Generator Fault Diagnosis Method and System Based on Vibration Trend Prediction

Abstract

The present application relates to the technical field of fault diagnosis, and discloses a generator fault diagnosis method and system based on vibration trend prediction. The method includes: training a prediction model for the regularized vibration sequence through an extreme learning machine algorithm to obtain a vibration trend prediction model, the model including an input layer, a hidden layer and an output layer; performing an adaptive adjustment process on the vibration trend prediction model through a sum of squared errors evaluation function to obtain an optimized prediction model, wherein when the error ratio of the number of cycles is greater than two, the regular function value is set to zero; when the error ratio of the number of cycles is less than or equal to two, the regular function value is equal to the product of the previous regular function value and the natural logarithm of the error ratio; comparing and analyzing the output result of the optimized prediction model with the real-time monitoring data to obtain a generator fault diagnosis result. The present application improves the efficiency and accuracy of generator fault diagnosis based on vibration trend prediction.

Description

Technical Field

[0001] This application relates to the field of fault diagnosis, and particularly to a generator fault diagnosis method and system based on vibration trend prediction. Background Art

[0002] As a core device in the power system, the operating state of a generator is directly related to the safety and reliability of power supply. Currently, the fault diagnosis of generators is mainly achieved by online monitoring of vibration data. Common diagnosis methods include spectrum analysis, wavelet transform, empirical mode decomposition, etc. These methods analyze the vibration signals in the time domain and frequency domain, extract characteristic parameters, and establish a fault diagnosis model. With the development of artificial intelligence technology, fault diagnosis methods based on machine learning have gradually been applied to generator condition monitoring. Algorithms such as support vector machines and neural networks are used for pattern recognition and fault classification of vibration signals.

[0003] However, the existing fault diagnosis methods have some limitations in processing generator vibration signals. Firstly, vibration signals often have non-linear and non-stationary characteristics, and traditional signal processing methods are difficult to effectively extract the fault features therein. Secondly, most diagnostic algorithms only focus on the current vibration state and lack the ability to predict the development trend of faults, resulting in the inability to detect potential fault hazards early. In addition, in practical applications, due to the change of equipment operating conditions and the interference of environmental noise, the reliability and accuracy of diagnostic results are also affected. Summary of the Invention

[0004] This application provides a generator fault diagnosis method and system based on vibration trend prediction, which is used to solve the technical problem in the prior art that there is a lack of predictive analysis ability for vibration trends and it is difficult to achieve early warning of faults.

[0005] First aspect, the present application provides a generator fault diagnosis method based on vibration trend prediction. The generator fault diagnosis method based on vibration trend prediction includes: performing real-time acquisition and processing on the generator vibration signal to obtain historical vibration signal sequence data; performing variational mode decomposition processing on the historical vibration signal sequence data to obtain a plurality of intrinsic mode function subsequences; performing regularization processing on the intrinsic mode function subsequences, wherein the subsequences are processed in a dual mode through a preset regularization function to obtain a regularized vibration sequence. The dual mode processing includes: when the absolute value of the time error is less than 5% of the regularization function value, the regularization function value is processed by multiplying the time error by the regularization function period value; when the absolute value of the time error is greater than or equal to 5% of the regularization function value, the regularization function value is processed by adding the time error to the regularization function period value; training a prediction model for the regularized vibration sequence through an extreme learning machine algorithm to obtain a vibration trend prediction model, wherein the prediction model includes an input layer, a hidden layer, and an output layer, and the weight values of the input layer and the neuron thresholds of the hidden layer are randomly generated; performing adaptive adjustment processing on the vibration trend prediction model through a sum of squared errors evaluation function to obtain an optimized prediction model, wherein when the error ratio of the number of cycles is greater than two, the regularization function value is set to zero; when the error ratio of the number of cycles is less than or equal to two, the regularization function value is equal to the product of the previous regularization function value and the natural logarithm of the error ratio; comparing and analyzing the output result of the optimized prediction model with the real-time monitoring data to obtain a generator fault diagnosis result.

[0006] Second aspect, the present application provides a generator fault diagnosis system based on vibration trend prediction. The generator fault diagnosis system based on vibration trend prediction includes:

[0007] An acquisition module, configured to perform real-time acquisition and processing on the generator vibration signal to obtain historical vibration signal sequence data;

[0008] A decomposition module, configured to perform variational mode decomposition processing on the historical vibration signal sequence data to obtain a plurality of intrinsic mode function subsequences;

[0009] A processing module, configured to perform regularization processing on the intrinsic mode function subsequences, wherein the subsequences are processed in a dual mode through a preset regularization function to obtain a regularized vibration sequence. The dual mode processing includes: when the absolute value of the time error is less than 5% of the regularization function value, the regularization function value is processed by multiplying the time error by the regularization function period value; when the absolute value of the time error is greater than or equal to 5% of the regularization function value, the regularization function value is processed by adding the time error to the regularization function period value;

[0010] A training module, configured to train a prediction model for the regularized vibration sequence through an extreme learning machine algorithm to obtain a vibration trend prediction model, where the prediction model includes an input layer, a hidden layer, and an output layer, and the weight values of the input layer and the neuron thresholds of the hidden layer are randomly generated;

[0011] An adjustment module, configured to adaptively adjust the vibration trend prediction model through a sum of squared errors evaluation function to obtain an optimized prediction model, where when the error ratio of the number of cycles is greater than two, the regular function value is set to zero; when the error ratio of the number of cycles is less than or equal to two, the regular function value is equal to the product of the previous regular function value and the natural logarithm of the error ratio;

[0012] A comparison module, configured to compare and analyze the output result of the optimized prediction model with real-time monitoring data to obtain a generator fault diagnosis result.

[0013] In the technical solution provided by this application, by collecting and processing the vibration signals of the generator in real time, complete historical vibration signal sequence data is obtained, ensuring the accuracy and timeliness of the original data. The variational mode decomposition processing method is used to decompose the historical vibration signal sequence data to obtain multiple intrinsic mode function subsequences, effectively separating different frequency components in the signal and enhancing the analysis ability of complex vibration signals. By regularizing the intrinsic mode function subsequences and introducing a dual-mode processing mechanism, different calculation methods are adopted according to the magnitude of the absolute value of the time error, effectively reducing the influence of non-periodic and non-linear characteristics on the prediction accuracy. The extreme learning machine algorithm is used to train the prediction model for the regularized vibration sequence. This algorithm simplifies the training process of the traditional neural network by randomly generating the weight values of the input layer and the neuron thresholds of the hidden layer, improving the calculation efficiency. The sum of squared errors evaluation function is introduced to adaptively adjust the vibration trend prediction model, and the regular function value is dynamically adjusted according to the magnitude of the error ratio of the number of cycles, enhancing the adaptability of the model to changes in the vibration trend. Finally, by comparing and analyzing the output result of the optimized prediction model with real-time monitoring data, abnormal changes in the operating state of the generator can be detected in a timely manner, the fault type can be accurately judged, early warning and precise diagnosis of generator faults are realized, and the safety and reliability of equipment operation are improved. Description of the Drawings

[0014] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0015] Figure 1Schematic diagram of an embodiment of the generator fault diagnosis method based on vibration trend prediction in the embodiments of the present application;

[0016] Figure 2 Schematic diagram of an embodiment of the generator fault diagnosis system based on vibration trend prediction in the embodiments of the present application. Detailed implementation manners

[0017] The embodiments of the present application provide a generator fault diagnosis method and system based on vibration trend prediction. Terms such as "first", "second", "third", "fourth", etc. (if any) in the specification, claims and above-mentioned drawings of the present application are used to distinguish similar objects, and do not necessarily describe a specific order or sequence. It should be understood that such data can be interchanged under appropriate circumstances so that the embodiments described here can be implemented in an order other than that illustrated or described here. In addition, the terms "comprising" or "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units does not necessarily limit to those clearly listed steps or units, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0018] For ease of understanding, the following describes the specific process of the embodiments of the present application. Please refer to Figure 1 An embodiment of the generator fault diagnosis method based on vibration trend prediction in the embodiments of the present application includes:

[0019] Step S101: Collect and process the vibration signals of the generator in real time to obtain historical vibration signal sequence data;

[0020] Step S102: Perform variational mode decomposition on the historical vibration signal sequence data to obtain multiple intrinsic mode function subsequences;

[0021] Step S103: Regularize the intrinsic mode function subsequences. Specifically, perform dual-mode processing on the subsequences through a preset regular function to obtain a regularized vibration sequence. The dual-mode processing includes: when the absolute value of the time error is less than 5% of the regular function value, the regular function value is equal to the product of the time error and the regular function period value for processing; when the absolute value of the time error is greater than or equal to 5% of the regular function value, the regular function value is equal to the sum of the time error and the regular function period value for processing;

[0022] Step S104: Train a prediction model for the regularized vibration sequence through the extreme learning machine algorithm to obtain a vibration trend prediction model. The prediction model includes an input layer, a hidden layer, and an output layer. The weight values of the input layer and the neuron thresholds of the hidden layer are randomly generated;

[0023] Step S105: The vibration trend prediction model is adaptively adjusted through the sum of squared errors evaluation function to obtain an optimized prediction model. Among them, when the error ratio of the number of cycles is greater than two, the value of the law function is set to zero; when the error ratio of the number of cycles is less than or equal to two, the value of the law function is equal to the product of the previous law function value and the natural logarithm of the error ratio.

[0024] Step S106: Compare and analyze the output result of the optimized prediction model with the real-time monitoring data to obtain the generator fault diagnosis result.

[0025] It can be understood that the execution subject of this application can be a generator fault diagnosis system based on vibration trend prediction, or a terminal or a server. Specifically, it is not limited here. In this embodiment of the application, the server is used as the execution subject for illustration.

[0026] Specifically, during the real-time acquisition of the generator vibration signal, a high-precision vibration sensor is used to collect the vibration signal generated during the operation of the generator. The sampling frequency is set to 10 kHz to ensure the acquisition accuracy of the signal. The collected original vibration signal is subjected to wavelet denoising processing to remove the influence of environmental noise and interference signals. Feature parameters such as the root mean square value, peak value, and kurtosis of the vibration signal are obtained through time-domain feature extraction. At the same time, frequency-domain conversion is performed to obtain a frequency-domain feature sequence. The time-domain and frequency-domain features are fused to form a complete historical vibration signal sequence data. When performing variational mode decomposition on the obtained historical vibration signal sequence data, first determine that the decomposition layer is 5 layers, and perform multi-scale decomposition on the signal. By constructing a quadratic penalty term and a Lagrange multiplier, an augmented Lagrangian expression is established, and the alternating direction multiplier method is used for iterative optimization to obtain vibration modes at different scales. Each mode represents an inherent oscillation component in the original signal, and these components are called intrinsic mode function subsequences. Hilbert transform is performed on each subsequence to obtain the instantaneous frequency feature, thereby realizing the complete decomposition of the vibration signal.

[0027] In the regularization processing stage, a regularization function is set for each intrinsic mode function subsequence, and a dual-mode processing method is used for regularization. When the absolute value of the calculated time error is less than 5% of the regularization function value, the regularization function value is set to the product of the time error and the regularization function period value; when the absolute value of the time error is less than or equal to 5%, the regularization function value is set to the sum of the time error and the regularization function period value. This dual-mode processing method effectively reduces the influence of non-periodic and non-linear characteristics on the prediction accuracy, and finally obtains a regularized vibration sequence. The extreme learning machine algorithm is used to process the regularized vibration sequence, and a prediction model including an input layer, a hidden layer, and an output layer is constructed. The number of nodes in the input layer is set to 12, corresponding to the feature dimension of the regularized vibration sequence, the number of nodes in the hidden layer is set to 24, and the number of nodes in the output layer is 1. The weight values of the input layer and the neuron thresholds of the hidden layer are initialized by random generation, and the range is set between [-1, 1]. The output of the hidden layer is non-linearly mapped through an activation function, and the weight matrix of the output layer is calculated. Finally, the vibration trend prediction model is constructed. In the adaptive adjustment stage of the prediction model, a sum of squared errors evaluation function is introduced to evaluate the prediction results of the model. Calculate the error between the predicted value and the actual value of the model, count the number of cycles, and calculate the error ratio. When the error ratio of the number of cycles is greater than 2, the regularization function value is directly set to 0; when the error ratio of the number of cycles is less than or equal to 2, the regularization function value is updated to the product of the previous regularization function value and the natural logarithm of the error ratio. This adaptive adjustment mechanism can adjust the model parameters in a timely manner according to the change of the prediction error, and obtain an optimized prediction model with better prediction performance. Compare the output result of the optimized prediction model with the real-time collected monitoring data, calculate the feature deviation, divide the deviation interval, perform fault type matching and verification, and finally obtain a diagnosis result including the fault type, fault probability, and severity level.

[0028] Taking a certain generator as an example, during normal operation, the amplitude of the original vibration signal collected by the vibration sensor fluctuates between 0.1 mm and 0.5 mm. After variational mode decomposition, 5 intrinsic mode function subsequences are obtained, and the main vibration components are concentrated in the second and third subsequences. These subsequences are regularized. When the absolute value of the time error at a certain moment is 0.03 (less than 5% of the regularization function value 0.8), the product mode is used for processing; when the absolute value of the time error is 0.04 (equal to 5%), the sum mode is used for processing. After training the prediction model through the extreme learning machine algorithm, the calculated error ratio of the number of cycles is 2.5, and at this time, the regularization function value is set to zero; when the error ratio drops to 1.8, if the previous regularization function value is 0.6, the new regularization function value is updated to 0.35. Finally, the deviation between the vibration trend output by the prediction model and the real-time monitoring data is within 3%, accurately identifying the misalignment fault of the generator rotor, with a fault probability of 85% and a medium severity level.

[0029] In the embodiment of the present application, by collecting and processing the vibration signals of the generator in real time, a complete historical vibration signal sequence data is obtained, ensuring the accuracy and timeliness of the original data. The variational mode decomposition processing method is used to decompose the historical vibration signal sequence data, and multiple intrinsic mode function subsequences are obtained, effectively separating different frequency components in the signal and enhancing the analysis ability of complex vibration signals. By regularizing the intrinsic mode function subsequences and introducing a dual-mode processing mechanism, different calculation methods are adopted according to the magnitude of the absolute value of the time error, effectively reducing the influence of non-periodic and non-linear characteristics on the prediction accuracy. The extreme learning machine algorithm is used to train the prediction model for the regularized vibration sequence. This algorithm simplifies the training process of the traditional neural network by randomly generating the input layer weight values and the hidden layer neuron thresholds, improving the calculation efficiency. An error sum of squares evaluation function is introduced to adaptively adjust the vibration trend prediction model, and the regular function value is dynamically adjusted according to the magnitude of the error ratio of the number of cycles, enhancing the adaptability of the model to changes in the vibration trend. Finally, by comparing and analyzing the output results of the optimized prediction model with the real-time monitoring data, abnormal changes in the operating state of the generator can be detected in a timely manner, the fault type can be accurately judged, early warning and precise diagnosis of generator faults are realized, and the safety and reliability of equipment operation are improved.

[0030] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0031] (1) Set the sampling frequency of the generator vibration signal to obtain the vibration signal sampling parameters, and perform the Nyquist sampling theorem test on the vibration signal sampling parameters to obtain the effective sampling parameters;

[0032] (2) Collect the vibration signal based on the effective sampling parameters to obtain the original vibration signal data, and perform wavelet denoising processing on the original vibration signal data to obtain the denoised vibration signal data;

[0033] (3) Extract the time-domain features from the denoised vibration signal data to obtain the time-domain feature sequence, and perform frequency-domain conversion processing on the time-domain feature sequence to obtain the frequency-domain feature sequence;

[0034] (4) Perform feature fusion processing on the time-domain feature sequence and the frequency-domain feature sequence to obtain the fusion feature sequence, and perform normalization processing on the fusion feature sequence to obtain the standardized feature sequence;

[0035] (5) Divide the standardized feature sequence by a sliding time window to obtain time series segments, and perform data integrity checks on the time series segments to obtain the historical vibration signal sequence data.

[0036] Specifically, the sampling frequency of the vibration sensor is reasonably set. According to the characteristics of the generator vibration signal, the sampling frequency is generally set between 1 kHz and 20 kHz. The specific setting needs to meet the requirements of the Nyquist sampling theorem, that is, the sampling frequency should be no less than twice the highest frequency of the signal. Through the spectral analysis of the vibration signal, it is found that the main vibration frequency components of the generator are distributed in the range of 10 Hz to 2 kHz. Therefore, choosing a sampling frequency of 10 kHz fully meets the requirements of the sampling theorem, ensuring the integrity and accuracy of the signal. When collecting vibration signals, a highly sensitive piezoelectric acceleration sensor is used, and the sensor is installed at key parts such as the generator bearing seat and the machine base. The sensor converts the collected mechanical vibration signal into an electrical signal, which is amplified and filtered through a signal conditioning circuit, and then undergoes analog-to-digital conversion through a high-precision data acquisition card to obtain the original digital vibration signal data. To eliminate the influence of measurement noise and environmental interference, the wavelet denoising processing method is used for the original signal. The db4 wavelet basis function is selected, the signal is decomposed into 4 layers, the wavelet coefficients are processed through a soft threshold function, and finally the denoised vibration signal data is reconstructed.

[0037] Time-domain feature extraction is performed on the denoised vibration signal data, and statistical feature parameters including root mean square value, peak value, kurtosis factor, margin factor, etc. are calculated. The root mean square value reflects the energy magnitude of the vibration, the peak value represents the maximum amplitude of the vibration, the kurtosis factor describes the pulse characteristics of the signal, and the margin factor reflects the steepness of the waveform. These time-domain features form a time-domain feature sequence, which can describe the characteristics of the vibration signal from different angles. At the same time, a fast Fourier transform is performed on the vibration signal to obtain a frequency-domain feature sequence, including feature parameters such as the main frequency component in the spectrum, the frequency band energy distribution, and the harmonic components. To comprehensively reflect the characteristics of the vibration signal, the time-domain feature sequence and the frequency-domain feature sequence need to be fused. The feature-level fusion method is adopted. First, weight allocation is performed on different feature parameters, with larger weights assigned to important features and smaller weights assigned to secondary features, and a fused feature sequence is obtained through weighted combination. Then, normalization processing is performed on the fused feature sequence to unify the feature parameters with different dimensions into the [0, 1] interval to obtain a standardized feature sequence, which is convenient for subsequent data processing and analysis.

[0038] The standardized feature sequence is divided using a sliding time window. The window length is set to 1 second, and the sliding step is 0.1 second, which not only ensures the continuity of the data but also can timely reflect the changes in vibration characteristics. Integrity checks are performed on the data within each time window, and the segments with missing or abnormal data are removed, and finally a complete and reliable historical vibration signal sequence data is obtained.

[0039] Taking a large steam turbine generator set as an example, when operating at the rated speed of 3000 rpm, vibration sensors are installed on the bearing pedestal for data acquisition. The sampling frequency is set at 10 kHz. Since the highest vibration frequency of the generator is about 2 kHz, the Nyquist sampling theorem is satisfied (10 kHz > 2×2 kHz). The amplitude of the collected original vibration signal is between 0.1 mm and 0.5 mm. After wavelet denoising, the noise amplitude is reduced by about 80%. The root mean square value of 0.25 mm, the peak value of 0.48 mm, the kurtosis factor of 3.2, and the margin factor of 4.5 are obtained by time-domain feature extraction; the frequency-domain analysis shows that the amplitude of the rotating frequency component (50 Hz) is 0.15 mm, and there are 2x and 3x frequency harmonic components. When fusing these characteristic parameters, a weight of 0.4 is assigned to the amplitude-related characteristics, and a weight of 0.6 is assigned to the frequency-related characteristics. After normalization, the standardized characteristic values are all between 0 and 1. A 1-second time window (including 10,000 sampling points) is used for division, and 9000 sampling points overlap between adjacent windows. The generated historical vibration signal sequence contains complete time-domain and frequency-domain characteristic information.

[0040] In a specific embodiment, the process of executing step S102 may specifically include the following steps:

[0041] (1) Perform data segmentation processing on the historical vibration signal sequence data to obtain multiple vibration signal sub-segments, and perform spectral analysis on the multiple vibration signal sub-segments to obtain spectral characteristic data;

[0042] (2) Calculate the number of variational mode decomposition layers for the spectral characteristic data to obtain the decomposition layer parameter, and estimate the central frequency of the decomposition layer parameter to obtain the central frequency values of each layer;

[0043] (3) Set the bandwidth parameter for each layer's central frequency value to obtain the decomposition bandwidth parameter, and construct the quadratic penalty term for the decomposition bandwidth parameter to obtain the penalty constraint condition;

[0044] (4) Construct the Lagrangian operator for the penalty constraint condition to obtain the augmented Lagrangian expression, and perform alternating direction multiplier iteration on the augmented Lagrangian expression to obtain the decomposition coefficient sequence;

[0045] (5) Perform Hilbert transform processing on the decomposition coefficient sequence to obtain the instantaneous frequency sequence, and perform subsequence reconstruction on the instantaneous frequency sequence to obtain multiple intrinsic mode function subsequences.

[0046] Specifically, the processing of vibration signal sequence data first requires data segmentation. The long-time sequence is divided into multiple sub-segments at fixed time intervals. The length of each sub-segment is set to 1024 sampling points, and 50% of the data points overlap between adjacent sub-segments to ensure signal continuity. Perform a fast Fourier transform on each vibration signal sub-segment to obtain spectral data, and extract characteristic parameters such as frequency components, amplitudes, and phases from it to form a complete spectral feature data set. Based on the obtained spectral feature data, calculate the optimal decomposition layer number of variational mode decomposition. By analyzing the number of main frequency components and energy distribution in the spectrum, initially determine the range of the decomposition layer number, generally set between 3 and 8 layers. Estimate the central frequency for each layer. The estimation method is based on the principle of spectral energy weighted average, and the frequency region with larger spectral energy is used as the central frequency of each layer. For example, for generator vibration signals, the central frequency of the first layer usually corresponds to the fundamental frequency (rotation frequency) component, and subsequent layers correspond to higher harmonic and other characteristic frequency components.

[0047] For the central frequency values of each layer, set the corresponding bandwidth parameters. The bandwidth parameters determine the frequency range of each layer of decomposition and directly affect the decomposition accuracy and effect. Usually, the bandwidth is set to 20%-30% of the central frequency, so that there is an appropriate frequency band overlap between adjacent layers while maintaining good resolution ability. After determining the bandwidth parameters, construct a quadratic penalty term to constrain the bandwidth of each modal component. The construction of the quadratic penalty term takes into account the constraints on the frequency axis and the smoothness requirements on the time axis to form a complete penalty constraint condition. Based on the penalty constraint condition, construct a Lagrangian operator. The Lagrangian operator transforms the original constrained optimization problem into an unconstrained optimization problem for easy solution. By introducing Lagrange multipliers, combine the original objective function and constraint conditions to form an augmented Lagrangian expression. Use the alternating direction multiplier method to iteratively solve the augmented Lagrangian expression. Update the modal function, central frequency, and Lagrange multiplier separately in each iteration until convergence to obtain the final decomposition coefficient sequence.

[0048] Perform Hilbert transform processing on the obtained decomposition coefficient sequence. This transform can extract the instantaneous frequency characteristics of the signal. The Hilbert transform first converts a real signal into an analytic signal and then calculates the derivative of the phase function to obtain the instantaneous frequency. According to the characteristics of the instantaneous frequency sequence, each subsequence is reconstructed, and finally a complete set of intrinsic mode function subsequences is obtained. Taking the vibration signal processing of a certain generator set as an example, the 10-second vibration data collected contains a total of 100,000 sampling points, which are divided into 195 sub-segments, each with 1024 points, and the adjacent sub-segments overlap by 512 points. Perform spectral analysis on the first sub-segment and find that the main frequency components are distributed at positions such as 50 Hz (fundamental frequency), 100 Hz (second harmonic), 150 Hz (third harmonic), 200 Hz (fourth harmonic), and 300 Hz (bearing fault characteristic frequency). Based on the spectral analysis results, set the number of layers of variational mode decomposition to 5 layers, and the center frequencies of each layer are estimated to be 50 Hz, 100 Hz, 150 Hz, 200 Hz, and 300 Hz respectively. When setting the bandwidth parameter, the bandwidth of the first layer (50 Hz) is set to 15 Hz, and the bandwidths of the other layers are 25 Hz, 35 Hz, 45 Hz, and 60 Hz respectively. After constructing the Lagrangian operator and performing alternating direction multiplier iteration (setting the maximum number of iterations to 500 times and the convergence threshold to 1e-6), 5 sets of decomposition coefficient sequences are obtained. Perform Hilbert transform on these coefficient sequences, and the calculated instantaneous frequencies show that the average frequency of the first intrinsic mode function subsequence is 49.8 Hz, which is basically consistent with the theoretical fundamental frequency of 50 Hz; the average frequency of the second subsequence is 99.5 Hz, corresponding to the second harmonic component; the main frequency of the third subsequence is around 151 Hz, characterizing the third harmonic feature; the fourth and fifth subsequences respectively reflect the fourth harmonic and bearing fault characteristic frequency components. These decomposition results clearly show each characteristic component in the generator vibration signal, providing an important basis for subsequent fault diagnosis.

[0049] In a specific embodiment, the process of performing step S103 may specifically include the following steps:

[0050] (1) Perform a sliding time window process on the intrinsic mode function subsequence to obtain a time window sequence, and calculate the time error of the time window sequence to obtain a time error sequence;

[0051] (2) Perform an absolute value operation on the time error sequence to obtain an error absolute value sequence, and perform a threshold comparison on the error absolute value sequence. Mark those greater than the threshold as the first type of error, and those less than or equal to the threshold as the second type of error to obtain an error classification result;

[0052] (3) Calculate the initial regular function value for the error classification result, and perform a period analysis on the initial regular function value to obtain a regular function period value;

[0053] (4) Perform a product operation on the sequence in the absolute error value sequence that is less than 5% of the regular function value to obtain the first-mode regular function value, and perform a summation operation on the sequence in the absolute error value sequence that is greater than or equal to 5% of the regular function value to obtain the second-mode regular function value;

[0054] (5) Perform a sequence merging process on the first-mode regular function value and the second-mode regular function value to obtain a dual-mode processing result, and perform a normalization process on the dual-mode processing result to obtain a regularized vibration sequence.

[0055] Specifically, the first step of regularizing the intrinsic mode function subsequence is the sliding time window process. The subsequence is divided into time windows of a fixed length (such as 1024 data points), the overlap rate between windows is set to 50%, and each time it slides 512 data points, thus obtaining a series of continuous time window sequences. For the data within each time window, calculate the time difference between it and the adjacent window to form a time error sequence, which reflects the change characteristics of the data in the time dimension. Perform an absolute value operation on the obtained time error sequence to uniformly convert positive and negative errors into positive values for subsequent threshold comparison. The threshold is set to 5% of the mean value of the data in the current processing window. Compare the absolute error value sequence with the threshold, mark the values greater than the threshold as the first type of error, and mark the values less than or equal to the threshold as the second type of error, thus obtaining a complete error classification result.

[0056] Based on the error classification result, calculate the regular function value for each type of error separately. The regular function is designed in the form of a piecewise continuous function, and different calculation methods are used for different types of errors. By performing Fourier analysis on the initial regular function value, identify the main periodic components to obtain the basic period value of the regular function, which reflects the inherent characteristic period of the vibration signal. According to the dual-mode processing strategy, perform a product operation on the sequence with an absolute error less than 5% of the regular function value, that is, multiply the time error by the regular function period value to obtain the regular function value in the first mode; for the sequence with an absolute error less than or equal to 5%, use a summation operation, add the time error to the regular function period value to obtain the regular function value in the second mode.

[0057] Merge the regular function values obtained in the two modes to form a complete dual-mode processing result. Perform a normalization process on the merged sequence to unify the numerical range into the interval [0, 1], and finally obtain a regularized vibration sequence.

[0058] Taking the vibration data processing of a certain generator as an example, an intrinsic mode function subsequence of 100,000 sampling points is selected and processed using a sliding window with a length of 1024 points. When processing the 50th time window, the calculated time error sequence values are [-0.03, 0.02, -0.04, 0.01, 0.05], and after absolute value operation, we get [0.03, 0.02, 0.04, 0.01, 0.05]. Setting the initial value of the regular function to 0.8, the 5% threshold is 0.04. Classify the error sequence. 0.03, 0.02, and 0.01 are less than 0.04, and the first mode is used for processing; 0.04 and 0.05 are greater than or equal to 0.04, and the second mode is used for processing.

[0059] Period analysis shows that the basic period value of this subsequence is 0.02 seconds. In the first mode, for the error value 0.03, calculate 0.03×0.02 = 0.0006 to obtain the first regular function value; in the second mode, for the error value 0.04, calculate 0.04 + 0.02 = 0.06 to obtain the second regular function value. Combine the processing results of the two modes and perform normalization processing. The finally obtained regularized vibration sequence can better reflect the periodic characteristics of the vibration signal. Through this dual-mode regularization processing method, the influence of non-periodic and non-linear characteristics on the prediction accuracy is effectively reduced, and the obtained regularized vibration sequence is more suitable for subsequent fault diagnosis and analysis. The processing results show that for vibration components with strong periodicity, the regularization effect is more significant, which is beneficial to extracting stable fault characteristics.

[0060] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0061] (1) Perform data standardization processing on the regularized vibration sequence to obtain neural network input data, and perform data format conversion on the neural network input data through the extreme learning machine algorithm to obtain input layer data;

[0062] (2) Construct a weight matrix for the input layer data to obtain input layer weight values, and assign random numbers to the input layer weight values to obtain a randomly generated weight matrix;

[0063] (3) Construct hidden layer nodes for the randomly generated weight matrix to obtain hidden layer structure parameters, and randomly generate neuron thresholds for the hidden layer structure parameters to obtain hidden layer neuron thresholds;

[0064] (4) Perform activation function mapping on the hidden layer neuron thresholds to obtain hidden layer output data, and calculate the connection weights of the output layer for the hidden layer output data to obtain an output layer parameter matrix;

[0065] (5) Construct a network topology for the output layer parameter matrix to obtain a vibration trend prediction model, where the vibration trend prediction model includes an input layer, a hidden layer, and an output layer.

[0066] Specifically, the regularized vibration sequence first needs to be subjected to data standardization processing. Using the Z-score standardization method, the data is transformed into a distribution with a mean of 0 and a standard deviation of 1. The standardized data is used as the input of the neural network and processed through the extreme learning machine algorithm. During the data format conversion process, the continuous time series data is reorganized into a format that meets the input requirements of the neural network. Each sample contains data for n time steps, and n is set to 12, indicating that the data of the first 12 time points is used to predict the value of the next time point. The weight matrix construction of the input layer data adopts a fully connected structure, and the matrix dimension is the input dimension × the number of hidden layer nodes. Taking 12 input nodes and 24 hidden layer nodes as an example, a 12×24 weight matrix is constructed. Each element in the weight matrix is initialized with a random number, and the value range is between [-1,1], randomly generated using a uniform distribution, ensuring the randomness and diversity of network training.

[0067] Based on the randomly generated weight matrix, construct the hidden layer node structure. The number of hidden layer nodes is set to twice the number of input layer nodes, that is, 24 nodes, and each node is connected to all nodes of the input layer. Set a threshold parameter for each hidden layer node, also randomly generated using a uniform distribution within the range of [-1,1], as the bias value of the neuron. The hidden layer uses the Sigmoid activation function for non-linear mapping, and the input data is converted into the output of the hidden layer after calculation by weights and thresholds. Based on the output data of the hidden layer, calculate the connection weights of the output layer. The optimal output layer weights are obtained by solving a linear equation system, and the solution of the equation system is solved using the generalized inverse matrix (Moore-Penrose pseudoinverse) method to obtain the parameter matrix of the output layer.

[0068] Connect the input layer, hidden layer, and output layer through the weight matrix to construct a complete network topology structure to form a vibration trend prediction model. The input layer of this model receives vibration data for 12 time steps, and through the non-linear transformation of 24 hidden layer nodes, finally outputs the predicted value of the next moment.

[0069] Taking the vibration trend prediction of a certain generator as an example, 1000 consecutive regular vibration sequence data points are processed. After data standardization, the mean changes from the original 0.45 to 0, and the standard deviation changes from 0.15 to 1. The data is reorganized into 988 training samples, each sample containing 12 input values and 1 target output value. An input weight matrix of 12×24 is constructed, and the randomly generated weight values are such as [-0.53, 0.86, -0.21], etc. The thresholds of 24 hidden layer nodes are randomly generated as values such as [-0.45, 0.32, 0.78], etc. The output of the hidden layer is calculated through the Sigmoid function, and then the output layer weight matrix of 24×1 is obtained through the calculation of the generalized inverse matrix. The root mean square error of the finally constructed prediction model on the test data is 0.08, showing good prediction performance. Through this method of constructing a prediction model based on the extreme learning machine, the characteristics of randomly generating the input weights and hidden layer thresholds are fully utilized, eliminating the cumbersome weight update process of traditional neural networks, and at the same time maintaining good prediction accuracy. The model structure is simple and clear, with high computational efficiency, and is suitable for real-time prediction of the vibration trend of generators.

[0070] In a specific embodiment, the process of executing step S105 may specifically include the following steps:

[0071] (1) Calculate the predicted value of the vibration trend prediction model to obtain the predicted sequence data, and compare the predicted sequence data with the real-time monitoring data to obtain the error sequence;

[0072] (2) Perform a square operation on the error sequence to obtain the error square sequence, and perform a summation calculation on the error square sequence to obtain the error square sum evaluation function value;

[0073] (3) Count the number of loops of the error square sum evaluation function value to obtain the loop number error sequence, and perform a ratio calculation on the loop number error sequence to obtain the loop number error ratio;

[0074] (4) Set the regular function value to zero for the data with a loop number error ratio greater than two to obtain the first type of regular function value, and perform a natural logarithm calculation on the data with a loop number error ratio less than or equal to two to obtain the logarithm sequence;

[0075] (5) Perform a product operation of the previous regular function value on the logarithm sequence to obtain the second type of regular function value, and merge the first type of regular function value and the second type of regular function value to obtain the optimized prediction model.

[0076] Specifically, when performing predictive calculations on the vibration trend prediction model, first input the vibration data at the current moment into the model to obtain the predicted value for the next moment, and continuously predict to form predicted sequence data. Compare the predicted sequence data point by point with the real-time collected monitoring data, record the difference between the predicted value and the actual value at each moment, and form a complete error sequence. Square each value in the error sequence to eliminate the influence of positive and negative errors, and obtain the error square sequence. Sum up all the values in the error square sequence to obtain the error sum of squares evaluation function value, which reflects the overall prediction accuracy of the prediction model.

[0077] For the error sum of squares evaluation function value, record the number of cycles of each model prediction to form a cycle number error sequence. Statistically analyze the error values within each cycle period, calculate the error ratio between adjacent two periods, and obtain the cycle number error ratio, which reflects the changing trend of the prediction error with the number of cycles.

[0078] According to the magnitude of the cycle number error ratio, adopt different processing strategies. When the error ratio is greater than 2, directly set the regular function value to 0 to obtain the first type of regular function value; when the error ratio is less than or equal to 2, calculate the natural logarithm of this ratio to form a logarithm sequence. This classification processing method can effectively cope with different degrees of prediction errors. Multiply the obtained logarithm sequence by the regular function value calculated in the previous time to obtain the second type of regular function value. Combine the first type of regular function value and the second type of regular function value to form a complete optimized prediction model.

[0079] Taking the vibration data of a certain generator as an example, 1000 data points are taken for prediction. The output sequence of the prediction model is [0.42, 0.45, 0.38, 0.41], and the corresponding actual monitoring values are [0.40, 0.43, 0.39, 0.40]. The calculated error sequence is [0.02, 0.02, -0.01, 0.01]. Squaring the error sequence gives [0.0004, 0.0004, 0.0001, 0.0001], and summing them up gives the error sum of squares as 0.001. In four consecutive cycle periods, the error sums of squares are [0.001, 0.0015, 0.0008, 0.0012] respectively, and the calculated cycle number error ratios are [1.5, 0.53, 1.5]. For the second ratio 0.53 (less than 2), calculating its natural logarithm gives -0.63. If the regular function value in the previous time is 0.8, then the second type of regular function value is -0.504; for the first ratio 1.5 (less than 2), the calculated second type of regular function value is -0.405; for the last ratio 1.5 (less than 2), the calculated second type of regular function value is -0.405.

[0080] Through this adaptive adjustment mechanism, the prediction model can dynamically adjust parameters according to real-time prediction errors and maintain good prediction accuracy. For example, in the subsequent prediction of 100 data points, the average error of the adjusted model is reduced to 0.008, showing good prediction performance.

[0081] In a specific embodiment, the process of executing step S106 may specifically include the following steps:

[0082] (1) Perform sequence data prediction on the optimized prediction model to obtain a prediction output result, and collect characteristic parameters of real-time monitoring data to obtain a real-time characteristic sequence;

[0083] (2) Calculate the numerical deviation between the prediction output result and the real-time characteristic sequence to obtain characteristic deviation data, and divide the characteristic deviation data into threshold ranges to obtain a deviation interval sequence;

[0084] (3) Match the deviation interval sequence with fault types to obtain initial fault characteristics, and verify the initial fault characteristics with historical data to obtain candidate fault types;

[0085] (4) Calculate the fault probability of the candidate fault types to obtain a fault probability distribution, and perform reliability analysis on the fault probability distribution to obtain a reliability index;

[0086] (5) Evaluate the fault severity of the reliability index to obtain a generator fault diagnosis result, where the generator fault diagnosis result includes the fault type, fault probability, and fault severity.

[0087] Specifically, the optimized prediction model predicts the vibration trend at fixed time intervals (such as 0.1 seconds) and continuously outputs a sequence of prediction results. At the same time, characteristic parameters are collected from the real-time operating state of the generator, including key indicators such as amplitude, frequency, and phase, to form a real-time characteristic sequence. The collection frequency of the characteristic parameters is synchronized with the prediction frequency to ensure the time correspondence of the data. Calculate the difference between the prediction output result at each moment and the corresponding real-time characteristic sequence data to obtain characteristic deviation data. According to historical operating experience and equipment specifications, the deviation values are divided into multiple grade intervals, such as a normal interval (deviation less than 5%), a slight anomaly interval (deviation 5%-15%), a moderate anomaly interval (deviation 15%-30%), and a severe anomaly interval (deviation greater than 30%) to form a complete deviation interval sequence.

[0088] Match the deviation interval sequence with a pre - established fault feature library to identify initial fault features. The fault feature library contains common generator fault types and their typical vibration feature patterns. Through comparison and verification with historical fault case data, screen out the most likely candidate fault types to improve the accuracy of diagnosis. For each candidate fault type, calculate its occurrence probability based on Bayesian probability theory. The probability calculation considers multiple factors, including feature matching degree, historical occurrence frequency, operating conditions, etc. At the same time, conduct a reliability analysis on the fault probability distribution to evaluate the credibility of the diagnosis results and obtain a reliability index reflecting the diagnosis accuracy.

[0089] According to the reliability index, combined with the impact degree of the fault on the equipment operation, conduct a hierarchical assessment of the severity of the fault, and finally generate a diagnosis result report including the fault type, fault probability, and severity.

[0090] Taking the fault diagnosis of a certain generator bearing as an example, the predicted vibration trend value output by the prediction model is 0.45 mm, the vibration value monitored in real - time is 0.52 mm, the calculated deviation is 0.07 mm, and the deviation percentage is 13.5%, belonging to the slight anomaly interval. This deviation feature is consistent with the typical feature of the inner ring fault of the bearing. Through verification with historical data, 85 out of 100 similar faults in the past were confirmed as inner ring faults of the bearing. Based on this, the calculated fault probability is 85% and the reliability index is 0.92. Considering the rapid development of the bearing fault and its potential harm to the equipment, diagnose this fault as medium severity and it needs to be repaired during the next planned shutdown. Through this systematic fault diagnosis process, accurately identify the equipment fault.

[0091] The above describes the generator fault diagnosis method based on vibration trend prediction in the embodiments of the present application. Next, describe the generator fault diagnosis system based on vibration trend prediction in the embodiments of the present application. Please refer to Figure 2 One embodiment of the generator fault diagnosis system based on vibration trend prediction in the embodiments of the present application includes:

[0092] An acquisition module 201, configured to perform real - time acquisition and processing on the generator vibration signal to obtain historical vibration signal sequence data;

[0093] A decomposition module 202, configured to perform variational mode decomposition processing on the historical vibration signal sequence data to obtain multiple intrinsic mode function subsequences;

[0094] A processing module 203 is configured to regularize the intrinsic mode function subsequence. Specifically, a dual-mode processing is performed on the subsequence through a preset regular function to obtain a regularized vibration sequence. The dual-mode processing includes: when the absolute value of the time error is less than 5% of the regular function value, the regular function value is set equal to the product of the time error and the regular function period value for processing; when the absolute value of the time error is greater than or equal to 5% of the regular function value, the regular function value is set equal to the sum of the time error and the regular function period value for processing.

[0095] A training module 204 is configured to train a prediction model for the regularized vibration sequence through an extreme learning machine algorithm to obtain a vibration trend prediction model. The prediction model includes an input layer, a hidden layer, and an output layer. The weight values of the input layer and the neuron thresholds of the hidden layer are randomly generated.

[0096] An adjustment module 205 is configured to adaptively adjust the vibration trend prediction model through a sum of squared errors evaluation function to obtain an optimized prediction model. Specifically, when the error ratio of the number of cycles is greater than two, the regular function value is set to zero; when the error ratio of the number of cycles is less than or equal to two, the regular function value is equal to the product of the previous regular function value and the natural logarithm of the error ratio.

[0097] A comparison module 206 is configured to compare and analyze the output result of the optimized prediction model with real-time monitoring data to obtain a generator fault diagnosis result.

[0098] Through the collaborative cooperation of the above-mentioned various components, by collecting and processing the vibration signals of the generator in real time, a complete historical vibration signal sequence data is obtained, ensuring the accuracy and timeliness of the original data. The variational mode decomposition processing method is used to decompose the historical vibration signal sequence data, and multiple intrinsic mode function subsequences are obtained, effectively separating different frequency components in the signal and enhancing the analysis ability of complex vibration signals. By regularizing the intrinsic mode function subsequences and introducing a dual-mode processing mechanism, different calculation methods are adopted according to the magnitude of the absolute value of the time error, effectively reducing the influence of non-periodic and non-linear characteristics on the prediction accuracy. The extreme learning machine algorithm is used to train the prediction model for the regularized vibration sequence. This algorithm simplifies the training process of the traditional neural network by randomly generating the input layer weight values and the hidden layer neuron thresholds, improving the calculation efficiency. The sum of squared errors evaluation function is introduced to adaptively adjust the vibration trend prediction model, and the regular function value is dynamically adjusted according to the magnitude of the error ratio of the number of cycles, enhancing the adaptability of the model to the changes in the vibration trend. Finally, by comparing and analyzing the output results of the optimized prediction model with the real-time monitoring data, the abnormal changes in the operating state of the generator can be detected in time, the fault type can be accurately judged, the early warning and precise diagnosis of the generator fault are realized, and the safety and reliability of the equipment operation are improved.

[0099] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, and are not intended to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the various embodiments of the present application.

Claims

1. A generator fault diagnosis method based on vibration trend prediction, characterized in that: The generator fault diagnosis method based on vibration trend prediction includes: Collect and process the generator vibration signal in real time to obtain historical vibration signal sequence data; Performing variational mode decomposition processing on the historical vibration signal sequence data to obtain a plurality of intrinsic mode function subsequences; The subsequence of the intrinsic modal function is subjected to regularization processing, wherein the subsequence is subjected to dual-mode processing by a preset regular function to obtain a regularized vibration sequence, wherein the dual-mode processing includes: when the absolute value of the time error is less than 5% of the regular function value, the regular function value is equal to the product of the time error and the regular function period value for processing; when the absolute value of the time error is greater than or equal to 5% of the regular function value, the regular function value is equal to the sum of the time error and the regular function period value for processing; The regularized vibration sequence is trained with an extreme learning machine algorithm to form a prediction model, where the prediction model includes an input layer, a hidden layer, and an output layer, and the weight value of the input layer and the neuron threshold of the hidden layer are randomly generated; The vibration trend prediction model is adaptively adjusted through an error square sum evaluation function to obtain an optimized prediction model, wherein when the cycle number error ratio is greater than two, the regular function value is set to zero; when the cycle number error ratio is less than or equal to two, the regular function value is equal to the product of the previous regular function value and the natural logarithm of the error ratio; The output result of the optimized prediction model is compared and analyzed with the real-time monitoring data to obtain the generator fault diagnosis result.

2. The generator fault diagnosis method based on vibration trend prediction according to claim 1 is characterized in that: The real-time collection and processing of the generator vibration signal to obtain historical vibration signal sequence data includes: The sampling frequency of the generator vibration signal is set to obtain the vibration signal sampling parameters, and the vibration signal sampling parameters are tested by Nyquist sampling theorem to obtain effective sampling parameters; Acquiring vibration signals on the effective sampling parameters to obtain original vibration signal data, and performing wavelet denoising on the original vibration signal data to obtain denoised vibration signal data; Performing time domain feature extraction on the noise-reduced vibration signal data to obtain a time domain feature sequence, and performing frequency domain conversion processing on the time domain feature sequence to obtain a frequency domain feature sequence; Performing feature fusion processing on the time domain feature sequence and the frequency domain feature sequence to obtain a fused feature sequence, and performing normalization processing on the fused feature sequence to obtain a standardized feature sequence; The standardized feature sequence is divided into sliding time windows to obtain time series segments, and the time series segments are checked for data integrity to obtain historical vibration signal sequence data.

3. The generator fault diagnosis method based on vibration trend prediction according to claim 1 is characterized in that: The variational modal decomposition process is performed on the historical vibration signal sequence data to obtain a plurality of intrinsic modal function subsequences, including: Performing data segmentation processing on the historical vibration signal sequence data to obtain a plurality of vibration signal sub-segments, and performing spectrum analysis on the plurality of vibration signal sub-segments to obtain spectrum feature data; Calculating the number of variational mode decomposition layers of the frequency spectrum feature data to obtain a decomposition layer number parameter, and estimating the center frequency of the decomposition layer number parameter to obtain a center frequency value of each layer; Setting bandwidth parameters for the center frequency values ​​of each layer to obtain decomposition bandwidth parameters, and constructing secondary penalty terms for the decomposition bandwidth parameters to obtain penalty constraint conditions; Constructing the penalty constraint condition using a Lagrangian operator to obtain an augmented Lagrangian expression, and performing alternating direction multiplier iteration on the augmented Lagrangian expression to obtain a decomposition coefficient sequence; The decomposition coefficient sequence is subjected to Hilbert transform processing to obtain an instantaneous frequency sequence, and the instantaneous frequency sequence is subjected to subsequence reconstruction to obtain a plurality of intrinsic mode function subsequences.

4. The generator fault diagnosis method based on vibration trend prediction according to claim 1 is characterized in that: The regularization processing is performed on the subsequence of the intrinsic modal function, wherein the subsequence is subjected to dual-mode processing by a preset regular function to obtain a regularized vibration sequence, wherein the dual-mode processing includes: when the absolute value of the time error is less than 5% of the regular function value, the regular function value is equal to the product of the time error and the regular function period value for processing; when the absolute value of the time error is greater than or equal to 5% of the regular function value, the regular function value is equal to the sum of the time error and the regular function period value for processing, including: Performing sliding time window processing on the intrinsic mode function subsequence to obtain a time window sequence, and performing time error calculation on the time window sequence to obtain a time error sequence; Performing an absolute value operation on the time error sequence to obtain an error absolute value sequence, and performing a threshold comparison on the error absolute value sequence, marking the error sequence greater than the threshold as a first type of error, and marking the error sequence less than or equal to the threshold as a second type of error, to obtain an error classification result; Performing regular function value calculation on the error classification result to obtain an initial regular function value, and performing period analysis on the initial regular function value to obtain a regular function period value; Performing a product operation on the sequence of the absolute value of the error that is less than 5% of the regular function value to obtain a first pattern regular function value, and performing a sum operation on the sequence of the absolute value of the error that is greater than or equal to 5% of the regular function value to obtain a second pattern regular function value; The first mode regular function value and the second mode regular function value are subjected to sequence merging processing to obtain a dual-mode processing result, and the dual-mode processing result is subjected to normalization processing to obtain a regularized vibration sequence.

5. The generator fault diagnosis method based on vibration trend prediction according to claim 1, characterized in that: The predictive model training is performed on the regularized vibration sequence by using an extreme learning machine algorithm to obtain a vibration trend prediction model, wherein the prediction model includes an input layer, a hidden layer and an output layer, and the weight value of the input layer and the neuron threshold of the hidden layer are randomly generated, including: Performing data standardization processing on the regularized vibration sequence to obtain neural network input data, and performing data format conversion on the neural network input data through an extreme learning machine algorithm to obtain input layer data; Constructing a weight matrix for the input layer data to obtain input layer weight values, and performing random number assignment on the input layer weight values ​​to obtain a randomly generated weight matrix; Constructing hidden layer nodes for the randomly generated weight matrix to obtain hidden layer structural parameters, and randomly generating neuron thresholds for the hidden layer structural parameters to obtain hidden layer neuron thresholds; Performing activation function mapping on the hidden layer neuron threshold to obtain hidden layer output data, and performing output layer connection weight calculation on the hidden layer output data to obtain an output layer parameter matrix; A network topology is constructed for the output layer parameter matrix to obtain a vibration trend prediction model, wherein the vibration trend prediction model includes the input layer, the hidden layer and the output layer.

6. The generator fault diagnosis method based on vibration trend prediction according to claim 1, characterized in that: The vibration trend prediction model is adaptively adjusted through the error square sum evaluation function to obtain an optimized prediction model, wherein when the cycle number error ratio is greater than two, the regular function value is set to zero; when the cycle number error ratio is less than or equal to two, the regular function value is equal to the product of the previous regular function value and the natural logarithm of the error ratio, including: Calculating predicted values ​​of the vibration trend prediction model to obtain predicted sequence data, and comparing the predicted sequence data with real-time monitoring data to obtain an error sequence; Performing a square operation on the error sequence to obtain an error square sequence, and performing a sum calculation on the error square sequence to obtain an error square sum evaluation function value; Performing cycle count on the error square sum evaluation function value to obtain a cycle error sequence, and performing ratio calculation on the cycle error sequence to obtain a cycle error ratio; Performing a regular function value zeroing process on the data whose cycle number error ratio is greater than two to obtain a first-type regular function value, and performing a natural logarithm calculation on the data whose cycle number error ratio is less than or equal to two to obtain a logarithmic sequence; The last regular function value product operation is performed on the logarithmic sequence to obtain a second type of regular function value, and the first type of regular function value and the second type of regular function value are combined to obtain the optimized prediction model.

7. The generator fault diagnosis method based on vibration trend prediction according to claim 1, characterized in that: The optimized prediction model output result is compared and analyzed with the real-time monitoring data to obtain the generator fault diagnosis result, including: Performing sequence data prediction on the optimized prediction model to obtain a prediction output result, and collecting characteristic parameters on the real-time monitoring data to obtain a real-time characteristic sequence; Performing numerical deviation calculation on the predicted output result and the real-time feature sequence to obtain feature deviation data, and dividing the feature deviation data into threshold ranges to obtain a deviation interval sequence; Performing fault type matching on the deviation interval sequence to obtain initial fault features, and performing historical data verification on the initial fault features to obtain candidate fault types; Calculating the failure probability of the candidate failure types to obtain a failure probability distribution, and performing reliability analysis on the failure probability distribution to obtain a reliability index; A fault severity evaluation is performed on the reliability index to obtain a generator fault diagnosis result, wherein the generator fault diagnosis result includes a fault type, a fault probability and a fault severity.

8. A generator fault diagnosis system based on vibration trend prediction, used to implement the generator fault diagnosis method based on vibration trend prediction as described in any one of claims 1 to 7, characterized in that: The generator fault diagnosis system based on vibration trend prediction includes: The acquisition module is used to collect and process the vibration signal of the generator in real time to obtain historical vibration signal sequence data; A decomposition module, used for performing variational modal decomposition processing on the historical vibration signal sequence data to obtain a plurality of intrinsic modal function subsequences; A processing module is used to perform regular processing on the subsequence of the intrinsic mode function, wherein the subsequence is subjected to dual-mode processing by a preset regular function to obtain a regular vibration sequence, wherein the dual-mode processing includes: when the absolute value of the time error is less than 5% of the regular function value, the regular function value is equal to the product of the time error and the regular function period value for processing; when the absolute value of the time error is greater than or equal to 5% of the regular function value, the regular function value is equal to the sum of the time error and the regular function period value for processing; A training module is used to perform prediction model training on the regularized vibration sequence through an extreme learning machine algorithm to obtain a vibration trend prediction model, wherein the prediction model includes an input layer, a hidden layer and an output layer, and the weight value of the input layer and the neuron threshold of the hidden layer are randomly generated; An adjustment module is used to adaptively adjust the vibration trend prediction model through an error square sum evaluation function to obtain an optimized prediction model, wherein when the cycle number error ratio is greater than two, the regular function value is set to zero; when the cycle number error ratio is less than or equal to two, the regular function value is equal to the product of the previous regular function value and the natural logarithm of the error ratio; The comparison module is used to compare and analyze the output result of the optimized prediction model with the real-time monitoring data to obtain the generator fault diagnosis result.

Images