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

Abstract

The application provides a thermal instrument fault diagnosis method and system based on vibration analysis in a thermal power plant, and relates to the technical field of fault diagnosis, which comprises the following steps: collecting and processing vibration signals of thermal instruments, extracting an optimized vibration feature vector, constructing a hybrid neural network model and training to obtain an optimized fault diagnosis model; performing fault diagnosis on the vibration feature vector to generate a fault type, a possibility and a confidence; performing fault risk assessment and case reasoning to generate a fault cause analysis report and maintenance suggestions.

Description

Technical Field

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

[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 techniques 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 make it difficult to fully explore the temporal dependencies and local spatial features in vibration signals. When dealing with high-dimensional nonlinear data, they are prone to underfitting or overfitting, resulting in limited diagnostic accuracy and generalization ability.

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

[0005] Therefore, a solution is urgently needed to address the problems existing in the current technology. Summary of the Invention

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

[0007] A first aspect of this invention provides a method for fault diagnosis of thermal instruments in a thermal power plant based on vibration analysis, comprising:

[0008] Vibration signals from thermal instruments in a thermal power plant are collected and converted into digital signals. Wavelet packet decomposition is performed on the digital signals to obtain sub-band wavelet coefficients. The energy distribution of each sub-band wavelet coefficient is calculated. Based on the energy distribution, a threshold is determined and soft thresholding and wavelet packet reconstruction are performed to obtain a denoised vibration signal. Time-frequency analysis is performed 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. These are combined to obtain the original feature vector. The maximum correlation minimum redundancy algorithm is used to select features to construct a feature subset, resulting in an optimized vibration feature vector.

[0009] A hybrid neural network model is constructed based on long short-term memory network and one-dimensional convolutional neural network. The temporal dependency, context information and local spatial features of the optimized vibration feature vector are extracted. The vibration mode corresponding to the fault type is learned. The features are weighted and fused by the attention mechanism to generate key features. The key features are labeled by the active learning-based intelligent labeling system to generate fault type labels and train the hybrid neural network model. The parameters are optimized by combining transfer learning and cross-validation to obtain the optimized fault diagnosis model.

[0010] Based on the optimized fault diagnosis model, fault diagnosis is performed on the optimized vibration feature vector, generating the corresponding fault type, fault probability, and confidence level of the thermal instrument. Combining the importance of the thermal instrument and historical fault situations, fault risk assessment is performed through Bayesian network reasoning to obtain the fault risk level. Case reasoning is performed based on historical fault situations to generate a fault cause analysis report and maintenance suggestions, which are then sent to equipment maintenance personnel.

[0011] In one alternative implementation,

[0012] Vibration signals from thermal instruments in a thermal power plant are collected and converted into digital signals. Wavelet packet decomposition is performed on the digital signals to obtain sub-band wavelet coefficients. The energy distribution of each sub-band wavelet coefficient is calculated. Based on the energy distribution, a threshold is determined, and soft thresholding and wavelet packet reconstruction are performed to obtain a denoised vibration signal. Time-frequency analysis is performed on the denoised vibration signal to obtain a time-frequency distribution map. Time-domain features, frequency-domain features, and time-frequency-domain features are extracted and combined to obtain the original feature vector. A feature subset is constructed using the maximum correlation minimum redundancy algorithm to obtain the optimized vibration feature vector, which includes:

[0013] The vibration signal of the thermal instrument in the power plant is obtained by the pre-set accelerometer, combined with the output voltage signal of the pre-set charge amplifier, the sampling frequency is set by the Nyquist sampling theorem, and the voltage signal is sampled and pre-processed by the pre-set sampler to generate an initial digital signal. The digital signal is obtained by removing the DC component and outliers.

[0014] The digital signal is decomposed into multi-level wavelet packet decomposition to scale space of different frequency bands to obtain wavelet coefficients corresponding to different sub-bands. The energy distribution corresponding to each sub-band is obtained through energy calculation. Soft thresholding is performed in combination with a pre-set threshold. Noise is removed by shrinking the amplitude of the wavelet coefficients. Wavelet packet reconstruction is performed on the wavelet coefficients after soft thresholding to obtain the denoised vibration signal.

[0015] The denoised vibration signal is segmented using a short-time Fourier transform. Within each segment, a Fourier transform is performed to obtain the spectral distribution and time-frequency representation. The time-frequency distribution map is then synthesized to extract the time-domain features of the vibration signal, including mean, peak value, kurtosis, and waveform factor. Frequency-domain features are obtained by determining the spectral center and the root mean square value of the spectrum. The instantaneous frequency and marginal spectrum are calculated based on the time-frequency representation to obtain the time-frequency domain features. These features are then 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. 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, resulting in optimized vibration feature vectors, which are then combined into a feature subset for output.

[0017] In one alternative implementation,

[0018] The first correlation is obtained by calculating the mutual information between each original feature vector and the fault type, 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, used to indicate the correlation, ψ() is the digamma function, representing the logarithmic reciprocal of the gamma function, ψ(k) represents the entropy related to the k-th nearest neighbor, N represents the number of samples, i represents the sample index, and n x (∈ i ) represents the sample x i Distance less than ∈ i The number of samples, n y (∈ i ) represents the sample y i Distance less than ∈ i The number of samples, ∈ i Indicates sample x i and y i The maximum distance to their respective k-th nearest neighbors, ψ(N), represents the entropy when the number of samples is N.

[0021] In one alternative implementation,

[0022] A hybrid neural network model is constructed based on Long Short-Term Memory (LSTM) networks and one-dimensional convolutional neural networks. The temporal dependencies, contextual information, and local spatial features of the optimized vibration feature vectors are extracted. Vibration patterns corresponding to fault types are learned. Features are weighted and fused using an attention mechanism to generate key features, which are then labeled using an active learning-based intelligent annotation system. Fault type labels are generated, and the hybrid neural network model is trained. Parameter optimization is performed using transfer learning and cross-validation to obtain an optimized fault diagnosis model, including:

[0023] A hybrid neural network model is constructed, consisting of a two-layer long short-term memory network and a three-layer one-dimensional convolutional neural network. Each layer of the long short-term memory network contains 128 neurons, and the kernel sizes of the three layers of the one-dimensional convolutional neural network are 32, 64, and 128, respectively, with a convolution stride of 1.

[0024] The optimized vibration feature vector is input into a hybrid neural network model. Temporal features are extracted through a long short-term memory network, and local spatial features are extracted through a one-dimensional convolutional neural network. The features are then input into a fully connected layer.

[0025] An attention mechanism is introduced, and 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 are calculated using the Softmax function. The attention weights are then weighted to obtain weighted features, and the sums are used to obtain fused features.

[0026] An intelligent annotation system based on active learning is constructed. An uncertainty sampling strategy is adopted to select the sample with the predicted probability closest to 0.5 as the sample to be annotated, and fault type labels are generated for the fused features.

[0027] The fusion features and fault type labels are paired to form a training dataset. 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.

[0028] 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 dataset is divided into five subsets, four subsets are randomly selected as the training set, and the selected subsets are used as the test set. Five models are trained, hyperparameters are optimized and performance is evaluated. The model with the best performance is selected as the optimized fault diagnosis model.

[0029] In one alternative implementation,

[0030] An uncertainty sampling strategy is adopted, and the sample with the predicted probability closest to 0.5 is selected as the sample to be labeled. The fault type label is generated for the fused features as shown in the following formula:

[0031]

[0032] Where u(x) represents the uncertainty fraction, Let represent the model's predicted probability of sample x belonging to class j, and C represent the total number of classes. This represents the maximum value in the model's probability prediction for sample x. Represents entropy measurement. Confidence level represents the difference between the most likely category and the next most likely category. This represents the model's predicted probability for sample x in category g.

[0033] In one alternative implementation,

[0034] Based on the optimized fault diagnosis model, fault diagnosis is performed on the optimized vibration feature vector to generate the corresponding fault type, fault probability, and confidence level for thermal instruments. Combining the importance of the thermal instruments and historical fault data, fault risk assessment is conducted through Bayesian network inference to obtain the fault risk level. Case-based reasoning is then performed based on historical fault data to generate a fault cause analysis report and maintenance recommendations, which are then sent to equipment maintenance personnel.

[0035] Based on the pre-built optimized fault diagnosis model, fault diagnosis is performed on the optimized vibration feature vector to generate the fault type and fault probability corresponding to 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 convolution and pooling operations. The multi-scale fault features are mapped to the fault type and fault probability. The confidence level is obtained by combining the logistic regression algorithm. Based on the output of the intermediate layer, a multi-classifier is trained for each fault type to output the probability distribution of the fault severity and obtain the fault diagnosis result.

[0036] Based on the probability of failure, combined with the importance of thermal instruments and historical failure data, the health status, failure status, environmental factors and causal relationships of thermal instruments are modeled by Bayesian network. The time series of risk level is dynamically inferred by combining the failure diagnosis results and historical operating data. The conditional probability is obtained and updated by parameter learning algorithm. Decision variables are introduced to obtain the failure risk level.

[0037] Historical fault data and expert knowledge are organized into a heterogeneous information network, mapped to a low-dimensional vector space using graph embedding technology, weights are learned through a graph attention network, neighbor information is aggregated to generate semantic representation vectors, deep semantic connections are generated by combining recurrent neural network encoding, 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, maintenance suggestions are generated and sent to equipment maintenance personnel.

[0038] In one alternative implementation,

[0039] By combining fault diagnosis results and historical operational data to dynamically infer the risk level over a time series, and using a parameter learning algorithm to obtain and update the conditional probability, and introducing decision variables, the fault risk level is obtained as shown in the following formula:

[0040]

[0041] Where R represents the overall risk level, indicating the risk level under given observation data Y. 1：T Under the condition that, the current state X tRisk assessment level, Xt represents the state variable at time point t, Y 1：T Let a represent the sequence of observed variables from time 1 to T. t (X t () represents state X at time t, given all prior observations and states. t The probability, β t (X t () represents state X at time t, given future observation data and the current state. t The probability, Risk(X) t U t ) indicates that in state X t and decision U t Risk measurement under U t This represents the decision made at time point t.

[0042] A second aspect of this invention provides a fault diagnosis system for thermal instruments in a thermal power plant based on vibration analysis, comprising:

[0043] The first unit is used to collect vibration signals from thermal instruments in thermal power plants and convert them into digital signals. Wavelet packet decomposition is performed on the digital signals to obtain sub-band wavelet coefficients. The energy distribution of each sub-band wavelet coefficient is calculated. Based on the energy distribution, a threshold is determined and soft thresholding and wavelet packet reconstruction are performed to obtain a denoised vibration signal. Time-frequency analysis is performed 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. These are combined to obtain the original feature vector. The maximum correlation minimum redundancy algorithm is used to select features to construct a feature subset and obtain the optimized vibration feature vector.

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

[0045] The third unit is used to diagnose faults based on the optimized vibration feature vectors according to the optimized fault diagnosis model, generate the fault type, fault probability and confidence level of the thermal instrument, and conduct fault risk assessment through Bayesian network reasoning in combination with the importance of the thermal instrument and historical faults to obtain the fault risk level. It also conducts case reasoning in combination with historical faults to generate a fault cause analysis report and maintenance suggestions, which are then sent to the equipment maintenance personnel.

[0046] A third aspect of the present invention,

[0047] An electronic device is provided, comprising:

[0048] processor;

[0049] Memory used to store processor-executable instructions;

[0050] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0051] Fourth aspect of the embodiments of the present invention,

[0052] A computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0053] In this invention, high-quality processing and feature representation of vibration signals are achieved 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 long short-term memory networks and one-dimensional convolutional neural networks, and the attention mechanism enables the collaborative extraction and adaptive fusion of temporal and spatial features. The active learning strategy improves the efficiency of sample labeling, and the application of transfer learning and cross-validation enhances the generalization ability of the model. By combining Bayesian network reasoning with case reasoning, an intelligent closed loop from fault diagnosis to risk assessment and maintenance decision-making is realized, which not only ensures the accuracy of diagnosis but also provides actionable maintenance suggestions. In summary, this invention realizes the intelligentization of the entire process of fault diagnosis and maintenance decision-making for thermal instruments. Through the organic combination of feature optimization, deep learning, risk assessment, and case reasoning, an accurate, reliable, practical, and efficient intelligent equipment operation and maintenance system is constructed, providing a complete technical solution for the preventive maintenance of thermal instruments in thermal power plants. Attached Figure Description

[0054] Figure 1 This is a flowchart illustrating the fault diagnosis method for thermal instruments in thermal power plants based on vibration analysis, according to an embodiment of the present invention.

[0055] Figure 2 This is a comparison of the time-domain waveforms after denoising in the vibration analysis-based fault diagnosis method for thermal instruments in thermal power plants according to an embodiment of the present invention.

[0056] Figure 3 This is a fault diagnosis performance thermogram of the vibration analysis-based fault diagnosis method for thermal instruments in thermal power plants, as described in an embodiment of the present invention.

[0057] Figure 4 This is a schematic diagram of the structure of a thermal power plant thermal instrument fault diagnosis system based on vibration analysis, according to an embodiment of the present invention. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

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

[0060] Figure 1 This is a flowchart illustrating the fault diagnosis method for thermal instruments in thermal power plants based on vibration analysis, as described in an embodiment of the present invention. Figure 1 As shown, the method includes:

[0061] S1. Collect vibration signals from thermal instruments in a 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 thresholding and wavelet packet reconstruction to obtain denoised vibration signals. Perform time-frequency analysis on the denoised vibration signals to obtain time-frequency distribution maps and extract time-domain features, frequency-domain features, and time-frequency-domain features. Combine these features to obtain the original feature vector. Select features using the maximum correlation minimum redundancy algorithm to construct a feature subset and obtain the optimized vibration feature vector.

[0062] The subband wavelet coefficients represent the signal in different frequency bands through wavelet transform, reflecting the signal's characteristics at various scales and facilitating the analysis of local changes and frequency components. The soft thresholding is a signal denoising technique that adjusts the wavelet coefficients by setting a threshold to reduce noise while preserving important signal features. The wavelet packet reconstruction is the process of recombining the subband coefficients obtained from wavelet packet transform into the original signal, providing finer frequency band control. The time-frequency distribution plot shows the joint representation of the signal in the time and frequency domains, helping to intuitively understand the signal's dynamic characteristics. The maximum correlation minimum redundancy algorithm is a feature selection method designed to select features that are highly correlated with the target variable but have low redundancy among themselves, thereby improving model performance.

[0063] In one alternative implementation,

[0064] Vibration signals from thermal instruments in a thermal power plant are collected and converted into digital signals. Wavelet packet decomposition is performed on the digital signals to obtain sub-band wavelet coefficients. The energy distribution of each sub-band wavelet coefficient is calculated. Based on the energy distribution, a threshold is determined, and soft thresholding and wavelet packet reconstruction are performed to obtain a denoised vibration signal. Time-frequency analysis is performed on the denoised vibration signal to obtain a time-frequency distribution map. Time-domain features, frequency-domain features, and time-frequency-domain features are extracted and combined to obtain the original feature vector. A feature subset is constructed using the maximum correlation minimum redundancy algorithm to obtain the optimized vibration feature vector, which includes:

[0065] The vibration signal of the thermal instrument in the power plant is obtained by the pre-set accelerometer, combined with the output voltage signal of the pre-set charge amplifier, the sampling frequency is set by the Nyquist sampling theorem, and the voltage signal is sampled and pre-processed by the pre-set sampler to generate an initial digital signal. The digital signal is obtained by removing the DC component and outliers.

[0066] The digital signal is decomposed into multi-level wavelet packet decomposition to scale space of different frequency bands to obtain wavelet coefficients corresponding to different sub-bands. The energy distribution corresponding to each sub-band is obtained through energy calculation. Soft thresholding is performed in combination with a pre-set threshold. Noise is removed by shrinking the amplitude of the wavelet coefficients. Wavelet packet reconstruction is performed on the wavelet coefficients after soft thresholding to obtain the denoised vibration signal.

[0067] The denoised vibration signal is segmented using a short-time Fourier transform. Within each segment, a Fourier transform is performed to obtain the spectral distribution and time-frequency representation. The time-frequency distribution map is then synthesized to extract the time-domain features of the vibration signal, including mean, peak value, kurtosis, and waveform factor. Frequency-domain features are obtained by determining the spectral center and the root mean square value of the spectrum. The instantaneous frequency and marginal spectrum are calculated based on the time-frequency representation to obtain the time-frequency domain features. These features are then combined into an original feature vector.

[0068] The mutual information between each original feature vector and the fault type is calculated to obtain the first correlation. The mutual information between different original feature vectors is calculated to obtain the redundancy. Feature selection is performed with the goal of maximizing correlation and minimizing redundancy, resulting in optimized vibration feature vectors, which are then combined into a feature subset for output. Kurtosis is a statistic describing the shape of a probability distribution, reflecting the sharpness of the distribution, and is used for signal feature extraction and anomaly detection. The spectral center is the centroid of the signal spectrum, representing the average position of the signal's frequency components, and is commonly used in audio signal processing. The marginal spectrum refers to the energy distribution of the signal within a specific frequency range; spectral characteristics are revealed by calculating the energy contribution of the signal at each frequency. Mutual information is a measure of the degree of interdependence between two random variables, quantifying the amount of information obtained about one variable through another. Redundancy describes the degree of information duplication in an information system; high redundancy indicates a large amount of repetitive information, which may lead to inefficiency.

[0069] The original vibration signal is obtained by an accelerometer. The accelerometer is installed in key positions of thermal instruments, such as bearing housings and pipe connections, to collect vibration signals generated during equipment operation. The accelerometer converts mechanical vibration into electrical signal output. Piezoelectric accelerometers are usually used, with a sensitivity of about 100mV / 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 to amplify 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 vibration signals of thermal instruments is usually 0-5kHz, the sampling frequency is set to 12.8kHz.

[0071] The sampler samples and preprocesses the voltage signal. It employs 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 with a quantization accuracy of 0.15mV. The sampler incorporates an anti-aliasing filter with a cutoff frequency of 5kHz. The initial digital signal obtained from sampling undergoes DC component removal and outlier removal to obtain the preprocessed digital signal. DC component removal uses a high-pass filter with a cutoff frequency set to 1Hz. Outlier removal uses the 3σ criterion, discarding sampling points with amplitudes exceeding the mean ± 3 standard deviations.

[0072] The preprocessed digital signal is subjected to multi-level wavelet packet decomposition using the db4 wavelet basis function, decomposing the signal into 5 levels. After 5 levels of wavelet packet decomposition, 32 sub-bands are obtained, each corresponding to a different frequency band. Taking a sampling frequency of 12.8kHz as an example, the frequency ranges of the 32 sub-bands are 0-200Hz, 200-400Hz, ..., 6200-6400Hz. The energy of the wavelet coefficients of each sub-band is calculated to obtain the energy distribution. The energy calculation adopts the method of sum of squares of wavelet coefficients.

[0073] The thresholds for each subband are determined based on the energy distribution and soft thresholding is performed. The VisuShrink method is used to select the thresholds, and the threshold size is related to the noise standard deviation and the signal length. The soft thresholding sets the coefficients smaller than the threshold to zero and shrinks the coefficients larger than the threshold. The processed wavelet coefficients are reconstructed by wavelet packets to obtain the denoised vibration signal. The reconstruction process is the reverse of the decomposition process. The inverse transform is used to reconstruct the processed wavelet coefficients into a time-domain signal.

[0074] Time-frequency analysis was performed on the denoised vibration signal using the short-time Fourier transform method. The Hanning window was selected as the window function, with a window length of 1024 points and an overlap rate of 50%. A 512-point FFT was performed on the signal within each time window to obtain a 128×257 time-frequency distribution matrix. Time-domain features, frequency-domain features, and time-frequency-domain features were extracted based on the time-frequency distribution.

[0075] Time-domain characteristics 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 impulse characteristics, and the waveform factor reflects the waveform complexity. Frequency-domain characteristics include the spectral center and the root mean square (RMS) value. The spectral center reflects the frequency location where energy is concentrated, and the RMS value reflects the dispersion of the spectrum. Time-frequency domain characteristics include instantaneous frequency and marginal spectrum. The instantaneous frequency reflects the frequency variation characteristics over time, and the marginal spectrum reflects the overall distribution of the spectrum.

[0076] The extracted features are combined into an original feature vector. The maximum correlation and minimum redundancy algorithm is used to select the original features. The mutual information between each feature and the fault type is calculated to obtain the correlation. The mutual information between different features is calculated to obtain the redundancy. With the goal of maximizing the correlation and minimizing the redundancy, the optimal feature subset is selected to obtain the optimized vibration feature vector.

[0077] In this embodiment, wavelet packet decomposition and soft thresholding denoising methods are used to effectively suppress noise in the vibration signal and improve the accuracy of subsequent feature extraction. Combining time-domain, frequency-domain, and time-frequency-domain features, the characteristics of the vibration signal are comprehensively characterized. Time-domain features reflect the statistical characteristics of the signal, frequency-domain features reflect the spectral distribution, and time-frequency-domain features reflect the frequency variation over 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, which removes redundant and irrelevant features and improves the discriminative 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 thermal instruments in thermal power plants, improving the accuracy and reliability of fault identification.

[0078] Figure 2 This is a comparison of the time-domain waveforms after denoising in the vibration analysis-based fault diagnosis method for thermal power plant instruments according to an embodiment of the present invention. It shows the time-domain waveforms of the vibration signals from thermal power plant instruments after processing with different denoising methods. The top image shows the original noisy signal with a signal-to-noise ratio of 10dB and a noise amplitude of 1.26V. The waveform exhibits significant random fluctuations and interference, particularly a peak of 2.82V at 0.3s, a stable region of 0.03V at 0.6s, and a peak of 2.76V at 0.9s.

[0079] The second layer shows the signal processed by this technical solution. The signal-to-noise ratio is improved to 18.76dB (an increase of 8.76dB), the root mean square error is only 0.0156, and the waveform smoothness is significantly improved. This technical solution uses the db4 wavelet basis function for 5-level wavelet packet decomposition, combined with soft thresholding, which can accurately preserve the amplitude of key feature points. For example, the peak value at 0.3s is 2.75V (only a 2.5% loss compared to the original signal), the stationary value at 0.6s is reduced to 0.01V (close to the ideal value of 0V), and the peak value at 0.9s is 2.71V (only a 1.8% loss). The waveform is smooth overall, without obvious spurious oscillations, while maintaining the continuity and original characteristics of the signal.

[0080] The third layer shows the signal after classical wavelet denoising (Donoho method), with a signal-to-noise ratio improved to 16.14 dB and a root mean square error of 0.0298. Although some noise is suppressed, the waveform exhibits certain distortion and discontinuities, with significant amplitude attenuation at feature points. The peak value at 0.3s drops to 2.58V (a loss of 8.5%), the stable region value at 0.6s is 0.09V (a significant deviation), and the peak value at 0.9s is 2.54V (a loss of 8.0%). This method uses a single threshold and fails to fully consider the energy distribution characteristics of different frequency bands, resulting in insufficient signal feature preservation.

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

[0082] The comparative results fully demonstrate the superiority of this technical solution, which integrates wavelet packet decomposition and soft thresholding. This solution not only performs best in improving the signal-to-noise ratio (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). Particularly in maintaining the integrity of signal features, this solution effectively suppresses noise while accurately preserving the key features of the vibration signal, which is crucial for subsequent feature extraction and fault diagnosis. The results show that using wavelet packet decomposition for multi-scale signal analysis and applying corresponding soft thresholding strategies in different sub-bands can simultaneously achieve noise suppression and feature preservation, significantly outperforming traditional signal processing methods. In one optional implementation,

[0083] The first correlation is obtained by calculating the mutual information between each original feature vector and the fault type, 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 correlation, ψ() is the digamma function, representing the logarithmic reciprocal of the gamma function, ψ(k) represents the entropy related to the k-th nearest neighbor, N represents the number of samples, i represents the sample index, and n x (∈ i ) represents the sample x i Distance less than ∈ i The number of samples, n y (∈ i ) represents the sample y i Distance less than ∈ i The number of samples, ∈ i Indicates sample x i and y i The maximum distance to their respective k-th nearest neighbors, ψ(N), represents the entropy when the number of samples is N.

[0086] In this embodiment, the correlation strength between the original feature vector and the fault type is quantitatively evaluated by using the mutual information calculation method. This ensures the numerical stability of the calculation and allows for dynamic adjustment based on the local density distribution characteristics of the samples. By calculating the mutual information as a correlation index, an important basis is provided for subsequent feature selection and fault diagnosis, which helps to improve the accuracy and efficiency of diagnosis. In summary, this embodiment can effectively identify the features that are most discriminative for fault diagnosis, avoid the difficulties of parameter estimation, and take into account nonlinear correlation, providing a reliable evaluation basis for subsequent feature selection.

[0087] S2. A hybrid neural network model is constructed based on long short-term memory network and one-dimensional convolutional neural network. The temporal dependency, context information and local spatial features of the optimized vibration feature vector are extracted. The vibration mode corresponding to the fault type is learned. The features are weighted and fused by the attention mechanism to generate key features. The key features are labeled by the intelligent labeling system based on active learning to generate fault type labels and train the hybrid neural network model. The parameters are optimized by combining transfer learning and cross-validation to obtain the optimized fault diagnosis model.

[0088] The one-dimensional convolutional neural network is a deep learning model specifically designed 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 dependency refers to the temporal order and correlation between data points, which is particularly important when processing time-series data and is used to capture the dynamic changes of signals. The intelligent labeling system assigns labels to data through automated algorithms, improving labeling efficiency and consistency, especially on large-scale datasets. The high-quality fault type label refers to accurate, detailed, and representative fault classification labels, which helps improve the performance of fault diagnosis models.

[0089] In one alternative implementation,

[0090] A hybrid neural network model is constructed based on Long Short-Term Memory (LSTM) networks and one-dimensional convolutional neural networks. The temporal dependencies, contextual information, and local spatial features of the optimized vibration feature vectors are extracted. Vibration patterns corresponding to fault types are learned. Features are weighted and fused using an attention mechanism to generate key features, which are then labeled using an active learning-based intelligent annotation system. Fault type labels are generated, and the hybrid neural network model is trained. Parameter optimization is performed using transfer learning and cross-validation to obtain an optimized fault diagnosis model, including:

[0091] A hybrid neural network model is constructed, consisting of a two-layer long short-term memory network and a three-layer one-dimensional convolutional neural network. Each layer of the long short-term memory network contains 128 neurons, and the kernel sizes of the three layers of the one-dimensional convolutional neural network are 32, 64, and 128, respectively, with a convolution stride of 1.

[0092] The optimized vibration feature vector is input into a hybrid neural network model. Temporal features are extracted through a long short-term memory network, and local spatial features are extracted through a one-dimensional convolutional neural network. The features are then input into a fully connected layer.

[0093] An attention mechanism is introduced, and 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 are calculated using the Softmax function. The attention weights are then weighted to obtain weighted features, and the sums are used to obtain fused features.

[0094] An intelligent annotation system based on active learning is constructed. An uncertainty sampling strategy is adopted to select the sample with the predicted probability closest to 0.5 as the sample to be annotated, and fault type labels are generated for the fused features.

[0095] The fusion features and fault type labels are paired to form a training dataset. 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.

[0096] 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 dataset is divided into five subsets, four subsets are randomly selected as the training set, and the selected subsets are used as the test set. Five models are trained, hyperparameters are optimized and performance is evaluated. The model with the best performance is selected as the optimized fault diagnosis model.

[0097] The uncertainty sampling strategy is an active learning method that selects samples with high model uncertainty for labeling to improve learning efficiency. The adaptive moment estimation optimization algorithm optimizes the model training process by dynamically adjusting the learning rate and other parameters, thereby improving convergence speed and accuracy. The five-fold cross-validation method is a model validation technique that divides the dataset into five subsets, uses four subsets for training in sequence, and uses the remaining subset for testing to evaluate the model's generalization ability.

[0098] A hybrid neural network model is constructed, consisting of a two-layer long short-term memory network and a three-layer one-dimensional convolutional neural network. Each layer of the long short-term memory network contains 128 neurons and is used to extract temporal features. The three convolutional kernels of the one-dimensional convolutional neural network have kernel sizes of 32, 64, and 128, respectively, and a convolution stride of 1, and are used to extract local spatial features.

[0099] The optimized vibration feature vector is input into the hybrid neural network model. Specifically, the vibration data matrix containing 1000 time steps and 128-dimensional features for each time step is input into the Long Short-Term Memory network to extract temporal features. At the same time, the vibration data matrix is ​​input into a 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] An attention mechanism is introduced to perform weighted fusion of 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. For example, the attention weights of the output features of the Long Short-Term Memory Network are [0.6, 0.3, 0.1], and the attention weights of the output features of the One-Dimensional Convolutional Neural Network are [0.5, 0.3, 0.2]. The two features are weighted based on the attention weights to obtain weighted features. The two weighted features are then added together to obtain a 256-dimensional fused feature vector.

[0101] An intelligent labeling system based on active learning is constructed, employing an uncertainty sampling strategy to predict unlabeled samples. The sample with the predicted probability closest to 0.5 is selected as the sample to be labeled. For example, if the system predicts that an unlabeled sample belongs to fault type A with a probability of 0.48, it is selected as the sample to be labeled. The system generates corresponding fault type labels for the fused features, such as "bearing inner ring fault" and "bearing outer ring fault".

[0102] The fusion features and fault type labels are paired to form a training dataset. A hybrid neural network model is trained using the cross-entropy loss function and the adaptive moment estimation optimization algorithm. The parameters are iteratively optimized until the loss function converges to obtain the initial fault diagnosis model.

[0103] The initial fault diagnosis model was optimized using transfer learning. Specifically, the parameters of the first few layers of the initial model were kept unchanged, while only the parameters of the later layers were fine-tuned to achieve knowledge transfer. A five-fold cross-validation method was used to evaluate the model performance. The dataset was divided into five subsets, and four subsets were randomly selected as the training set each time, with the remaining subset used as the test set. Five models were trained. The hyperparameters of each model were repeatedly optimized and their performance evaluated, such as adjusting the learning rate from 0.001 to 0.0001 and the dropout rate from 0.5 to 0.3. Finally, the model with the highest accuracy on the test set was selected as the optimized fault diagnosis model.

[0104] In this embodiment, by fusing a long short-term memory network and a one-dimensional convolutional neural network, the temporal dependencies and local spatial features of vibration signals can be captured simultaneously, improving the comprehensiveness and effectiveness of feature extraction. An attention mechanism is introduced to adaptively weight and fuse different types of features, highlighting the importance of key features and enhancing the model's feature representation ability and diagnostic accuracy. A combination of active learning and transfer learning is employed, reducing manual annotation costs while improving the model's generalization ability, enabling it to adapt to fault diagnosis tasks under different operating conditions. In summary, this embodiment achieves automatic extraction of fault features and continuous optimization of the diagnostic model, featuring high diagnostic accuracy, strong generalization ability, and high annotation efficiency, providing reliable technical support for fault diagnosis of thermal instruments in power plants.

[0105] Figure 3 This is a heatmap illustrating the fault diagnosis performance of a vibration-based fault diagnosis method for thermal instruments in power plants, as described in this invention. It shows the performance of different fault diagnosis methods under various signal-to-noise ratio environments; darker colors indicate higher accuracy. Figure 3 It is clear that as the noise level increases (signal-to-noise ratio decreases), the accuracy of all methods decreases, but the degree of decrease varies significantly. Our proposed solution exhibits the strongest noise resistance; even under harsh conditions with a low signal-to-noise ratio (10dB), its accuracy remains as high as 95.2%, only 4.3 percentage points lower than in a noise-free environment. In contrast, the accuracy of the traditional FFT method plummeted from 86.8% to 53.4% ​​under the same conditions, a decrease of 33.4 percentage points; the SVM method decreased by 25.4 percentage points to 67.2%; and the CNN and LSTM methods decreased to 84.3% and 80.4% respectively, performing relatively better but still significantly weaker than our proposed solution.

[0106] Long Short-Term Memory (LSTM) networks effectively capture long-term dependencies in time-series signals, reducing the impact of random noise. One-dimensional convolutional neural networks extract robust spatial features through local receptive fields and weight sharing mechanisms. Attention mechanisms further enhance key features in the signal while suppressing noise. Furthermore, the active learning strategy ensures the model learns from the most representative samples, enhancing its generalization ability under varying noise levels. In summary, this technical solution demonstrates high practical value and adaptability in harsh industrial environments, providing a reliable guarantee for the industrial implementation of fault diagnosis technology.

[0107] In one alternative implementation,

[0108] An uncertainty sampling strategy is adopted, and the sample with the predicted probability closest to 0.5 is selected as the sample to be labeled. The fault type label is generated for the fused features as shown in the following formula:

[0109]

[0110] Where u(x) represents the uncertainty fraction, Let represent the model's predicted probability of sample x belonging to class j, and C represent the total number of classes. This represents the maximum value in the model's probability prediction for sample x. Represents entropy measurement. Confidence level represents the difference between the most likely category and the next most likely category. This represents the model's predicted probability for sample x in category g.

[0111] In this embodiment, the uncertainty score of a sample is calculated by comprehensively considering the difference between prediction entropy and confidence level. The entropy measure reflects the dispersion of the prediction probability distribution, while the confidence level characterizes the reliability of the optimal prediction category. The dual evaluation mechanism can more accurately identify samples with large uncertainty in the model prediction, thereby prioritizing the selection of the most valuable samples for manual annotation. Prioritizing the annotation of uncertain samples can quickly improve model performance, enabling the model to achieve better results with less labeled data and accelerating the model iteration speed. In summary, this embodiment improves annotation efficiency, ensures annotation quality, and effectively reduces the cost of manual annotation.

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

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

[0114] In one alternative implementation,

[0115] Based on the optimized fault diagnosis model, fault diagnosis is performed on the optimized vibration feature vector to generate the corresponding fault type, fault probability, and confidence level for thermal instruments. Combining the importance of the thermal instruments and historical fault data, fault risk assessment is conducted through Bayesian network inference to obtain the fault risk level. Case-based reasoning is then performed based on historical fault data to generate a fault cause analysis report and maintenance recommendations, which are then sent to equipment maintenance personnel.

[0116] Based on the pre-built optimized fault diagnosis model, fault diagnosis is performed on the optimized vibration feature vector to generate the fault type and fault probability corresponding to 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 convolution and pooling operations. The multi-scale fault features are mapped to the fault type and fault probability. The confidence level is obtained by combining the logistic regression algorithm. Based on the output of the intermediate layer, a multi-classifier is trained for each fault type to output the probability distribution of the fault severity and obtain the fault diagnosis result.

[0117] Based on the probability of failure, combined with the importance of thermal instruments and historical failure data, the health status, failure status, environmental factors and causal relationships of thermal instruments are modeled by Bayesian network. The time series of risk level is dynamically inferred by combining the failure diagnosis results and historical operating data. The conditional probability is obtained and updated by parameter learning algorithm. Decision variables are introduced to obtain the failure risk level.

[0118] Historical fault data and expert knowledge are organized into a heterogeneous information network, mapped to a low-dimensional vector space using graph embedding technology, weights are learned through a graph attention network, neighbor information is aggregated to generate semantic representation vectors, deep semantic connections are generated by combining recurrent neural network encoding, 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, maintenance suggestions are generated and sent to equipment maintenance personnel.

[0119] The logistic regression algorithm is a statistical method for binary classification problems. By fitting the linear relationship of 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 classification performance. It is often used to handle multi-class classification problems. The discrete time refers to the method of data collection and processing within a specific time interval. It is often used in time series analysis. The time slice divides continuous time into several small intervals to facilitate segmented analysis of data. The heterogeneous information network refers to a network composed of different types of nodes and edges. It is used to represent multiple relationships and information in complex systems. The template-based natural language generation is a method of generating natural language text through predefined templates and rules. It is often used in automated reporting and dialogue systems.

[0120] The collected vibration signals are preprocessed, including denoising and filtering, and time-domain, frequency-domain, and time-frequency-domain features are extracted, such as root mean square value, peak value, and power spectral density, to form vibration feature vectors. Principal component analysis and other methods are used to reduce the dimensionality and optimize the feature vectors.

[0121] A convolutional neural network (CNN) is constructed, comprising 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 then mapped to fault types and probabilities via fully connected layers. Simultaneously, a logistic regression model is trained, and the CNN outputs are fused to obtain the confidence score.

[0122] For each type of fault, such as bearing inner race fault or outer race fault, a multi-classifier is trained, taking the intermediate layer features of the CNN as input and outputting the probability distribution of fault severity, such as minor, moderate, severe, etc.

[0123] A Bayesian network is constructed, with nodes including device health status, fault status, ambient temperature, humidity, etc. A conditional probability table is learned using historical data, and a decision variable is introduced in each time slice to indicate whether maintenance should be performed. Based on real-time observation data and diagnostic results, the risk level is obtained through probabilistic reasoning.

[0124] A heterogeneous information network is established, including node types such as equipment, fault, cause, and handling. Graph embedding and graph attention networks are used to map fault nodes into vector representations, and recurrent neural networks are used to encode related nodes to capture deep semantics.

[0125] The most relevant historical cases are retrieved by vector similarity. Using a predefined template, the retrieved case information is filled into the template to generate a natural language report. Based on the report content and preset rules, the optimal maintenance decision is determined, such as replacing parts or adjusting parameters.

[0126] In this embodiment, the accuracy and reliability of fault diagnosis are improved by integrating multiple deep learning methods. A Bayesian network is used for dynamic risk assessment, enabling real-time early warning of fault risks. Based on heterogeneous information networks and case reasoning, analysis reports and maintenance suggestions are automatically generated, improving diagnostic efficiency. The multi-layered intelligent analysis framework automates the entire process from fault diagnosis to risk assessment and maintenance decision-making, significantly enhancing the scientific nature and timeliness of equipment maintenance. In summary, this embodiment achieves accurate diagnosis and scientific maintenance of thermal instrument faults, featuring accurate diagnosis, controllable risks, and reliable decision-making, providing an intelligent solution for equipment maintenance management.

[0127] In one alternative implementation,

[0128] By combining fault diagnosis results and historical operational data to dynamically infer the risk level over a time series, and using a parameter learning algorithm to obtain and update the conditional probability, and introducing decision variables, the fault risk level is obtained as shown in the following formula:

[0129]

[0130] Where R represents the overall risk level, indicating the risk level under given observation data Y. 1：T Under the condition that, the current state X t Risk assessment level, Xt represents the state variable at time point t, Y 1：T Let a represent the sequence of observed variables from time 1 to T. t (X t () represents state X at time t, given all prior observations and states. t The probability, β t (X t () represents state X at time t, given future observation data and the current state. t The probability, Risk(X) t U t ) indicates that in state X t and decision U t Risk measurement under U t This represents the decision made at time point t.

[0131] In this embodiment, by integrating fault diagnosis results and historical operating data, a comprehensive assessment of equipment risk is achieved, improving the accuracy and reliability of risk assessment. Employing a state-space model and probabilistic reasoning algorithms effectively handles the uncertainty of equipment states, making it suitable for risk assessment of complex dynamic systems. Introducing decision variables combines risk assessment with decision optimization, providing a basis for preventative maintenance and optimized operation of equipment. In summary, this embodiment achieves dynamic and accurate assessment of equipment risk levels. By considering the combined impact of historical, current, and future states, and integrating the role of maintenance decisions, it provides more scientific and reliable risk assessment results, offering strong data support for equipment maintenance decisions.

[0132] Figure 4 This is a schematic diagram of the structure of a thermal power plant thermal instrumentation fault diagnosis system based on vibration analysis, as described in an embodiment of the present invention. Figure 4 As shown, the system includes:

[0133] The first unit is used to collect vibration signals from thermal instruments in thermal power plants and convert them into digital signals. Wavelet packet decomposition is performed on the digital signals to obtain sub-band wavelet coefficients. The energy distribution of each sub-band wavelet coefficient is calculated. Based on the energy distribution, a threshold is determined and soft thresholding and wavelet packet reconstruction are performed to obtain a denoised vibration signal. Time-frequency analysis is performed 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. These are combined to obtain the original feature vector. The maximum correlation minimum redundancy algorithm is used to select features to construct a feature subset and obtain the optimized vibration feature vector.

[0134] The second unit is used to construct a hybrid neural network model based on long short-term memory network and one-dimensional convolutional neural network, extract the temporal dependency, context information and local spatial features of the optimized vibration feature vector, learn the vibration mode corresponding to the fault type, combine attention mechanism to perform weighted fusion of features, generate key features and combine them with an intelligent labeling system based on active learning to label, generate fault type labels and train the hybrid neural network model, combine transfer learning and cross-validation to optimize parameters and obtain an optimized fault diagnosis model;

[0135] The third unit is used to diagnose faults based on the optimized vibration feature vectors according to the optimized fault diagnosis model, generate the fault type, fault probability and confidence level of the thermal instrument, and conduct fault risk assessment through Bayesian network reasoning in combination with the importance of the thermal instrument and historical faults to obtain the fault risk level. It also conducts case reasoning in combination with historical faults to generate a fault cause analysis report and maintenance suggestions, which are then sent to the equipment maintenance personnel.

[0136] A third aspect of the present invention,

[0137] An electronic device is provided, comprising:

[0138] processor;

[0139] Memory used to store processor-executable instructions;

[0140] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0141] Fourth aspect of the embodiments of the present invention,

[0142] A computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0143] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0144] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for fault diagnosis of thermal instruments in thermal power plants based on vibration analysis, characterized in that, include: Vibration signals from thermal instruments in a power plant are collected and converted into digital signals. Wavelet packet decomposition is performed on the digital signals to obtain sub-band wavelet coefficients. The energy distribution of each sub-band wavelet coefficient is calculated. Based on the energy distribution, a threshold is determined, and soft thresholding and wavelet packet reconstruction are performed to obtain a denoised vibration signal. Time-frequency analysis is performed on the denoised vibration signal to obtain a time-frequency distribution map. Time-domain features, frequency-domain features, and time-frequency-domain features are extracted and combined to obtain the original feature vector. A feature subset is constructed using the maximum correlation minimum redundancy algorithm to obtain the optimized vibration feature vector, including: The vibration signal of the thermal instrument in the power plant is obtained by the pre-set accelerometer, combined with the output voltage signal of the pre-set charge amplifier, the sampling frequency is set by the Nyquist sampling theorem, and the voltage signal is sampled and pre-processed by the pre-set sampler to generate an initial digital signal. The digital signal is obtained by removing the DC component and outliers. The digital signal is decomposed into multi-level wavelet packet decomposition to scale space of different frequency bands to obtain wavelet coefficients corresponding to different sub-bands. The energy distribution corresponding to each sub-band is obtained through energy calculation. Soft thresholding is performed in combination with a pre-set threshold. Noise is removed by shrinking the amplitude of the wavelet coefficients. Wavelet packet reconstruction is performed on the wavelet coefficients after soft thresholding to obtain the denoised vibration signal. The denoised vibration signal is segmented using a short-time Fourier transform. Within each segment, a Fourier transform is performed to obtain the spectral distribution and time-frequency representation. The time-frequency distribution map is then synthesized to extract the time-domain features of the vibration signal, including mean, peak value, kurtosis, and waveform factor. Frequency-domain features are obtained by determining the spectral center and the root mean square value of the spectrum. The instantaneous frequency and marginal spectrum are calculated based on the time-frequency representation to obtain the time-frequency domain features. These features are then combined into an original feature vector. The mutual information between each original feature vector and the fault type is calculated to obtain the first correlation. 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. The optimized vibration feature vectors are then combined into a feature subset for output. The first correlation is obtained by calculating the mutual information between each original feature vector and the fault type, as shown in the following formula: Where I(X;Y) represents the mutual information between the original feature vector and different fault types, used to indicate the correlation, ψ() is the digamma function, representing the logarithmic reciprocal of the gamma function, ψ(k) represents the entropy related to the k-th nearest neighbor, N represents the number of samples, i represents the sample index, and n x (∈ i ) represents the sample x i Distance less than ∈ i The number of samples, n y (∈ i ) represents the sample y i Distance less than ∈ i The number of samples, ∈ i Indicates sample x i and y i The maximum distance to their respective k-th nearest neighbors, ψ(N) represents the entropy when the number of samples is N; A hybrid neural network model is constructed based on long short-term memory network and one-dimensional convolutional neural network. The temporal dependency, context information and local spatial features of the optimized vibration feature vector are extracted. The vibration mode corresponding to the fault type is learned. The features are weighted and fused by the attention mechanism to generate key features. The key features are labeled by the active learning-based intelligent labeling system to generate fault type labels and train the hybrid neural network model. The parameters are optimized by combining transfer learning and cross-validation to obtain the optimized fault diagnosis model. Based on the optimized fault diagnosis model, fault diagnosis is performed on the optimized vibration feature vector, generating the corresponding fault type, fault probability, and confidence level of the thermal instrument. Combining the importance of the thermal instrument and historical fault situations, fault risk assessment is performed through Bayesian network reasoning to obtain the fault risk level. Case reasoning is performed based on historical fault situations to generate a fault cause analysis report and maintenance suggestions, which are then sent to equipment maintenance personnel.

2. The method according to claim 1, characterized in that, A hybrid neural network model is constructed based on Long Short-Term Memory (LSTM) networks and one-dimensional convolutional neural networks. The temporal dependencies, contextual information, and local spatial features of the optimized vibration feature vectors are extracted. Vibration patterns corresponding to fault types are learned. Features are weighted and fused using an attention mechanism to generate key features, which are then labeled using an active learning-based intelligent annotation system. Fault type labels are generated, and the hybrid neural network model is trained. Parameter optimization is performed using transfer learning and cross-validation to obtain an optimized fault diagnosis model, including: A hybrid neural network model is constructed, consisting of a two-layer long short-term memory network and a three-layer one-dimensional convolutional neural network. Each layer of the long short-term memory network contains 128 neurons, and the kernel sizes of the three layers of the one-dimensional convolutional neural network are 32, 64, and 128, respectively, with a convolution stride of 1. The optimized vibration feature vector is input into a hybrid neural network model. Temporal features are extracted through a long short-term memory network, and local spatial features are extracted through a one-dimensional convolutional neural network. The features are then input into a fully connected layer. An attention mechanism is introduced, and 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 are calculated using the Softmax function. The attention weights are then weighted to obtain weighted features, and the sums are used to obtain fused features. An intelligent annotation system based on active learning is constructed. An uncertainty sampling strategy is adopted to select the sample with the predicted probability closest to 0.5 as the sample to be annotated, and fault type labels are generated for the fused features. The fusion features and fault type labels are paired to form a training dataset. 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 dataset is divided into five subsets, four subsets are randomly selected as the training set, and the selected subsets are used as the test set. Five models are trained, hyperparameters are optimized and performance is evaluated. The model with the best performance is selected as the optimized fault diagnosis model.

3. The method according to claim 2, characterized in that, An uncertainty sampling strategy is adopted, and the sample with the predicted probability closest to 0.5 is selected as the sample to be labeled. The fault type label is generated for the fused features as shown in the following formula: Where u(x) represents the uncertainty fraction, Let represent the model's predicted probability of sample x belonging to class j, and C represent the total number of classes. This represents the maximum value in the model's probability prediction for sample x. Represents entropy measurement. Confidence level represents the difference between the most likely category and the next most likely category. This represents the model's predicted probability for sample x in category g.

4. The method according to claim 1, characterized in that, Based on the optimized fault diagnosis model, fault diagnosis is performed on the optimized vibration feature vector to generate the corresponding fault type, fault probability, and confidence level for thermal instruments. Combining the importance of the thermal instruments and historical fault data, fault risk assessment is conducted through Bayesian network inference to obtain the fault risk level. Case-based reasoning is then performed based on historical fault data to generate a fault cause analysis report and maintenance recommendations, which are then sent to equipment maintenance personnel. Based on the pre-built optimized fault diagnosis model, fault diagnosis is performed on the optimized vibration feature vector to generate the fault type and fault probability corresponding to 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 convolution and pooling operations. The multi-scale fault features are mapped to the fault type and fault probability. The confidence level is obtained by combining the logistic regression algorithm. Based on the output of the intermediate layer, a multi-classifier is trained for each fault type to output the probability distribution of the fault severity and obtain the fault diagnosis result. Based on the probability of failure, combined with the importance of thermal instruments and historical failure data, the health status, failure status, environmental factors and causal relationships of thermal instruments are modeled by Bayesian network. The time series of risk level is dynamically inferred by combining the failure diagnosis results and historical operating data. The conditional probability is obtained and updated by parameter learning algorithm. Decision variables are introduced to obtain the failure risk level. Historical fault data and expert knowledge are organized into a heterogeneous information network, mapped to a low-dimensional vector space using graph embedding technology, weights are learned through a graph attention network, neighbor information is aggregated to generate semantic representation vectors, deep semantic connections are generated by combining recurrent neural network encoding, 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, maintenance suggestions are generated and sent to equipment maintenance personnel.

5. The method according to claim 4, characterized in that, By combining fault diagnosis results and historical operational data to dynamically infer the risk level over a time series, and using a parameter learning algorithm to obtain and update the conditional probability, and introducing decision variables, the fault risk level is obtained as shown in the following formula: Where R represents the overall risk level, indicating the risk level under given observation data Y. 1：T Under the condition that, the current state X t Risk assessment level, Xt represents the state variable at time point t, Y 1：T Let a represent the sequence of observed variables from time 1 to T. t (X t () represents state X at time t, given all prior observations and states. t The probability, β t (X t () represents state X at time t, given future observation data and the current state. t The probability, Risk(X) t U t ) indicates that in state X t and decision U t Risk measurement under U t This represents the decision made at time point t.

6. A vibration analysis-based fault diagnosis system for thermal instruments in thermal power plants, used to implement the method described in any one of claims 1-5, characterized in that, include: The first unit is used to collect vibration signals from thermal instruments in thermal power plants and convert them into digital signals. Wavelet packet decomposition is performed on the digital signals to obtain sub-band wavelet coefficients. The energy distribution of each sub-band wavelet coefficient is calculated. Based on the energy distribution, a threshold is determined and soft thresholding and wavelet packet reconstruction are performed to obtain a denoised vibration signal. Time-frequency analysis is performed 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. These are combined to obtain the original feature vector. The maximum correlation minimum redundancy algorithm is used to select features to construct a feature subset and obtain the optimized vibration feature vector. The second unit is used to construct a hybrid neural network model based on long short-term memory network and one-dimensional convolutional neural network, extract the temporal dependency, context information and local spatial features of the optimized vibration feature vector, learn the vibration mode corresponding to the fault type, combine attention mechanism to perform weighted fusion of features, generate key features and combine them with an intelligent labeling system based on active learning to label, generate fault type labels and train the hybrid neural network model, combine transfer learning and cross-validation to optimize parameters and obtain an optimized fault diagnosis model; The third unit is used to diagnose faults based on the optimized vibration feature vectors according to the optimized fault diagnosis model, generate the fault type, fault probability and confidence level of the thermal instrument, and conduct fault risk assessment through Bayesian network reasoning in combination with the importance of the thermal instrument and historical faults to obtain the fault risk level. It also conducts case reasoning in combination with historical faults to generate a fault cause analysis report and maintenance suggestions, which are then sent to the equipment maintenance personnel.

7. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 5.

8. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 5.

Images