A Fault Warning Method for Coal Mill Based on Attention Mechanism

Abstract

The present invention discloses a coal mill fault warning method based on an attention mechanism, belonging to the technical field of coal mill condition monitoring and fault warning. It includes Step 1: Obtain the dynamic change data of all parameters during the historical normal operation time of the coal mill; Step 2: Analyze the causes and phenomena of coal mill coal breakage, coal blockage, and coal spontaneous combustion faults, and select the key measuring point parameters representing the coal mill state as modeling variables; Step 3: Perform data preprocessing on the data set; Step 4: Establish a Transformer prediction model based on the attention mechanism and conduct online training; Step 5: Obtain the prediction deviation degree of the Transformer prediction model, and use this prediction deviation to characterize the health degree of the coal mill equipment; Step 6: Process the prediction deviation degree sequence through the sliding window method, calculate the fault warning threshold, and conduct fault warning. The numerical values of the key measuring point parameters selected in the present invention will have a large deviation from the model prediction values, realizing the early fault warning of the equipment.

Description

Technical Field

[0001] The present invention relates to the technical field of coal mill condition monitoring and fault warning, and particularly to a coal mill fault warning method based on an attention mechanism. Background Art

[0002] The coal mill is an important auxiliary equipment in thermal power plants, and its condition directly affects the stability and safety of unit operation. When a fault occurs inside the coal mill, causing a decrease in the pulverized coal flow rate, resulting in a decrease in the heat generated by boiler combustion, and the energy contained in the superheated steam cannot meet the energy balance required by the steam turbine, the unit load will rapidly decrease. In severe cases of faults, the boiler combustion state is affected, leading to boiler flameout and unit shutdown. The normal and stable operation of the coal mill is an important guarantee for the safe and economic operation of the power plant. Therefore, by monitoring the operation status of the coal mill and performing timely maintenance on the equipment when early faults occur, it is of practical and important significance to avoid the further development of faults.

[0003] In Agrawal, V., B. K. Panigrahi and P. M. V. Subbarao, Review of control and fault diagnosis methods applied to coal mills[J]. Journal of process control, 2015. 32: p. 138 - 153., the fault warning methods are divided into three categories: model-based, signal-based, and data-based.

[0004] Model-based methods mainly analyze the complex operating principles of equipment, establish differential equations that conform to the physical properties of the equipment, and judge the equipment condition by comparing the deviation between the model output and the measured value. Odgaard, P.F. and B. Mataji, Observer-based fault detection and moisture estimating in coal mills[J]. Control Engineering Practice, 2008. 16(8): p. 909-921. uses a simplified energy balance equation to monitor and diagnose abnormal energy flow in coal mills. Guo, S., et al., A new model-based approach for power plant Tube-ball mill condition monitoring and fault detection[J]. Energy Conversion and Management, 2014. 80: p. 10-19. and Wei, J., J. Wang and Q.H. Wu, Development of a Multisegment Coal Mill Model Using an Evolutionary Computation Technique[J]. IEEE transactions on energy conversion, 2007. 22(3): p. 718-727. establish a multi-segment coal mill model and monitor the state of the coal mill equipment by monitoring the anomalies of model variables. Model-based methods rely on the established mathematical models, which have clear physical meanings and strong interpretability. However, due to the complexity of actual industrial processes and numerous interference factors, it is difficult to establish a high-precision mathematical model, which limits the application of this type of method.

[0005] Signal-based methods are widely used in fault diagnosis where high-frequency signals exist, such as the current of coal mills, the differential pressure at the inlet and outlet of coal mills, and the vibration of coal mills. Liu Jiwei, Multiscale State Monitoring Method and Application Based on Big Data [D]. Beijing: North China Electric Power University, 2013. Time-frequency analysis of the operating signals of coal mills is carried out through wavelet decomposition to filter out the noise components of the signals and assist in monitoring the operating conditions of the equipment. Su, Z., et al., Experimental investigation of vibration signal of an industrial tubular ball mill: Monitoring and diagnosing [J]. Minerals Engineering, 2008. 21(10): p. 699-710. A system for recording the vibration signals of coal mills is designed, and it is transformed into energy amplitudes through wavelet packet decomposition. Through this model and method, the operating mode of the coal mill can be diagnosed. Signal-based methods avoid establishing complex mathematical models and only monitor the equipment status by analyzing the operating data of the equipment and extracting fault features.

[0006] The method based on historical data judges the equipment status by comparing the deviation between normal data and fault data. Zhang Wentao, et al., Fault Diagnosis of Coal Mills Based on Particle Swarm Optimization Kernel Extreme Learning Machine [J]. Electric Power Science and Engineering, 2018. 34(09): pp. 54-58. A black-box model of the coal mill is established through the kernel extreme learning machine, and the model parameters are optimized through the particle swarm algorithm. Hong, X., Z. Xu and Z. Zhang, Abnormal Condition Monitoring and Diagnosis for Coal Mills Based on Support Vector Regression [J]. IEEE access, 2019. 7: p. 170488-170499. Using the support vector machine, a model is established based on the operating data of the coal mill, and the deviation between the model prediction and the actual value is compared. When the deviation is too large, it is judged that a fault has occurred.

[0007] With the development of deep learning, the fault diagnosis method of neural networks has been widely applied in industry, but it is less applied in the field of fault early warning of thermal power plant equipment. Wen, L., et al., A New Convolutional NeuralNetwork-Based Data-Driven Fault Diagnosis Method[J]. IEEE Transactions onIndustrial Electronics, 2018. 65(7): p. 5990-5998., Xie, T., X. Huang and S. Choi, Intelligent Mechanical Fault Diagnosis Using Multisensor Fusion andConvolution Neural Network[J]. IEEE Transactions on Industrial Informatics, 2022. 18(5): p. 3213-3223., and Chen, H., et al., A deep convolutional neural networkbased fusion method of two-direction vibration signal data for health stateidentification of planetary gearboxes[J]. Measurement, 2019. 146: p. 268-278. utilize data fusion and convolutional neural network technology to mine the information in the vibration data of gearboxes and establish a gearbox fault diagnosis model. Wen, L., X. Li and L. Gao, A New Two-Level Hierarchical Diagnosis Network Based onConvolutional Neural Network[J]. IEEE transactions on instrumentation andmeasurement, 2020. 69(2): p. 330-338. establish a two-level convolutional neural network model to judge the fault type and severity of gears simultaneously.Hu, Y., et al., Modeling of Coal Mill System Used for Fault Simulation[J]. Energies (Basel), 2020. 13(7): p. 1784. and Hu, Y., et al., Research on fault diagnosis of coal mill system based on the simulated typical fault samples[J]. Measurement, 2020. 161: p. 107864. By building a mechanism model of the coal mill to simulate fault data, the state of the coal mill is judged using a stacked autoencoder with an encoding-decoding structure. Equipment such as wind power gearboxes and coal mills has a delay characteristic. The input and output of the equipment are not only determined by the current moment, but the historical state also affects the variables. The special structure of the recurrent neural network can consider the relationship between different variables while considering the dependence relationship of the variables in terms of time sequence. Yong Bin, et al., A method for early warning of the state of a wind power gearbox based on a gated recurrent network integrating multi-source data[J]. Acta Energiae Solaris Sinica, 2021. 42(08): pp. 421-425, Huang Rongzhou, et al., State monitoring of a wind power gearbox based on a long short-term memory network integrating SCADA data[J]. Acta Energiae Solaris Sinica, 2021. 42(01): pp. 235-239, and Wang Chao, Li Dazhong, Bearing fault early warning of a wind turbine gearbox based on an LSTM network[J]. Electric Power Science and Engineering, 2020. 36(09): pp. 40-45. Use RNN, LSTM, and GRU to establish a black-box model of the wind power gearbox, and diagnose the equipment condition through the change trend of the model prediction error. Currently, the main deep learning-based fault diagnosis methods are CNN and RNN. However, RNN uses the state calculated at the previous moment as the input for the next moment, and this inherent order limits the parallelization of the model. Due to the limited receptive field of the convolutional kernel in CNN, when the spatial distance between features increases, the number of operations also increases. To solve the above problems, Ashish Vaswani, N.S.N.P., Attention is All You Need[C]. Conference on Neural Information Processing Systems, Long Beach, 2017. proposed a Transformer network structure that only contains an attention mechanism, and its powerful performance has been proven in the fields of natural language processing and computer vision.

[0008] In view of the shortcomings of the existing technology, there is still a need for a coal mill fault warning method based on the attention mechanism, which does not rely on accurate mechanism models and complete expert knowledge, but uses advanced models in the field of deep learning to build a Transformer prediction model based only on the attention mechanism. Summary of the invention

[0009] The purpose of the present invention is to propose a coal mill fault early warning method based on attention mechanism, characterized in that the method comprises the following steps:

[0010] Step 1: Obtain the dynamic change data of all parameters of the coal mill of the DCS system of the thermal power plant during the historical normal operation time to form a data set; the time interval of the dynamic change data is at most 1 minute;

[0011] Step 2: Analyze the causes and phenomena of coal mill failures such as coal shortage, coal blockage, and coal spontaneous combustion, and select key measurement point parameters representing the coal mill status as modeling variables;

[0012] Step 3: Preprocess the data set in step 1;

[0013] Step 4: Using the key measurement point parameters as input variables, establish a Transformer prediction model based on the attention mechanism and perform online training;

[0014] Step 5: Obtain the prediction deviation of the Transformer prediction model in step 4, and use this prediction deviation to characterize the health of the coal mill equipment;

[0015] Step 6: Process the prediction deviation sequence through the sliding window method, calculate the fault warning threshold, and perform fault warning.

[0016] The matrix of the key measurement point parameters in step 2 is expressed as:

[0017]

[0018] Among them, X1 is the coal mill current, X2 is the instantaneous coal feed rate, X3 is the coal mill outlet air-powder mixture pressure, X4 is the coal mill outlet air-powder mixture temperature, X5 is the coal mill inlet primary air pressure, X6 is the coal mill inlet primary air temperature, X7 is the coal mill inlet primary air volume, X ij Represents the jth sample point data of the i-th measurement point parameter.

[0019] The step 3 includes the following sub-steps:

[0020] Step 31: Perform data validity check on the data set column by column, and remove abnormal timestamp data and abnormal numerical data;

[0021] Step 32: Perform wavelet packet denoising on the data after data cleaning in Step 31 column by column;

[0022] Step 33: Normalize the data after denoising in Step 32;

[0023] Step 34: Convert the dataset into supervised data to train the Transformer model.

[0024] The formula for wavelet packet denoising in Step 32 is as follows:

[0025]

[0026]

[0027] T(j,n) = σ / 2 / σ x (4)

[0028]

[0029] where f(t) is the data sequence of the measurement point parameters in a certain column of the matrix of key measurement point parameters, ψ(τ,a) is the wavelet basis function, τ is the translation parameter, t - τ controls the translation distance of the wavelet basis function along the time axis, a is the scaling parameter, WT(a,τ) is the wavelet coefficient obtained after decomposition, P i k (x) is the wavelet coefficient of the k-th node in the i-th layer, is the Shannon entropy of the wavelet coefficient of the k-th node in the i-th layer, T is the denoising threshold, σ x is the variance of the original data f(t), σ 2 is the variance of the wavelet coefficient of the high-frequency subband , is the n-th data point in the wavelet coefficient sequence of the j-th node in the l-th layer.

[0030] Specifically, Step 34 is as follows:

[0031] Use the window method to sample the time series {X i (1), X i (2),..., X i (n), X i (n + 1),...} of the i-th key measurement point parameter, and use as the input data of the Transformer model to predict the parameter value at time t, where Input tThe input sample for the model established by the sliding window at the t-th time; using X(t) = {X1(t), X2(t),..., X7(t)} as the true value to correct the Transformer model, and creating data pairs through the sliding window to obtain the training dataset of the Transformer model.

[0032] Step 4 includes the following sub-steps:

[0033] Step 41: Construct a Transformer prediction model including a Transformer network, an activation function, and initial weights;

[0034] Step 42: Compile the Transformer prediction model, and select an optimization algorithm, a learning objective, and an evaluation metric;

[0035] Step 43: Divide the dataset into 80% as the training set and 20% as the test set to train the Transformer prediction model, and set the training batch to 512 and the number of training epochs to 200 for hyperparameter tuning.

[0036] Specifically, Step 41 includes:

[0037] First, regard the input data established by the sliding window as a 3×7 matrix, and the position encoding calculation formula is:

[0038]

[0039] In the formula, pos represents the row index of the 3×7 matrix representing the chronological order, i represents the column index of the 3×7 matrix representing the positions of different measurement point parameters, and d model represents the input dimension of the connected multi-head attention module;

[0040] Then, superimpose the calculated position encoding information on the original input, calculate the attention evaluation metric of the input through the multi-head attention module, and obtain the correlation information between the input data and the predicted value. The calculation formula is:

[0041]

[0042] In the formula, X is the data input to the attention module, and W Q , W K , W V is the weight of the fully connected layer that maps the input X, and W o is the fusion matrix that fuses the calculation results of multiple independent attention modules;

[0043] After that, a residual connection is made between the input and the output of the attention module to optimize the gradient of forward propagation; the data is normalized by the normalization module; the fully connected layer and the normalization module are connected to continue extracting the deep features of the data, forming an encoder; multiple encoders are stacked to comprehensively extract the features of the input data;

[0044] Finally, through the real future moment data, after position encoding and calculation by the attention module, the attention evaluation index is calculated again with the calculation result of the encoder to obtain the importance evaluation information between the input and the predicted value, forming a decoder; multiple decoders are stacked to extract deep features; a fully connected layer with a dimension of 1×7 is connected behind the decoder, and the Sigmoid activation function is selected to predict the future moment value of the measuring point parameters of the coal mill.

[0045] The optimization algorithm in step 42 selects the Adam algorithm, and the parameters are set as lr = 0.002, beta_1 = 0.9, beta_2 = 0.999, epsilon = 1*10 -8 , where lr is the learning rate, beta_1 is the exponential decay rate of the first moment estimate, beta_2 is the exponential decay rate of the second moment estimate, and epsilon is the parameter to prevent division by zero;

[0046] The learning objective selects the mean square error function MSE:

[0047]

[0048] In the formula, is the true value, y is the model predicted value, and n is the number of input samples;

[0049] The evaluation indexes select the mean absolute error MAE, the mean percentage error MAPE, and the correlation coefficient R 2 :

[0050]

[0051]

[0052]

[0053] The calculation formula for the prediction deviation degree in step 5 is:

[0054]

[0055] In the formula, is the true value, and y is the model predicted value.

[0056] The calculation method for the fault warning threshold in step 6 is as follows:

[0057] After processing the prediction deviation sequence by the sliding window method, the average prediction deviation is obtained. And the probability density function of the average prediction deviation is calculated by the kernel density estimation method. The calculation formula is:

[0058]

[0059] In the formula, is the probability density function, t is the element in the sample sequence, t k is the average value of the sample sequence elements, h is the bandwidth, m is the length of the sample sequence, and σ is the variance of the sample sequence;

[0060] Then solve the cumulative integral function of the prediction deviation sequence to obtain the upper limit value t of the sequence with a 99% confidence level d as the fault warning threshold, and the calculation formula is:

[0061]

[0062] The beneficial effects of the present invention are as follows:

[0063] 1. When the coal mill fails, the numerical values of the key measurement point parameters selected by the present invention will deviate greatly from the model prediction values, realizing the early warning of equipment failures.

[0064] 2. The multi-variable prediction model of the Transformer coal mill established by the present invention fully explores the complex causal relationships between different variables, and realizes the mining of time-dependent relationships through position encoding, fully learning the characteristics of the measurement point data in the time series, avoiding the possibility of a large amount of missing effective information, and the results are relatively more reasonable and accurate.

[0065] 3. The model prediction accuracy of the present invention is higher than that of traditional models such as CNN and RNN. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 is the flow chart of the fault warning of the present invention;

[0067] Figure 2 is the effect diagram of wavelet packet noise reduction;

[0068] Figure 3 is the structure diagram of the Transformer model;

[0069] Figure 4 is the probability density function of the prediction deviation;

[0070] Figure 5 is the prediction deviation curve of the Transformer model;

[0071] Figure 6Box plot of prediction errors for each model;

[0072] Figure 7(a) shows the prediction deviation and threshold curves of CNN and Transformer;

[0073] Figure 7(b) shows the prediction deviation and threshold curves of LSTM and Transformer;

[0074] Figure 7(c) shows the prediction deviation and threshold curves of CNN+LSTM and Transformer;

[0075] Figure 7(d) shows the online warning results of each model. Specific implementation mode

[0076] The present invention proposes a fault warning method for coal mills based on an attention mechanism. The following further describes the present invention with reference to the accompanying drawings and specific embodiments.

[0077] Figure 1 The flow chart of the fault warning method for coal mills based on the attention mechanism includes the following steps:

[0078] Step 1: Obtain the measured point parameter data of the historical normal working state of the coal mill to form a data set;

[0079] Step 2: Analyze the causes and phenomena of coal mill faults, and screen the key measured point parameters representing the state of the coal mill as modeling variables. Specifically:

[0080] Collect the operation data of a 660MW power station coal mill for a one-week span time with a sampling period of 1 minute. Select the variables related to the coal blockage fault as modeling variables, establish the data matrix as shown in Equation (1), and the selected modeling variables are shown in Table 1:

[0081] Table 1 Variable names related to coal blockage faults

[0082]

[0083] Step 3: Perform data preprocessing on the historical data set, including the following sub-steps:

[0084] Step S31: Perform data validity check and outlier removal on the matrix data obtained in Step 2 column by column, and remove abnormal timestamp data (duplicate timestamps, incorrect sampling intervals) and abnormal numerical data (data exceeding the upper and lower limits of the normal state, equipment missing values);

[0085] Step S32: Perform wavelet packet denoising processing on the training data set after data cleaning according to Equations (2)-(5), and the denoising effect is as Figure 2 shown;

[0086] Step S33: Normalize the denoised data according to the following formula to eliminate the influence of numerical dimensions on prediction;

[0087]

[0088] In the formula, X i ={X i1 , X i2 ,..., X in} is the parameter data of the i-th measurement point, X min is the minimum value in X i , and X max is the maximum value in X i .

[0089] Step S34: Convert the data set into supervised data for training the Transformer network model;

[0090] Convert the time series data into a data set that can train the model. Specifically, record the time series of the parameter of the i-th measurement point as {X i (1), X i (2),..., X i (n), X i (n + 1),...}. Sample the sequence through the window method to establish the input data matrix Input t . Predict the parameter value at time t, and use X(t) = {X1(t), X2(t),..., X7(t)} as the true value to correct the model. Through the sliding window, create multiple data pairs to obtain the training data set of the Transformer model. Through training, establish a multi-input and multi-output prediction model of the coal mill based on Transformer.

[0091] Step 4: Use the measurement point parameters related to the fault as input variables to establish a prediction model Transformer based only on the attention mechanism and perform online training, including the following sub-steps:

[0092] Step S41: Construct the structure of the Transformer prediction model. The model structure is as Figure 3 shown, including the design of the Transformer network structure, the selection of activation functions, and the initialization of model weights; specifically including:

[0093] Send the input data established by the sliding window into the model. Predict the future data based on the data of three historical moments. Then regard all the measurement point parameter sequences as a 3×7 matrix, and perform positional encoding on the input data according to Equation (6);

[0094] Then, the calculated position encoding information is superimposed on the original input. After that, according to Equation (7), the attention evaluation index of the input is calculated through the multi-head attention module to obtain the correlation information between the input data and the predicted value;

[0095] After that, a residual connection is made between the input and the output of the attention module to optimize the gradient of the forward propagation; then the data is normalized through the normalization module; then the fully connected layer and the normalization module are connected to continue extracting the deep features of the data. The above steps constitute an encoder; finally, multiple encoders are stacked to comprehensively extract the features of the input data;

[0096] Through the real future moment data, after position encoding and calculation by the attention module, the attention evaluation index is calculated again with the calculation result of the encoder to obtain the importance evaluation information between the input and the predicted value. The above steps constitute a decoder; multiple decoders are stacked to extract deep features. A fully connected layer with a dimension of 1×7 is connected behind the decoder to predict the future moment values of the measuring point parameters of the coal mill, and the activation function is selected as Sigmoid.

[0097] Step S42, Compiling the Transformer prediction model, including setting the learning objective, optimization algorithm, and model evaluation index; specifically including:

[0098] Select the Adam algorithm as the optimization algorithm, and the algorithm parameters are set to lr = 0.002, beta_1 = 0.9, beta_2 = 0.999, epsilon = 1*10 -8 , where lr is the learning rate, beta_1 is the exponential decay rate of the first moment estimate, beta_2 is the exponential decay rate of the second moment estimate, and epsilon is the parameter to prevent division by zero;

[0099] Express the training index of the Transformer prediction model as the mean square error function MSE, and calculate according to Equation (8);

[0100] Calculate the evaluation indexes of the Transformer prediction model, the mean absolute error MAE, the mean percentage error MAPE, and the correlation coefficient R according to Equations (9)-(11) 2 .

[0101] Step S43, Training the Transformer prediction model and debugging the hyperparameters, including the division of the training set, test set, and validation set, and the adjustment of the network hyperparameters; divide the data set into 80% as the training set and 20% as the test set; set the network hyperparameters as: the training batch is 512, and the number of training epochs is 200.

[0102] Step 5: Obtain the prediction deviation obtained by the Transformer prediction model to characterize the health degree of the coal mill equipment; the prediction value vector composed of the parameters of each measuring point at each moment obtained by the Transformer prediction model Calculate the prediction deviation degree with the actual value vector y = {y1,..., y7} to characterize the operation status of the model. When the prediction deviation degree increases, it indicates that the equipment deviates from the normal operation state and potential faults occur. Calculate the prediction deviation degree index according to Equation (12).

[0103] Step 6: Process the comprehensive residual sequence by the sliding window method, calculate the fault warning threshold, and conduct fault warning. Specifically: the prediction deviation degree sequence obtained by Equation (12), after being processed by the sliding window method, the average prediction deviation degree is obtained The average prediction deviation degree curve is as Figure 5 shown; and by the kernel density estimation method, calculate the probability density function and warning threshold of the average prediction deviation degree according to Equations (13) - (14), and the probability density function is as shown; Figure 4 shown;

[0104] When the coal mill operates normally, the model can accurately predict the variables, and the prediction deviation degree is lower than the warning threshold. When the prediction deviation degree continuously exceeds the warning threshold, it is considered that the coal mill has a fault. If the prediction deviation degree continuously increases, it is considered that the coal mill fault is in the developing stage, and the staff needs to check and repair the equipment; after repair, if the prediction deviation degree decreases, it is considered that the fault is eliminated and the system can operate normally.

[0105] Establish CNN, LSTM, CNN + LSTM, and Transformer coal mill prediction models respectively. Calculate the prediction accuracy evaluation indexes of the models, and the results are shown in the table.

[0106] Prediction accuracy indexes of each model

[0107]

[0108]

[0109] By comparing with the baseline model, it can be seen that the prediction accuracy of the Transformer model is better than that of CNN, LSTM, and CNN + LSTM, and a high-precision coal mill prediction model can be established.

[0110] The prediction deviation degree box plot is as Figure 6 shown.

[0111] As can be seen from the figure, the median of the prediction deviation of the Transformer model under normal operating conditions (the position of the red line in the box plot) is smaller, and the upper limit of the error (the black line extending upward in the box plot) is lower than that of the CNN, LSTM, and CNN+LSTM models. This proves that the model has higher prediction accuracy and a smaller fluctuation range of prediction deviation.

[0112] The comparison of the warning results of each model is as Figure 7(a) , 7(b) , 7(c), and 7(d) show.

[0113] By analyzing Figure 7(d), an early coal blockage fault occurred at point 2094 in the coal mill. All types of models can issue alarms in the early stage of the coal blockage fault. Among them, the Transformer model can issue an alarm at point 1679, the CNN+LSTM model issues an alarm at point 1696, and the CNN has false alarm situations. Through the online warning experiment, it is verified that the fault warning strategy of wavelet packet denoising-attention mechanism-kernel density estimation can issue alarms in a timely manner in the early stage of the coal mill fault.

[0114] In summary, the present invention proposes to model the coal mill data through deep learning. When there are potential fault hazards in the equipment, the prediction deviation of the model will increase. By comparing it with the calculated fault threshold, early warning of faults is realized, and the prediction accuracy is higher than that of traditional models such as CNN and RNN. The deep learning method can fully mine the massive historical data of the coal mill equipment and establish an efficient and practical model to detect and warn the real-time state of the coal mill.

Claims

1. A fault warning method for coal mills based on the attention mechanism, characterized in that The method comprises the following steps, Step 1: Obtain the dynamic change data of all parameters of the coal mill of the DCS system of the thermal power plant during the historical normal operation time to form a data set; the time interval of the dynamic change data is at most 1 minute; Step 2: Analyze the causes and phenomena of coal mill failures such as coal shortage, coal blockage, and coal spontaneous combustion, and select key measurement point parameters representing the state of the coal mill as modeling variables; the matrix of the key measurement point parameters in step 2 is expressed as: Among them, X1 is the current of the coal mill, X2 is the instantaneous coal feeding amount, X3 is the pressure of the pulverized coal-air mixture at the outlet of the coal mill, X4 is the temperature of the pulverized coal-air mixture at the outlet of the coal mill, X5 is the primary air pressure at the inlet of the coal mill, X6 is the primary air temperature at the inlet of the coal mill, X7 is the primary air volume at the inlet of the coal mill, and X ij represents the data of the j-th sample point of the parameter of the i-th measuring point; Step 3: Preprocess the data set in step 1; step 3 includes the following sub-steps: Step 31: Perform data validity check on the data set column by column, and remove abnormal timestamp data and abnormal numerical data; Step 32: performing wavelet packet denoising on the data after data cleaning in step 31 by column; Step 33: normalize the data after noise reduction in step 32; Step 34: Convert the dataset into supervised data to train the Transformer model; Step 4: Using the key measurement point parameters as input variables, establish a Transformer prediction model based on the attention mechanism and perform online training; Step 5: Obtain the prediction deviation of the Transformer prediction model in step 4, and use this prediction deviation to characterize the health of the coal mill equipment; Step 6: Process the prediction deviation sequence through the sliding window method, calculate the fault warning threshold, and perform fault warning; The calculation method of the fault warning threshold in step 6 is as follows: After processing the prediction deviation sequence with the sliding window method, the average prediction deviation is obtained And calculate the probability density function of the average prediction deviation by the kernel density estimation method The calculation formula is as follows: h = 1.06σm -0.2 In the formula, is the probability density function, t is an element in the sample sequence, and t k is the average value of the sample sequence elements, h is the bandwidth, m is the length of the sample sequence, and σ is the variance of the sample sequence; Solve the cumulative integral function of the prediction deviation degree sequence again to obtain the upper limit value t of the sequence with a 99% confidence level d As the fault warning threshold, the calculation formula is as follows:

2. The method for early warning of coal mill faults based on the attention mechanism according to claim 1, wherein, The formula for wavelet packet denoising in step 32 is as follows: T(j,n) = σ 2 / σ x Among them, f(t) is the data sequence of a certain column of measurement point parameters in the matrix of key measurement point parameters, ψ(τ,a) is the wavelet basis function, τ is the translation parameter, t - τ controls the translation distance of the wavelet basis function along the time axis, a is the scaling parameter, and WT(a,τ) is the wavelet coefficient obtained after decomposition. is the wavelet coefficient of the k-th node in the i-th layer. is the Shannon entropy of the wavelet coefficient of the k-th node in the i-th layer, T is the noise reduction threshold, and σ x is the variance of the original data f(t), and σ 2 is the variance of the wavelet coefficients in the high-frequency sub-band. is the n-th data point in the wavelet coefficient sequence of the j-th node in the l-th layer.

3. The method for early warning of coal mill faults based on the attention mechanism according to claim 1, wherein The step 34 is specifically as follows: Using the window method to sample the time series {X i (1), X i (2),..., X i (n), X i (n + 1),...} of the parameters of the i-th key measurement point, and using as the input data of the Transformer model to predict the parameter value at time t, where Input t is the input sample of the model established by the t-th sliding window; using X(t) = {X1(t), X2(t),..., X7(t)} as the true value to correct the Transformer model, and creating data pairs through the sliding window to obtain the training data set of the Transformer model.

4. The method for early warning of coal mill faults based on the attention mechanism according to claim 1, wherein, The step 4 includes the following sub-steps: Step 41: Construct a Transformer prediction model including a Transformer network, an activation function, and initialization weights; Step 42: Compile the Transformer prediction model and select the optimization algorithm, learning objectives, and evaluation indicators; Step 43: Divide the dataset into 80% as a training set and 20% as a test set to train the Transformer prediction model, and set the training batch to 512 and the training round to 200 for hyperparameter adjustment.

5. The method for early warning of coal mill faults based on the attention mechanism according to claim 4, characterized in that The step 41 specifically includes: First, the input data established by the sliding window is regarded as a 3×7 matrix, and the position encoding calculation formula is: where pos represents the row index of the 3×7 matrix indicating the chronological order, i represents the column index of the 3×7 matrix indicating the positions of different measuring point parameters, and d model represents the input dimension of the connected multi-head attention module; Then, the calculated position encoding information is superimposed on the original input, and the attention evaluation index of the input is calculated through the multi-head attention module to obtain the correlation information between the input data and the predicted value. The calculation formula is: Where X is the data input to the attention module, and W Q , W K , W V are the weights of the fully connected layers that map the input X, and W o is the fusion matrix that fuses the calculation results of multiple independent attention modules; After that, the input and the output of the attention module are residually connected to optimize the gradient of the forward propagation; the data is standardized through the normalization module; the fully connected layer and the normalization module are connected to continue to extract the deep features of the data to form an encoder; multiple encoders are stacked to fully extract the features of the input data; Finally, through the real future moment data, after position encoding and calculation by the attention module, the attention evaluation index is calculated again with the calculation result of the encoder to obtain the importance evaluation information between the input and the predicted value, forming a decoder; multiple decoders are stacked to extract deep features; a fully connected layer with a dimension of 1×7 is connected behind the decoder, and the Sigmoid activation function is selected to predict the future moment value of the measuring point parameters of the coal mill.

6. The method for early warning of coal mill faults based on the attention mechanism according to claim 4, wherein The optimization algorithm in step 42 selects the Adam algorithm, and the parameters are set as lr = 0.002, beta_1 = 0.9, beta_2 = 0.999, epsilon = 1*10 -8 , where lr is the learning rate, beta_1 is the exponential decay rate of the first moment estimate, beta_2 is the exponential decay rate of the second moment estimate, and epsilon is a parameter to prevent division by zero; The mean square error function MSE is selected as the learning target: where is the true value, y is the model prediction value, and n is the number of input samples; The evaluation metrics selected are the Mean Absolute Error (MAE), the Mean Absolute Percentage Error (MAPE), and the correlation coefficient R 2 :

7. The method for early warning of coal mill faults based on the attention mechanism according to claim 1, characterized in that, The calculation formula for the prediction deviation in step 5 is: Wherein, is the true value and y is the model predicted value.

Images