Large thermal power generating unit blower surge fault early warning method based on multi-head attention generative adversarial network

Abstract

The invention provides a large-scale thermal power generating unit blower surge fault early warning method based on a multi-head attention generative adversarial network, and the method comprises the steps: data collection, data preprocessing, model training and reasoning, fault evaluation, and fault early warning, and employs a multi-head attention mechanism to capture the global features of input data; a multi-head attention layer is used to replace a traditional convolutional layer, a feature fusion module is introduced, outputs of different levels and different attention heads are fused, a multi-head attention mechanism further improves feature extraction capability, diversity of generated samples and alleviates the problem of mode collapse, and a method of combining confrontation loss and feature matching loss is adopted to improve the feature extraction capability and the diversity of generated samples. The quality of the generated sample is improved, the parameters of the generator and the discriminator are updated by adopting an optimization method, the generated sample is evaluated through multiple evaluation indexes, and finally the accuracy and timeliness of fault early warning are improved. The early warning system supports integration to a third-party monitoring system, and has strong versatility and expansibility.

Description

Technical Field

[0001] The present invention relates to the technical field of fault warning for blowers of large thermal power units, and in particular to a surge fault warning method for blowers of large thermal power units based on a multi-head attention generative adversarial network. Background Art

[0002] The blower in a large thermal power unit is an important device to ensure the normal operation of the boiler. Its main function is to provide sufficient air for boiler combustion to ensure the smooth progress of the combustion process. Its operating state directly affects the safety and stability of the thermal power unit. Surge failure is a common and serious type of failure, which may cause equipment damage, production stagnation and safety hazards. Surge is a phenomenon that occurs when the blower is under unstable working conditions. Its main feature is that the flow rate and pressure of the blower fluctuate greatly. During normal operation, there is a relatively stable relationship between the pressure and flow rate of the blower, and the operating point is located in the stable area of ​​the performance curve. When the blower is close to the surge state, the flow rate will decrease and the pressure will undergo pulsating and drastic changes. In this alternating stress environment for a long time, fatigue cracks are likely to occur in the blades, and the bearings of the blower will also be seriously affected. When the blower surges, the amount and pressure of air sent to the boiler are unstable. Some coal powder may not be completely burned due to insufficient air, thereby reducing the thermal efficiency of the boiler. Severe surge may also cause unstable combustion of the boiler or even flameout. The vibration caused by the surge of the blower will be transmitted to other equipment through connecting parts such as pipes and air ducts.

[0003] Benefiting from the rapid development of computer technology and artificial intelligence, data-driven methods are gradually applied to surge fault diagnosis of blowers. Data-driven methods require a large number of data samples under fault conditions for training. However, in actual blower fault diagnosis scenarios, it is often difficult to obtain enough fault samples. In actual industrial scenarios, most induced draft fans operate in a healthy state, and the operating frequency under fault conditions is very low. Therefore, there are fewer fault sample data, and there is an obvious data imbalance problem between normal samples and fault samples. It is difficult to train a good fault diagnosis model, which will seriously affect the accuracy of fault diagnosis. The operating data signal of the blower contains rich fault information, such as frequency components, amplitude changes, etc. The fault model needs to be able to automatically mine the deep fault features in these signals, unlike the traditional method that requires manually designed feature extractors. The fault data of the blower usually presents complex nonlinear relationships. The model needs to be able to handle nonlinear data well. The model needs to capture the nonlinear relationship between these parameters through its powerful nonlinear modeling capabilities, so as to more accurately identify faults. Therefore, when faced with new and unseen wind turbine operation data, the model needs to have strong generalization capabilities and be able to better perform fault diagnosis through data enhancement and better feature extraction.

[0004] Compared with the prior art, the differences are as follows:

[0005] Technical comparison with patent CN114962317B "Anti-surge control method for toothed single-stage or multi-stage centrifugal compressor"

[0006] CN114962317B uses differential pressure and current models to perform anti-surge control, focusing on real-time monitoring and adjusting the compressor operating status; while this patent uses a multi-head attention generative adversarial network and deep learning technology for fault warning, focusing on the learning and generation of data features. There are essential differences between the two in terms of control mechanism and methodology.

[0007] CN114962317B emphasizes solving the control failure problem caused by sensor failure, with the goal of improving the safety and reliability of the compressor; this patent focuses on improving the accuracy and timeliness of fault warnings, and captures fault characteristics through advanced model training and reasoning processes. There are obvious differences between the two in terms of the focus of technical implementation.

[0008] CN114962317B is mainly applicable to anti-surge control of centrifugal compressors, focusing on real-time control of mechanical equipment; while this patent focuses on fault warning of blowers of large thermal power units, involving a wider range of fault types and prediction capabilities. There are essential differences in the application fields and target equipment of the two.

[0009] Technical comparison with patent CN115616398A "Intelligent monitoring method for the operating status of blast furnace TRT generator set"

[0010] CN115616398A uses vibration sensors and signal reconstruction technology to monitor operating conditions, emphasizing the identification of fault features and limit warnings; this patent uses a multi-head attention mechanism to extract data features and generate samples, focusing on improving the accuracy of fault warnings and the diversity of generated samples. The two are essentially different in data processing and monitoring methods.

[0011] CN115616398A performs simple threshold warning by setting indicator limits, focusing on abnormal monitoring status; while this patent combines adversarial loss and feature matching loss, uses deep learning methods to generate high-quality warning information, aiming to achieve more complex and accurate fault prediction. There are significant differences between the two in the complexity and accuracy of the warning mechanism.

[0012] The intelligent monitoring method of CN115616398A mainly relies on the integration of existing sensor data and focuses on real-time status monitoring; this patent supports integration into third-party monitoring systems, has stronger versatility and scalability, and emphasizes the flexibility and applicability of the model. There is an essential difference between the two in terms of system integration capabilities and application flexibility.

[0013] In the prior art, the monitoring and early warning methods for blower surge faults are not perfect enough. The traditional fault early warning methods have problems such as low accuracy and delayed early warning time, and lack of effective early warning mechanism. Therefore, a more accurate and timely blower surge fault early warning method and system are needed. Summary of the invention

[0014] In order to solve the above technical problems, the present invention proposes a surge fault warning method for the fan of a large thermal power unit based on a multi-head attention generative adversarial network. By real-time monitoring of the operating status data of the fan, the advantages of the multi-head attention mechanism and the generative adversarial network are utilized to accurately identify potential faults and give early warnings, so as to improve the accuracy and timeliness of fault warnings.

[0015] To achieve the above object, the technical solution adopted by the present invention is:

[0016] A surge fault warning method for the fan of a large thermal power unit based on a multi-head attention generative adversarial network monitors the operating status data of the fan in real time, takes advantage of the multi-head attention mechanism and the generative adversarial network, accurately identifies potential faults and gives early warnings to improve the accuracy and timeliness of fault warnings, including data collection, data preprocessing, model training and reasoning, fault assessment, and fault warning: including the following steps:

[0017] S101: Collect historical operation data of the fan of a large thermal power plant, including but not limited to the fan air volume, fan opening, unit load, vibration signal, temperature signal and current signal;

[0018] S102: using a data preprocessing module to preprocess the collected historical operation data of the induced draft fan, including data cleaning and feature extraction, firstly screening the steady-state data, then performing correlation analysis, dimensionality reduction, denoising, normalization, data cleaning, labeling and feature engineering, to provide a data set for model training;

[0019] S103: Construct a fault diagnosis model based on a multi-head attention generative adversarial network. The generator uses a multi-head attention mechanism to capture the global features of the input data. The multi-head attention layer replaces the traditional convolutional layer, allowing the generator to learn different feature representations on different attention heads. The discriminator uses multi-head attention to improve the ability to distinguish between generated samples and real samples. Under the multi-head attention mechanism, each attention head calculates the attention score of the input and generates the final output by weighted average.

[0020] In the generator and discriminator, a feature fusion module is introduced to fuse the outputs of different levels and different attention heads to enhance the expressiveness of the model. The quality of generated samples is improved by combining adversarial loss with feature matching loss. Feature matching loss is calculated by comparing the feature representations of generated samples and real samples in the discriminator. Stochastic gradient descent SGD or adaptive moment estimation optimization algorithm Adam is used to update the parameters of the generator and discriminator. The generated samples are evaluated through evaluation indicators to ensure that the improved model is superior to the baseline model in terms of generation quality and diversity. Through continuous adversarial training, the parameters of the generator and discriminator are optimized and adjusted, so that the pseudo data generated by the generator is closer and closer to the real data, and the discriminator's discriminative ability is stronger and stronger, and finally the target fault diagnosis model is obtained;

[0021] S104: inputting the real-time collected operation data of the blower into the trained fault diagnosis model, diagnosing and predicting the fault of the blower through the synergy of the generator and the discriminator, judging whether the blower has a fault according to the output result of the discriminator, and determining the type and severity of the fault;

[0022] S105: When the early warning model determines that the blower is surging, the early warning signal will be promptly notified to the power station staff so that corresponding preventive measures can be taken. At the same time, the early warning information will be uploaded to the monitoring system of the power station to integrate the early warning system with the monitoring system of the power station so that the operation and maintenance personnel can conduct remote monitoring and decision-making.

[0023] As a further improvement of the present invention, the data preprocessing module in step S102 performs denoising, smoothing and standardization on the collected data, extracts features, uses Python or R or MATLAB, and utilizes SciPy and Pandas databases for data processing and feature extraction; during the real-time data acquisition, LabVIEW, MATLAB or custom-developed acquisition software is used to transmit data through sensors and data acquisition cards. The sensors are used to collect the operating status data of the blower, including vibration sensors, temperature sensors and pressure sensors. The data acquisition card converts the sensor signals into digital signals and performs data acquisition. The custom-developed acquisition software collects the operating data of the blower of a large thermal power plant through a DCS / SIS system in the form of a data communication interface.

[0024] As a further improvement of the present invention, the step S103 uses the adaptive moment estimation optimization algorithm Adam to update the parameters of the generator and the discriminator, as follows:

[0025] The calculation formula of the Adam optimization algorithm is:

[0026]

[0027] Among them, mt and v t denote the first-order moment estimate and the second-order moment estimate, β and β2 denote the exponential decay rate of the moment estimate, ▽ θ L(θ t-1 ) indicates that the loss function is in the parameter θ t-1 The gradient at and They represent the modified first-order moment estimate and second-order moment estimate respectively, α represents the learning rate, represents a very small constant, and ∈ is used to prevent the denominator from being zero.

[0028] As a further improvement of the present invention, the multi-head attention mechanism of the generator in step S103 is as follows: attention weight calculation is performed, and for the input sequence Assume that its dimension is n×d, where n is the number of samples, i.e. the length of the sequence, and d is the feature dimension;

[0029] First, the input data is mapped into query, key, and value through linear transformation. The formula is as follows:

[0030] Q=XW Q ,K=XW K ,V=XW V

[0031] in is the learnable weight matrix, d k is the dimension of the attention head;

[0032] Then use the soft max function to calculate the attention weight, and calculate the attention score by weighted summing of the values ​​​​through the attention weight:

[0033]

[0034] Here k is the dimension of the key, used for scaling;

[0035] Then extend the above process to multiple heads h:

[0036] MultiHead(Q,K,V)=Concat(head1,...,head h )W O

[0037] The calculation for each head is:

[0038] head i =Attention(QW i Q ,KW i K ,VW i V )

[0039] Here W i Q ,W i K ,W i V is the linear transformation weight of each head.

[0040] As a further improvement of the present invention, in the generator loss function of step S103, regarding the adversarial loss: the goal of the generator is to maximize the loss of the discriminator, and the loss function is expressed as:

[0041]

[0042] Where D is the discriminator, G is the generator, and z is the input noise;

[0043] By comparing the features of the generated samples and the real samples in the discriminator, the feature matching loss is defined as:

[0044]

[0045] Where φ is the feature extraction layer of the discriminator, and x is the real sample;

[0046] The final generator loss function is:

[0047]

[0048] Here λ is the weight parameter;

[0049] The loss function of the discriminator is defined as:

[0050]

[0051] The training goal is to minimize the loss of the generator and the discriminator:

[0052]

[0053] As a further improvement of the present invention, in step S103, a feature fusion module is introduced into the generator and the discriminator to fuse the outputs of different levels and different attention heads to enhance the expression ability of the model, and feature fusion is performed by feature concatenation or linear transformation fusion;

[0054] The feature splicing combines features of different levels or different attention heads by splicing;

[0055] Assume there are multiple feature outputs F1, F2, ..., F n , whose dimensions are all d, then the concatenated features are expressed as:

[0056] Fconcat =Concat(F1,F2,...,F n )

[0057] Its dimension will become n×d if the number of features is n;

[0058] The linear transformation fusion maps the features to the same dimension through linear transformation, and then adds or concatenates them;

[0059] Assume that each feature is passed through the weight matrix W i Perform a linear transformation:

[0060] F i '=F i W i

[0061] Then you can choose concatenation or weighted summation:

[0062] F fused =Concat(F 1′ ,F 2′ ,...,F n′ )

[0063] or

[0064]

[0065] As a further improvement of the present invention, the step S103 adopts the following steps of optimizing by stochastic gradient descent:

[0066] Randomly initialize the generator parameters θ G and the discriminator parameters θ D

[0067] Calculate the loss of the generator and the discriminator, and calculate the gradient of the loss of the generator and the discriminator with respect to their parameters:

[0068]

[0069] S4: Update the parameters of the generator and discriminator:

[0070]

[0071] where η is the learning rate.

[0072] As a further improvement of the present invention, step S103 evaluates the quality of data generated by the generator through evaluation indicators, evaluates the generated samples, and ensures that the improved model is superior to the baseline model in terms of generation quality and diversity. The specific model is as follows:

[0073] Specifically, accuracy, recall and F1 score are used to evaluate model performance.

[0074]

[0075] Where TP, FP, and FN are the number of true positives, false positives, and false negatives, respectively.

[0076] Beneficial effects: The present invention has the following advantages:

[0077] The present invention proposes a method and system for surging fault warning of a large-scale thermal power unit fan based on a multi-head attention generative adversarial network, which specifically includes the steps of data collection and preprocessing, construction of an improved generative adversarial network model, model training, fault diagnosis and result output. The multi-head attention mechanism can simultaneously focus on information at different positions, improve the model's feature extraction capability for fan operation data, and thus improve the accuracy of fault warning. The generative adversarial network can continuously improve the performance of the model through adversarial training, so that the data generated by the generator is closer to the real data. The generative adversarial network can enhance the feature extraction capability of the fan operation data through the multi-head attention mechanism, and enhance the feature representation capability: improve the diversity of generated samples, alleviate the problem of mode collapse, and thus improve the accuracy and timeliness of fault warning. Combination of data preprocessing and real-time monitoring: By combining real-time monitoring and data preprocessing, the hidden faults of the fan can be discovered in time, and the timeliness of fault warning can be improved. The whole process automatically analyzes and diagnoses the operation data of the fan without manual intervention, thereby improving the efficiency of fault diagnosis. The present invention can be applied to different types of large-scale thermal power plant blowers, has strong versatility and expansibility, and provides powerful technical support for surge warning and state monitoring of blowers in thermal power plants. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] Figure 1 It is a schematic flow chart of the induced draft fan surge fault warning method in the present invention;

[0079] Figure 2 It is a schematic diagram of the training process of the generator and the discriminator in the present invention. DETAILED DESCRIPTION

[0080] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments:

[0081] The present invention provides a method and system for early warning of surge fault of a large-scale thermal power unit blower based on a multi-head attention generative adversarial network, such as Figure 1 As shown, the following steps are included:

[0082] By establishing a fault diagnosis process, including data collection, data preprocessing, model training and reasoning, fault assessment, and fault warning, real-time monitoring and fault prediction of fan surge can be achieved.

[0083] S101: Collect historical operation data of the blower of a large thermal power plant, including but not limited to the blower air volume, blower opening, unit load, vibration signal, temperature signal, current signal, etc.

[0084] S102: Preprocess the collected historical operation data of the induced draft fan, including data cleaning and feature extraction. First, the steady-state data is screened, and then correlation analysis, dimensionality reduction, denoising, normalization, data cleaning, labeling and feature engineering are performed to provide a high-quality data set for model training.

[0085] S103: Construct a fault diagnosis model based on a multi-head attention generative adversarial network. The generator uses a multi-head attention mechanism to capture the global features of the input data. The multi-head attention layer replaces the traditional convolutional layer, allowing the generator to learn different feature representations on different attention heads. The discriminator uses multi-head attention to improve the ability to distinguish between generated samples and real samples. Under the multi-head attention mechanism, each attention head calculates the attention score of the input and generates the final output by weighted average. In the generator and discriminator, a feature fusion module is introduced to fuse the outputs of different levels and different attention heads to enhance the expression ability of the model. The adversarial loss and feature matching loss are combined to improve the quality of generated samples. The feature matching loss is calculated by comparing the feature representations of generated samples and real samples in the discriminator. Stochastic gradient descent SGD or Adam optimizer is used to update the parameters of the generator and discriminator, and the generated samples are evaluated by multiple evaluation indicators to ensure that the improved model is superior to the baseline model in terms of generation quality and diversity. Through continuous adversarial training, the parameters of the generator and discriminator are optimized and adjusted, so that the pseudo data generated by the generator is closer and closer to the real data, and the discriminator's discriminative ability is stronger and stronger, and finally the target fault diagnosis model is obtained.

[0086] S104: Input the real-time collected fan operation data into the trained fault diagnosis model, and diagnose and predict the fan fault through the synergy of the generator and the discriminator. According to the output result of the discriminator, it can be judged whether the fan has a fault, and the type and severity of the fault can be determined.

[0087] S105: When the early warning model determines that the blower is surging, the early warning signal will be promptly notified to the power station staff so that corresponding preventive measures can be taken. At the same time, the early warning information can also be uploaded to the monitoring system of the power station, so that the early warning system can be integrated with the monitoring system of the power station for remote monitoring and decision-making by operation and maintenance personnel.

[0088] The data preprocessing module performs denoising, smoothing and standardization on the collected data and extracts features. Python, R or MATLAB can be used to perform data processing and feature extraction using libraries such as SciPy and Pandas.

[0089] During the real-time data collection, LabVIEW, MATLAB or custom-developed collection software is used to transmit data through sensors and data acquisition cards. The sensors are used to collect the operating status data of the blower, such as vibration sensors, temperature sensors, pressure sensors, etc. The data acquisition card converts the sensor signal into a digital signal and performs data collection. The custom-developed collection software can collect the operating data of the blower of a large thermal power plant through the DCS / SIS system in the form of a data communication interface.

[0090] The feature engineering performs feature selection and extraction to construct a feature vector including time series features, statistical features and frequency domain features.

[0091] The generative adversarial network consists of a generator and a discriminator. The generator adopts a deep neural network structure, including an input layer, a hidden layer, and an output layer. The input layer receives the preprocessed fan operation data, the hidden layer extracts features from the input data through a multi-head attention mechanism, and the output layer generates pseudo data similar to the real data. When training the model, the preprocessed fan operation data is divided into a training set and a test set, and the training set is used to train the improved generative adversarial network model. During the training process, the model gradually converges by alternately optimizing the parameters of the generator and the discriminator.

[0092] The multi-head attention mechanism is as follows: attention weight calculation is performed for the input sequence Assume that its dimension is n×d, where n is the number of samples, i.e., the sequence length, and d is the feature dimension.

[0093] First, the input data is mapped into query, key, and value through linear transformation. The formula is as follows:

[0094] Q=XW Q ,K=XW K ,V=XW V

[0095] in is the learnable weight matrix, d k is the dimension of the attention head.

[0096] Then use the soft max function to calculate the attention weight, and calculate the attention score by weighted summing of the values ​​​​through the attention weight:

[0097]

[0098] Here k is the dimension of the key, used for scaling.

[0099] Then extend the above process to multiple heads h:

[0100] MultiHead(Q,K,V)=Concat(head1,...,head h )W O

[0101] The calculation for each head is:

[0102] head i =Attention(QW i Q ,KW i K ,VW i V )

[0103] Here W i Q ,W i K ,W i V is the linear transformation weight of each head.

[0104] The generator receives a random noise vector z. After the initial linear transformation and multi-head attention mechanism, the output vector will pass through a series of generation layers to gradually convert the features into the format of the target data. Finally, the generator outputs a fake data sample G(z). The distribution of this sample should be as close as possible to the distribution of the real data. In the generator, random noise z is input to generate fake data G(z):

[0105] G(z)=W G ·z+b G

[0106] The discriminator receives two types of inputs: one is a sample x from the real data distribution, and the other is a fake sample G(z) generated by the generator. Similar to the generator, the discriminator can also contain a multi-head attention mechanism layer. Its calculation process is similar to the multi-head attention layer in the generator, which is used to capture the correlation between different parts of the input data and enhance the discriminator's understanding of the data structure. For the input data (which can be a real sample or a generated sample), it is divided into multiple heads, and the query, key, value, attention score and output of each head are calculated, and then the outputs of all heads are concatenated to obtain the feature representation after multi-head attention. After the previous attention layer, the discriminant result is output through a fully connected layer and an activation function (such as a sigmoid function). For the input x, the discriminator outputs a probability value D(x), which indicates the probability that x is real data. If D(x) is close to 1, the discriminator believes that x is real data; if D(x) is close to 0, it believes that x is fake data generated by the generator. The goal of the discriminator is to distinguish between real data x and generated data G(z):

[0107] D(x)=σ(W D ·x+b D )

[0108] In the generator loss function, about the adversarial loss: the goal of the generator is to maximize the loss of the discriminator. The loss function can be expressed as:

[0109]

[0110] Where D is the discriminator, G is the generator, and z is the input noise.

[0111] By comparing the features of the generated samples and the real samples in the discriminator, the feature matching loss can be defined as:

[0112]

[0113] Where φ is the feature extraction layer of the discriminator and x is the real sample.

[0114] The final generator loss function is:

[0115]

[0116] Here λ is the weight parameter.

[0117] The loss function of the discriminator is defined as:

[0118]

[0119] The training goal is to minimize the loss of the generator and the discriminator:

[0120]

[0121] In the generator and discriminator, a feature fusion module is introduced to fuse the outputs of different levels and different attention heads to enhance the expressiveness of the model. Feature fusion is performed by feature concatenation or linear transformation fusion.

[0122] The feature splicing combines features of different levels or different attention heads by splicing.

[0123] Assume there are multiple feature outputs F1, F2, ..., F n , whose dimensions are all d, then the concatenated features can be expressed as:

[0124] F concat =Concat(F1,F2,...,F n )

[0125] Its dimension will become n×d (if the number of features is n).

[0126] The linear transformation fusion maps features to the same dimension through linear transformation and then performs addition or concatenation.

[0127] Assume that each feature is passed through the weight matrix W i Perform a linear transformation:

[0128] F i '=F i W i

[0129] Then you can choose concatenation or weighted summation:

[0130] F fused =Concat(F 1′ ,F 2′ ,...,F n′ )

[0131] or

[0132]

[0133] The optimization algorithm is used to optimize the parameters of the generator and discriminator models to improve the performance and convergence speed of the model. Stochastic gradient descent (SGD) or adaptive moment estimation optimization algorithm (Adam) is further used to update the generator and discriminator parameters. Preferably, the Adam optimization algorithm is used. The Adam optimization algorithm is a gradient-based optimization algorithm that can adaptively adjust the learning rate to improve the convergence speed and performance of the model. The calculation formula of the Adam optimization algorithm is:

[0134]

[0135] Among them, m t and v t denote the first-order moment estimate and the second-order moment estimate, β and β2 denote the exponential decay rate of the moment estimate, ▽ θ L(θ t-1 ) indicates that the loss function is in the parameter θ t-1 The gradient at and They represent the modified first-order moment estimate and second-order moment estimate respectively, α represents the learning rate, represents a very small constant, and ∈ is used to prevent the denominator from being zero.

[0136] In each training iteration, the parameters of the generator and discriminator are updated by calculating the loss, gradient, and using the SGD or Adam optimization algorithm. SGD is simple and direct, while Adam accelerates convergence by using estimates of the first and second order moments and performs better when dealing with sparse gradients. Choosing the right optimizer and hyperparameters is crucial to the training effect.

[0137] Preferably, the Adam optimizer is used to update the generator and discriminator parameters, and optionally, stochastic gradient descent (SGD) is used for optimization.

[0138] The steps for optimization using stochastic gradient descent (SGD) are as follows

[0139] S1: Randomly initialize the generator parameters θ G and the discriminator parameters θ D

[0140] S2: Calculate the loss of the generator and the discriminator,

[0141] S3: Calculate the gradient of the loss of the generator and discriminator with respect to their parameters:

[0142]

[0143] S4: Update the parameters of the generator and discriminator:

[0144]

[0145] where η is the learning rate.

[0146] The detailed steps of using Adam optimizer to update the parameters of generator and discriminator are as follows

[0147] S1: Initialize the generator parameters θ G and the discriminator parameters θ D , as well as momentum and variance terms

[0148] m G =0,v G =0,m D =0,vD =0

[0149] Set hyperparameters, including learning rate α, first-order moment estimate decay rate β1, second-order moment estimate decay rate β2, small constant ξ

[0150] S2: Calculate the loss of the generator and the discriminator,

[0151] S3: Calculate the gradients of the generator and discriminator:

[0152]

[0153] S4: Update first and second order moments

[0154] For the generator, the calculation formula is as follows

[0155] m G ←β1m G +(1-β1)g G

[0156]

[0157] For the discriminator, the calculation formula is as follows

[0158] m D ←β1m D +(1-β1)g D

[0159]

[0160] S5: Corrected first and second order moments, the formula is as follows

[0161]

[0162] S6: Update the parameters of the generator and discriminator. The formula is as follows:

[0163]

[0164] Use appropriate evaluation metrics to evaluate the quality of data generated by the generator, and evaluate the generated samples using a variety of evaluation metrics to ensure that the improved model is superior to the baseline model in terms of generation quality and diversity.

[0165] Specifically, accuracy, recall and F1 score are used to evaluate model performance.

[0166]

[0167] Where TP, FP, and FN are the number of true positives, false positives, and false negatives, respectively.

[0168] Use a step-by-step training strategy to first train the generator and then the discriminator, or alternate training in each epoch to ensure a dynamic balance between the two. Tune hyperparameters such as learning rate, batch size, and number of attention heads to find the best training configuration.

[0169] The training process is as follows Figure 2 As shown below

[0170] S1: Fixed generator, trained discriminator

[0171] Sample a batch of real samples x from the real data distribution, and sample a batch of noise vectors z from the noise distribution, and get the generated samples through the generator. Calculate the loss of the discriminator for the real samples and the loss for the generated samples.

[0172] The total discriminator loss is The discriminator parameters are updated through the back-propagation algorithm to minimize this loss. This step enables the discriminator to better distinguish between real data and generated data.

[0173] S2: Fixed discriminator, training generator

[0174] Sample a batch of noise vectors z from the noise distribution again, and obtain the generated samples G(z) through the generator. Calculate the loss L of the generator G =-E z～p(z) [logD(G(z))], updates the parameters of the generator through the back-propagation algorithm to minimize this loss. This step enables the generator to generate more realistic data to deceive the discriminator.

[0175] S3: Repeat the training process until convergence.

[0176] Alternately train the discriminator and generator, repeat the forward propagation, loss calculation and parameter update steps of the generator and discriminator, and after multiple iterations, the performance of the generator and discriminator will gradually improve. The data generated by the generator can "cheat" the discriminator with higher quality, and the data generated by the generator will be closer and closer to the distribution of real data. The discriminator will also become more accurate in distinguishing the authenticity of the data. Until it reaches a relatively stable state. In this state, the generator can generate high-quality fake data with a similar distribution to the real data, and the discriminator will find it difficult to accurately distinguish between real data and fake data. During the training process, the performance of the model can be monitored based on the verification data set.

[0177] The multi-head attention generative adversarial network model is trained using preprocessed real data. By continuously adjusting the model parameters, the data generated by the generator is made as close to the real data as possible, while the discriminator can accurately distinguish between real data and generated data.

[0178] Furthermore, the detailed design of the generator structure design and the discriminator structure of this embodiment is as follows:

[0179] (1) Generator structure design

[0180] The main goal of the generator is to generate fan fault data samples that can deceive the discriminator, thereby helping the model to better learn fault features. The following is the structural design of the generator:

[0181] Input layer: Receives random noise vectors (such as random noise or sensor data from actual equipment) and possible partial real fault data features as input. The random noise vector can be generated by normal distribution, etc. The dimension of the input layer is consistent with the feature dimension of the fault data. For example, it can be set to a 20-dimensional vector.

[0182] Multi-head attention mechanism: Use multiple attention heads to capture the dependencies between different features. The dimension of each attention head is set to 64 and the number of heads is set to 8.

[0183] Intermediate layer: It consists of multiple fully connected layers, which are used to further process and transform the output of the multi-head attention layer. The fully connected layer can gradually adjust the feature dimension to a size suitable for generating fault data samples. For example, three fully connected layers can be set to connect the attention output to multiple fully connected layers (3 layers), and each layer outputs 128 features. The intermediate layer can also introduce activation functions to increase the nonlinear expression ability of the model. The ReLU activation function is used in each layer, and the tanh activation function is used in the output layer to generate normalized data. Use the deconvolution (transposed convolution) layer to map the features back to the original data space, use 3 convolution layers, the filter size is 3×3, the step size is 1, and the output dimension is 1 (the final generated data).

[0184] Output layer: Generates synthetic fault data samples with similar distribution to the real fault data. The dimension of the output layer is consistent with the feature dimension of the real fault data. For example, if the real fault data has 20 features and the dimension of each feature is 16, then the dimension of the output layer is 20×16.

[0185] (2) Discriminator structure design

[0186] The discriminator is responsible for determining whether the input data is from the real blower status or the synthetic fault data sample generated by the generator. The following is the structural design of the discriminator:

[0187] Input layer: receives real fault data samples or synthetic fault data samples generated by the generator as input. The dimension of the input layer is consistent with the feature dimension of the fault data.

[0188] Convolutional layer: Multiple convolutional layers can be used to extract local features of the input data. Convolutional layers can effectively capture the spatial structure information in the fault data. For example, three convolutional layers can be set, and the number of filters in each convolutional layer gradually increases to extract richer features. The first convolutional layer can have 32 filters, the second convolutional layer can have 64 filters, and the third convolutional layer can have 128 filters, with a filter size of 3×3 and a stride of 2 to increase the feature extraction capability. The convolutional layer uses the Leaky ReLU activation function

[0189] Pooling layer: A maximum pooling layer is added after every two convolutional layers to reduce the feature map size.

[0190] Multi-head attention mechanism: Similar to the multi-head attention layer in the generator, the multi-head attention layer in the discriminator is also used to capture the dependencies between different positions. Here, multi-head attention can help the discriminator better understand the global characteristics of the fault data, so as to more accurately judge the authenticity of the data. Two attention heads are used, and the dimension is set to 128.

[0191] The purpose is to further extract important features and reduce the impact of noise.

[0192] Middle layer: It is composed of multiple fully connected layers, which are used to further process and integrate the output of the multi-head attention layer. The fully connected layer can gradually compress the feature dimension to a smaller dimension for the final classification judgment. For example, two fully connected layers can be set. The first fully connected layer compresses the feature dimension from 128 to 64, and the second fully connected layer compresses it to 32. After two fully connected layers, a binary classification value (real or generated) is output.

[0193] Output layer: Outputs a single scalar value, indicating the probability that the input data is real fault data. The output layer uses the sigmoid activation function, which limits the output value to between 0 and 1 for binary classification. If the output value is close to 1, it means that the discriminator believes that the input data is real fault data; if the output value is close to 0, it means that the discriminator believes that the input data is synthetic fault data generated by the generator.

[0194] (3) Training process

[0195] The generator and discriminator are trained alternately to optimize the loss functions of both. The Adam optimizer is used and the learning rate is set to 0.0002.

[0196] (4) Effect evaluation

[0197] The cross-validation method is used to evaluate the accuracy and robustness of the model, and the distribution of the generated data and the real data is compared to ensure that the generator can effectively capture the fault characteristics of the blower.

[0198] The above description is only a preferred embodiment of the present invention and does not constitute any other form of limitation to the present invention. Any modification or equivalent change made based on the technical essence of the present invention still falls within the scope of protection required by the present invention.

Claims

1. A surge fault warning method for a large-scale thermal power unit fan based on a multi-head attention generative adversarial network, characterized by: The steps include: S101: Collect historical operation data of the fan of a large thermal power plant, including but not limited to the fan air volume, fan opening, unit load, vibration signal, temperature signal and current signal; S102: using a data preprocessing module to preprocess the collected historical operation data of the induced draft fan, including data cleaning and feature extraction, firstly screening the steady-state data, then performing correlation analysis, dimensionality reduction, denoising, normalization, data cleaning, labeling and feature engineering, to provide a data set for model training; S103: Construct a fault diagnosis model based on a multi-head attention generative adversarial network. The generator uses a multi-head attention mechanism to capture the global features of the input data. The multi-head attention layer replaces the traditional convolutional layer, allowing the generator to learn different feature representations on different attention heads. The discriminator uses multi-head attention to improve the ability to distinguish between generated samples and real samples. Under the multi-head attention mechanism, each attention head calculates the attention score of the input and generates the final output by weighted average. In the generator and discriminator, a feature fusion module is introduced to fuse the outputs of different levels and different attention heads to enhance the expressiveness of the model. The quality of generated samples is improved by combining adversarial loss with feature matching loss. Feature matching loss is calculated by comparing the feature representations of generated samples and real samples in the discriminator. Stochastic gradient descent SGD or adaptive moment estimation optimization algorithm Adam is used to update the parameters of the generator and discriminator. The generated samples are evaluated through evaluation indicators to ensure that the improved model is superior to the baseline model in terms of generation quality and diversity. Through continuous adversarial training, the parameters of the generator and discriminator are optimized and adjusted, so that the pseudo data generated by the generator is closer and closer to the real data, and the discriminator's discriminative ability is stronger and stronger, and finally the target fault diagnosis model is obtained; S104: inputting the real-time collected operation data of the blower into the trained fault diagnosis model, diagnosing and predicting the fault of the blower through the synergy of the generator and the discriminator, judging whether the blower has a fault according to the output result of the discriminator, and determining the type and severity of the fault; S105: When the early warning model determines that the blower is surging, the early warning signal will be promptly notified to the power station staff so that corresponding preventive measures can be taken. At the same time, the early warning information will be uploaded to the monitoring system of the power station to integrate the early warning system with the monitoring system of the power station so that the operation and maintenance personnel can conduct remote monitoring and decision-making.

2. According to claim 1, the method for early warning of surge fault of blower fan of large thermal power unit based on multi-head attention generative adversarial network is characterized in that: The data preprocessing module in step S102 performs denoising, smoothing and standardization on the collected data, extracts features, uses Python or R or MATLAB, and utilizes SciPy and Pandas databases for data processing and feature extraction; during the real-time data collection, LabVIEW, MATLAB or custom-developed collection software is used to transmit data through sensors and data acquisition cards. The sensors are used to collect the operating status data of the blower, including vibration sensors, temperature sensors and pressure sensors. The data acquisition card converts the sensor signals into digital signals and performs data collection. The custom-developed collection software collects the operating data of the blower of a large thermal power plant through a DCS / SIS system in the form of a data communication interface.

3. According to claim 1, the method for early warning of surge fault of blower fan of large thermal power plant based on multi-head attention generative adversarial network is characterized in that: The step S103 uses the adaptive moment estimation optimization algorithm Adam to update the parameters of the generator and the discriminator, as follows: The calculation formula of the Adam optimization algorithm is: Among them, m t and v t Represent the first-order moment estimate and the second-order moment estimate respectively, β and β2 represent the exponential decay rate of the moment estimate, It means that the loss function is in the parameter θ t-1 The gradient at and They represent the modified first-order moment estimate and second-order moment estimate respectively, α represents the learning rate, represents a very small constant, and ∈ is used to prevent the denominator from being zero.

4. According to claim 1, the method for early warning of surge fault of blower fan of large thermal power unit based on multi-head attention generative adversarial network is characterized in that: The multi-head attention mechanism of the generator in step S103 is as follows: attention weight calculation is performed for the input sequence Assume that its dimension is n×d, where n is the number of samples, i.e. the length of the sequence, and d is the feature dimension; First, the input data is mapped into query, key, and value through linear transformation. The formula is as follows: Q=XW Q ,K=XW K ,V=XW V in is the learnable weight matrix, d k is the dimension of the attention head; Then use the soft max function to calculate the attention weight, and calculate the attention score by weighted summing of the values ​​​​through the attention weight: Here k is the dimension of the key, used for scaling; Then extend the above process to multiple heads h: MultiHead(Q,K,V)=Concat(head1,...,head h )W O The calculation for each head is: head i =Attention(QW i Q ,KW i K ,VW i V ) Here W i Q ,W i K ,W i V is the linear transformation weight of each head.

5. The method for early warning of surge fault of blower fan of large thermal power unit based on multi-head attention generative adversarial network according to claim 1 is characterized in that: In the step S103, regarding the generator loss function, the goal of the generator is to maximize the loss of the discriminator, and the loss function is expressed as: Where D is the discriminator, G is the generator, and z is the input noise; By comparing the features of the generated samples and the real samples in the discriminator, the feature matching loss is defined as: Where φ is the feature extraction layer of the discriminator, and x is the real sample; The final generator loss function is: Here λ is the weight parameter; The loss function of the discriminator is defined as: The training goal is to minimize the loss of the generator and the discriminator:

6. The method for early warning of surge fault of blower fan of large thermal power plant based on multi-head attention generative adversarial network according to claim 1 is characterized by: In step S103, a feature fusion module is introduced into the generator and the discriminator to fuse the outputs of different levels and different attention heads to enhance the expression ability of the model, and feature fusion is performed by feature concatenation or linear transformation fusion; The feature splicing combines features of different levels or different attention heads by splicing; Assume there are multiple feature outputs F1, F2, ..., F n , whose dimensions are all d, then the concatenated features are expressed as: F concat =Concat(F1,F2,...,F n ) Its dimension will become n×d if the number of features is n; The linear transformation fusion maps the features to the same dimension through linear transformation, and then adds or concatenates them; Assume that each feature is passed through the weight matrix W i Perform a linear transformation: F i '=F i W i Then you can choose concatenation or weighted summation: F fused =Concat(F 1′ ,F 2′ ,...,F n′ ) or 7. The method for early warning of surge fault of blower fan of large thermal power plant based on multi-head attention generative adversarial network according to claim 1 is characterized by: The steps of optimizing by stochastic gradient descent in step S103 are as follows: Randomly initialize the generator parameters θ G and the discriminator parameters θ D Calculate the loss of the generator and the discriminator, and calculate the gradient of the loss of the generator and the discriminator with respect to their parameters: S4: Update the parameters of the generator and discriminator: where η is the learning rate.

8. The method for early warning of surge fault of blower fan of large thermal power plant based on multi-head attention generative adversarial network according to claim 1 is characterized by: The step S103 evaluates the data quality generated by the generator through evaluation indicators, evaluates the generated samples, and ensures that the improved model is superior to the baseline model in terms of generation quality and diversity. The specific model is as follows: Specifically, accuracy, recall and F1 score are used to evaluate model performance. Where TP, FP, and FN are the number of true positives, false positives, and false negatives, respectively.

Images