Thermal power plant thermal instrument fault diagnosis method and system based on vibration analysis

Abstract

The invention provides a thermal power plant thermal instrument fault diagnosis method and system based on vibration analysis, and relates to the technical field of fault diagnosis, and the method comprises the steps: collecting and processing a vibration signal of a thermal instrument, extracting an optimized vibration feature vector, constructing a hybrid neural network model, and training the hybrid neural network model to obtain an optimized fault diagnosis model; performing fault diagnosis on the vibration feature vector, and generating a fault type, possibility and confidence; and performing fault risk assessment and case reasoning, and generating a fault reason analysis report and a maintenance suggestion.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and in particular, to a fault diagnosis method and system for thermal instruments in thermal power plants based on vibration analysis. Background Art

[0002] Traditional vibration signal processing methods are difficult to effectively remove noise interference under complex working conditions, resulting in inaccurate extracted feature information. Existing signal denoising technologies often use fixed thresholds, which cannot adapt to the noise characteristics of different frequency bands and are prone to loss of useful information;

[0003] Most existing fault diagnosis models are based on shallow neural networks or traditional machine learning algorithms, which are difficult to fully mine the temporal dependence relationship and local spatial features in vibration signals, and are prone to underfitting or overfitting problems when dealing with high-dimensional non-linear data, with limited diagnostic accuracy and generalization ability;

[0004] Current fault diagnosis systems often lack further analysis and decision support for diagnosis results. Most systems only give the fault type, but fail to provide more comprehensive diagnostic information such as fault risk assessment, cause analysis, and maintenance suggestions, making it difficult to provide effective guidance for equipment maintenance personnel;

[0005] Therefore, there is an urgent need for a solution to solve the problems existing in the prior art. Summary of the Invention

[0006] Embodiments of the present invention provide a fault diagnosis method and system for thermal instruments in thermal power plants based on vibration analysis, which can at least solve some problems existing in the prior art.

[0007] In a first aspect of the embodiments of the present invention, a fault diagnosis method for thermal instruments in thermal power plants based on vibration analysis is provided, including:

[0008] Collect the vibration signal of the thermal instrument in the thermal power plant and convert it into a digital signal, perform wavelet packet decomposition on the digital signal to obtain sub-band wavelet coefficients, calculate the energy distribution of each sub-band wavelet coefficient, determine a threshold based on the energy distribution and perform soft threshold processing and wavelet packet reconstruction to obtain a denoised vibration signal, perform time-frequency analysis on the denoised vibration signal to obtain a time-frequency distribution diagram and extract time-domain features, frequency-domain features, and time-frequency domain features, combine them to obtain an original feature vector, and select features through the maximum correlation minimum redundancy algorithm to construct a feature subset to obtain an optimized vibration feature vector;

[0009] Construct a hybrid neural network model based on the long short-term memory network and the one-dimensional convolutional neural network, extract the temporal dependence, context information, and local spatial features of the optimized vibration feature vector, learn the vibration patterns corresponding to the fault types, combine the attention mechanism to weight and fuse the features, generate key features, and combine with the intelligent annotation system based on active learning for annotation, generate fault type labels and train the hybrid neural network model, and optimize the parameters by combining transfer learning and cross-validation to obtain an optimized fault diagnosis model;

[0010] Perform fault diagnosis on the optimized vibration feature vector according to the optimized fault diagnosis model, generate the fault type, fault possibility, and confidence level corresponding to the thermal instrument, combine the importance of the thermal instrument and the historical fault situation, and perform fault risk assessment through Bayesian network reasoning to obtain the fault risk level, and perform case reasoning in combination with the historical fault situation to generate a fault cause analysis report and maintenance suggestions and send them to the equipment maintenance personnel.

[0011] In an optional implementation manner,

[0012] Collect the vibration signals of the thermal instruments in the thermal power plant and convert them into digital signals, perform wavelet packet decomposition on the digital signals to obtain sub-band wavelet coefficients, calculate the energy distribution of each sub-band wavelet coefficient, determine the threshold based on the energy distribution and perform soft threshold processing and wavelet packet reconstruction to obtain the denoised vibration signals, perform time-frequency analysis on the denoised vibration signals, obtain the time-frequency distribution diagram and extract the time-domain features, frequency-domain features, and time-frequency domain features, combine them to obtain the original feature vector, and select features through the maximum correlation minimum redundancy algorithm to construct a feature subset, and the optimized vibration feature vector obtained includes:

[0013] Obtain the vibration signals of the thermal instruments in the thermal power plant according to the pre-set acceleration sensor, combine the output voltage signals of the pre-set charge amplifier, set the sampling frequency through the Nyquist sampling theorem, and combine the pre-set sampler to sample and pre-process the voltage signals to generate the initial digital signals and obtain the digital signals by removing the DC component and abnormal points;

[0014] Perform multi-layer wavelet packet decomposition on the digital signals to different scale spaces of different frequency bands to obtain the wavelet coefficients corresponding to different sub-bands, calculate the energy distribution corresponding to each sub-band through energy calculation, perform soft threshold processing in combination with the pre-set threshold, remove the noise by shrinking the amplitude of the wavelet coefficients, and perform wavelet packet reconstruction on the wavelet coefficients after soft threshold processing to obtain the denoised vibration signals;

[0015] The denoised vibration signal is segmented by short-time Fourier transform, and Fourier transform is performed within each segment of the signal to obtain the spectral distribution and time-frequency representation. The time-frequency distribution diagram is comprehensively obtained, and the time-domain features corresponding to the vibration signal are extracted, including the mean, peak value, kurtosis, and waveform factor. The frequency-domain features are obtained by determining the spectral center and the root mean square value of the spectrum. The time-frequency domain features are obtained by calculating the instantaneous frequency and the marginal spectrum based on the time-frequency representation. The features are combined into an original feature vector;

[0016] The mutual information between each original feature vector and the fault type is calculated to obtain the first correlation, and the mutual information between different original feature vectors is calculated to obtain the redundancy. Feature selection is performed with the goal of maximizing the correlation and minimizing the redundancy, and the optimized vibration feature vectors are obtained and combined into a feature subset for output.

[0017] In an alternative embodiment,

[0018] The mutual information between each original feature vector and the fault type is calculated to obtain the first correlation as shown in the following formula:

[0019]

[0020] where I(X; Y) represents the mutual information between the original feature vector and different fault types, which is used to indicate the degree of correlation. ψ() is the digamma function, which represents the logarithmic reciprocal of the gamma function. ψ(k) represents the entropy quantity related to the k-th nearest neighbor. N represents the number of samples, i represents the sample index, and n x (∈ i ) represents the number of samples whose distance from the sample x i is less than ∈ i . n y (∈ i ) represents the number of samples whose distance from the sample y i is less than ∈ i . ∈ i represents the maximum distance from the sample x i and y i to their respective k-th nearest neighbors, and ψ(N) represents the entropy quantity when the number of samples is N.

[0021] In an alternative embodiment,

[0022] A hybrid neural network model is constructed based on the long short-term memory network and the one-dimensional convolutional neural network to extract the temporal dependence relationship, context information, and local spatial features of the optimized vibration feature vector, learn the vibration patterns corresponding to the fault types, combine the attention mechanism to perform weighted fusion on the features, generate key features and combine with an intelligent annotation system based on active learning for annotation, generate fault type labels and train the hybrid neural network model, and combine transfer learning and cross-validation for parameter optimization to obtain an optimized fault diagnosis model including:

[0023] Construct a hybrid neural network model, including two layers of long short-term memory networks and three layers of one-dimensional convolutional neural networks. Each layer of the long short-term memory network contains 128 neurons. The convolutional kernel sizes of the three layers of the one-dimensional convolutional neural network are 32, 64, and 128 respectively, and the convolutional stride is 1;

[0024] Input the optimized vibration feature vector into the hybrid neural network model. Extract temporal features through the long short-term memory network and local spatial features through the one-dimensional convolutional neural network, and input the features into the fully connected layer;

[0025] Introduce an attention mechanism. Use the Softmax function to calculate the attention weights of the output features of the long short-term memory network and the output features of the one-dimensional convolutional neural network, perform weighting based on the attention weights to obtain weighted features, and add them to get the fused features;

[0026] Construct an intelligent annotation system based on active learning. Adopt an uncertainty sampling strategy, select the samples with the prediction probability closest to 0.5 as the samples to be annotated, and generate fault type labels for the fused features;

[0027] Pair the fused features and the fault type labels to form a training data set. Use the cross-entropy loss function and the adaptive moment estimation optimization algorithm to train the hybrid neural network model to obtain an initial fault diagnosis model;

[0028] Realize knowledge transfer by reusing the network layers of the trained model. Adopt a five-fold cross-validation method to evaluate the model performance. Divide the data set into five subsets, randomly select four subsets as the training set, and the selected subset as the test set. Train five models, optimize the hyperparameters and evaluate the performance, and select the model with the best performance as the optimized fault diagnosis model.

[0029] In an alternative embodiment,

[0030] Adopt an uncertainty sampling strategy, select the samples with the prediction probability closest to 0.5 as the samples to be annotated, and generate fault type labels for the fused features as shown in the following formula:

[0031]

[0032] Where, u(x) represents the uncertainty score, represents the prediction probability of the model for sample x in class j, C represents the total number of classes, represents the maximum value in the model probability prediction for sample x, represents the entropy measure, is the confidence, representing the gap between the most likely class and the second most likely class, represents the prediction probability of the model for sample x in class g.

[0033] In an alternative embodiment,

[0034] Perform fault diagnosis on the optimized vibration feature vector according to the optimized fault diagnosis model, generate the fault type, fault probability, and confidence level corresponding to the thermal instrument, combine the importance and historical fault conditions of the thermal instrument, perform fault risk assessment through Bayesian network reasoning to obtain the fault risk level, combine with historical fault conditions for case-based reasoning, generate a fault cause analysis report and maintenance suggestions, and send them to the equipment maintenance personnel, including:

[0035] Based on the pre-constructed optimized fault diagnosis model, perform fault diagnosis on the optimized vibration feature vector to generate the fault type and fault probability corresponding to the thermal instrument. Input the vibration signal in time series through a convolutional neural network, extract multi-scale fault features through convolutional operations and pooling operations, map the multi-scale fault features to the fault type and fault probability, combine with the logistic regression algorithm to obtain the confidence level, train a multi-classifier for each fault type based on the output of the intermediate layer, output the probability distribution of the fault severity, and obtain the fault diagnosis result;

[0036] Based on the fault probability, combine the importance and historical fault conditions of the thermal instrument, model the health state, fault state, environmental factors, and causal relationships of the thermal instrument through a Bayesian network, combine the fault diagnosis result and historical operation data to dynamically infer the time series of the risk level, obtain the conditional probability through the parameter learning algorithm and update it, introduce decision variables, and obtain the fault risk level;

[0037] Organize historical fault conditions and expert knowledge into a heterogeneous information network, map it to a low-dimensional vector space through graph embedding technology, learn weights through a graph attention network, aggregate neighbor information to generate a semantic representation vector, combine with a recurrent neural network for encoding to generate deep semantic connections, retrieve relevant historical cases, generate a fault cause analysis report through template-based natural language generation, determine the optimal maintenance decision, generate maintenance suggestions, and send them to the equipment maintenance personnel.

[0038] In an alternative embodiment,

[0039] Combine the fault diagnosis result and historical operation data to dynamically infer the time series of the risk level, obtain the conditional probability through the parameter learning algorithm and update it, introduce decision variables, and obtain the fault risk level as shown in the following formula:

[0040]

[0041] where R is the overall risk level, indicating the current state X 1：T under the condition of the given observed data Y tThe risk assessment level, Xt represents the state variable at time point t, and Y 1：T represents the sequence of observed variables from time 1 to T, and a t (X t ) represents the probability of state X t at time t given all previous observations and states, and β t (X t ) represents the probability of state X t at time t given future observation data and states, and Risk(X t , U t ) represents the risk measure under state X t and decision U t , and U t represents the decision taken at time point t.

[0042] In the second aspect of the embodiments of the present invention, a fault diagnosis system for thermal instrumentation in a thermal power plant based on vibration analysis is provided, including:

[0043] The first unit is used to collect the vibration signals of the thermal instrumentation in the thermal power plant and convert them into digital signals, perform wavelet packet decomposition on the digital signals to obtain sub-band wavelet coefficients, calculate the energy distribution of each sub-band wavelet coefficient, determine a threshold based on the energy distribution and perform soft threshold processing and wavelet packet reconstruction to obtain the denoised vibration signal, perform time-frequency analysis on the denoised vibration signal to obtain a time-frequency distribution diagram and extract time-domain features, frequency-domain features, and time-frequency domain features, combine them to obtain an original feature vector, select features through the maximum correlation minimum redundancy algorithm to construct a feature subset, and obtain an optimized vibration feature vector;

[0044] The second unit is used to construct a hybrid neural network model based on a long short-term memory network and a one-dimensional convolutional neural network, extract the temporal dependence relationship, context information, and local spatial features of the optimized vibration feature vector, learn the vibration patterns corresponding to the fault types, perform weighted fusion on the features in combination with an attention mechanism, generate key features and combine them with an intelligent annotation system based on active learning for annotation, generate fault type labels and train the hybrid neural network model, and optimize the parameters in combination with transfer learning and cross-validation to obtain an optimized fault diagnosis model;

[0045] The third unit is used to perform fault diagnosis on the optimized vibration feature vector according to the optimized fault diagnosis model, generate the fault type, fault probability, and confidence level corresponding to the thermal instrumentation, combine the importance and historical fault conditions of the thermal instrumentation, perform fault risk assessment through Bayesian network inference to obtain the fault risk level, perform case reasoning in combination with historical fault conditions, generate a fault cause analysis report and maintenance suggestions, and send them to the equipment maintenance personnel.

[0046] In the third aspect of the embodiments of the present invention,

[0047] Provided is an electronic device, including:

[0048] A processor;

[0049] A memory for storing instructions executable by the processor;

[0050] Wherein, the processor is configured to call the instructions stored in the memory to execute the foregoing method.

[0051] In the fourth aspect of the embodiments of the present invention,

[0052] Provided is a computer-readable storage medium, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the foregoing method is implemented.

[0053] In the present invention, high-quality processing and feature expression of vibration signals are realized through wavelet packet decomposition and multi-domain feature extraction. The introduction of the maximum correlation minimum redundancy algorithm ensures the effectiveness and non-redundancy of features. A hybrid model is constructed by combining a long short-term memory network and a one-dimensional convolutional neural network, and the attention mechanism is used to realize the collaborative extraction and adaptive fusion of temporal features and spatial features. The sample annotation efficiency is improved through an active learning strategy. The application of transfer learning and cross-validation enhances the generalization ability of the model. The combination of Bayesian network reasoning and case-based reasoning realizes an intelligent closed-loop from fault diagnosis to risk assessment and then to maintenance decision-making, which not only ensures the accuracy of diagnosis but also provides operable maintenance suggestions. In summary, the present invention realizes the full-process intelligence of thermal instrument fault diagnosis and maintenance decision-making in thermal power plants. Through the organic combination of feature optimization, deep learning, risk assessment, and case-based reasoning, an accurate, reliable, practical, and efficient equipment intelligent operation and maintenance system is constructed, providing a complete technical solution for the preventive maintenance of thermal instruments in thermal power plants. Description of the Drawings

[0054] Figure 1 It is a schematic flow chart of the method for fault diagnosis of thermal instruments in thermal power plants based on vibration analysis according to the embodiments of the present invention;

[0055] Figure 2 It is a comparison diagram of time-domain waveforms after denoising of the method for fault diagnosis of thermal instruments in thermal power plants based on vibration analysis according to the embodiments of the present invention;

[0056] Figure 3 It is a heat map of the fault diagnosis performance of the method for fault diagnosis of thermal instruments in thermal power plants based on vibration analysis according to the embodiments of the present invention;

[0057] Figure 4 It is a schematic structural diagram of the system for fault diagnosis of thermal instruments in thermal power plants based on vibration analysis according to the embodiments of the present invention. Detailed Embodiments

[0058] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part rather than all of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0059] The technical solutions of the present invention will be described in detail below with specific embodiments. These specific embodiments may be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0060] Figure 1 The flow chart of the fault diagnosis method for thermal instrumentation in thermal power plants based on vibration analysis according to the embodiments of the present invention is as shown Figure 1 and the method includes:

[0061] S1. Collect the vibration signals of the thermal instrumentation in the thermal power plant and convert them into digital signals. Perform wavelet packet decomposition on the digital signals to obtain sub-band wavelet coefficients, calculate the energy distribution of each sub-band wavelet coefficient, determine the threshold based on the energy distribution and perform soft threshold processing and wavelet packet reconstruction to obtain the denoised vibration signals. Perform time-frequency analysis on the denoised vibration signals to obtain the time-frequency distribution diagram and extract time-domain features, frequency-domain features, and time-frequency domain features, and combine them to obtain the original feature vector. Select features through the maximum correlation minimum redundancy algorithm to construct a feature subset and obtain the optimized vibration feature vector;

[0062] The sub-band wavelet coefficients are the representations of the signals in different frequency bands obtained through wavelet transform, which reflect the characteristics of the signals at various scales, facilitating the analysis of the local changes and frequency components of the signals. The soft threshold processing is a signal denoising technique that adjusts the wavelet coefficients by setting a threshold to reduce noise and retain important signal features. The wavelet packet reconstruction is the process of recombining the sub-band coefficients obtained through wavelet packet transform into the original signal, providing finer frequency band control for the signal. The time-frequency distribution diagram shows the joint representation of the signal in the time and frequency domains, helping to intuitively understand the dynamic characteristics of the signal. The maximum correlation minimum redundancy algorithm is a feature selection method aimed at selecting features that are highly correlated with the target variable but have low redundancy with each other to improve the model performance.

[0063] In an alternative embodiment,

[0064] Collect the vibration signals of the thermal instruments in the thermal power plant and convert them into digital signals. Perform wavelet packet decomposition on the digital signals to obtain sub-band wavelet coefficients. Calculate the energy distribution of each sub-band wavelet coefficient. Determine the threshold based on the energy distribution and perform soft threshold processing and wavelet packet reconstruction to obtain the denoised vibration signals. Conduct time-frequency analysis on the denoised vibration signals to obtain the time-frequency distribution diagram and extract time-domain features, frequency-domain features, and time-frequency domain features. Combine them to obtain the original feature vector. Select features through the maximum correlation minimum redundancy algorithm to construct a feature subset, and obtain the optimized vibration feature vector including:

[0065] Obtain the vibration signals of the thermal instruments in the thermal power plant according to the pre-set acceleration sensor, combine the output voltage signals of the pre-set charge amplifier, set the sampling frequency through the Nyquist sampling theorem, and sample and preprocess the voltage signals in combination with the pre-set sampler to generate the initial digital signal and obtain the digital signal by removing the DC component and outliers;

[0066] Perform multi-layer wavelet packet decomposition on the digital signals to different scale spaces of different frequency bands to obtain wavelet coefficients corresponding to different sub-bands. Calculate the energy distribution corresponding to each sub-band through energy calculation, perform soft threshold processing in combination with the pre-set threshold, remove noise by shrinking the amplitude of the wavelet coefficients, and perform wavelet packet reconstruction on the wavelet coefficients after soft threshold processing to obtain the denoised vibration signals;

[0067] Segment the denoised vibration signals through short-time Fourier transform, perform Fourier transform within each segment of the signal to obtain the frequency spectrum distribution and time-frequency representation, comprehensively obtain the time-frequency distribution diagram, extract the time-domain features corresponding to the vibration signals, including mean, peak value, kurtosis, and waveform factor, obtain the frequency-domain features by determining the spectral center and root mean square value of the frequency spectrum, calculate the instantaneous frequency and marginal spectrum based on the time-frequency representation to obtain the time-frequency domain features, and combine the features into the original feature vector;

[0068] Calculate the mutual information between each original feature vector and the fault type to obtain the first correlation, and calculate the mutual information between different original feature vectors to obtain the redundancy. Feature selection is carried out with the goal of maximizing the correlation and minimizing the redundancy, and the optimized vibration feature vectors are obtained and combined into a feature subset for output. The kurtosis is a statistic that describes the shape of the probability distribution, reflecting the sharpness of the distribution, and is used for signal feature extraction and anomaly detection. The spectral centroid is the centroid of the signal spectrum, representing the average position of the signal frequency components, and is usually used in audio signal processing. The marginal spectrum refers to the energy distribution of the signal within a specific frequency range, and reveals the spectral characteristics by calculating the energy contribution of the signal at each frequency. The mutual information is a measure of the mutual dependence between two random variables, quantifying the amount of information obtained about one variable through the other variable. The redundancy describes the degree of repetition of information in the information system. High redundancy indicates the existence of a large amount of duplicate information, which may lead to low efficiency.

[0069] The original vibration signal is acquired through an acceleration sensor, which is installed at key positions of the thermal instrument, such as the bearing housing, pipeline connection points, etc., for collecting the vibration signals generated during the operation of the equipment. The acceleration sensor converts mechanical vibration into an electrical signal for output. Usually, a piezoelectric acceleration sensor is used, and its sensitivity is about 100 mV / g.

[0070] The acquired original vibration signal is amplified by a charge amplifier and output as a voltage signal. The gain of the charge amplifier is set to 100, amplifying the weak charge signal into a voltage signal with an amplitude in the range of ±5V. According to the Nyquist sampling theorem, the sampling frequency is set to 2.56 times the highest frequency of the signal to avoid aliasing. Considering that the frequency range of the thermal instrument vibration signal is usually 0 - 5 kHz, the sampling frequency is set to 12.8 kHz.

[0071] The sampler samples and preprocesses the voltage signal. The sampler uses a 16-bit A / D converter to convert the analog voltage signal into a digital signal. Each sampling point is represented by a 16-bit binary number, and the quantization accuracy is 0.15 mV. The sampler is built-in with an anti-aliasing filter with a cut-off frequency of 5 kHz. The initial digital signal obtained by sampling undergoes DC component removal and outlier rejection processing to obtain the preprocessed digital signal. The DC component removal uses a high-pass filtering method with a cut-off frequency set to 1 Hz, and the outlier rejection uses the 3σ criterion to remove the sampling points with amplitudes exceeding the mean ±3 times the standard deviation.

[0072] Perform multi-level wavelet packet decomposition on the preprocessed digital signal. Using the db4 wavelet basis function, decompose the signal to 5 levels. After 5-level wavelet packet decomposition, 32 subbands are obtained. Each subband corresponds to a different frequency band. Taking a sampling frequency of 12.8 kHz as an example, the frequency ranges of the 32 subbands are 0 - 200 Hz, 200 - 400 Hz,..., 6200 - 6400 Hz. Calculate the energy of the wavelet coefficients for each subband to obtain the energy distribution. The energy calculation uses the method of the sum of the squares of the wavelet coefficients.

[0073] Determine the threshold for each subband based on the energy distribution and perform soft threshold processing. The threshold selection uses the VisuShrink method. The threshold size is related to the noise standard deviation and the signal length. In soft threshold processing, the coefficients smaller than the threshold are set to zero, and the coefficients larger than the threshold are subjected to shrinkage processing. The processed wavelet coefficients are reconstructed by wavelet packet to obtain the denoised vibration signal. The reconstruction process is the opposite of the decomposition process, and the inverse transform is used to reconstruct the processed wavelet coefficients into a time-domain signal.

[0074] Perform time-frequency analysis on the denoised vibration signal. Using the short-time Fourier transform method, select the Hanning window function, with a window length of 1024 points and an overlap rate of 50%. Perform 512-point FFT on the signal within each time window to obtain a 128×257 time-frequency distribution matrix. Extract time-domain features, frequency-domain features, and time-frequency domain features based on the time-frequency distribution.

[0075] The time-domain features include mean, peak value, kurtosis, and waveform factor. The mean reflects the DC component of the signal, the peak value reflects the maximum amplitude, the kurtosis reflects the pulse characteristics, and the waveform factor reflects the waveform complexity. The frequency-domain features include the spectral center and the root mean square value of the spectrum. The spectral center reflects the frequency position where the energy is concentrated, and the root mean square value of the spectrum reflects the degree of dispersion of the spectrum. The time-frequency domain features include instantaneous frequency and marginal spectrum. The instantaneous frequency reflects the variation characteristics of the frequency over time, and the marginal spectrum reflects the overall distribution of the spectrum.

[0076] Combine the extracted features into an original feature vector. Use the maximum correlation minimum redundancy algorithm to select the original features. Calculate the mutual information between each feature and the fault type to obtain the correlation, and calculate the mutual information between different features to obtain the redundancy. With the goal of maximizing the correlation and minimizing the redundancy, select the optimal feature subset to obtain the optimized vibration feature vector.

[0077] In this embodiment, the wavelet packet decomposition and soft threshold denoising method are adopted to effectively suppress the noise in the vibration signal, improve the accuracy of subsequent feature extraction, and comprehensively characterize the characteristics of the vibration signal by combining time-domain, frequency-domain, and time-frequency-domain features. The time-domain features reflect the statistical characteristics of the signal, the frequency-domain features reflect the frequency spectrum distribution, and the time-frequency-domain features reflect the variation law of frequency with time. The fusion of multi-domain features can more comprehensively characterize the vibration state of the equipment. The maximum correlation minimum redundancy algorithm is used for feature selection to remove redundant and irrelevant features and improve the discriminant ability of the features. In summary, this embodiment effectively extracts and optimizes the time-domain, frequency-domain, and time-frequency-domain features of the vibration signal of the thermal instrumentation in thermal power plants, and improves the accuracy and reliability of fault identification.

[0078] Figure 2 This is the comparison diagram of the time-domain waveform after denoising for the fault diagnosis method of thermal instrumentation in thermal power plants based on vibration analysis in the embodiment of the present invention, which shows the comparison of the time-domain waveforms of the vibration signals of thermal instrumentation in thermal power plants processed by different denoising methods. The original noisy signal is shown at the top, with a signal-to-noise ratio of 10 dB and a noise amplitude of 1.26 V. The waveform shows significant random fluctuations and interference, especially the peak value at 0.3 s is 2.82 V, the stable region value at 0.6 s is 0.03 V, and the peak value at 0.9 s is 2.76 V.

[0079] The second layer shows the signal processed by the technical solution of the present invention, with the signal-to-noise ratio increased to 18.76 dB (an increase of 8.76 dB), and the root mean square error is only 0.0156. The waveform smoothness is significantly improved. The technical solution of the present invention uses the db4 wavelet basis function for 5-layer wavelet packet decomposition and combines soft threshold processing, which can accurately retain the amplitudes of key feature points. For example, the peak value at 0.3 s is 2.75 V (only a loss of 2.5% compared with the original signal), the stable region value at 0.6 s drops to 0.01 V (close to the ideal value of 0 V), and the peak value at 0.9 s is 2.71 V (a loss of only 1.8%). The overall waveform is smooth, without obvious pseudo-oscillations, and at the same time, the continuity and original features of the signal are maintained.

[0080] The third layer shows the signal processed by the classical wavelet denoising (Donoho method), with the signal-to-noise ratio increased to 16.14 dB and the root mean square error of 0.0298. Although part of the noise is suppressed, there are certain distortions and discontinuities in the waveform, and the amplitudes of the feature points decay significantly. The peak value at 0.3 s drops to 2.58 V (a loss of 8.5%), the stable region value at 0.6 s is 0.09 V (a large deviation), and the peak value at 0.9 s is 2.54 V (a loss of 8.0%). This method uses a single threshold processing and fails to fully consider the energy distribution characteristics of different frequency bands, resulting in insufficient signal feature retention ability.

[0081] The bottom layer is the signal after EMD denoising. The signal-to-noise ratio is increased to 16.96 dB, and the root mean square error is 0.0243. Its performance lies between the previous two. EMD denoising performs well in maintaining the amplitude of feature points. The peak value at 0.3 s is 2.65 V (a loss of 6.0%), and the peak value at 0.9 s is 2.62 V (a loss of 5.1%). However, there are local unevenness in the waveform, and the value in the stable region at 0.6 s is 0.06 V, indicating that its noise suppression effect in the smooth region is not as good as that of this technical solution. Although the EMD method can adaptively decompose signals, it is vulnerable to endpoint effects and mode mixing when dealing with complex noises.

[0082] The comparison results fully demonstrate the superiority of this technical solution integrating wavelet packet decomposition and soft threshold processing. This technical solution not only performs best in terms of signal-to-noise ratio improvement (2.62 dB higher than classical wavelet denoising and 1.8 dB higher than EMD denoising), but also significantly outperforms other methods in terms of root mean square error (47.7% lower than classical wavelet denoising and 35.8% lower than EMD denoising). Especially in maintaining the integrity of signal features, this technical solution can not only effectively suppress noise but also accurately retain the key features of vibration signals, which is crucial for subsequent feature extraction and fault diagnosis. The results show that using wavelet packet decomposition for multi-scale analysis of signals and adopting corresponding soft threshold processing strategies in different sub-bands can take into account both noise suppression and feature preservation, significantly outperforming traditional signal processing methods. In an optional implementation mode,

[0083] Calculate the mutual information between each original feature vector and the fault type to obtain the first correlation as shown in the following formula:

[0084]

[0085] where I(X; Y) represents the mutual information between the original feature vector and different fault types, used to indicate the degree of correlation, ψ() is the digamma function, representing the logarithmic reciprocal of the gamma function, ψ(k) represents the entropy quantity related to the k-th nearest neighbor, N represents the number of samples, i represents the sample index, n x (∈ i ) represents the number of samples whose distance from sample x i is less than ∈ i , n y (∈ i ) represents the number of samples whose distance from sample y i is less than ∈ i , ∈ i represents the maximum distance from sample x i and y i to their k-th nearest neighbors, and ψ(N) represents the entropy quantity when the number of samples is N.

[0086] In this embodiment, the mutual information calculation method is used to quantitatively evaluate the correlation strength between the original feature vector and the fault type, which not only ensures the numerical stability of the calculation but also can be dynamically adjusted according to the local density distribution characteristics of the samples. By calculating the mutual information as the correlation index, it provides an important basis for subsequent feature selection and fault diagnosis, helping to improve the accuracy and efficiency of diagnosis. In summary, this embodiment can effectively identify the most discriminative features for fault diagnosis, avoid the difficulties of parameter estimation, and consider the non-linear correlation, providing a reliable evaluation basis for subsequent feature selection.

[0087] S2. Construct a hybrid neural network model based on the long short-term memory network and the one-dimensional convolutional neural network to extract the temporal dependence, context information, and local spatial features of the optimized vibration feature vector, learn the vibration patterns corresponding to the fault types, combine the attention mechanism to perform weighted fusion on the features, generate key features, and combine them with an intelligent annotation system based on active learning for annotation, generate fault type labels and train the hybrid neural network model, and combine transfer learning and cross-validation to optimize the parameters to obtain an optimized fault diagnosis model;

[0088] The one-dimensional convolutional neural network is a deep learning model specifically for processing one-dimensional data (such as time series signals). It extracts local features through convolutional layers and is suitable for classification and regression tasks. The temporal dependence refers to the time order and correlation between data points, which is particularly important when processing time series data and is used to capture the dynamic changes of the signal. The intelligent annotation system assigns labels to data through automated algorithms to improve the annotation efficiency and consistency, especially for large-scale data sets. The high-quality fault type labels refer to accurate, detailed, and representative fault classification labels, which help to improve the performance of the fault diagnosis model.

[0089] In an alternative embodiment,

[0090] Constructing a hybrid neural network model based on the long short-term memory network and the one-dimensional convolutional neural network to extract the temporal dependence, context information, and local spatial features of the optimized vibration feature vector, learn the vibration patterns corresponding to the fault types, combine the attention mechanism to perform weighted fusion on the features, generate key features, and combine them with an intelligent annotation system based on active learning for annotation, generate fault type labels and train the hybrid neural network model, and combine transfer learning and cross-validation to optimize the parameters to obtain an optimized fault diagnosis model includes:

[0091] Construct a hybrid neural network model, including two layers of long short-term memory networks and three layers of one-dimensional convolutional neural networks. Each layer of the long short-term memory network contains 128 neurons, and the sizes of the convolutional kernels of the three layers of the one-dimensional convolutional neural network are 32, 64, and 128 respectively, and the convolutional stride is 1;

[0092] Input the optimized vibration feature vector into the hybrid neural network model. Extract temporal features through the long short-term memory network and local spatial features through the one-dimensional convolutional neural network, and then input the features into the fully connected layer;

[0093] Introduce the attention mechanism. Use the Softmax function to calculate the attention weights of the output features of the long short-term memory network and the one-dimensional convolutional neural network, perform weighting based on the attention weights to obtain weighted features, and add them to get the fused features;

[0094] Construct an intelligent annotation system based on active learning. Adopt the uncertainty sampling strategy, select the samples with the prediction probability closest to 0.5 as the samples to be annotated, and generate fault type labels for the fused features;

[0095] Pair the fused features and the fault type labels to form a training data set. Use the cross-entropy loss function and the adaptive moment estimation optimization algorithm to train the hybrid neural network model to obtain the initial fault diagnosis model;

[0096] Realize knowledge transfer by reusing the network layers of the trained model. Adopt the five-fold cross-validation method to evaluate the model performance. Divide the data set into five subsets, randomly select four subsets as the training set, and the selected subset as the test set. Train five models, optimize the hyperparameters and evaluate the performance, and select the model with the best performance as the optimized fault diagnosis model.

[0097] The uncertainty sampling strategy is an active learning method. By selecting samples with higher model uncertainty for annotation, the learning efficiency is improved. The adaptive moment estimation optimization algorithm optimizes the model training process by dynamically adjusting the learning rate and other parameters, improving the convergence speed and accuracy. The five-fold cross-validation method is a model validation technique. The data set is divided into five subsets, and four of them are used for training in turn, and the remaining one subset is used for testing to evaluate the generalization ability of the model.

[0098] Construct a hybrid neural network model, including two layers of long short-term memory networks and three layers of one-dimensional convolutional neural networks. Each layer of the long short-term memory network contains 128 neurons for extracting temporal features. The convolutional kernel sizes of the three layers of the one-dimensional convolutional neural network are 32, 64, and 128 respectively, and the convolutional stride is 1 for extracting local spatial features.

[0099] Input the optimized vibration feature vector into the hybrid neural network model. Specifically, input the vibration data matrix containing 1000 time steps and 128-dimensional features at each time step into the long short-term memory network to extract temporal features. At the same time, input the vibration data matrix into the one-dimensional convolutional neural network to extract local spatial features. The long short-term memory network outputs a 128-dimensional temporal feature vector, and the one-dimensional convolutional neural network outputs a 128-dimensional spatial feature vector.

[0100] The attention mechanism is introduced to perform weighted fusion on features. The Softmax function is used to calculate the attention weights of the output features of the long short-term memory network and the output features of the one-dimensional convolutional neural network respectively. For example, the attention weight of the output features of the long short-term memory network is [0.6, 0.3, 0.1], and the attention weight of the output features of the one-dimensional convolutional neural network is [0.5, 0.3, 0.2]. Based on the attention weights, the two types of features are weighted to obtain weighted features, and the two weighted features are added together to obtain a 256-dimensional fused feature vector.

[0101] An intelligent annotation system based on active learning is constructed. The uncertainty sampling strategy is adopted to predict unannotated samples, and the sample with the prediction probability closest to 0.5 is selected as the sample to be annotated. For example, for an unannotated sample, if the system predicts that the probability of it belonging to fault type A is 0.48, then it is selected as the sample to be annotated. The system generates corresponding fault type labels for the fused features, such as "inner ring fault of bearing", "outer ring fault of bearing", etc.

[0102] The fused features and the fault type labels are paired to form a training data set. The cross-entropy loss function and the adaptive moment estimation optimization algorithm are used to train the hybrid neural network model, and the parameters are iteratively optimized until the loss function converges to obtain an initial fault diagnosis model.

[0103] The transfer learning method is used to optimize the initial fault diagnosis model. Specifically, the network parameters of the first few layers of the initial model are kept unchanged, and only the network parameters of the subsequent layers are fine-tuned to achieve knowledge transfer. The five-fold cross-validation method is used to evaluate the model performance. The data set is evenly divided into 5 subsets. Each time, 4 subsets are randomly selected as the training set, and the remaining 1 subset is used as the test set to train 5 models. The hyperparameters of each model are repeatedly optimized and the performance is evaluated, such as adjusting the learning rate from 0.001 to 0.0001, the dropout rate from 0.5 to 0.3, etc. Finally, the model with the highest accuracy on the test set is selected as the optimized fault diagnosis model.

[0104] In this embodiment, by integrating the long short-term memory network and the one-dimensional convolutional neural network, the temporal dependence relationship and local spatial features of vibration signals can be captured simultaneously, improving the comprehensiveness and effectiveness of feature extraction. The attention mechanism is introduced to adaptively weight and fuse different types of features, highlighting the importance of key features and enhancing the feature expression ability and diagnostic accuracy of the model. The method combining active learning and transfer learning reduces the cost of manual annotation while improving the generalization ability of the model, enabling it to adapt to fault diagnosis tasks under different working conditions. In summary, this embodiment realizes the automatic extraction of fault features and the continuous optimization of the diagnostic model, featuring high diagnostic accuracy, strong generalization ability, and high annotation efficiency, providing reliable technical support for the fault diagnosis of thermal instruments in thermal power plants.

[0105] Figure 3 It is the heat map of the fault diagnosis performance of the fault diagnosis method for thermal instruments in thermal power plants based on vibration analysis according to the embodiment of the present invention, showing the performance of different fault diagnosis methods in various signal-to-noise ratio environments. The darker the color, the higher the accuracy. From Figure 3 it can be clearly seen that as the noise level increases (the signal-to-noise ratio decreases), the accuracy of all methods shows a downward trend, but the degree of decrease varies significantly. This technical solution exhibits the strongest anti-noise performance. Even under the harsh condition of a low signal-to-noise ratio (10 dB), its accuracy is still as high as 95.2%, only decreasing by 4.3 percentage points compared to the noise-free environment. In contrast, the accuracy of the traditional FFT method drops sharply from 86.8% to 53.4% under the same conditions, a decrease of 33.4 percentage points; the SVM method decreases by 25.4 percentage points to 67.2%; the CNN method and the LSTM method drop to 84.3% and 80.4% respectively, showing relatively good performance but still significantly weaker than this technical solution.

[0106] The long short-term memory network can effectively capture the long-term dependence relationship in temporal signals and reduce the influence of random noise; the one-dimensional convolutional neural network extracts robust spatial features through the local receptive field and weight sharing mechanism; the attention mechanism further strengthens the key features in the signal while suppressing the noise influence. In addition, the active learning strategy ensures that the model can learn from the most representative samples, enhancing its generalization ability in different noise environments. In summary, this technical solution has high practical value and adaptability in harsh industrial environments, providing reliable guarantee for the industrial implementation of fault diagnosis technology.

[0107] In an alternative embodiment,

[0108] The uncertainty sampling strategy is adopted to select the samples with the prediction probability closest to 0.5 as the samples to be annotated, and the fault type labels for the fused features are generated as shown in the following formula:

[0109]

[0110] Among them, u(x) represents the uncertainty score, represents the predicted probability of the model for sample x in class j, C represents the total number of classes, represents the maximum value in the model probability prediction for sample x, represents the entropy measure, is the confidence level, representing the gap between the most likely class and the second most likely class, represents the predicted probability of the model for sample x in class g.

[0111] In this embodiment, the uncertainty score of the sample is calculated by comprehensively considering the prediction entropy and the confidence level difference. Among them, the entropy measure reflects the dispersion degree of the predicted probability distribution, while the confidence level characterizes the reliability of the optimal predicted class. The dual evaluation mechanism can more accurately identify the samples with large uncertainties in the model prediction, so as to preferentially select the most valuable samples for manual annotation. Preferentially annotating uncertain samples can quickly improve the model performance, enabling the model to achieve good results with less labeled data, accelerating the model iteration speed. In summary, this embodiment not only improves the annotation efficiency but also ensures the annotation quality, effectively reducing the manual annotation cost.

[0112] S3. Perform fault diagnosis on the optimized vibration feature vector according to the optimized fault diagnosis model, generate the corresponding fault type, fault possibility, and confidence level of the thermal instrument, combine the importance and historical fault conditions of the thermal instrument, perform fault risk assessment through Bayesian network reasoning to obtain the fault risk level, combine the historical fault conditions for case reasoning, generate a fault cause analysis report and maintenance suggestions, and send them to the equipment maintenance personnel.

[0113] The fault possibility refers to the probability that a device or system fails under specific conditions, usually estimated through historical data and statistical analysis. The fault risk level is an assessment of the fault possibility and its potential impact to determine the risk degree of the fault to the system or operation.

[0114] In an alternative embodiment,

[0115] Performing fault diagnosis on the optimized vibration feature vector according to the optimized fault diagnosis model, generating the corresponding fault type, fault possibility, and confidence level of the thermal instrument, combining the importance and historical fault conditions of the thermal instrument, performing fault risk assessment through Bayesian network reasoning to obtain the fault risk level, combining the historical fault conditions for case reasoning, generating a fault cause analysis report and maintenance suggestions, and sending them to the equipment maintenance personnel includes:

[0116] Based on a pre-constructed optimized fault diagnosis model, fault diagnosis is carried out on the optimized vibration feature vector to generate the corresponding fault types and fault probabilities of the thermal instrument. The vibration signal is input as a time series through a convolutional neural network, and multi-scale fault features are extracted through convolutional operations and pooling operations. The multi-scale fault features are mapped into fault types and fault probabilities, and the confidence level is obtained by combining the logistic regression algorithm. A multi-classifier is trained for each fault type based on the output of the intermediate layer, and the probability distribution of the fault severity is output to obtain the fault diagnosis result;

[0117] Based on the fault probability, combined with the importance and historical fault conditions of the thermal instrument, the health status, fault status, environmental factors and causal relationships of the thermal instrument are modeled through a Bayesian network. Combining the fault diagnosis result and historical operation data, the time series of the risk level is dynamically inferred, and the conditional probability is obtained and updated through the parameter learning algorithm. Decision variables are introduced to obtain the fault risk level;

[0118] The historical fault conditions and expert knowledge are organized into a heterogeneous information network, which is mapped to a low-dimensional vector space through graph embedding technology. The weights are learned through a graph attention network, and the neighbor information is aggregated to generate a semantic representation vector. Combining with a recurrent neural network for encoding to generate deep semantic connections, relevant historical cases are retrieved, and a fault cause analysis report is generated through template-based natural language generation. The optimal maintenance decision is determined, and maintenance suggestions are generated and sent to the equipment maintenance personnel.

[0119] The logistic regression algorithm is a statistical method for binary classification problems. By fitting the linear relationship of the data, it predicts the probability that a sample belongs to a certain category. The multi-classifier refers to a method that combines multiple classification models to improve the classification performance, and is often used to handle multi-class classification problems. The discrete time refers to the way of data collection and processing within a specific time interval, and is often used in time series analysis. The time slice is to divide the continuous time into several small intervals to facilitate the segmented analysis of data. The heterogeneous information network refers to a network composed of different types of nodes and edges, which is used to represent various relationships and information in a complex system. The template-based natural language generation is a method of generating natural language text through predefined templates and rules, and is often used in automated reports and dialogue systems.

[0120] The collected vibration signal is preprocessed, including denoising, filtering, etc., and time domain, frequency domain and time-frequency domain features such as root mean square value, peak value, power spectral density, etc. are extracted to form a vibration feature vector. Methods such as principal component analysis are used to reduce the dimension and optimize the feature vector.

[0121] Build a convolutional neural network, which includes multiple convolutional layers, pooling layers and fully connected layers. Taking the time series of vibration signals as input, multi-scale features are extracted through convolution and pooling operations. The extracted features are mapped to fault types and probabilities through the fully connected layer. At the same time, a logistic regression model is trained to fuse the CNN output to obtain the confidence level.

[0122] For each fault type, such as inner race fault and outer race fault of bearings, etc., train a multi-classifier. Taking the features of the middle layer of the CNN as input, the probability distribution of the fault severity is output, such as slight, medium, severe, etc.

[0123] Build a Bayesian network. The nodes include the device health status, fault status, environmental temperature, humidity, etc. Use historical data to learn the conditional probability table. Introduce a decision variable at each time slice to represent whether maintenance is to be carried out. According to the real-time observation data and diagnostic results, the risk level is obtained through probability inference.

[0124] Establish a heterogeneous information network, which includes node types such as devices, faults, causes, and dispositions. Using graph embedding and graph attention networks, map the fault nodes into vector representations, and use recurrent neural networks to encode relevant nodes to capture deep semantics.

[0125] Retrieve the most relevant historical cases through vector similarity. Using a predefined template, fill the retrieved case information into the template to generate a natural language report. According to the content of the report, combined with preset rules, determine the optimal maintenance decision, such as replacing parts, adjusting parameters, etc.

[0126] In this embodiment, through the fusion of multiple deep learning methods, the accuracy and reliability of fault diagnosis are improved. The Bayesian network is used for dynamic risk assessment to realize real-time early warning of fault risks. Based on the heterogeneous information network and case-based reasoning, an analysis report and maintenance suggestions are automatically generated, improving the diagnostic efficiency. The multi-level intelligent analysis framework realizes the full process automation from fault diagnosis to risk assessment and then to maintenance decision-making, significantly enhancing the scientificity and timeliness of equipment maintenance. In summary, this embodiment realizes the accurate diagnosis and scientific maintenance of thermal instrument faults, with the characteristics of accurate diagnosis, controllable risk, and reliable decision-making, providing an intelligent solution for equipment maintenance management.

[0127] In an alternative embodiment,

[0128] Combine the fault diagnosis results and historical operation data to dynamically infer the time series of the risk level. Obtain the conditional probability through the parameter learning algorithm and update it. Introduce a decision variable to obtain the fault risk level as shown in the following formula:

[0129]

[0130] Among them, R is the overall risk level, indicating the risk assessment level of the current state X under the condition of the given observed data Y 1：T ; Xt represents the state variable at time point t, and Y t represents the sequence of observed variables from time 1 to T. a 1：T (X t ) represents the probability of the state X t at time t given all previous observations and states. β t (X t ) represents the probability of the state X t at time t given future observed data and states. Risk(X t , U t ) represents the risk metric under the state X t and the decision U t . U t represents the decision taken at time point t. t

[0131] In this embodiment, by fusing the fault diagnosis results and historical operation data, a comprehensive assessment of the equipment risk is realized, the accuracy and reliability of the risk assessment are improved. By using the state space model and probability inference algorithm, the uncertainty of the equipment state can be effectively processed, which is applicable to the risk assessment of complex dynamic systems. By introducing decision variables, the combination of risk assessment and decision optimization is realized, providing a basis for the preventive maintenance and optimized operation of the equipment. In summary, this embodiment realizes the dynamic and accurate assessment of the equipment risk level. By considering the comprehensive influence of historical, current and future states and combining the role of maintenance decisions, a more scientific and reliable risk assessment result is provided, providing strong data support for the equipment maintenance decision

[0132] Figure 4 is the structural schematic diagram of the thermal instrument fault diagnosis system for thermal power plants based on vibration analysis according to the embodiment of the present invention. As Figure 4 shown, the system includes:

[0133] The first unit is used to collect the vibration signals of the thermal instruments in the thermal power plant and convert them into digital signals, perform wavelet packet decomposition on the digital signals to obtain sub-band wavelet coefficients, calculate the energy distribution of each sub-band wavelet coefficient, determine the threshold based on the energy distribution and perform soft threshold processing and wavelet packet reconstruction to obtain the denoised vibration signals, perform time-frequency analysis on the denoised vibration signals to obtain the time-frequency distribution diagram and extract time-domain features, frequency-domain features and time-frequency domain features, combine them to obtain the original feature vector, and select features through the maximum correlation minimum redundancy algorithm to construct a feature subset to obtain the optimized vibration feature vector;

[0134] A second unit, which is used to construct a hybrid neural network model based on a long short-term memory network and a one-dimensional convolutional neural network, extract the temporal dependence, context information, and local spatial features of the optimized vibration feature vector, learn the vibration patterns corresponding to the fault types, combine the attention mechanism to perform weighted fusion on the features, generate key features, and combine with an active learning-based intelligent annotation system for annotation, generate fault type labels, and train the hybrid neural network model, and perform parameter optimization by combining transfer learning and cross-validation to obtain an optimized fault diagnosis model;

[0135] A third unit, which is used to perform fault diagnosis on the optimized vibration feature vector according to the optimized fault diagnosis model, generate the fault type, fault possibility, and confidence level corresponding to the thermal instrument, combine the importance and historical fault conditions of the thermal instrument, perform fault risk assessment through Bayesian network reasoning to obtain the fault risk level, perform case reasoning in combination with the historical fault conditions, generate a fault cause analysis report and maintenance suggestions, and send them to the equipment maintenance personnel.

[0136] In the third aspect of the embodiments of the present invention,

[0137] A kind of electronic device is provided, including:

[0138] A processor;

[0139] A memory for storing instructions executable by the processor;

[0140] Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0141] In the fourth aspect of the embodiments of the present invention,

[0142] A computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0143] The present invention can be a method, device, system, and / or computer program product. The computer program product can include a computer-readable storage medium, on which computer-readable program instructions for executing various aspects of the present invention are uploaded.

[0144] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: it is still possible to modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A fault diagnosis method for thermal instrumentation in thermal power plants based on vibration analysis, characterized in that, Including: Collect the vibration signals of the thermal instruments in the thermal power plant and convert them into digital signals. Perform wavelet packet decomposition on the digital signals to obtain sub-band wavelet coefficients. Calculate the energy distribution of each sub-band wavelet coefficient. Determine the threshold based on the energy distribution and perform soft threshold processing and wavelet packet reconstruction to obtain the denoised vibration signals. Perform time-frequency analysis on the denoised vibration signals to obtain the time-frequency distribution diagram and extract time-domain features, frequency-domain features, and time-frequency domain features. Combine them to obtain the original feature vector. Select features through the maximum correlation minimum redundancy algorithm to construct a feature subset and obtain the optimized vibration feature vector; Construct a hybrid neural network model based on the long short-term memory network and the one-dimensional convolutional neural network. Extract the temporal dependence relationship, context information, and local spatial features of the optimized vibration feature vector. Learn the vibration patterns corresponding to the fault types. Combine the attention mechanism to perform weighted fusion on the features, generate key features, and combine with the intelligent annotation system based on active learning for annotation. Generate fault type labels and train the hybrid neural network model. Combine transfer learning and cross-validation to optimize the parameters and obtain the optimized fault diagnosis model; Perform fault diagnosis on the optimized vibration feature vector according to the optimized fault diagnosis model to generate the fault type, fault probability, and confidence level corresponding to the thermal instrument. Combine the importance of the thermal instrument and the historical fault situation, perform fault risk assessment through Bayesian network inference to obtain the fault risk level. Combine the historical fault situation to perform case reasoning, generate a fault cause analysis report and maintenance suggestions, and send them to the equipment maintenance personnel.

2. The method according to claim 1, wherein Collect the vibration signals of the thermal instruments in the thermal power plant and convert them into digital signals. Perform wavelet packet decomposition on the digital signals to obtain sub-band wavelet coefficients. Calculate the energy distribution of each sub-band wavelet coefficient. Determine the threshold based on the energy distribution and perform soft threshold processing and wavelet packet reconstruction to obtain the denoised vibration signals. Perform time-frequency analysis on the denoised vibration signals to obtain the time-frequency distribution diagram and extract time-domain features, frequency-domain features, and time-frequency domain features. Combine them to obtain the original feature vector. Select features through the maximum correlation minimum redundancy algorithm to construct a feature subset and obtain the optimized vibration feature vector including: Obtain the vibration signals of the thermal instruments in the thermal power plant according to the pre-set acceleration sensor. Combine the output voltage signal of the pre-set charge amplifier. Set the sampling frequency according to the Nyquist sampling theorem. Combine the pre-set sampler to sample and preprocess the voltage signal to generate the initial digital signal and obtain the digital signal by removing the DC component and outliers; Perform multi-layer wavelet packet decomposition on the digital signal to different scale spaces of different frequency bands to obtain the wavelet coefficients corresponding to different sub-bands. Calculate the energy distribution corresponding to each sub-band through energy calculation. Combine the pre-set threshold for soft threshold processing. Remove noise by shrinking the amplitude of the wavelet coefficients. Perform wavelet packet reconstruction on the wavelet coefficients after soft threshold processing to obtain the denoised vibration signal; The denoised vibration signal is segmented by short-time Fourier transform, and Fourier transform is performed within each segment of the signal to obtain the spectral distribution and time-frequency representation. The time-frequency distribution map is comprehensively obtained, and the time-domain features corresponding to the vibration signal are extracted, including mean, peak value, kurtosis, and waveform factor. The frequency-domain features are obtained by determining the spectral center and the root mean square value of the spectrum. The time-frequency domain features are obtained by calculating the instantaneous frequency and marginal spectrum based on the time-frequency representation. The features are combined into the original feature vector; The mutual information between each original feature vector and the fault type is calculated to obtain the first correlation, and the mutual information between different original feature vectors is calculated to obtain the redundancy. Feature selection is performed with the goal of maximizing the correlation and minimizing the redundancy, and the optimized vibration feature vectors are combined into a feature subset for output.

3. The method according to claim 2, wherein The mutual information between each original feature vector and the fault type is calculated to obtain the first correlation as shown in the following formula: Among them, I(X; Y) represents the mutual information between the original feature vector and different fault types, which is used to indicate the degree of correlation. ψ() is the digamma function, representing the logarithmic reciprocal of the gamma function. ψ(k) represents the entropy quantity related to the k-th nearest neighbor. N represents the number of samples, i represents the sample index, and n x (∈ i ) represents the number of samples whose distance from the sample x i is less than ∈ i , and n y (∈ i ) represents the number of samples whose distance from the sample y i is less than ∈ i . ∈ i represents the maximum distance from the sample x i and y i to their respective k-th nearest neighbors. ψ(N) represents the entropy quantity when the number of samples is N.

4. The method according to claim 1, characterized in that A hybrid neural network model is constructed based on the long short-term memory network and the one-dimensional convolutional neural network. The temporal dependence, context information, and local spatial features of the optimized vibration feature vector are extracted, the vibration patterns corresponding to the fault types are learned, the features are weighted and fused by combining the attention mechanism, key features are generated and labeled by combining the active learning-based intelligent annotation system, fault type labels are generated and the hybrid neural network model is trained, and parameter optimization is performed by combining transfer learning and cross-validation to obtain the optimized fault diagnosis model, including: A hybrid neural network model is constructed, including two layers of long short-term memory networks and three layers of one-dimensional convolutional neural networks. Each layer of the long short-term memory network contains 128 neurons, and the convolutional kernel sizes of the three layers of the one-dimensional convolutional neural network are 32, 64, and 128 respectively, and the convolutional stride is 1; The optimized vibration feature vector is input into the hybrid neural network model. The temporal features are extracted by the long short-term memory network, and the local spatial features are extracted by the one-dimensional convolutional neural network. The features are input into the fully connected layer; The attention mechanism is introduced, and the Softmax function is used to calculate the attention weights of the output features of the long short-term memory network and the output features of the one-dimensional convolutional neural network. Weighting is performed based on the attention weights to obtain the weighted features, and the fused features are obtained by adding them up; An active learning-based intelligent annotation system is constructed. The uncertainty sampling strategy is adopted, and the samples with the prediction probability closest to 0.5 are selected as the samples to be annotated, and the fault type labels are generated for the fused features; The fused features and the fault type labels are paired to form a training data set, and the hybrid neural network model is trained using the cross-entropy loss function and the adaptive moment estimation optimization algorithm to obtain the initial fault diagnosis model; Knowledge transfer is achieved by reusing the network layers of the trained model. The five-fold cross-validation method is used to evaluate the model performance. The data set is divided into five subsets, four subsets are randomly selected as the training set, and the selected subset is used as the test set. Five models are trained, the hyperparameters are optimized and the performance is evaluated, and the model with the best performance is selected as the optimized fault diagnosis model.

5. The method according to claim 4, wherein The uncertainty sampling strategy is adopted, and the samples with the prediction probability closest to 0.5 are selected as the samples to be annotated, and the fault type labels are generated for the fused features as shown in the following formula: Among them, \(u(x)\) represents the uncertainty score, represents the predicted probability of the model for the sample \(x\) in class \(j\), \(C\) represents the total number of classes, represents the maximum value in the model probability prediction for the sample \(x\), represents the entropy measure, is the confidence, representing the gap between the most likely class and the second most likely class, represents the predicted probability of the model for the sample \(x\) in class \(g\).

6. The method according to claim 1, characterized in that, Fault diagnosis is performed on the optimized vibration feature vector according to the optimized fault diagnosis model to generate the corresponding fault type, fault possibility and confidence level of the thermal instrument. Combining the importance and historical fault conditions of the thermal instrument, fault risk assessment is carried out through Bayesian network reasoning to obtain the fault risk level. Combining the historical fault conditions, case reasoning is performed to generate a fault cause analysis report and maintenance suggestions and send them to the equipment maintenance personnel, including: Based on the pre-constructed optimized fault diagnosis model, fault diagnosis is performed on the optimized vibration feature vector to generate the corresponding fault type and fault possibility of the thermal instrument. The vibration signal is input in time series through a convolutional neural network, and multi-scale fault features are extracted through convolutional operations and pooling operations. The multi-scale fault features are mapped into the fault type and fault possibility, and the confidence level is obtained by combining the logistic regression algorithm. A multi-classifier is trained for each fault type based on the output of the intermediate layer, and the probability distribution of the fault severity is output to obtain the fault diagnosis result; Based on the fault possibility, combining the importance and historical fault conditions of the thermal instrument, the health state, fault state and environmental factors and causal relationships of the thermal instrument are modeled through a Bayesian network. Combining the fault diagnosis result and historical operation data, the time series of the risk level is dynamically inferred. The conditional probability is obtained and updated through the parameter learning algorithm, and decision variables are introduced to obtain the fault risk level; Organize the historical fault conditions and expert knowledge into a heterogeneous information network, map it to a low-dimensional vector space through graph embedding technology, learn weights through a graph attention network, aggregate neighbor information to generate a semantic representation vector, combine a recurrent neural network to encode to generate deep semantic connections, retrieve relevant historical cases, and generate a fault cause analysis report through template-based natural language generation, determine the optimal maintenance decision, generate maintenance suggestions and send them to the equipment maintenance personnel.

7. The method according to claim 6, characterized in that, Combining the fault diagnosis result and historical operation data, the time series of the risk level is dynamically inferred. The conditional probability is obtained and updated through the parameter learning algorithm, and decision variables are introduced to obtain the fault risk level as shown in the following formula: where R is the overall risk level, representing the risk assessment level of the current state X 1：T under the condition of the given observed data Y t , Xt represents the state variable at time point t, and Y 1：T represents the sequence of observed variables from time 1 to T, a t (X t ) represents the probability of the state X t at time t given all previous observations and states, β t (X t ) represents the probability of the state X t at time t given future observed data and states, Risk(X t , U t ) represents the risk measure under the state X t and the decision U t , and U t represents the decision taken at time point t.

8. A fault diagnosis system for thermal instruments in a thermal power plant based on vibration analysis, which is used to implement the method described in any one of the foregoing claims 1-7, characterized in that, Including: The first unit is used to collect the vibration signal of the thermal instrument in the thermal power plant and convert it into a digital signal, perform wavelet packet decomposition on the digital signal to obtain sub-band wavelet coefficients, calculate the energy distribution of each sub-band wavelet coefficient, determine the threshold based on the energy distribution and perform soft threshold processing and wavelet packet reconstruction to obtain the denoised vibration signal. Perform time-frequency analysis on the denoised vibration signal to obtain a time-frequency distribution map and extract time-domain features, frequency-domain features and time-frequency domain features, combine them to obtain an original feature vector, and select features through the maximum correlation minimum redundancy algorithm to construct a feature subset to obtain the optimized vibration feature vector; The second unit is used to construct a hybrid neural network model based on a long short-term memory network and a one-dimensional convolutional neural network, extract the temporal dependence, context information, and local spatial features of the optimized vibration feature vector, learn the vibration patterns corresponding to the fault types, combine the attention mechanism to perform weighted fusion on the features, generate key features, and combine with an active learning-based intelligent annotation system for annotation, generate fault type labels, and train the hybrid neural network model, and perform parameter optimization by combining transfer learning and cross-validation to obtain an optimized fault diagnosis model; The third unit is used to perform fault diagnosis on the optimized vibration feature vector according to the optimized fault diagnosis model, generate the fault type, fault probability, and confidence level corresponding to the thermal instrument, combine the importance of the thermal instrument and the historical fault situation, perform fault risk assessment through Bayesian network inference to obtain the fault risk level, perform case reasoning by combining the historical fault situation, generate a fault cause analysis report and maintenance suggestions, and send them to the equipment maintenance personnel.

9. An electronic device, characterized in that, Comprising: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, the method according to any one of claims 1 to 7 is implemented.

Images