An artificial intelligence-based fault diagnosis method for thermal power equipment

Abstract

The application discloses a kind of based on artificial intelligence's thermal power equipment fault diagnosis method, it is related to artificial intelligence technical field, thermal power equipment training data acquisition, thermal power equipment fault diagnosis, thermal power equipment training data labeling, thermal power equipment fault diagnosis;Adopt the comprehensive method of band filter+wavelet packet transformation to carry out vibration signal preprocessing, avoid the information loss caused by single method, can extract key fault characteristics in a variety of frequency bands and time domain range, solved the problem that traditional filtering method is easily failed under the multi-working condition coupling vibration environment, improved the complete reservation ability of fault mode;Adopt the dynamic penalty factor based on misclassification rate and iterative process, reduce penalty in early training, avoid convergence difficulty, gradually strengthen penalty in later period, improve the correction ability of misclassification sample, solved the problem that traditional fixed penalty coefficient or simple attenuation strategy is difficult to adapt to different training stage needs, make model training more stable, efficient.

Description

Technical Field

[0001] This invention relates specifically to the field of artificial intelligence technology, and more specifically to an artificial intelligence-based method for diagnosing faults in thermal power equipment. Background Technology

[0002] Thermal power equipment is a crucial component of energy production. Key components such as steam turbines, generators, and transformers are susceptible to various factors during long-term operation, including high temperatures, high-speed rotation, mechanical wear, and electromagnetic interference. This can lead to common faults such as bearing damage, rotor imbalance, bearing loosening, and gearbox malfunctions. If these faults are not detected and addressed promptly, they can not only reduce equipment operating efficiency but also potentially cause serious safety accidents and even large-scale downtime losses. Therefore, establishing efficient and accurate fault diagnosis methods is crucial for the safety and economic efficiency of thermal power equipment. A search revealed that Chinese invention patent CN115163426B proposes a wind turbine fault detection method and system based on AI auscultation, as well as a wind turbine safety system. The method involves collecting a third sound signal during wind turbine blade operation, performing digital-to-analog conversion and noise reduction on the third sound signal to obtain sound signal C, extracting the actual features of sound signal C, and comparing it with standard features to calculate a first similarity. If the first similarity is less than or equal to a first preset value, the actual feature is compared with any feature in the fault feature set, and a second similarity is calculated for each. The second similarities are sorted in descending order, and the top M fault features corresponding to the second similarity are selected to obtain the corresponding fault type or a combination of M fault types. Upon fault detection, an alarm is triggered, and fault diagnosis opinions and repair plans are provided. This solution utilizes AI auscultation combined with deep learning technology to extract and compare acoustic features of real-time collected operating noise to detect whether blade faults have occurred in real time. Chinese invention patent CN119596909A discloses a method, system, storage medium, and electronic device for diagnosing equipment faults, relating to the field of industrial automation technology. The method for diagnosing equipment faults includes: determining the target service instance distributed when completing equipment registration service for the current device in the basic service module; analyzing the equipment data of the current device using the target service module to which the service instance belongs; and determining whether the current device has a fault based on the analysis results. The target service module belongs to the upper-level service module, and the upper-level service module includes at least one of the following: an equipment status acquisition service module, an asynchronous message service module, an equipment trend service module, a historical data persistence service module, and an equipment status analysis service module. This technical solution solves the technical problem of low efficiency in equipment fault diagnosis. Chinese invention patent CN119596126A discloses a high-voltage switch fault diagnosis system based on artificial intelligence audio recognition technology. The system includes at least one audio acquisition device for collecting audio signals generated during the operation of the high-voltage switch; a data preprocessing module for preprocessing the audio signals; and a feature extraction module for extracting feature vectors from the preprocessed audio signals. This high-voltage switch fault diagnosis system, using a high-sensitivity MEMS microphone and advanced audio processing technology, can accurately capture subtle abnormal sounds during the operation of the high-voltage switch, improving the accuracy of fault diagnosis. By combining data from multiple sensors, such as temperature and vibration sensors, it enhances the accuracy and reliability of fault diagnosis. The system can automatically adjust the sensitivity parameters of the model according to the ambient noise level, reducing false alarms. When nylon filaments are stretched mechanically, only a certain length of filament can be stretched. This means that the continuous transmission of nylon filaments cannot be effectively guaranteed during the stretching process, which reduces the working efficiency of the stretching device. The most significant problem with existing technologies is that traditional machine learning methods based on neural network models use activation functions that are difficult to adapt to complex nonlinear fault characteristics. Existing fault diagnosis models often employ single activation functions such as ReLU, Sigmoid, and Tanh, which struggle to preserve complex vibration signal characteristics in high-dimensional spaces, leading to feature aliasing between different fault modes. Furthermore, existing technologies require further solutions to the following issues: Current methods often employ single filters (low-pass, band-pass, or high-pass filters) or simple time-frequency transformations (Fourier transforms, etc.) for signal preprocessing, but under complex operating conditions, this can easily lead to signal distortion or changes in fault characteristics. Feature loss; traditional fault diagnosis models generally use a fixed learning rate or a learning rate that decays over time, but this method is difficult to dynamically adjust according to the needs of different training stages, which can easily lead to early convergence difficulties or insufficient optimization in the later stages; existing fault diagnosis methods usually use cross-entropy loss + L2 regularization, but L2 regularization only constrains the model complexity and is difficult to improve the sparsity and fault discrimination ability of the equipment fault diagnosis model; existing fault diagnosis methods usually use fixed step size gradient descent or momentum optimization algorithms (such as Adam, SGD, etc.), but lack optimization for parameter distribution and gradient history information, and are prone to getting trapped in local optima. Summary of the Invention

[0003] The purpose of this invention is to provide an artificial intelligence-based fault diagnosis method for thermal power equipment. In this device, a two-stage activation function (LeakyReLU + hyperbolic tangent function) is used in the hidden layer of the probabilistic neural network. This not only preserves negative gradient information and prevents feature loss, but also enhances the ability to locate transient impacts. This solves the problem that traditional single activation functions are difficult to capture high-dimensional nonlinear fault features, and improves the model's ability to distinguish complex vibration signals, thereby addressing the problems mentioned in the background technology.

[0004] To achieve the above objectives, the present invention provides the following technical solution: An artificial intelligence-based fault diagnosis method for thermal power equipment specifically includes the following steps: S1. Training data acquisition for thermal power equipment: Data is acquired through vibration sensors installed on key parts of the equipment. These sensors can monitor and record vibration signals during equipment operation in real time. The acquired data is recorded in time series format, with each data point including a timestamp and vibration intensity value. The storage format is a structured database format, where each row represents a record at a time point, and the columns represent the timestamp and vibration data from different sensors, respectively. S2. Initial screening of thermal power equipment data; removal of shutdown data, elimination of outliers, and data balancing. S3. Training data annotation for thermal power equipment; The training data is annotated manually by combining equipment operation and maintenance records with expert experience to manually annotate vibration signal data. S4. Fault diagnosis of thermal power equipment; using probabilistic neural networks for fault diagnosis of thermal power equipment. As a further technical solution of the present invention, the removal of shutdown data in S2 is as follows: Since the vibration signal of thermal power equipment is close to zero when it is shut down, these data have no practical value for the fault diagnosis model. Therefore, it is necessary to remove all data collected during the shutdown period; Remove outliers: Statistical methods are used to detect and remove abnormal vibration signal data to avoid noise data interfering with model training; Data balancing: In actual data collection, normal operation data is often much more than fault data, which may cause the model training to be biased towards the normal category. Undersampling or oversampling methods are used to make the number of samples of each category relatively balanced. As a further technical solution of the present invention, as described in S4; the training process of the probabilistic neural network is as follows: preprocessing the vibration signal data of thermal power equipment; the method of preprocessing the vibration signal data of thermal power equipment is expressed as follows: ; In the formula, For the first The feature vectors obtained after preprocessing represent the effective features of the vibration signal data of thermal power equipment after denoising and interference removal. For the first A raw vibration signal sample, representing the raw time-series or spectral data acquired from the sensor; This is a preprocessing and feature extraction function, representing the operations of filtering, transforming, and feature aggregation on the original signal; It is a positive integer; By mapping the original signal into denoised features through preprocessing and feature extraction functions, it can adapt to the complex distribution of vibration signals of thermal power equipment and retain complete fault modes at multiple scales. The preprocessing and feature extraction functions are calculated using a combination of bandpass filtering and wavelet packet transform to avoid signal distortion and loss of fault features caused by a single method. The preprocessing method for vibration signal data of thermal power equipment is expressed as follows: ; In the formula, This represents a bandpass filter function, which characterizes the initial filtering and frequency band extraction of noise interference from the original vibration signal using a specified bandpass frequency range; This represents the wavelet packet transform function, which extracts more subtle and precise fault feature information from the vibration signal by adaptively decomposing and reconstructing the filtered signal at multiple scales.

[0005] As a further technical solution of the present invention, the forward propagation activation function of the probabilistic neural network is defined as follows: Since vibration signals of thermal power equipment often exhibit high-dimensional and nonlinear distributions in fault diagnosis scenarios, relying solely on linear models is insufficient for accurate classification. By employing a two-stage activation function in the hidden layer of the probabilistic neural network to achieve adaptive reconstruction of the feature space, the intrinsic distribution relationship of the vibration signals of thermal power equipment can be more effectively explored. The calculation method of the forward propagation activation function of the probabilistic neural network is expressed as follows: ; In the formula, The input data features characterize the vector of preprocessed vibration signal from thermal power equipment. It is a one-stage activation function, and the LeakyReLU activation function is used to prevent the vanishing of negative gradients and preserve the integrity of fault features; It is a two-stage activation function, which is further nonlinearly mapped through the hyperbolic tangent function. The hyperbolic tangent function has a steep decay characteristic, which can enhance the ability to locate transient impacts. These are the weights of the probabilistic neural network; This is the bias of the probabilistic neural network; Use the LeakyReLU activation function; It is the hyperbolic tangent activation function; For offset features; In the above formula, the first-stage activation function and the second-stage activation function are used together as the activation function of neurons in the probabilistic neural network, which enables the intermediate feature vectors processed by the probabilistic neural network to maintain good separability in high-dimensional space. The offset feature is used to adaptively adjust the distribution center position of the neuron's output features. By batch statistically analyzing the feature centers, the network achieves adaptive drift and localization in the high-dimensional feature space, enhancing the network's sensitivity to distinguishing different fault modes. The calculation method is expressed as follows: ; In the formula, For the training set Features of each sample; The total number of samples in the entire training set; This study analyzes the deep analytical capability of the two-stage activation function for vibration signal features through visualization analysis of the three-dimensional feature space distribution. Different points in the three-dimensional feature space represent the distribution of vibration signal features under different fault modes after activation function processing. Specifically, blue and orange points (in the conventional activation function sub-plot) represent the original signal features of two different fault modes (e.g., bearing fault and imbalance fault); green and red points (in the two-stage activation sub-plot) correspond to similar fault features after processing by the two-stage activation function, but the distribution is reconstructed through nonlinear mapping. Experiments compare the feature mapping effects of the conventional activation function and the two-stage activation mechanism of this invention. The results show that in the three-dimensional space composed of "time-amplitude-activation value," the feature points of different fault modes exhibit a mixed and overlapping state in the traditional method. In contrast, this technique, through two-stage nonlinear transformation and feature offset adjustment, enables feature points of different categories to form clustered regions with clear intervals, verifying the advantages of two-stage activation in feature space reconstruction and demonstrating its ability to more effectively distinguish easily confused fault modes. As a further technical solution of the present invention, a dynamic penalty factor is calculated during the forward propagation process of the probabilistic neural network: In the classification task of fault diagnosis, different fault categories have different misclassification costs, and the model is unstable in the early stage of training, requiring a relatively small penalty, while in the later stage, it is necessary to strengthen the correction of misclassification; by adopting a dynamic penalty factor, the penalty increases smoothly and stably with the iteration process, avoiding the convergence risk in the early stage of probabilistic neural network training, and also fully strengthening the correction of fault misclassification in the later stage of probabilistic neural network training. The calculation method is expressed as follows: ; In the formula, For the first The dynamic penalty factor in each iteration represents the intensity of the penalty for misclassified samples; The initial penalty factor represents the base penalty level at the beginning of the iteration; A numerical value representing the current iteration number; For attenuation parameters; It is a logarithmic function with a default base of 10; Indicates the first The number of misclassified samples in the next iteration; The number of samples input into the probabilistic neural network in the current training batch; preferably, Set to 0.1; By adaptively adjusting the penalty intensity by combining the number of iterations and the misclassification ratio, the model achieves a more lenient penalty in the early stage of training and gradually strengthens the error correction capability in the later stage, effectively improving classification stability and accuracy. This paper analyzes the misclassification penalty strategies during training and compares the effects of a fixed penalty coefficient, a simple decay strategy, and the dynamic penalty factor of this invention. Experimental results show that traditional strategies are prone to convergence difficulties in the early stages of training due to excessive penalty, while insufficient penalty in the later stages makes it difficult to correct stubborn misclassified samples. In contrast, this invention, by combining iterative progress with a dynamic adjustment mechanism of real-time misclassification ratio, maintains appropriate penalty in the early stages of training to avoid oscillations and continuously strengthens error correction capabilities in the later stages, resulting in a steady decrease in the number of misclassified samples and significantly improving the model's sensitivity to distinguishing different fault types.

[0006] The dynamic penalty factor refers to the parameter used to dynamically adjust the weighted penalty for misclassified samples during the training process of the probabilistic neural network. As a further technical solution of the present invention, the classification result of the forward propagation of the probabilistic neural network is calculated as follows: The classification result of the forward propagation of the probabilistic neural network is calculated by performing linear combination and probability normalization on the output of the two-stage activation function to obtain the classification prediction result of the sample, which is expressed as follows: ; In the formula, Here, represents the weight matrix of the output layer of the probabilistic neural network, and represents the training parameters, which are updated during training using gradient descent. Here, represents the bias of the output layer of the probabilistic neural network, and represents the training parameters, which are updated during training using gradient descent. It is the Softmax probability normalization function; For the last layer (classification layer) of the probabilistic neural network, for the first... The input features of the nth sample represent the nth... The feature vector is obtained by calculating the original input through a multi-layer activation function of a probabilistic neural network; For the first The predicted output of each sample represents the probabilistic neural network's prediction of the fault type of the sample. As a further technical solution of the present invention, the loss function of the current iteration of the probabilistic neural network is calculated as follows: Since fault diagnosis of thermal power equipment has high requirements for accuracy, sparsity, and model complexity, it is necessary to avoid the network's dependence on invalid parameters while ensuring high accuracy. By combining cross-entropy, L1 regularization, and dynamic penalty factors, the classification accuracy of fault samples is focused on during the training process of the probabilistic neural network, and unnecessary redundant weights are sparsified, thereby balancing accuracy and model lightweighting. The loss function of the probabilistic neural network is expressed as follows: ; In the formula, For probabilistic neural networks, this is the loss function, which represents a comprehensive measure of classification error and network complexity. For the first The true labels of each sample represent the actual categories of thermal power equipment failures; For the first The predicted output of each sample represents the probabilistic neural network's prediction of the fault type of the sample. It is an exponential function; This represents the number of samples input into the probabilistic neural network in the current training batch. For regularization terms of the weights in a probabilistic neural network, this term represents an indicator that controls the complexity of the network. is the L1 norm of the weights of the probabilistic neural network, which characterizes the sparsity of the constrained network parameters; To adjust the parameters, characterizing the influence of the L1 regularization term on the overall loss; For the first The dynamic penalty factor for each iteration. Preferably, Set to 0.3 By employing a combination of cross-entropy, L1 regularization, and a dynamic penalty factor, the probabilistic neural network improves classification accuracy while achieving sparsity of model weights and adaptive control over misclassified samples, effectively avoiding overfitting and reducing network complexity. The regularization term of the probabilistic neural network weights is calculated by combining the L2 norm and a weight smoothing term to balance the magnitude and spatial smoothness of network weights, thereby improving the model's generalization performance. The calculation method is expressed as follows: ; In the formula, The square of the L2 norm of the weight matrix of the probabilistic neural network is represented by the following expression: The mean of the weight matrix is... The weight matrix is ​​the first Line 1 Column elements, Adjust the parameters for the first weight; This is the second weighting adjustment parameter; preferably, Set to 0.3, Set to 0.7; As a further technical solution of the present invention, backpropagation and parameter updating of a probabilistic neural network are performed. Since the learning of complex fault characteristics in the vibration signal of thermal power equipment requires fine control of parameter updates, too large or too small learning step size may lead to convergence difficulties or getting trapped in local optima. By using a correction adjustment function to adjust the parameter update amplitude of the probabilistic neural network in each update, the update amount of weights and biases not only depends on the gradient, but also makes additional corrections based on the current parameters, so as to better dynamically control the convergence process of different fault characteristics of thermal power equipment. The backpropagation and parameter update methods of a probabilistic neural network are represented as follows: ; In the formula, is the weight of the updated probabilistic neural network, representing the final weight parameters after this round of iteration, which is also the initial weight of the probabilistic neural network in the next round of iteration; is the bias of the updated probabilistic neural network, representing the final bias parameters after this round of iteration, which is also the initial bias of the probabilistic neural network in the next round of iteration. The learning rate of the probabilistic neural network represents the adjustment step size for each parameter update. This represents the gradient of the loss function of a probabilistic neural network with respect to the weights. This represents the gradient of the loss function of a probabilistic neural network with respect to the bias. To modify the adjustment function, additional terms are used to adjust the current parameters. Preferably, Set to 0.01 By introducing an adaptive adjustment term for the current parameters through the adjustment function, the parameter updates are ensured to follow the gradient direction while preventing over-updates, effectively alleviating oscillations or convergence stagnation during network training and improving training stability. The correction adjustment function combines the gradient history information and parameter value distribution characteristics of the current parameters. Through parameter standardization and dynamic correction of gradient momentum, it makes parameter updates smoother and more efficient, achieving dynamic correction control. The calculation method is expressed as follows:

[0007] In the formula It is the variance function; The historical gradient momentum term is a parameter; The mean of the weight matrix; The mean of the bias vector; Adjust the hyperparameters for the first correction; Adjust the hyperparameters for the second correction; To prevent small constants with denominators of zero, preferably, Set to 0.3, Set to 0.5 The historical gradient momentum term of the parameters is calculated using an exponentially weighted moving average of historical gradients. This method records the historical trend of parameter updates, ensuring the smoothness and stability of the update direction and effectively avoiding training oscillations or local optima caused by gradient fluctuations. The calculation method is expressed as follows:

[0008] In the formula, The gradient momentum term of the weight parameters represents the combined trend of the current weight gradient and the historical gradient. The gradient momentum term of the bias parameter represents the combined trend of the current bias gradient and the historical gradient. This represents the historical gradient momentum value of the weight parameters in the previous iteration (set to 0 by default in the first iteration). This represents the historical gradient momentum value of the bias parameter in the previous iteration (set to 0 by default in the first iteration). This represents the gradient of the loss function of a probabilistic neural network with respect to the weights. This represents the gradient of the loss function of a probabilistic neural network with respect to the bias. This is the momentum smoothing coefficient, and its value ranges from [value range missing]. Between these, it is used to suppress gradient abrupt changes and alleviate parameter oscillations caused by complex fault characteristics in the vibration signals of thermal power equipment; This is the average weight from the previous iteration; This is the mean of the bias from the previous iteration; This is a sign function, and the output parameter's mean changes in the direction (+1 or -1). This indicates element-wise multiplication. Preferably, Set to 0.9; As a further technical solution of the present invention, a training termination condition is set; the training process of the probabilistic neural network follows the gradient descent training method of conventional neural networks, that is, the forward propagation and error backpropagation process of the probabilistic neural network are repeatedly iterated until the classification on the training set is less than a preset threshold, then the iteration stops, indicating that the model training is complete; the training set refers to a set of vibration signal samples of thermal power equipment with known categories, which is used to guide the probabilistic neural network training process to continuously optimize parameters until the classification error converges within the preset accuracy requirement; As a further technical solution of the present invention, fault diagnosis of thermal power equipment; after the probabilistic neural network is trained, its inference process is used to realize fault diagnosis of thermal power equipment, specifically including the following steps: 1) Input the vibration signal data of actual thermal power equipment into the trained probabilistic neural network model. These data are first processed by preprocessing and feature extraction functions to ensure that the format and features of the input data are consistent with those during training. 2) The input feature data is propagated forward through the multi-layer structure of the network. Each layer applies corresponding weights, biases and activation functions to gradually transform the input data into a high-level feature representation. It should be noted that the two-stage activation function defined in the probabilistic neural network enhances the model's ability to capture and classify fault features. 3) In the last layer of the probabilistic neural network, the Softmax function is used to transform the final feature vector into a probability distribution. These probabilities represent various possible fault types, and the category with the highest probability is selected as the predicted fault type. 4) Provide the prediction results to the maintenance team or automated system to make corresponding maintenance decisions or make automatic adjustments, which may include adjusting operating parameters, arranging repairs or replacing damaged parts; Compared with the prior art, the beneficial effects of the present invention are: This invention employs a combined approach of bandpass filtering and wavelet packet transform for vibration signal preprocessing, avoiding information loss caused by a single method. It can extract key fault features across multiple frequency bands and time domains, solving the problem of traditional filtering methods failing under multi-condition coupled vibration environments and improving the ability to retain fault modes completely. This invention employs a dynamic penalty factor based on the misclassification rate and iteration process. The penalty is reduced in the early stage of training to avoid convergence difficulties, and the penalty is gradually increased in the later stage to improve the ability to correct misclassified samples. This solves the problem that traditional fixed penalty coefficients or simple decay strategies are difficult to adapt to the needs of different training stages, making model training more stable and efficient. This invention, based on cross-entropy loss, employs L1 regularization (sparses model parameters) and a dynamic penalty factor (increases the penalty for misclassification), balancing classification accuracy and model lightweightness. It overcomes the limitation of traditional L2 regularization in solely controlling model complexity, improves the model's generalization ability, and reduces the risk of overfitting. This invention employs a modified adjustment function in the gradient descent process of a probabilistic neural network. By combining historical gradient information with the mean distribution of parameters, the weights and biases are dynamically adjusted, making parameter updates smoother and avoiding oscillations or local optima. This solves the problem that fixed learning rate or simple decay strategies are difficult to adapt to gradient changes during training, thus improving convergence speed and stability. Attached Figure Description

[0009] Figure 1 This is a time-domain waveform diagram of the vibration signal of the thermal power equipment according to the present invention.

[0010] Figure 2 This is a comparison chart of the feature retention capabilities of different preprocessing methods of the present invention.

[0011] Figure 3 These are the feature space distribution map of the conventional activation function and the feature space reconstruction map of the two-stage activation function in this invention.

[0012] Figure 4 This is a graph showing the change in the number of misclassifications under different penalty strategies in this invention. Detailed Implementation

[0013] 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.

[0014] Please see Figure 1-4 In this embodiment of the invention, a fault diagnosis method for thermal power equipment based on artificial intelligence specifically includes the following steps: S1. Training data acquisition for thermal power equipment: Data is acquired through vibration sensors installed on key parts of the equipment. These sensors can monitor and record vibration signals during equipment operation in real time. The acquired data is recorded in time series format, with each data point including a timestamp and vibration intensity value. The storage format is a structured database format, where each row represents a record at a time point, and the columns represent the timestamp and vibration data from different sensors, respectively. S2. Initial screening of thermal power equipment data; removal of shutdown data, elimination of outliers, and data balancing. S3. Training data annotation for thermal power equipment; The training data is annotated manually by combining equipment operation and maintenance records with expert experience to manually annotate vibration signal data. S4. Fault diagnosis of thermal power equipment; using probabilistic neural networks for fault diagnosis of thermal power equipment. The removal of shutdown data in S2 is as follows: Since the vibration signal of thermal power equipment is close to zero when it is shut down, these data have no practical value for the fault diagnosis model. Therefore, it is necessary to remove all data collected during shutdown. Outlier removal: Statistical methods are used to detect and remove abnormal vibration signal data to avoid noise data interfering with model training. Data balancing: In actual data collection, normal operation data is often much more than fault data, which may cause the model training to be biased towards the normal category. Undersampling or oversampling methods are used to make the number of samples of each category relatively balanced. As a further technical solution of the present invention, as described in S4; the training process of the probabilistic neural network is as follows: preprocessing the vibration signal data of thermal power equipment; the method of preprocessing the vibration signal data of thermal power equipment is expressed as follows: In the formula, For the first The feature vectors obtained after preprocessing represent the effective features of the vibration signal data of thermal power equipment after denoising and interference removal. For the first A raw vibration signal sample, representing the raw time-series or spectral data acquired from the sensor; This is a preprocessing and feature extraction function, representing the operations of filtering, transforming, and feature aggregation on the original signal; It is a positive integer; By mapping the original signal into denoised features through preprocessing and feature extraction functions, it can adapt to the complex distribution of vibration signals of thermal power equipment and retain complete fault modes at multiple scales. The preprocessing and feature extraction functions are calculated using a combination of bandpass filtering and wavelet packet transform to avoid signal distortion and loss of fault features caused by a single method. The preprocessing method for vibration signal data of thermal power equipment is expressed as follows:

[0015] In the formula, This represents a bandpass filter function, which characterizes the initial filtering and frequency band extraction of noise interference from the original vibration signal using a specified bandpass frequency range; This represents the wavelet packet transform function, which extracts more subtle and precise fault feature information from the vibration signal by adaptively decomposing and reconstructing the filtered signal at multiple scales.

[0016] By adopting the above technical solution, a comprehensive method of bandpass filtering and wavelet packet transform is used for vibration signal preprocessing. This avoids information loss caused by a single method and can extract key fault features in multiple frequency bands and time domains. It solves the problem that traditional filtering methods are prone to failure in multi-condition coupled vibration environments and improves the ability to completely preserve fault modes. In this embodiment, as a further technical solution of the present invention, the forward propagation activation function of the probabilistic neural network is defined as follows: Since the vibration signals of thermal power equipment often exhibit high-dimensional and nonlinear distributions in fault diagnosis scenarios, relying solely on linear models is insufficient for accurate classification. By employing a two-stage activation function in the hidden layer of the probabilistic neural network to achieve adaptive reconstruction of the feature space, the inherent distribution relationship of the vibration signals of thermal power equipment can be more effectively explored. The calculation method of the forward propagation activation function of the probabilistic neural network is expressed as follows: ; In the formula, The input data features characterize the vector of preprocessed vibration signal from thermal power equipment. It is a one-stage activation function, and the LeakyReLU activation function is used to prevent the vanishing of negative gradients and preserve the integrity of fault features; It is a two-stage activation function, which is further nonlinearly mapped through the hyperbolic tangent function. The hyperbolic tangent function has a steep decay characteristic, which can enhance the ability to locate transient impacts. These are the weights of the probabilistic neural network; This is the bias of the probabilistic neural network; Use the LeakyReLU activation function; It is the hyperbolic tangent activation function; For offset features; In the above formula, the first-stage activation function and the second-stage activation function are used together as the activation function of neurons in the probabilistic neural network, which enables the intermediate feature vectors processed by the probabilistic neural network to maintain good separability in high-dimensional space. The offset feature is used to adaptively adjust the distribution center position of the neuron's output features. By batch statistically analyzing the feature centers, the network achieves adaptive drift and localization in the high-dimensional feature space, enhancing the network's sensitivity to distinguishing different fault modes. The calculation method is expressed as follows: ; In the formula, For the training set Features of each sample; The total number of samples in the entire training set; This study analyzes the deep analytical capability of the two-stage activation function for vibration signal features through visualization analysis of the three-dimensional feature space distribution. Different points in the three-dimensional feature space represent the distribution of vibration signal features under different fault modes after activation function processing. Specifically, blue and orange points (in the conventional activation function sub-plot) represent the original signal features of two different fault modes (e.g., bearing fault and imbalance fault); green and red points (in the two-stage activation sub-plot) correspond to similar fault features after processing by the two-stage activation function, but the distribution is reconstructed through nonlinear mapping. Experiments compare the feature mapping effects of the conventional activation function and the two-stage activation mechanism of this invention. The results show that in the three-dimensional space composed of "time-amplitude-activation value," the feature points of different fault modes exhibit a mixed and overlapping state in the traditional method. In contrast, this technique, through two-stage nonlinear transformation and feature offset adjustment, enables feature points of different categories to form clustered regions with clear intervals, verifying the advantages of two-stage activation in feature space reconstruction and demonstrating its ability to more effectively distinguish easily confused fault modes. In this embodiment, as a further technical solution of the present invention, a dynamic penalty factor is calculated during the forward propagation process of the probabilistic neural network: In the classification task of fault diagnosis, different fault categories have different misclassification costs, and the model is unstable in the early stage of training, requiring a relatively small penalty, while in the later stage, it is necessary to strengthen the correction of misclassification; by adopting a dynamic penalty factor, the penalty increases smoothly and stably with the iteration process, avoiding the convergence risk in the early stage of probabilistic neural network training, and also fully strengthening the correction of fault misclassification in the later stage of probabilistic neural network training. The calculation method is expressed as follows: ; In the formula, For the first The dynamic penalty factor in each iteration represents the intensity of the penalty for misclassified samples; The initial penalty factor represents the base penalty level at the beginning of the iteration; A numerical value representing the current iteration number; For attenuation parameters; It is a logarithmic function with a default base of 10; Indicates the first The number of misclassified samples in the next iteration; This represents the number of samples input into the probabilistic neural network in the current training batch. Preferably, Set to 0.1; By adaptively adjusting the penalty intensity by combining the number of iterations and the misclassification ratio, the model achieves a more lenient penalty in the early stage of training and gradually strengthens the error correction capability in the later stage, effectively improving classification stability and accuracy. This paper analyzes the misclassification penalty strategies during training and compares the effects of a fixed penalty coefficient, a simple decay strategy, and the dynamic penalty factor of this invention. Experimental results show that traditional strategies are prone to convergence difficulties in the early stages of training due to excessive penalty, while insufficient penalty in the later stages makes it difficult to correct stubborn misclassified samples. In contrast, this invention, by combining iterative progress with a dynamic adjustment mechanism of real-time misclassification ratio, maintains appropriate penalty in the early stages of training to avoid oscillations and continuously strengthens error correction capabilities in the later stages, resulting in a steady decrease in the number of misclassified samples and significantly improving the model's sensitivity to distinguishing different fault types.

[0017] The dynamic penalty factor refers to the parameter used to dynamically adjust the weighted penalty for misclassified samples during the training process of the probabilistic neural network. As a further technical solution of the present invention, the classification result of the forward propagation of the probabilistic neural network is calculated as follows: The classification result of the forward propagation of the probabilistic neural network is calculated by performing linear combination and probability normalization on the output of the two-stage activation function to obtain the classification prediction result of the sample, which is expressed as follows: ; In the formula, Here, represents the weight matrix of the output layer of the probabilistic neural network, and represents the training parameters, which are updated during training using gradient descent. Here, represents the bias of the output layer of the probabilistic neural network, and represents the training parameters, which are updated during training using gradient descent. It is the Softmax probability normalization function; For the last layer (classification layer) of the probabilistic neural network, for the first... The input features of the nth sample represent the nth... The feature vector is obtained by calculating the original input through a multi-layer activation function of a probabilistic neural network; For the first The predicted output of each sample represents the probabilistic neural network's prediction of the fault type of the sample. By adopting the above technical solution, a dynamic penalty factor based on the misclassification rate and iteration process is used. The penalty is reduced in the early stage of training to avoid convergence difficulties, and the penalty is gradually increased in the later stage to improve the correction ability of misclassified samples. This solves the problem that traditional fixed penalty coefficients or simple decay strategies are difficult to adapt to the needs of different training stages, making model training more stable and efficient. Furthermore, as a further technical solution of the present invention, the loss function of the current iteration of the probabilistic neural network is calculated as follows: Since fault diagnosis of thermal power equipment has high requirements for accuracy, sparsity, and model complexity, it is necessary to avoid the network's dependence on invalid parameters while ensuring high accuracy. By combining cross-entropy, L1 regularization, and dynamic penalty factors, the classification accuracy of fault samples is focused on during the training process of the probabilistic neural network, and unnecessary redundant weights are sparsified, thereby balancing accuracy and model lightweighting. The loss function of the probabilistic neural network is expressed as follows: ; In the formula, For probabilistic neural networks, this is the loss function, which represents a comprehensive measure of classification error and network complexity. For the first The true labels of each sample represent the actual categories of thermal power equipment failures; For the first The predicted output of each sample represents the probabilistic neural network's prediction of the fault type of the sample. It is an exponential function; This represents the number of samples input into the probabilistic neural network in the current training batch. For regularization terms of the weights in a probabilistic neural network, this term represents an indicator that controls the complexity of the network. is the L1 norm of the weights of the probabilistic neural network, which characterizes the sparsity of the constrained network parameters; To adjust the parameters, characterizing the influence of the L1 regularization term on the overall loss; For the first The dynamic penalty factor for each iteration. Preferably, Set to 0.3 By employing a combination of cross-entropy, L1 regularization, and a dynamic penalty factor, the probabilistic neural network improves classification accuracy while achieving sparsity of model weights and adaptive control over misclassified samples, effectively avoiding overfitting and reducing network complexity. The regularization term of the probabilistic neural network weights is calculated by combining the L2 norm and a weight smoothing term to balance the magnitude and spatial smoothness of network weights, thereby improving the model's generalization performance. The calculation method is expressed as follows: ; In the formula, The square of the L2 norm of the weight matrix of the probabilistic neural network is represented by the following expression: The mean of the weight matrix is... The weight matrix is ​​the first Line 1 Column elements, Adjust the parameters for the first weight; This is the second weighting adjustment parameter. Preferably, Set to 0.3, Set to 0.7; In this embodiment, as a further technical solution of the present invention, backpropagation and parameter updating of the probabilistic neural network are performed. Since the learning of complex fault characteristics in the vibration signal of thermal power equipment requires fine control of parameter updates, too large or too small learning step size may lead to convergence difficulties or getting trapped in local optima. By using a correction adjustment function to adjust the parameter update amplitude of the probabilistic neural network in each update, the update amount of weights and biases not only depends on the gradient, but also makes additional corrections based on the current parameters, so as to better dynamically control the convergence process of different fault characteristics of thermal power equipment. The backpropagation and parameter update methods of a probabilistic neural network are represented as follows: In the formula, is the weight of the updated probabilistic neural network, representing the final weight parameters after this round of iteration, which is also the initial weight of the probabilistic neural network in the next round of iteration; is the bias of the updated probabilistic neural network, representing the final bias parameters after this round of iteration, which is also the initial bias of the probabilistic neural network in the next round of iteration. The learning rate of the probabilistic neural network represents the adjustment step size for each parameter update. This represents the gradient of the loss function of a probabilistic neural network with respect to the weights. This represents the gradient of the loss function of a probabilistic neural network with respect to the bias. To modify the adjustment function, additional terms are used to adjust the current parameters. Preferably, Set to 0.01 By introducing an adaptive adjustment term for the current parameters through the adjustment function, the parameter updates are ensured to follow the gradient direction while preventing over-updates, effectively alleviating oscillations or convergence stagnation during network training and improving training stability. The correction adjustment function combines the gradient history information and parameter value distribution characteristics of the current parameters. Through parameter standardization and dynamic correction of gradient momentum, it makes parameter updates smoother and more efficient, achieving dynamic correction control. The calculation method is expressed as follows:

[0018] In the formula It is the variance function; The historical gradient momentum term is a parameter; The mean of the weight matrix; The mean of the bias vector; Adjust the hyperparameters for the first correction; Adjust the hyperparameters for the second correction; To prevent small constants with denominators of zero, preferably, Set to 0.3, Set to 0.5 The historical gradient momentum term of the parameters is calculated using an exponentially weighted moving average of historical gradients. This method records the historical trend of parameter updates, ensuring the smoothness and stability of the update direction and effectively avoiding training oscillations or local optima caused by gradient fluctuations. The calculation method is expressed as follows:

[0019] In the formula, The gradient momentum term of the weight parameters represents the combined trend of the current weight gradient and the historical gradient. The gradient momentum term of the bias parameter represents the combined trend of the current bias gradient and the historical gradient. This represents the historical gradient momentum value of the weight parameters in the previous iteration (set to 0 by default in the first iteration). This represents the historical gradient momentum value of the bias parameter in the previous iteration (set to 0 by default in the first iteration). This represents the gradient of the loss function of a probabilistic neural network with respect to the weights. This represents the gradient of the loss function of a probabilistic neural network with respect to the bias. This is the momentum smoothing coefficient, and its value ranges from [value range missing]. Between these, it is used to suppress gradient abrupt changes and alleviate parameter oscillations caused by complex fault characteristics in the vibration signals of thermal power equipment; This is the average weight from the previous iteration; This is the mean of the bias from the previous iteration; This is a sign function, and the output parameter's mean changes in the direction (+1 or -1). This indicates element-wise multiplication. Preferably, Set to 0.9; As a further technical solution of the present invention, a training termination condition is set; the training process of the probabilistic neural network follows the gradient descent training method of conventional neural networks, that is, the forward propagation and error backpropagation process of the probabilistic neural network are repeatedly iterated until the classification on the training set is less than a preset threshold, then the iteration stops, indicating that the model training is complete; the training set refers to a set of vibration signal samples of thermal power equipment with known categories, which is used to guide the probabilistic neural network training process to continuously optimize parameters until the classification error converges within the preset accuracy requirement; By adopting the above technical solution, based on cross-entropy loss, L1 regularization (sparses model parameters) and dynamic penalty factor (increases misclassification penalty) are used to balance classification accuracy and model lightweighting. This solves the limitation of traditional L2 regularization in controlling model complexity, improves the model's generalization ability, and reduces the risk of overfitting. In this embodiment, for fault diagnosis of thermal power equipment, after the probabilistic neural network is trained, its inference process is used to realize fault diagnosis of thermal power equipment, specifically including the following steps: 1) Input the vibration signal data of actual thermal power equipment into the trained probabilistic neural network model. These data are first processed by preprocessing and feature extraction functions to ensure that the format and features of the input data are consistent with those during training. 2) The input feature data is propagated forward through the multi-layer structure of the network. Each layer applies corresponding weights, biases and activation functions to gradually transform the input data into a high-level feature representation. It should be noted that the two-stage activation function defined in the probabilistic neural network enhances the model's ability to capture and classify fault features. 3) In the last layer of the probabilistic neural network, the Softmax function is used to transform the final feature vector into a probability distribution. These probabilities represent various possible fault types, and the category with the highest probability is selected as the predicted fault type. 4) Provide the prediction results to the maintenance team or automated system to make corresponding maintenance decisions or make automatic adjustments, which may include adjusting operating parameters, arranging repairs or replacing damaged parts; By adopting the above technical solution, a correction adjustment function is used in the gradient descent process of probabilistic neural networks. Combining historical gradient information and the mean distribution of parameters, the weights and biases are dynamically adjusted, making parameter updates smoother and avoiding oscillations or local optima. This solves the problem that fixed learning rate or simple decay strategies are difficult to adapt to gradient changes during training, and improves convergence speed and stability.

[0020] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0021] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for fault diagnosis of thermal power equipment based on artificial intelligence, characterized in that: Specifically, the following steps are included: S1. Training data acquisition for thermal power equipment: Data is acquired through vibration sensors installed on key parts of the equipment. These sensors can monitor and record vibration signals during equipment operation in real time. The acquired data is recorded in time series format, with each data point including a timestamp and vibration intensity value. The storage format is a structured database format, where each row represents a record at a time point, and the columns represent the timestamp and vibration data from different sensors, respectively. S2. Initial screening of thermal power equipment data; removal of shutdown data, elimination of outliers, and data balancing. S3. Training data annotation for thermal power equipment; The training data is annotated manually by combining equipment operation and maintenance records with expert experience to manually annotate vibration signal data. S4. Fault diagnosis of thermal power equipment: Input vibration signal data into a probabilistic neural network and use the probabilistic neural network to diagnose faults in thermal power equipment. The training process for a probabilistic neural network is as follows: Preprocessing of vibration signal data from thermal power equipment; the method for preprocessing the vibration signal data from thermal power equipment is as follows: ； In the formula, For the first The feature vectors obtained after preprocessing represent the effective features of the vibration signal data of thermal power equipment after denoising and interference removal. For the first A raw vibration signal sample, representing the raw time-series or spectral data acquired from the sensor; This is a preprocessing and feature extraction function, representing the operations of filtering, transforming, and feature aggregation on the original signal; It is a positive integer; The original signal is mapped into denoised features through preprocessing and feature extraction functions; The preprocessing and feature extraction functions are calculated using a combination of bandpass filtering and wavelet packet transform. The preprocessing method for the vibration signal data of thermal power equipment is expressed as follows: ； In the formula, This represents a bandpass filter function, which characterizes the initial filtering and frequency band extraction of noise interference from the original vibration signal using a specified bandpass frequency range; Represents the wavelet packet transform function; Define the forward propagation activation function for a probabilistic neural network. A two-stage activation function is used in the hidden layers of the probabilistic neural network. The calculation method for the forward propagation activation function of a probabilistic neural network is expressed as follows: ； In the formula, The input data features characterize the vector of preprocessed vibration signal from thermal power equipment. It is a one-stage activation function. It is a two-stage activation function; These are the weights of the probabilistic neural network; This is the bias of the probabilistic neural network; Use the LeakyReLU activation function; It is the hyperbolic tangent activation function; For offset features; In the above formula, the first-stage activation function and the second-stage activation function are used together as the activation function of neurons in the probabilistic neural network, which enables the intermediate feature vectors processed by the probabilistic neural network to maintain good separability in high-dimensional space. The offset feature is used to adaptively adjust the distribution center position of the neuron's output features. By batch statistically analyzing the feature centers, the network achieves adaptive drift and localization in the high-dimensional feature space, enhancing the network's sensitivity to distinguishing different fault modes. The calculation method is expressed as follows: ； In the formula, For the training set Features of each sample; The total number of samples in the entire training set; The dynamic penalty factor in the forward propagation process of a probabilistic neural network is calculated as follows: ； In the formula, For the first The dynamic penalty factor for each iteration; As the initial penalty factor; A numerical value representing the current iteration number; For attenuation parameters; It is a logarithmic function with a default base of 10; Indicates the first The number of misclassified samples in the next iteration; This represents the number of samples input into the probabilistic neural network in the current training batch. The dynamic penalty factor refers to the parameter used to dynamically adjust the weighted penalty for misclassified samples during the training process of the probabilistic neural network. Calculate the loss function of the current iteration of the probabilistic neural network; the loss function of the probabilistic neural network is expressed as follows: ； In the formula, For probabilistic neural networks, this is the loss function, which represents a comprehensive measure of classification error and network complexity. For the first The true labels of each sample represent the actual categories of thermal power equipment failures; For the first The predicted output of each sample represents the probabilistic neural network's prediction of the fault type of the sample. It is an exponential function; This represents the number of samples input into the probabilistic neural network in the current training batch. For regularization terms of the weights in a probabilistic neural network, this term represents an indicator that controls the complexity of the network. is the L1 norm of the weights of the probabilistic neural network, which characterizes the sparsity of the constrained network parameters; To adjust the parameters, characterizing the influence of the L1 regularization term on the overall loss; For the first The dynamic penalty factor for each iteration; The regularization term of the probabilistic neural network weights is calculated by combining cross-entropy, L1 regularization, and a dynamic penalty factor, and is expressed as follows: ； In the formula, The square of the L2 norm of the weight matrix of the probabilistic neural network is represented by the following expression: The mean of the weight matrix is... The weight matrix is ​​the first Line 1 Column elements, Adjust the parameters for the first weight; Adjust the parameters for the second weight; Perform backpropagation and parameter update of the probabilistic neural network; The backpropagation and parameter update methods of a probabilistic neural network are represented as follows: ； In the formula, The weights of the updated probabilistic neural network represent the final weight parameters after this round of iteration, which are also the initial weights of the probabilistic neural network in the next round of iteration. The learning rate of the probabilistic neural network represents the adjustment step size for each parameter update. This represents the gradient of the loss function of a probabilistic neural network with respect to the weights. This is the bias for the updated probabilistic neural network. This represents the gradient of the loss function of a probabilistic neural network with respect to the bias. For adjusting the function, it represents additional terms that modify the current parameters; The correction adjustment function combines the gradient history information and parameter value distribution characteristics of the current parameters. Through parameter standardization and dynamic correction of gradient momentum, it makes parameter updates smoother and more efficient, achieving dynamic correction control. The calculation method is expressed as follows: ； In the formula It is the variance function; The historical gradient momentum term is a parameter; The mean of the weight matrix; The mean of the bias vector; Adjust the hyperparameters for the first correction; Adjust the hyperparameters for the second correction; To prevent small constants with a denominator of zero; The historical gradient momentum term of the parameter is calculated using an exponentially weighted moving average of the historical gradient. This method records the historical trend of parameter updates, ensuring the smoothness and stability of the update direction. The calculation method is expressed as follows: ； In the formula, The gradient momentum term of the weight parameters represents the combined trend of the current weight gradient and the historical gradient. The gradient momentum term of the bias parameter represents the combined trend of the current bias gradient and the historical gradient. This represents the historical gradient momentum value of the weight parameters during the previous iteration; This represents the historical gradient momentum value of the bias parameter during the previous iteration; This represents the gradient of the loss function of a probabilistic neural network with respect to the weights. This represents the gradient of the loss function of a probabilistic neural network with respect to the bias. This is the momentum smoothing coefficient, and its value ranges from [value range missing]. between; This is the average weight from the previous iteration; This is the mean of the bias from the previous iteration; It is a symbolic function; This indicates element-wise multiplication.

2. The method for fault diagnosis of thermal power equipment based on artificial intelligence according to claim 1, characterized in that: The classification result of the forward propagation of the probabilistic neural network is calculated by performing a linear combination and probability normalization on the output of the two-stage activation function to obtain the classification prediction result of the sample, expressed as: ； In the formula, This is the weight matrix of the output layer of the probabilistic neural network; This is the bias of the output layer of the probabilistic neural network; It is the Softmax probability normalization function; For the last layer of a probabilistic neural network, for the first... The input features of the nth sample represent the nth... The feature vector is obtained by calculating the original input through a multi-layer activation function of a probabilistic neural network; For the first The predicted output of each sample represents the probabilistic neural network's prediction of the fault type of the sample.

3. The method for fault diagnosis of thermal power equipment based on artificial intelligence according to claim 1, characterized in that: Set the training termination condition; the training set refers to the set of vibration signal samples of thermal power equipment of known categories, which is used to guide the probabilistic neural network training process to continuously optimize parameters until the classification error converges within the preset accuracy requirements.

4. The method for fault diagnosis of thermal power equipment based on artificial intelligence according to claim 1, characterized in that: Fault diagnosis of thermal power equipment: After the probabilistic neural network is trained, its inference process is used to realize fault diagnosis of thermal power equipment, specifically including the following steps: 1) Input the vibration signal data of actual thermal power equipment into the trained probabilistic neural network model. These data are first processed by preprocessing and feature extraction functions to ensure that the format and features of the input data are consistent with those during training. 2) The input feature data is propagated forward through the multi-layer structure of the network. Each layer applies corresponding weights, biases and activation functions to gradually transform the input data into a high-level feature representation. 3) In the last layer of the probabilistic neural network, the Softmax function is used to transform the final feature vector into a probability distribution. These probabilities represent various possible fault types, and the category with the highest probability is selected as the predicted fault type. 4) Provide the prediction results to the maintenance team or automated system to make corresponding maintenance decisions or make automatic adjustments, including adjusting operating parameters, arranging repairs or replacing damaged parts.

Images