preprocessing the original vibration data to remove noise interference components is the key to improving the accuracy of subsequent feature extraction. Preprocessing methods include filtering, de-meaning, and normalization. For example, a low-pass filter is used to remove high-frequency noise, a high-pass filter is used to remove low-frequency drift, the DC component of the signal is eliminated through de-meaning, and normalization scales the signal amplitude to a unified range for subsequent processing. Extracting time-domain features, frequency-domain features, and time-frequency domain features to construct a vibration event feature vector is the core of identifying different vibration events. Time-domain features such as mean, variance, and kurtosis reflect the statistical characteristics of the signal; frequency-domain features such as spectral center and spectral entropy reveal the frequency distribution of the signal; time-frequency domain features such as wavelet transform coefficients depict the local characteristics of the signal in the time-frequency domain. For example, the time-domain mean of a normal vibration event is relatively low and stable, the spectral entropy of a loose vibration event is relatively high, and the wavelet transform coefficients of a fatigue vibration event change significantly in a specific frequency band. Based on the constructed vibration event feature vector, the vibration data is labeled to form a labeled dataset. The labeling process requires domain knowledge. For example, through expert experience or existing fault records, each vibration signal is labeled as normal, loose, fatigued, or resonant. The quality of the labeled dataset directly affects the training effect of the model. Using a convolutional neural network (CNN) model for training is an effective method for vibration event recognition using deep learning technology. CNN can automatically extract the deep features of the signal through multiple layers of convolution and pooling operations. For example, setting 3 convolutional layers with convolution kernel sizes of 3×3, 5×5, and 7×7 respectively, and using the max-pooling method to gradually extract the local and global features of the signal. The network parameters are adjusted through the backpropagation algorithm to minimize the loss of the model on the training data. Using a certain amount of labeled data as the test set to test the trained model and evaluate its performance. Evaluation metrics include accuracy, recall, F1 value, etc. For example, the test set contains 100 samples, and the model correctly identifies 90 of them, with an accuracy of 90%; among all the samples that are actually loose, the model identifies 80% of them, with a recall of 80%; the F1 value is the harmonic mean of the accuracy and recall, reflecting the comprehensive performance of the model. If the model performance does not reach the preset threshold, it is necessary to return to the training step for adjustment. For example, increasing the number of network layers to improve the model complexity, adjusting the convolution kernel size to capture more detailed features, and increasing the training data to improve the generalization ability of the model. Through multiple iterations of optimization until the model performance meets the requirements.The trained model is applied to the actual vibration data analysis to realize the automatic identification of vibration events. For example, the vibration signal of the wind turbine is monitored in real time and input into the trained CNN model, and the model outputs the category of vibration events, such as "normal", "loose", "fatigue" or "resonance". Through automatic identification, potential faults can be discovered in time, equipment damage can be avoided, and operation and maintenance efficiency can be improved. The process of feature extraction and model training aims to improve the accuracy and robustness of vibration event identification. The fusion of time domain, frequency domain and time-frequency domain features enables the model to comprehensively characterize vibration signals from multiple angles; the deep feature extraction capability of CNN further enhances the discrimination of the model. The quality of the labeled data set and the optimization of network parameters are the key factors to improve the performance of the model. In practical applications, this multi-step, multi-level feature processing and model training method can significantly improve the performance of the vibration event identification system, reduce the misdiagnosis rate, and improve the maintenance efficiency and reliability of the equipment. By continuously optimizing the feature processing and model training strategies, the intelligence level of the system can be further improved to achieve more efficient and accurate equipment health management.

[0072] The trained vibration event recognition model is applied to the actual vibration data analysis. According to the input vibration data, the trained model is used to predict it and identify the vibration event category to which the vibration data belongs.

[0073] Obtain a set of vibration data actually collected as the input data of the model. Preprocess the input vibration data, including denoising, normalization and other operations, to obtain standardized vibration data. Extract the time domain, frequency domain and other characteristic parameters based on the preprocessed vibration data. Input the extracted characteristic parameters into the pre-trained vibration event recognition model. The vibration event recognition model calculates the event category to which the vibration data is most likely to belong based on the input characteristic parameters through the internal classification algorithm. Compare the event category result output by the model with the preset event category threshold. If it is greater than the threshold, it is determined to be this category, otherwise it is determined to be an unknown category. According to the determined event category, output the vibration event recognition result corresponding to the vibration data to complete the analysis and recognition of the actual vibration data.

[0074] Specifically, a set of vibration data actually collected is obtained as the input data of the model. For example, during the operation of the wind turbine, the vibration signal is collected in real time by the acceleration sensor installed in the cabin. These signals contain the vibration information of the wind turbine under different working conditions, such as normal operation, bearing looseness, blade fatigue, and structural resonance. The input vibration data is preprocessed, including denoising, normalization and other operations to obtain standardized vibration data. Denoising can be achieved through wavelet transform or filter. For example, a low-pass filter is used to remove high-frequency noise interference and retain low-frequency vibration signals. Normalization is to adjust the amplitude range of the vibration signal to a uniform interval, such as between 0 and 1, to facilitate subsequent feature extraction and model input. Assuming that the amplitude range of the original vibration signal is -5 to 5, it is normalized to the interval of 0 to 1 through linear transformation. According to the preprocessed vibration data, its time domain, frequency domain and other characteristic parameters are extracted. Time domain features include mean, variance, peak value, etc., and frequency domain features are obtained through fast Fourier transform (FFT), such as main frequency, frequency band energy, etc. For example, for a preprocessed vibration signal, its time domain features are calculated: mean is 0.3, variance is 0.1, and peak is 0.8; the frequency domain features obtained by FFT transformation are: main frequency is 50Hz, and the energy of 0-100Hz frequency band accounts for 70%. The extracted feature parameters are input into the pre-trained vibration event recognition model. Assume that the model is a convolutional neural network (CNN) trained with a large amount of labeled data, and its input layer accepts time domain and frequency domain feature vectors. For example, the above extracted feature vector [0.3, 0.1, 0.8, 50, 70] is input into the input layer of the CNN model. Based on the input feature parameters, the vibration event recognition model calculates the event category to which the vibration data is most likely to belong through the internal classification algorithm. After multiple layers of convolution, pooling and full connection operations, the model finally outputs the probability distribution of each category. For example, the model output results are: normal vibration event probability 0.2, loose vibration event probability 0.5, fatigue vibration event probability 0.1, and resonant vibration event probability 0.2. According to the maximum probability principle, it is determined that the vibration data is most likely to belong to a loose vibration event. Compare the event category result output by the model with the preset event category threshold. If it is greater than the threshold, it is determined to be this category, otherwise it is determined to be an unknown category. The preset threshold can be 0.4, that is, only when the probability of a certain category is greater than 0.4, it is determined to be this category. In the above example, the probability of a loose vibration event is 0.5, which is greater than the threshold value of 0.4, so it is determined to be a loose vibration event. According to the determined event category, the vibration event identification result corresponding to the vibration data is output to complete the analysis and identification of the actual vibration data. For example, the system outputs "A loose vibration event is detected, and it is recommended to check the tightness of the bearing", thereby providing maintenance personnel with specific fault diagnosis information. Through the above steps, automatic identification from raw vibration data to vibration event categories is achieved.Denoising and normalization ensure the purity and consistency of the data, feature extraction captures the key information of the vibration signal, the CNN model uses this information for efficient classification, and the threshold judgment increases the reliability of the recognition results. The whole process is closely linked to ensure the accuracy and practicality of vibration event recognition. In practical applications, such a vibration event recognition system can significantly improve the efficiency and accuracy of equipment maintenance. For example, in the operation and maintenance of wind turbines, timely detection and repair of loose bearings can avoid more serious failures caused by bearing failure, extend the service life of equipment, and reduce maintenance costs. Through real-time monitoring and intelligent analysis of vibration data, accurate control of equipment status is achieved, and the reliability and safety of equipment operation are improved. In addition, during the feature extraction and model training process, adjustments and optimizations can be made according to actual needs. For example, for specific types of vibration events, more sophisticated feature extraction methods can be designed, or the network structure of the CNN model can be adjusted to improve the accuracy and robustness of recognition. Through continuous iteration and optimization, the performance of the vibration event recognition system will continue to improve, and better serve the intelligent operation and maintenance of various types of equipment. Through the above-mentioned multi-faceted analysis and examples, the specific implementation methods and technical effects of the vibration event recognition system are demonstrated. Each link supports each other and together constitutes a complete and efficient vibration event identification process, providing strong technical support for practical applications.

[0075] S106, inputting the vibration signal into the vibration event recognition model in the form of a stream, dynamically updating the vibration event recognition model parameters by using a sliding window and an incremental learning mechanism, constructing an online recognition framework, and realizing real-time recognition and early warning of vibration events.

[0076] The real-time vibration signal data collected by the vibration sensor is obtained and converted into a data format and feature representation suitable for input into the vibration event recognition model; the sliding window mechanism is used to divide the continuous vibration signal into data segments of fixed length, and each data segment corresponds to one input and recognition of the model; according to the characteristics of each data segment, the pre-trained vibration event recognition model is used for reasoning to identify the vibration event type and occurrence probability corresponding to the segment; if the identified vibration event type is a warning event and the probability of occurrence exceeds the preset threshold, a warning signal is triggered to notify relevant personnel to deal with it; during the recognition process, the incremental learning mechanism is used to dynamically update the model parameters, and the recognition results of each data segment are used as new training samples to fine-tune the model parameters; through continuous incremental learning, the vibration event recognition model can adapt to changes in the environment and equipment, and improve recognition accuracy and robustness; an online recognition framework is constructed to integrate modules such as data acquisition, feature extraction, event recognition, and incremental learning into a real-time processing pipeline to achieve real-time recognition and warning of vibration events.

[0077] Specifically, obtaining real-time vibration signal data collected by the vibration sensor is the basis of the entire recognition process. Vibration sensors are usually installed in key parts of mechanical equipment, such as bearings, motors, etc., and can monitor the vibration of the equipment in real time. For example, a centrifuge in a factory is equipped with a vibration sensor. The sensor collects vibration data at a sampling frequency of 1000 Hz per second. These data are stored in the form of time series and contain important information about the operating status of the equipment. Converting the collected vibration signal into a data format and feature representation suitable for the input model is a key step to ensure accurate model recognition. Specifically, the original vibration signal can be normalized so that its value range is between 0 and 1 to eliminate the influence of different sensor sensitivity differences. At the same time, the time domain features (such as mean, variance), frequency domain features (such as spectral energy distribution) and time-frequency domain features (such as wavelet transform coefficients) of the signal are extracted to construct a multidimensional feature vector. For example, for a 10-second vibration signal, the features of its mean of 0.5, variance of 0.2, and spectral energy mainly concentrated near 50 Hz are extracted to form a vector containing multiple eigenvalues. The sliding window mechanism is used to divide the continuous vibration signal into data segments of fixed length, and each segment corresponds to the input and recognition of the model once. Assuming that the window length is set to 1 second, the vibration data of each second is processed as an independent data segment. In this way, the continuous vibration signal is divided into multiple overlapping segments, ensuring the continuity and integrity of the signal. The pre-trained vibration event recognition model is used for reasoning to identify the vibration event type and occurrence probability corresponding to each data segment. For example, the model may output that the probability of a certain segment being a "loose vibration event" is 0.8, and the probability of being a "normal vibration event" is 0.2. The output of the model includes not only the event type, but also its probability of occurrence, which is convenient for subsequent decision-making. If the identified vibration event type is a warning event and the probability of occurrence exceeds the preset threshold, the warning signal is triggered. Assuming the preset threshold is 0.7, when a segment is identified as a "fatigue vibration event" with a probability of 0.8, the system will automatically trigger the warning signal, and notify the relevant personnel to deal with it through SMS, email, or sound and light alarm, so as to avoid equipment failure in time. During the recognition process, the incremental learning mechanism is used to dynamically update the model parameters, and the recognition results of each data segment are used as new training samples to fine-tune the model parameters. For example, a newly collected vibration data segment is identified as a "resonant vibration event". The system adds the segment and its label "resonance" to the training set, and updates the model parameters through a small number of iterations to gradually adapt the model to the new data distribution. Through continuous incremental learning, the vibration event recognition model can adapt to changes in the environment and equipment, and improve recognition accuracy and robustness. Assuming that after the equipment has been running for a period of time, the vibration characteristics change due to wear and tear. The incremental learning mechanism can continuously update the model so that it still maintains a high recognition rate in the new environment.An online recognition framework is constructed to integrate modules such as data acquisition, feature extraction, event recognition, and incremental learning into a real-time processing pipeline to achieve real-time recognition and early warning of vibration events. For example, a real-time processing system is designed in which the vibration sensor data is transmitted to the feature extraction module in real time through the data acquisition module, the extracted feature vector is input into the event recognition module for reasoning, and the recognition result is used to update the model parameters through the incremental learning module. The entire process is completed within milliseconds, ensuring real-time and accuracy. Through the above steps, the vibration event recognition system can efficiently and accurately identify abnormal vibrations in equipment operation, issue early warnings in a timely manner, avoid equipment failures, and improve production efficiency and safety. Each link cooperates with each other to form a complete and dynamically updated vibration event recognition and early warning system.

[0078] On the other hand, this embodiment also provides a cable vibration event recognition system based on a deep neural network, including:

[0079] A data acquisition module is used to acquire an original vibration signal sequence of the cable, and preprocess the original vibration signal sequence to obtain a preprocessed vibration signal sequence;

[0080] A feature set module, used to perform wavelet transform processing on the preprocessed vibration signal sequence, extract the time domain features, frequency domain features and time-frequency domain features of the preprocessed vibration signal sequence, and construct a feature set;

[0081] A fusion module is used to perform adaptive signal decomposition on the pre-processed vibration signal sequence by using empirical mode decomposition to obtain a plurality of intrinsic mode function components, extract local features at different scales from the intrinsic mode function components, and obtain a multi-scale feature set; wherein the intrinsic mode function components represent local features at different scales;

[0082] A model building module, used for fusing the feature set with the multi-scale feature set to form a comprehensive feature set, building a training data set based on the comprehensive feature set, and inputting the training data set into a convolutional neural network model for training to obtain a vibration event recognition model;

[0083] The identification module is used to identify the cable vibration event based on the vibration event identification model.

[0084] On the other hand, this embodiment also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the cable vibration event identification method based on a deep neural network when executing the computer program.

[0085] On the other hand, this embodiment further provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements a cable vibration event identification method based on a deep neural network.

[0086] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A cable vibration event recognition method based on deep neural network, characterized in that: include: Acquire an original vibration signal sequence of the cable, and preprocess the original vibration signal sequence to obtain a preprocessed vibration signal sequence; Performing wavelet transform processing on the preprocessed vibration signal sequence, extracting time domain features, frequency domain features and time-frequency domain features of the preprocessed vibration signal sequence, and constructing a feature set; Adaptively decomposing the preprocessed vibration signal sequence by empirical mode decomposition to obtain a plurality of intrinsic mode function components, extracting local features at different scales from the intrinsic mode function components to obtain a multi-scale feature set; wherein the intrinsic mode function components represent local features at different scales; The feature set is integrated with the multi-scale feature set to form a comprehensive feature set, a training data set is constructed based on the comprehensive feature set, and the training data set is input into a convolutional neural network model for training to obtain a vibration event recognition model; Cable vibration events are identified based on the vibration event identification model.

2. The method according to claim 1, characterized in that The process of obtaining the pre-processed vibration signal sequence comprises: The original vibration signal sequence of the cable is obtained. Aiming at the noise interference problem existing in the original vibration signal sequence, a denoising algorithm is used to perform denoising on the original vibration signal sequence, thereby removing the high-frequency noise components therein and obtaining a denoised vibration signal sequence. According to the actual working environment characteristics of the cable, a digital filtering algorithm is used to filter the denoised vibration signal sequence, and the filter parameters are designed to perform frequency domain filtering on the denoised vibration signal sequence to remove useless frequency components and improve the signal-to-noise ratio of the vibration signal to obtain the preprocessed vibration signal sequence.

3. The method according to claim 1, characterized in that The process of constructing the feature set includes: Obtaining a preprocessed vibration signal sequence, performing wavelet transform processing on the sequence, and obtaining a wavelet coefficient matrix; According to the wavelet coefficient matrix, the time domain characteristics of the vibration signal are extracted, including the mean, variance and peak statistical characteristics; According to the wavelet coefficient matrix, the frequency domain characteristics of the vibration signal are extracted, including the morphological characteristics of the spectrum center and spectrum width; According to the wavelet coefficient matrix, the time-frequency domain characteristics of the vibration signal are extracted, including the energy characteristics of the wavelet energy distribution; The extracted time domain features, frequency domain features and time-frequency domain features are fused to construct a feature set including statistical features, morphological features and energy features.

4. The method according to claim 1, characterized in that: The process of obtaining the multi-scale feature set includes: The empirical mode decomposition method is used to perform adaptive signal decomposition on the preprocessed vibration signal sequence and decompose it into multiple eigenmode function components; Extracting local features at different scales based on the multiple eigenmode function components, and constructing a multi-scale feature matrix based on the local features; The principal component analysis method is used to reduce the dimension of the multi-scale feature matrix to obtain the multi-scale feature set.

5. The method according to claim 1, characterized in that: The process of fusing the feature set with the multi-scale feature set to form a comprehensive feature set includes: Obtaining a given feature set and the multi-scale feature set, and analyzing a feature attribute of each feature therein; Using feature selection algorithms, we can quantitatively score each feature by evaluating its contribution and relevance to the target task. According to the scoring results of the feature selection algorithm, the features are sorted according to their importance; The scale threshold of the optimal feature subset is determined, and the sorted features are included in the optimal feature subset according to the importance from high to low, and are outputted after reaching the scale threshold to obtain the comprehensive feature set.

6. The method according to claim 1, characterized in that The process of obtaining the vibration event recognition model comprises: A convolutional neural network model is adopted, and the comprehensive feature set is used as training data. The convolutional neural network model is trained by setting the number of network layers, convolution kernel size and pooling method to obtain a trained vibration event recognition model. The trained vibration event recognition model is tested, and the performance of the trained vibration event recognition model is evaluated by calculating the accuracy, recall rate and F1 value of the model on the test set. If the performance of the vibration event recognition model reaches a preset performance threshold, it is saved as the final vibration event recognition model. Otherwise, the training step is returned to retrain the model by adjusting the network parameters and adding training data.

7. The method according to claim 1, characterized in that The process of identifying cable vibration events based on the vibration event identification model includes: Acquire the real-time vibration signal data collected by the vibration sensor and convert it into a data format and feature representation suitable for input into the vibration event recognition model; The sliding window mechanism is used to divide the continuous vibration signal into data segments of fixed length, and each data segment corresponds to one input and recognition of the model; According to the characteristics of each data segment, reasoning is performed based on the vibration event recognition model to identify the type and occurrence probability of the vibration event corresponding to the segment; If the identified vibration event type is a warning event and the probability of occurrence exceeds the preset threshold, a warning signal is triggered to notify relevant personnel to handle the situation; During the recognition process, the incremental learning mechanism is used to dynamically update the model parameters, and the recognition results of each data segment are used as new training samples to fine-tune the model parameters.

8. A cable vibration event recognition system based on deep neural network, characterized in that: include: A data acquisition module is used to acquire an original vibration signal sequence of the cable, and preprocess the original vibration signal sequence to obtain a preprocessed vibration signal sequence; A feature set module, used to perform wavelet transform processing on the preprocessed vibration signal sequence, extract the time domain features, frequency domain features and time-frequency domain features of the preprocessed vibration signal sequence, and construct a feature set; A fusion module is used to perform adaptive signal decomposition on the pre-processed vibration signal sequence by using empirical mode decomposition to obtain a plurality of intrinsic mode function components, extract local features at different scales from the intrinsic mode function components, and obtain a multi-scale feature set; wherein the intrinsic mode function components represent local features at different scales; A model building module, used for fusing the feature set with the multi-scale feature set to form a comprehensive feature set, building a training data set based on the comprehensive feature set, and inputting the training data set into a convolutional neural network model for training to obtain a vibration event recognition model; The identification module is used to identify the cable vibration event based on the vibration event identification model.

9. An electronic device comprising a memory, a processor, and a computing program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the method described in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method described in any one of claims 1 to 7 is implemented.

Images