Thermal power equipment fault diagnosis method based on artificial intelligence

Abstract

The invention discloses a thermal power equipment fault diagnosis method based on artificial intelligence, and relates to the technical field of artificial intelligence. The method comprises the steps of thermal power equipment training data acquisition, thermal power equipment fault diagnosis, thermal power equipment training data labeling and thermal power equipment fault diagnosis. A comprehensive method of band-pass filtering and wavelet packet transformation is adopted for vibration signal preprocessing, information loss caused by a single method is avoided, key fault features can be extracted in multiple frequency bands and time domain ranges, the problem that a traditional filtering method is prone to failure in a multi-working-condition coupling vibration environment is solved, and the complete retention capacity of a fault mode is improved; the dynamic penalty factor based on the misclassification rate and the iteration process is adopted, penalty is reduced in the early stage of training, convergence difficulty is avoided, penalty is gradually enhanced in the later stage, the correction capacity of misclassification samples is improved, the problem that a traditional fixed penalty coefficient or a simple attenuation strategy is difficult to adapt to requirements of different training stages is solved, and model training is more stable and efficient.

Description

Technical Field

[0001] The present invention specifically relates to the technical field of artificial intelligence, and specifically to a method for fault diagnosis of thermal power equipment based on artificial intelligence. Background Art

[0002] Thermal power equipment is an important part of energy production. Its key equipment such as steam turbines, generators, and transformers are prone to being affected by various factors such as high temperature, high-speed rotation, mechanical wear, and electromagnetic interference during long-term operation, resulting in common faults such as bearing damage, rotor imbalance, bearing looseness, and gearbox failures. If these faults cannot be detected and handled in a timely manner, it will not only reduce the operating efficiency of the equipment, but may also lead to serious safety accidents and even cause large-scale shutdown losses. Therefore, establishing an efficient and accurate fault diagnosis method is crucial for the safety and economy of thermal power equipment;

[0003] After retrieval: The Chinese invention patent with the publication number CN115163426B proposes a method and system for fan fault detection and a fan safety system based on AI auscultation. It collects the third sound signal during the operation of the fan blades, performs digital-to-analog conversion and noise reduction processing on the third sound signal to obtain the sound signal C, extracts the actual features of the sound signal C, and compares them with the standard feature reference to calculate the first similarity; in the case where the first similarity is less than or equal to the first preset value, the actual features are compared with any one of the fault feature sets, and the second similarity is calculated respectively. The second similarities are arranged in descending order, and the first M fault features corresponding to the second similarities are taken to obtain the corresponding fault type or a combination of M fault types; after detecting a fault, an alarm is given and a fault diagnosis opinion and a maintenance plan are provided. The above solution can use AI auscultation combined with deep learning technology to extract and compare acoustic features of real-time collected working noise to detect whether the blades have faults in real time;

[0004] The Chinese invention patent with the publication number CN119596909A proposes a method, system, storage medium, and electronic device for diagnosing equipment faults, which relates to the technical field of industrial automation. The method for diagnosing equipment faults includes: determining the target service instance distributed when the equipment registration service is completed for the current equipment in the basic service module; using the target service module to which the service instance belongs to analyze the equipment data of the current equipment, and determining whether the current equipment has a fault according to the analysis result of the analysis. The target service module belongs to the upper service module, and the upper service module at least includes one of the following: equipment status acquisition service module, asynchronous message service module, equipment trend service module, historical data persistence service module, equipment status analysis service module. By adopting the above technical solution, the technical problem of low efficiency of equipment fault diagnosis is solved;

[0005] The Chinese invention patent with the publication number CN119596126A proposes a high-voltage switch fault diagnosis system based on artificial intelligence audio recognition technology, including 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 based on artificial intelligence audio recognition technology can accurately capture weak abnormal sounds during the operation of the high-voltage switch by using high-sensitivity MEMS microphones and advanced audio processing technology, improve the accuracy of fault diagnosis, combine multiple sensor data, such as temperature sensors and vibration sensors, to improve the accuracy and reliability of fault diagnosis. The system can automatically adjust the sensitivity parameters of the model according to the ambient noise level to reduce false alarms;

[0006] When stretching nylon filaments mechanically, only filaments of a certain length can be stretched effectively. In this way, during the stretching process, the continuous transmission effect of nylon filaments cannot be effectively guaranteed, which will reduce the working efficiency of the stretching device;

[0007] The most main problems existing in the prior art are as follows: The activation functions adopted by traditional machine learning methods based on neural network models are difficult to adapt to complex non-linear fault characteristics. Existing fault diagnosis models mostly adopt single activation functions such as ReLU, Sigmoid, Tanh, etc., and it is difficult to retain complex vibration signal characteristics in high-dimensional space, resulting in feature aliasing between different fault modes; There are also the following problems in the prior art that need to be further solved: Existing methods mostly use single filters (low-pass, band-pass or high-pass filtering) or simple time-frequency transforms (Fourier transform, etc.) for signal preprocessing, but in complex working conditions, it is easy to cause signal distortion or loss of fault characteristics; Traditional fault diagnosis models generally adopt a fixed learning rate or a learning rate that decays with time, but this method is difficult to dynamically adjust according to the needs of different training stages, and it is easy to cause difficulties in early convergence or insufficient optimization in the later stage; Existing fault diagnosis methods usually adopt 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 gradient descent or momentum optimization algorithms (such as Adam, SGD, etc.), but lack optimization for parameter distribution and gradient history information, and are easy to fall into local optima. Summary of the Invention

[0008] The purpose of the present invention is to provide a thermal power equipment fault diagnosis method based on artificial intelligence. In this device, a two-stage activation function (LeakyReLU+hyperbolic tangent function) is used in the hidden layer of the probabilistic neural network, which not only retains the negative gradient information and prevents feature loss, but also enhances the ability to locate transient impacts, solves the problem that the traditional single activation function is difficult to capture high-dimensional nonlinear fault characteristics, and improves the model's ability to resolve complex vibration signals, so as to solve the problems of the above-mentioned background technology.

[0009] To achieve the above object, the present invention provides the following technical solutions:

[0010] A method for diagnosing thermal power equipment faults based on artificial intelligence specifically comprises the following steps:

[0011] S1. Thermal power equipment training data collection: collected through vibration sensors installed at key parts of the equipment. These sensors can monitor and record vibration signals of the equipment in real time. The collected data is recorded in the form of time series. Each data point includes a timestamp and a 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 timestamps and vibration data of different sensors respectively.

[0012] S2. Preliminary screening of thermal power equipment data; removal of shutdown data, elimination of outliers, and data balancing;

[0013] S3. Labeling of thermal power equipment training data. The labeling method of training data is manual labeling. By combining equipment operation and maintenance records with expert experience, the vibration signal data is manually labeled.

[0014] S4. Fault diagnosis of thermal power equipment; use probabilistic neural network to diagnose faults of thermal power equipment;

[0015] As a further technical solution of the present invention, the shutdown data in S2 is removed: since the vibration signal of the thermal power equipment in the shutdown state is close to zero, these data have no practical value to the fault diagnosis model, so it is necessary to remove all the collected data during the shutdown period; remove outliers: use statistical methods to detect and remove abnormal vibration signal data to avoid noise data interfering with model training: data balance processing: in actual data collection, normal operation data is often far 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;

[0016] 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 the thermal power equipment; the method of preprocessing the vibration signal data of the thermal power equipment is expressed as:

[0017] x i =f(si )

[0018] where x i is the i-th feature vector obtained after preprocessing, representing the effective features of the vibration signal data of thermal power equipment after denoising and interference removal; s i is the i-th original vibration signal sample, representing the original time series or spectral data collected from the sensor; f(·) is the preprocessing and feature extraction function, representing the operations of filtering, transforming, and feature aggregation on the original signal; i is a positive integer;

[0019] Mapping the original signal into denoised features through the preprocessing and feature extraction function can adapt to the complex distribution of the vibration signals of thermal power equipment and retain complete fault patterns at multiple scales;

[0020] The preprocessing and feature extraction function adopts the combined processing of a band-pass filter and wavelet packet transform for calculation to avoid signal distortion and loss of fault features caused by a single method. The way to preprocess the vibration signal data of thermal power equipment is expressed as:

[0021] f(s i ) = WPT(BPF(s i ));

[0022] where BPF(·) represents the band-pass filter function, representing the preliminary filtering of noise interference and frequency band extraction of the original vibration signal using the specified band-pass frequency range; WPT(·) represents the wavelet packet transform function, which extracts more subtle and accurate fault feature information in the vibration signal through adaptive multi-scale decomposition and reconstruction of the filtered signal.

[0023] As a further technical solution of the present invention, define the forward propagation activation function of the probabilistic neural network: Since in the fault diagnosis scenario, the vibration signals of thermal power equipment often show high-dimensional and non-linear distributions, it is difficult to accurately classify solely relying on linear models; by adopting a two-stage activation function in the hidden layer of the probabilistic neural network, the feature space can be adaptively reconstructed, and the internal distribution relationship of the vibration signals of thermal power equipment can be more effectively mined: the calculation method of the forward propagation activation function of the probabilistic neural network is expressed as:

[0024] H (1) (x) = LeRe(Wx + b)

[0025] H (2) (x) = tanh(W·∥H (1) (x) - x py ∥ + b);

[0026] where x is the input data feature, representing the vector after preprocessing of the vibration signal of the thermal power equipment; H (1)(·) is a one-stage activation function. The LeakyReLU activation function is adopted to prevent the disappearance of negative gradients and preserve the integrity of fault features; H (2) (·) is a two-stage activation function. The hyperbolic tangent function is used for further non-linear mapping. The hyperbolic tangent function has a steep decay characteristic and can enhance the positioning ability for transient impacts; W is the weight of the probabilistic neural network; b is the bias of the probabilistic neural network; LeRe(·) is the LeakyReLU activation function; tanh(·) is the hyperbolic tangent activation function; x py is the offset feature;

[0027] In the above formula, the one-stage activation function and the two-stage activation function jointly serve as the activation function of neurons in the probabilistic neural network, enabling the intermediate feature vectors processed by the probabilistic neural network to maintain good separability in the high-dimensional space;

[0028] The offset feature is used to adaptively adjust the distribution center position of the neuron output features. By statistically calculating the feature center in batches, the network realizes adaptive drift and positioning in the high-dimensional feature space, enhancing the sensitivity of the network to distinguish different fault modes. The calculation method is expressed as:

[0029]

[0030] In the formula, x j is the feature of the j-th sample in the training set; N is the total number of samples in the entire training set;

[0031] Through the visual analysis of the three-dimensional feature space distribution, the deep analysis ability of the two-stage activation function for vibration signal features is analyzed. Different points in the three-dimensional feature space diagram represent the distribution states of vibration signal features under different fault modes after being processed by the activation function. Specifically, the blue and orange points (in the sub-diagram of the conventional activation function) represent the original signal features of two different fault modes (such as bearing faults and imbalance faults); the green and red points (in the sub-diagram of the two-stage activation) correspond to the same type of fault features after being processed by the two-stage activation function, but the distribution is reconstructed through non-linear mapping. The experiment compares the feature mapping effects of the conventional activation function and the two-stage activation mechanism of the present invention. The results show that in the three-dimensional space composed of "time - amplitude - activation value" of the traditional method, the feature points of different fault modes show a mixed and overlapping state, while through the two-stage non-linear transformation and feature offset adjustment of this technology, the feature points of different categories form aggregation regions with clear intervals, verifying the advantage of the two-stage activation in feature space reconstruction and being able to more effectively distinguish easily confused fault modes;

[0032] As a further technical solution of the present invention, calculate the dynamic penalty factor during the forward propagation of the probabilistic neural network: In the classification task of fault diagnosis, there are differences in the misclassification costs of different fault categories for diagnosis, and the model is not yet stable in the early stage of training and requires relatively small penalties, while in the later stage, it is necessary to strengthen the correction of misclassifications; by adopting a dynamic penalty factor, the penalty increases smoothly and steadily with the iteration process, avoiding the convergence risk in the initial stage of the probabilistic neural network training, and also being able to fully strengthen the correction of fault misclassifications in the later stage of the probabilistic neural network training. The calculation method is expressed as:

[0033]

[0034] In the formula, γ(t) is the dynamic penalty factor of the t-th iteration, representing the penalty intensity for misclassified samples; γ0 is the initial penalty factor, representing the basic penalty level in the initial stage of iteration; int(t) represents the value of the current iteration number; δ is the attenuation parameter; log(·) is the logarithmic function, with the default base of 10; N mis (t) represents the number of misclassified samples in the t-th iteration; n is the number of samples input to the probabilistic neural network in the current training batch; preferably, δ is set to 0.1;

[0035] By adaptively adjusting the penalty intensity by combining the iteration number and the misclassification ratio, the effect of milder penalty in the initial stage of model training and gradually strengthening the error correction ability in the later stage is achieved, effectively improving the classification stability and accuracy;

[0036] Analyze the misclassification penalty strategy during the training process. The experimental results compare the adjustment effects of the fixed penalty coefficient, the simple attenuation strategy and the dynamic penalty factor of the present invention. The experimental results show that the traditional strategy is prone to convergence difficulties due to excessive penalty in the initial stage of training, and it is difficult to correct stubborn misclassified samples due to insufficient penalty in the later stage. However, the present invention combines the dynamic adjustment mechanism of the iteration process and the real-time misclassification ratio, maintains a moderate penalty in the early stage of training to avoid oscillations, and continuously strengthens the error correction ability in the later stage, so that the number of misclassified samples shows a steady downward trend, significantly improving the discrimination sensitivity of the model to different fault types.

[0037] The dynamic penalty factor refers to a parameter used to dynamically adjust the differential weighted penalty for misclassified samples during the training process of the probabilistic neural network;

[0038] As a further technical solution of the present invention, calculate the classification result of the forward propagation of the probabilistic neural network: Calculate the classification result of the forward propagation of the probabilistic neural network. By performing a linear combination and probability normalization on the output of the two-stage activation function, the classification prediction result of the sample is obtained, which is expressed as:

[0039] y i =sof(W last ·H(2) (x lasti ) + b last );

[0040] Wherein, W last is the weight matrix of the output layer of the probabilistic neural network, which is a training parameter and is updated by the gradient descent method during the training process; b last is the bias of the output layer of the probabilistic neural network, which is a training parameter and is updated by the gradient descent method during the training process; sof(·) is the Softmax probability normalization function; x lasti is the input feature of the i-th sample for the last layer (classification layer) of the probabilistic neural network, representing the feature vector calculated by the multi-layer activation function of the probabilistic neural network for the i-th original input; y i is the predicted output of the i-th sample, representing the fault type prediction of the probabilistic neural network for the sample;

[0041] As a further technical solution of the present invention, calculate the loss function of the current iteration of the probabilistic neural network: Since the 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 a dynamic penalty factor, focus on the classification accuracy of fault samples during the training process of the probabilistic neural network, and sparsify unnecessary redundant weights, so as to balance accuracy and model lightweight; the calculation method of the loss function of the probabilistic neural network is expressed as:

[0042]

[0043] Wherein, L is the loss function of the probabilistic neural network, representing the comprehensive measurement of classification error and network complexity; y i is the true label of the i-th sample, representing the actual category of the thermal power equipment fault; is the predicted output of the i-th sample, representing the fault type prediction of the probabilistic neural network for the sample; exp(·) is the exponential function; n is the number of samples input to the probabilistic neural network in the current training batch; R(W) is the regularization term of the weights of the probabilistic neural network, representing the index for controlling the network complexity; ∥W∥1 is the L1 norm of the weights of the probabilistic neural network, representing the degree of sparsity of the network parameters; η cs is the adjustment parameter, representing the influence degree of the L1 regularization term in the overall loss; γ(t) is the dynamic penalty factor for the t-th iteration. Preferably, η cs is set to 0.3

[0044] By adopting the combined action of cross-entropy, L1 regularization and dynamic penalty factor, the probabilistic neural network improves the classification accuracy while realizing the sparsity of model weights and the adaptive control of misclassified samples, effectively avoiding overfitting and reducing the network complexity; the regularization term of the probabilistic neural network weights is calculated by mixing the L2 norm and the weight smoothing term to take into account both the magnitude and spatial smoothing characteristics of the network weights, improving the generalization performance of the model. The calculation method is expressed as:

[0045]

[0046] In the formula, represents the square of the L2 norm of the probabilistic neural network weight matrix, and μ W is the mean of the weight matrix, W i,j is the element in the i-th row and j-th column of the weight matrix, and α rw is the first weight adjustment parameter; β rw is the second weight adjustment parameter; preferably, α rw is set to 0.3 and β rw is set to 0.7;

[0047] As a further technical solution of the present invention, the backpropagation and parameter update of the probabilistic neural network are performed; due to the learning of complex fault features in the vibration signals of thermal power equipment, fine control of parameter update is required. Too large or too small learning step sizes may both lead to convergence difficulties or getting stuck in local optima; by using a correction adjustment function to adjust the parameter update amplitude of the probabilistic neural network at each update, the update amounts of weights and biases not only depend on the gradient but also make additional corrections according to the current parameters, better dynamically controlling the convergence process for different fault features of thermal power equipment;

[0048] The way of backpropagation and parameter update of the probabilistic neural network is expressed as:

[0049]

[0050] In the formula, W new is the weight of the updated probabilistic neural network, representing the final weight parameter after this round of iteration, that is, 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 parameter after this round of iteration, that is, the initial bias of the probabilistic neural network in the next round of iteration; η is the learning rate of the probabilistic neural network, representing the adjustment step size ratio for each parameter update; represents the gradient of the probabilistic neural network loss function with respect to the weight; represents the gradient of the probabilistic neural network loss function with respect to the bias; σ(·) is the correction adjustment function, representing the additional term for correcting the current parameters. Preferably, η is set to 0.01

[0051] By introducing an adaptive correction term for the current parameter through the correction adjustment function, it is ensured that the parameter update not only follows the gradient direction but also prevents excessive update, effectively alleviating the oscillation or convergence stagnation during network training and improving training stability.

[0052] The correction adjustment function combines the gradient historical information of the current parameter with the parameter value distribution characteristics, and through parameter standardization and gradient momentum dynamic correction, makes the parameter update smoother and more efficient, realizing dynamic correction control. The calculation method is expressed as:

[0053]

[0054] In the formula, Var(·) is the variance function; m fs (·) is the historical gradient momentum term of the parameter; μ W is the mean of the weight matrix; μ b is the mean of the bias vector; λ sg is the first correction adjustment hyperparameter; θ sg is the second correction adjustment hyperparameter; □ is a small constant to prevent the denominator from being zero. Preferably, λ sg is set to 0.3, and θ sg is set to 0.5

[0055] The historical gradient momentum term of the parameter is calculated by the exponential weighted moving average of the historical gradient, which is used to record the historical trend of parameter update, ensure the smoothness and stability of the update direction, and effectively avoid the training oscillation or local optimum problem caused by gradient fluctuation. The calculation method is expressed as:

[0056]

[0057] In the formula, m fs (μ W ) is the gradient momentum term of the weight parameter, which represents the comprehensive trend of the current weight gradient and the historical gradient; m fs (μ b ) is the gradient momentum term of the bias parameter, which represents the comprehensive trend of the current bias gradient and the historical gradient; represents the historical gradient momentum value of the weight parameter at the previous iteration (default set to 0 at the first iteration); represents the historical gradient momentum value of the bias parameter at the previous iteration (default set to 0 at the first iteration); represents the gradient of the probability neural network loss function with respect to the weight; represents the gradient of the probability neural network loss function with respect to the bias; β m is the momentum smoothing coefficient, and its value range is between (0, 1), which is used to suppress gradient mutation and alleviate the parameter oscillation caused by complex fault characteristics in the vibration signal of thermal power equipment; μ W,previs the weight mean of the previous iteration; μ b,prev is the bias mean of the previous iteration; sign(·) is the sign function, which outputs the change direction of the parameter mean (+1 or -1); ⊙ represents element-wise multiplication. Preferably, β m is set to 0.9;

[0058] As a further technical solution of the present invention, a training end condition is set; the training process of the probabilistic neural network follows the training method of the gradient descent method of the conventional neural network, that is, the forward propagation and error backpropagation processes of the probabilistic neural network are repeatedly iterated until the classification on the training set during the training process is less than a preset threshold, then the iteration is stopped, indicating that the model training is completed; the training set refers to a set of vibration signal samples of thermal power equipment with known categories, which is used to guide the continuous optimization of parameters in the training process of the probabilistic neural network until the classification error converges within the preset accuracy requirement;

[0059] As a further technical solution of the present invention, fault diagnosis of thermal power equipment; after the probabilistic neural network training is completed, its inference process is used to realize fault diagnosis of thermal power equipment, which specifically includes the following steps:

[0060] 1) Input the vibration signal data of the actual thermal power equipment into the trained probabilistic neural network model. These data are first processed by the preprocessing and feature extraction function to ensure that the format and features of the input data are consistent with those during training;

[0061] 2) The input feature data undergoes forward propagation through the multi-layer structure of the network. Each layer applies the corresponding weights, biases, and activation functions to gradually convert the input data into a high-level feature representation. It should be noted that through the two-stage activation function defined in the probabilistic neural network, the model's ability to capture and classify fault features is enhanced;

[0062] 3) At the last layer of the probabilistic neural network, the Softmax function is used to convert 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;

[0063] 4) Provide the prediction result to the maintenance team or the automation system for making corresponding maintenance decisions or automatic adjustments, which may include adjusting operation parameters, arranging repairs, or replacing damaged components;

[0064] Compared with the prior art, the beneficial effects of the present invention are:

[0065] 1. The present invention uses a comprehensive method of band - pass filtering + wavelet packet transform for pre - processing vibration signals, avoiding information loss caused by a single method, being able to extract key fault features in a variety of frequency bands and time domains, solving the problem that traditional filtering methods are prone to failure in a multi - condition coupled vibration environment, and improving the ability to completely retain fault modes;

[0066] 2. The present invention uses a dynamic penalty factor based on misclassification rate and iteration process, reducing the penalty in the initial stage of training to avoid convergence difficulties, and gradually increasing the penalty in the later stage to improve the correction ability of misclassified samples, solving the problem that traditional fixed penalty coefficients or simple attenuation strategies are difficult to adapt to the requirements of different training stages, making the model training more stable and efficient;

[0067] 3. The present invention, based on cross - entropy loss, uses L1 regularization (sparsifying model parameters) and a dynamic penalty factor (enhancing misclassification penalty), taking into account both classification accuracy and model lightweighting, solving the limitation of traditional L2 regularization in solely controlling model complexity, improving the generalization ability of the model, and reducing the risk of overfitting;

[0068] 4. The present invention, during the gradient descent process of the probabilistic neural network, uses a modified adjustment function, combines gradient historical information and parameter mean distribution, and dynamically adjusts weights and biases, making parameter updates smoother, avoiding oscillations or local optima, solving the problem that fixed learning rates or simple attenuation strategies are difficult to adapt to gradient changes during training, and improving the convergence speed and stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 is the time - domain waveform diagram of the vibration signal of the thermal power equipment of the present invention.

[0070] Figure 2 is the comparison diagram of the feature retention capabilities of different pre - processing methods of the present invention.

[0071] Figure 3 is the feature space distribution diagram of the conventional activation function and the feature space reconstruction diagram of the two - stage activation function in the present invention.

[0072] Figure 4 is the diagram of the change in the number of misclassified samples with different penalty strategies in the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0073] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0074] See also Figures 1-4 In an embodiment of the present invention, a method for diagnosing a thermal power equipment fault based on artificial intelligence specifically comprises the following steps:

[0075] S1. Thermal power equipment training data collection: collected through vibration sensors installed at key parts of the equipment. These sensors can monitor and record vibration signals of the equipment in real time. The collected data is recorded in the form of time series. Each data point includes a timestamp and a 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 timestamps and vibration data of different sensors respectively.

[0076] S2. Preliminary screening of thermal power equipment data; removal of shutdown data, elimination of outliers, and data balancing;

[0077] S3. Labeling of thermal power equipment training data. The labeling method of training data is manual labeling. By combining equipment operation and maintenance records with expert experience, the vibration signal data is manually labeled.

[0078] S4. Fault diagnosis of thermal power equipment; use probabilistic neural network to diagnose faults of thermal power equipment;

[0079] Removing shutdown data in S2: Since the vibration signal of thermal power equipment in the shutdown state approaches zero, these data have no practical value to the fault diagnosis model, so it is necessary to remove all the collected data during the shutdown period; Removing outliers: Using statistical methods to detect and remove abnormal vibration signal data to avoid noise data interfering with model training: Data balancing processing: 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;

[0080] 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 the thermal power equipment; the method of preprocessing the vibration signal data of the thermal power equipment is expressed as:

[0081] x i =f(s i );where x i is the feature vector obtained after the i-th preprocessing, representing the effective features of the vibration signal data of thermal power equipment after denoising and interference removal; s i is the i-th original vibration signal sample, representing the original time series or spectrum data collected from the sensor; f(·) is the preprocessing and feature extraction function, representing the operation of filtering, transforming and feature aggregation of the original signal; i is a positive integer;

[0082] The original signal is mapped into denoised features through the preprocessing and feature extraction function, which can adapt to the complex distribution of vibration signals of thermal power equipment and retain the complete fault patterns at multiple scales;

[0083] The preprocessing and feature extraction function uses a combination of band-pass filters and wavelet packet transforms for calculation to avoid signal distortion and loss of fault features caused by a single method. The way to preprocess the vibration signal data of thermal power equipment is expressed as:

[0084] f(s i )=WPT(BPF(s i )); In the formula, BPF(·) represents the band-pass filter function, which characterizes the preliminary filtering of noise interference and frequency band extraction of the original vibration signal using the specified band-pass frequency range; WPT(·) represents the wavelet packet transform function, which extracts more subtle and accurate fault feature information in the vibration signal through adaptive multi-scale decomposition and reconstruction of the filtered signal.

[0085] By adopting the above technical solution, a comprehensive method of band-pass filtering + wavelet packet transform is used for vibration signal preprocessing to avoid information loss caused by a single method, and key fault features can be extracted in a variety of frequency bands and time domains, solving the problem that traditional filtering methods are prone to failure in a multi-condition coupled vibration environment and improving the ability to completely retain fault patterns;

[0086] In this embodiment, as a further technical solution of the present invention, the forward propagation activation function of the probabilistic neural network is defined: Since in the fault diagnosis scenario, the vibration signals of thermal power equipment often exhibit high-dimensional and non-linear distributions, it is difficult to accurately classify simply relying on a linear model; by adopting a two-stage activation function in the hidden layer of the probabilistic neural network, the feature space can be adaptively reconstructed, and the internal distribution relationship of the vibration signals of thermal power equipment can be more effectively mined: the calculation method of the forward propagation activation function of the probabilistic neural network is expressed as:

[0087] H (1) (x)=LeRe(Wx + b)

[0088] H (2) (x)=tanh(W·∥H (1) (x)-x py ∥ + b);

[0089] In the formula, x is the input data feature, which characterizes the vector after preprocessing of the vibration signal of the thermal power equipment; H (1) (·) is the first-stage activation function, which prevents the disappearance of negative gradients through the LeakyReLU activation function and retains the integrity of fault features; H (2)(·) is a two-stage activation function that performs a second non-linear mapping through the hyperbolic tangent function. The hyperbolic tangent function has a steep decay characteristic, which can enhance the ability to locate transient shocks; W is the weight of the probabilistic neural network; b is the bias of the probabilistic neural network; LeRe(·) is the LeakyReLU activation function; tanh(·) is the hyperbolic tangent activation function; x py is the offset feature;

[0090] In the above formula, the one-stage activation function and the two-stage activation function together serve as the activation function of the neurons in the probabilistic neural network, which can make the intermediate feature vectors processed by the probabilistic neural network maintain good separability in the high-dimensional space;

[0091] The offset feature is used to adaptively adjust the distribution center position of the neuron output features. By batch-statistically calculating the feature center, the network realizes adaptive drift and positioning in the high-dimensional feature space, enhancing the sensitivity of the network to distinguish different fault modes. The calculation method is expressed as:

[0092]

[0093] In the formula, x j is the feature of the j-th sample in the training set; N is the total number of samples in the entire training set;

[0094] Through the visualization analysis of the three-dimensional feature space distribution, the deep analysis ability of the two-stage activation function for vibration signal features is analyzed. Different points in the three-dimensional feature space diagram represent the distribution states of the vibration signal features under different fault modes after being processed by the activation function. Specifically, the blue and orange points (in the conventional activation function sub-diagram) represent the original signal features of two different fault modes (such as bearing fault and imbalance fault); the green and red points (in the two-stage activation sub-diagram) correspond to the same type of fault features after being processed by the two-stage activation function, but the distribution is reconstructed through non-linear mapping. The experiment compared the feature mapping effects of the conventional activation function and the two-stage activation mechanism of the present invention. The results show that in the three-dimensional space composed of "time - amplitude - activation value", the feature points of different fault modes in the traditional method show a mixed and overlapping state, while this technology through two-stage non-linear transformation and feature offset adjustment makes the feature points of different categories form aggregation regions with clear intervals, verifying the advantage of two-stage activation in feature space reconstruction and being able to more effectively distinguish easily confused fault modes;

[0095] In this embodiment, as a further technical solution of the present invention, a dynamic penalty factor is calculated during the forward propagation of the probabilistic neural network: in the classification task of fault diagnosis, there are differences in the misclassification costs of different fault categories for diagnosis, and the model is not yet stable in the early stage of training and requires relatively small penalties, while in the later stage, the correction of misclassifications needs to be strengthened; by adopting a dynamic penalty factor, the penalty increases smoothly and steadily with the iteration process, avoiding the convergence risk in the initial stage of the probabilistic neural network training, and also being able to fully strengthen the correction of fault misclassifications in the later stage of the probabilistic neural network training. The calculation method is expressed as:

[0096]

[0097] In the formula, γ(t) is the dynamic penalty factor for the t-th iteration, representing the penalty intensity for misclassified samples; γ0 is the initial penalty factor, representing the basic penalty level in the initial stage of iteration; int(t) represents the value of the current iteration number; δ is the attenuation parameter; log(·) is the logarithmic function, with the default base of 10; N mis (t) represents the number of misclassified samples in the t-th iteration; n is the number of samples input to the probabilistic neural network in the current training batch. Preferably, δ is set to 0.1;

[0098] By adaptively adjusting the penalty intensity by combining the iteration number and the misclassification ratio, the effect of milder penalty in the initial stage of model training and gradually strengthening the error correction ability in the later stage is achieved, effectively improving the classification stability and accuracy;

[0099] The misclassification penalty strategy during the training process is analyzed. The adjustment effects of a fixed penalty coefficient, a simple attenuation strategy, and the dynamic penalty factor of the present invention are experimentally compared. The experimental results show that the traditional strategy is prone to convergence difficulties due to excessive penalty in the initial stage of training, and it is difficult to correct stubborn misclassified samples due to insufficient penalty in the later stage. However, the present invention, through a dynamic adjustment mechanism that combines the iteration process and the real-time misclassification ratio, maintains a moderate penalty in the early stage of training to avoid oscillations, and continuously strengthens the error correction ability in the later stage, making the number of misclassified samples show a steady downward trend, significantly improving the discrimination sensitivity of the model to different fault types.

[0100] The dynamic penalty factor refers to a parameter used to dynamically adjust the differential weighted penalty for misclassified samples during the training process of the probabilistic neural network;

[0101] As a further technical solution of the present invention, the classification result of the forward propagation of the probabilistic neural network is calculated: 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, the classification prediction result of the sample is obtained, which is expressed as:

[0102] y i =sof(W last ·H(2) (x lasti ) + b last );

[0103] Wherein, W last is the weight matrix of the output layer of the probabilistic neural network, which is a training parameter and is updated by the gradient descent method during the training process; b last is the bias of the output layer of the probabilistic neural network, which is a training parameter and is updated by the gradient descent method during the training process; sof(·) is the Softmax probability normalization function; x lasti is the input feature of the i-th sample for the last layer (classification layer) of the probabilistic neural network, representing the feature vector calculated by the multi-layer activation function of the probabilistic neural network for the i-th original input; y i is the predicted output of the i-th sample, representing the fault type prediction of the probabilistic neural network for the sample;

[0104] By adopting the above technical solution, a dynamic penalty factor based on the misclassification rate and the iteration process is adopted. At the initial stage of training, the penalty is reduced to avoid convergence difficulties. In the later stage, the penalty is gradually increased to improve the correction ability of misclassified samples, solving the problem that the traditional fixed penalty coefficient or simple attenuation strategy is difficult to meet the requirements of different training stages, making the model training more stable and efficient;

[0105] Furthermore, as a further technical solution of the present invention, calculate the loss function of the current iteration of the probabilistic neural network: Since the fault diagnosis of thermal power equipment has high requirements for accuracy, sparsity and model complexity, it is necessary to avoid the dependence of the network on invalid parameters while ensuring high accuracy; By combining cross-entropy, L1 regularization and a dynamic penalty factor, the classification accuracy of fault samples is focused on during the training process of the probabilistic neural network, and unnecessary redundant weights are sparsified, so as to balance accuracy and model lightweight; The calculation method of the loss function of the probabilistic neural network is expressed as:

[0106]

[0107] Wherein, L is the loss function of the probabilistic neural network, representing the comprehensive measurement of classification error and network complexity; y i is the true label of the i-th sample, representing the actual category of the thermal power equipment fault; y i is the predicted output of the i-th sample, representing the fault type prediction of the probabilistic neural network for the sample; exp(·) is the exponential function; n is the number of samples input to the probabilistic neural network in the current training batch; R(W) is the regularization term of the weights of the probabilistic neural network, representing the index for controlling the network complexity; ∥W∥1 is the L1 norm of the weights of the probabilistic neural network, representing the degree of sparsity of the network parameters; η csis a tuning parameter that characterizes the influence of the L1 regularization term on the overall loss; γ(t) is the dynamic penalty factor at the t-th iteration. Preferably, η cs is set to 0.3

[0108] By adopting the combined action of cross-entropy, L1 regularization, and dynamic penalty factor, the probabilistic neural network can improve the classification accuracy while realizing the sparsity of model weights and the adaptive control of misclassified samples, effectively avoiding overfitting and reducing network complexity; the regularization term of the probabilistic neural network weights is calculated by mixing the L2 norm and the weight smoothing term to take into account both the magnitude and spatial smoothing characteristics of the network weights, improving the generalization performance of the model. The calculation method is expressed as:

[0109]

[0110] In the formula, represents the square of the L2 norm of the probabilistic neural network weight matrix, μ W is the mean of the weight matrix, W i,j is the element in the i-th row and j-th column of the weight matrix, α rw is the first weight adjustment parameter; β rw is the second weight adjustment parameter. Preferably, α rw is set to 0.3, and β rw is set to 0.7;

[0111] In this embodiment, as a further technical solution of the present invention, backpropagation and parameter update of the probabilistic neural network are performed; due to the learning of complex fault features in the vibration signals of thermal power equipment, fine control of parameter update is required. Excessive or too small learning steps 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 at each update, the update amounts of weights and biases not only depend on the gradient but also make additional corrections based on the current parameters, better dynamically controlling the convergence process for different fault features of thermal power equipment;

[0112] The way of backpropagation and parameter update of the probabilistic neural network is expressed as:

[0113]

[0114] In the formula,

[0115] W newis the weight of the updated probabilistic neural network, representing the final weight parameters after this round of iteration, that is, 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, that is, the initial bias of the probabilistic neural network in the next round of iteration; η is the learning rate of the probabilistic neural network, representing the adjustment step size ratio for each parameter update; represents the gradient of the loss function of the probabilistic neural network with respect to the weight; represents the gradient of the loss function of the probabilistic neural network with respect to the bias; σ(·) is the correction adjustment function, representing the additional term for correcting the current parameters. Preferably, η is set to 0.01

[0116] By introducing an adaptive correction term for the current parameters through the correction adjustment function, it is ensured that the parameter update not only follows the gradient direction but also prevents over-updating, effectively alleviating the oscillation or convergence stagnation during network training and improving training stability;

[0117] The correction adjustment function combines the gradient historical information and the parameter value distribution characteristics of the current parameters, and through parameter standardization and gradient momentum dynamic correction, makes the parameter update smoother and more efficient, realizing dynamic correction control. The calculation method is expressed as:

[0118]

[0119] In the formula, Var(·) is the variance function; m fs (·) is the historical gradient momentum term of the parameter; μ W is the mean of the weight matrix; μ b is the mean of the bias vector; λ sg is the first correction adjustment hyperparameter; θ sg is the second correction adjustment hyperparameter; □ is a small constant to prevent the denominator from being zero. Preferably, λ sg is set to 0.3, θ sg is set to 0.5

[0120] The historical gradient momentum term of the parameter is calculated by the exponentially weighted moving average of the historical gradients, which is used to record the historical trend of parameter updates, ensuring the smoothness and stability of the update direction, and effectively avoiding training oscillation or local optimum problems caused by gradient fluctuations. The calculation method is expressed as:

[0121]

[0122] In the formula, m fs (μ W ) is the gradient momentum term of the weight parameter, representing the comprehensive trend of the current weight gradient and the historical gradient; m fs (μ b) is the gradient momentum term of the bias parameter, representing the comprehensive trend of the current bias gradient and the historical gradient; represents the historical gradient momentum value of the weight parameter in the previous iteration (default set to 0 in the first iteration); represents the historical gradient momentum value of the bias parameter in the previous iteration (default set to 0 in the first iteration); represents the gradient of the probability neural network loss function with respect to the weight; represents the gradient of the probability neural network loss function with respect to the bias; β m is the momentum smoothing coefficient, with a value range between (0, 1), used to suppress gradient mutations and alleviate parameter oscillations caused by complex fault features in the vibration signals of thermal power equipment; μ W,prev is the weight mean of the previous iteration; μ b,prev is the bias mean of the previous iteration; sign(·) is the sign function, outputting the direction of change of the parameter mean (+1 or -1); ⊙ represents element-wise multiplication. Preferably, β m is set to 0.9;

[0123] As a further technical solution of the present invention, set the training end condition; the training process of the probability neural network follows the gradient descent training method of the conventional neural network, that is, repeat the forward propagation and error backpropagation processes of the probability neural network until the classification on the training set during the training process is less than the preset threshold, then stop the iteration, indicating that the model training is completed; the training set refers to the set of vibration signal samples of thermal power equipment with known categories, used to guide the continuous optimization of parameters in the training process of the probability neural network until the classification error converges within the preset accuracy requirements;

[0124] By adopting the above technical solutions, based on the cross-entropy loss, L1 regularization (sparsifying model parameters) and dynamic penalty factors (enhancing misclassification penalties) are adopted, taking into account both classification accuracy and model lightweight, solving the limitation of the traditional L2 regularization in solely controlling the model complexity, improving the generalization ability of the model, and reducing the risk of overfitting;

[0125] In this embodiment; for the fault diagnosis of thermal power equipment; after the probability neural network training is completed, use its inference process to achieve the fault diagnosis of thermal power equipment, specifically including the following steps:

[0126] 1), Input the vibration signal data of the actual thermal power equipment into the trained probability neural network model. These data are first processed by the preprocessing and feature extraction function to ensure that the format and features of the input data are consistent with those during training;

[0127] 2) The input feature data propagates forward through the multi-layer structure of the network. In each layer, corresponding weights, biases, and activation functions are applied to gradually transform the input data into high-level feature representations. It should be noted that through the two-stage activation function defined in the probabilistic neural network, the model's ability to capture and classify fault features is enhanced;

[0128] 3) At the last layer of the probabilistic neural network, the Softmax function is used to convert 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;

[0129] 4) The prediction results are provided to the maintenance team or automated system for making corresponding maintenance decisions or automatic adjustments, which may include adjusting operating parameters, scheduling repairs, or replacing damaged components;

[0130] By adopting the above technical solution, during the gradient descent process of the probabilistic neural network, a correction adjustment function is adopted, which combines the gradient historical information and the parameter mean distribution to dynamically adjust the weights and biases, making the parameter update smoother, avoiding oscillations or local optima, solving the problem that fixed learning rates or simple decay strategies are difficult to adapt to gradient changes during training, and improving the convergence speed and stability.

[0131] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in the present invention. Any reference signs in the claims should not be regarded as limiting the claims involved.

[0132] In addition, it should be understood that although this specification is described according to embodiments, not every embodiment only contains an independent technical solution. This narrative way of the specification is only for clarity. Those skilled in the art should regard 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 fault diagnosis method for thermal power equipment based on artificial intelligence, characterized in that: The specific steps include: S1. Thermal power equipment training data collection: collected through vibration sensors installed at key parts of the equipment. These sensors can monitor and record vibration signals of the equipment in real time. The collected data is recorded in the form of time series. Each data point includes a timestamp and a 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 timestamps and vibration data of different sensors respectively. S2. Preliminary screening of thermal power equipment data; removal of shutdown data, elimination of outliers, and data balancing; S3. Labeling of thermal power equipment training data. The labeling method of training data is manual labeling. By combining equipment operation and maintenance records with expert experience, the vibration signal data is manually labeled. S4. Fault diagnosis of thermal power equipment: Use probabilistic neural network to diagnose faults of thermal power equipment.

2. The method for diagnosing faults of thermal power equipment based on artificial intelligence according to claim 1, wherein: In the step S2, shutdown data is removed: since the vibration signal of thermal power equipment is close to zero when it is shut down, these data have no practical value to the fault diagnosis model, so it is necessary to eliminate all data collected during the shutdown period; eliminate outliers: use statistical methods to detect and remove abnormal vibration signal data to avoid noise data interfering with model training; data balancing processing: in actual data collection, normal operating data is often far 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.

3. The method for diagnosing faults of thermal power equipment based on artificial intelligence according to claim 1, wherein: In the step S2, the training process of the probabilistic neural network is as follows: preprocessing the vibration signal data of the thermal power equipment; the method of preprocessing the vibration signal data of the thermal power equipment is expressed as: x i = f(s i ) The original signal is mapped into denoised features through preprocessing and feature extraction functions, which can adapt to the complex distribution of vibration signals of thermal power equipment and retain the complete fault mode at multiple scales; The preprocessing and feature extraction function is calculated by using a bandpass filter and a wavelet packet transform to avoid signal distortion and loss of fault features caused by a single method. The preprocessing method of the vibration signal data of the thermal power equipment is expressed as follows: f(s i ) = WPT(BPF(s i ).

4. The method for fault diagnosis of thermal power equipment based on artificial intelligence according to claim 3, characterized in that: Define the forward propagation activation function of the probabilistic neural network: In the fault diagnosis scenario, the vibration signals of thermal power equipment often present high dimensions and nonlinear distribution, and it is difficult to accurately classify them by simply relying on linear models; by using 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 mined: The forward propagation activation function of the probabilistic neural network is calculated as follows: H (1) f(x) = LeRe(Wx + b) H (2) (x) = tanh(W · ∥H (1) (x) - x py ∥ + b); In the above formula, the one-stage activation function and the two-stage activation function are used together as the activation functions of neurons in the probabilistic neural network, which can make the intermediate feature vectors processed by the probabilistic neural network maintain good separability in high-dimensional space; The offset feature is used to adaptively adjust the distribution center position of the neuron output feature. By batch counting the feature center, the network can be adaptively drifted and positioned in the high-dimensional feature space, and the sensitivity of the network to distinguish different fault modes can be enhanced. The calculation method is expressed as:

5. The method for diagnosing faults of thermal power equipment based on artificial intelligence according to claim 3, characterized in that: Calculate the dynamic penalty factor during the forward propagation of the probabilistic neural network: In the classification task of fault diagnosis, there are differences in the misclassification costs for different fault categories in diagnosis. Moreover, in the early stage of training, the model is still unstable and requires relatively small penalties, while in the later stage, it is necessary to strengthen the correction of misclassifications. By adopting a dynamic penalty factor, the penalty increases smoothly and steadily with the iteration process, avoiding the convergence risk in the initial stage of probabilistic neural network training and fully strengthening the correction of fault misclassifications in the later stage of probabilistic neural network training. The calculation method is expressed as: By combining the number of iterations and the misclassification ratio to adaptively adjust the penalty intensity, it achieves a milder penalty in the initial stage of model training and gradually strengthens the error correction ability in the later stage, effectively improving the classification stability and accuracy. The dynamic penalty factor refers to the parameter used to dynamically adjust the differential weighted penalty for misclassified samples during the training process of the probabilistic neural network.

6. The method for diagnosing faults of thermal power equipment based on artificial intelligence according to claim 3, wherein: Calculate the classification result of the forward propagation of the probabilistic neural network: Calculate the classification result of the forward propagation of the probabilistic neural network. Through linear combination and probability normalization of the output of the two-stage activation function, the classification prediction result of the sample is obtained, which is expressed as: y i = sof(W last ·H (2) (x lasti ) + b last )。 7. The method for diagnosing faults of thermal power equipment based on artificial intelligence according to claim 3, wherein: Calculate the loss function of the current iteration of the probabilistic neural network: Since the 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 a dynamic penalty factor, during the training process of the probabilistic neural network, focus on the classification accuracy of fault samples and sparsify unnecessary redundant weights, thus taking into account both accuracy and model lightweighting. The calculation method of the loss function of the probabilistic neural network is expressed as: By jointly using cross-entropy, L1 regularization, and a dynamic penalty factor, the probabilistic neural network realizes the sparsification of model weights and the adaptive control of misclassified samples while improving the classification accuracy, effectively avoiding overfitting and reducing the network complexity. The regularization term of the weights of the probabilistic neural network is calculated by mixing the L2 norm and the weight smoothing term to take into account the magnitude and spatial smoothing characteristics of the network weights and improve the generalization performance of the model. The calculation method is expressed as: R(W) = α rw · ∥W∥ 2 2 + β rw Σ i,j (W i,j - μ W ) 2 。 8. The method for fault diagnosis of thermal power equipment based on artificial intelligence according to claim 3, characterized in that: Perform the backpropagation and parameter update of the probabilistic neural network; Since learning the complex fault characteristics in the vibration signals of thermal power equipment requires fine control of parameter updates, too large or too small learning step sizes may both lead to convergence difficulties or getting stuck in local optima. By using a correction adjustment function to adjust the parameter update amplitude of the probabilistic neural network at each update, the update amounts of the weights and biases not only depend on the gradient but also make additional corrections based on the current parameters, better dynamically controlling the convergence process for different fault characteristics of thermal power equipment. The method of backpropagation and parameter update of the probabilistic neural network is expressed as: By additionally introducing an adaptive correction term for the current parameters through the correction adjustment function, it ensures that the parameter update follows the gradient direction and prevents excessive updates, effectively alleviating the oscillation or convergence stagnation during the network training process and improving the training stability. The correction and adjustment function combines the gradient historical information of the current parameters with the parameter value distribution characteristics, and through parameter standardization and gradient momentum dynamic correction, makes the parameter update smoother and more efficient, realizes dynamic correction control, and the calculation method is expressed as: The historical gradient momentum term of the parameter is calculated by the exponential weighted moving average of the historical gradients, which is used to record the historical trend of parameter update, ensure the smoothness and stability of the update direction, and effectively avoid the training oscillation or local optimum problem caused by gradient fluctuation. The calculation method is expressed as:

9. The method for diagnosing faults of thermal power equipment based on artificial intelligence according to claim 3, wherein: Set the training end condition; the training set refers to the set of vibration signal samples of thermal power equipment with known categories, which is used to guide the continuous optimization of parameters in the training process of the probabilistic neural network until the classification error converges within the preset accuracy requirements.

10. A fault diagnosis method for thermal power equipment based on artificial intelligence according to claim 3, characterized in that: Fault diagnosis of thermal power equipment; after the training of the probabilistic neural network is completed, its inference process is used to realize the fault diagnosis of thermal power equipment, which specifically includes the following steps: 1). Input the vibration signal data of the actual thermal power equipment into the trained probabilistic neural network model. These data are first processed by the preprocessing and feature extraction function to ensure that the format and features of the input data are consistent with those during training; 2). The input feature data propagates forward through the multi-layer structure of the network. Each layer applies the corresponding weights, biases, and activation functions to gradually convert the input data into a high-level feature representation. It should be noted that through the two-stage activation function defined in the probabilistic neural network, the model's ability to capture and classify fault features is enhanced; 3). At the last layer of the probabilistic neural network, the Softmax function is used to convert 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 result to the maintenance team or the automation system for making corresponding maintenance decisions or automatic adjustments, which may include adjusting operation parameters, arranging repairs, or replacing damaged components.

Images