随着信息技术和自动化技术的快速发展，现代工业系统的集成度和复杂度越来越

With the rapid development of information technology and automation technology, the integration and complexity of modern industrial systems are getting higher and higher. The mutual influence between various parts is becoming more and more complex, leading to the increasing probability of system failures and functional failures, and once the failures occur, the harmful effects will be great, and the serious will lead to the failure and paralysis of the entire system. Therefore, with the improvement of system reliability requirements, fault prediction technology has received key attention in industry and academia. Failure prediction refers to predicting the time of failure or judging whether the system will fail at a certain point in the future based on the past and current operating status of the system that can be obtained. Online failure prediction is also called system life prediction. Its purpose is to predict the remaining service life (RUL) before failure and process collapse, given the current machine state and past operating conditions. <br>Inspired by the deep structure network, this paper aims at the high-dimensional and non-linear characteristics of complex industrial process data, which causes the system operation to have problems such as large calculation amount and long operation time. The fault prediction method with time convolutional network as the core is studied. The improvement of time convolutional network and the combination of attention mechanism improve the accuracy of model prediction and complete online fault prediction. The main researches are as follows: <br>(1) Implement a fault monitoring based on random forest-stack noise reduction auto-encoding. <br>In some fault situations, not all process variables are disturbed, and may not contain meaningful faults. Information, so an important issue is to eliminate redundant features. Aiming at the nonlinear and high-dimensional characteristics of complex industrial processes, the random forest algorithm (RF) is used to screen out important feature variables as input data to achieve dimensionality reduction of the input data. Then, stack multiple denoising autoencoders to construct a deep learning network structure; use Stack Denoising Auto Encoder (SDAE) algorithm to reconstruct the original features, and construct squared prediction error according to the reconstruction error. , SPE) as the monitoring statistics of the failure to predict the failure. <br>(2) Research a fault prediction method based on SFTCN to predict process fault status<br>Research a failure prediction method based on SFTCN to predict the failure state of the process. Aiming at the problem that the ReLU activation function in the residual module of the temporal convolutional network is zero in the negative interval of x, the negative gradient will be set to zero, and the neuron may not be activated. The Swish activation function is introduced and a design based on Swish is introduced. Activation function and FRN (Filter Response Normalization) normalized temporal convolutional network (Swish and FRN-based Temporal Convolutional Network, SFTCN). Furthermore, the obtained SPE is formed into a time series, and the forecasting model of SFTCN is used to realize the state trend prediction of SPE. <br>(3) A multi-variable attention-SFTCN method is proposed to realize the trend prediction <br>of the fault status of complex industrial processes. Because the fault is affected by multiple variables, the original features are combined to form a new data set, and a multi-variable attention-SFTCN method is proposed. The attention mechanism is used to weight the TCN output to improve the prediction accuracy of the network. <br>(4) Experimental research on actual data <br>The method proposed in the article is applied to a certain actual production data, and experiments are carried out to test the rationality and effectiveness of the failure prediction method studied. The results show that the method proposed in the article can effectively extract the deep features of variable data, significantly improve the accuracy of fault prediction, ensure the rapidity of monitoring, but also ensure its superior predictive ability, and can help operators find abnormalities in time and effectively prevent accidents. occur.

With the rapid development of information technology and automation technology, the integration and complexity of modern industrial systems are getting higher and higher. The interaction between the various parts is also becoming more and more complex, resulting in the failure of the system and the probability of functional failure gradually increased, and once the failure occurs, the harmful effect is great, serious will lead to the failure and paralysis of the entire system. Therefore, with the improvement of system reliability requirements, fault prediction technology has been paid more attention in industry and academia. Failure prediction refers to predicting when a failure occurs or judging whether the system will fail at some point in the future, based on the past and present operating conditions of the acquired system. Online fault prediction, also known as system life prediction, is designed to predict the remaining useful life (RUL) before a failure occurs and a process crashes, given the current machine state and past health.<br> Inspired by the deep structure network, this paper aims at the high-dimensional and nonlinear characteristics of complex industrial process data, which leads to the problems of large computation and long computation time in system operation, studies the fault prediction method with time-volume network as the core, and improves the time-volume network and combined with attention mechanism to improve the accuracy of model prediction and complete the on-line prediction of failure. The following aspects of research have been done:<br> (1) A fault monitoring based on random forest-stack noise reduction self-coding is implemented<br> In some failure cases, not all process variables are disturbed and may not contain meaningful information about the failure, so an important issue is troubleshooting redundant features. For the nonlinear and high-dimensional characteristics of complex industrial processes, the random forest algorithm (RF) is used to filter out important feature variables as input data and realize the degradation of input data. Multiple noise-cancelling auto-encoders are then stacked to build a deep learning network structure, and the fault is predicted using the Stack Denoising Auto Encoder (SDAE) algorithm to reconstruct the original features and construct square prediction errors (squared prediction errors, SPEs) as monitoring statistics for failures based on refactoring errors. <br> (2) A fault prediction method based on SFTCN is studied to predict the failure status of the process<br> A fault prediction method based on SFTCN is studied to predict the failure status of the process. For the problem that the ReLU activation function in the time constricted network residual module has zero conductors in the x negative interval, which causes the negative gradient to be set to zero, and the neurons may not be activated, introduce the Swish activation function and design a time convex network based on the Swish activation function and FRN Normalization (Swish FR and NN-base Temporal). Convolutional Network, SFTCN）。 Then the SPEs will form a time series, and the SFTCN prediction model will be used to realize the state trend prediction of SPEs.<br> (3) A multivariate attention-SFTCN method is proposed to predict the failure state trend of complex industrial processes<br> Because the fault is affected by many variables, it combines the original characteristics to form a new data set and proposes a multivariable attention-SFTCN method. Weighted distribution of TCN output by attention mechanism is used to improve the prediction accuracy of the network.<br> (4) Experimental study of actual data<br> The methods mentioned in this paper are applied to an actual production data, and the rationality and validity of the fault prediction methods studied are tested experimentally. The results show that the method mentioned in this paper can effectively extract the deep characteristics of variable data, obviously improve the accuracy of fault prediction, ensure the speed of monitoring and ensure its superior prediction ability, and can detect anomalies and effectively prevent accidents by snob operators.

With the rapid development of information technology and automation technology, the integration and complexity of modern industrial system is becoming higher and higher. The interaction between the various parts is becoming more and more complex, leading to the probability of system failure and functional failure is gradually increasing, and once the failure occurs, the harm is great, serious will lead to the failure and paralysis of the whole system. Therefore, with the improvement of system reliability requirements, fault prediction technology has been focused on in industry and academia. Fault prediction is to predict the time of fault occurrence or judge whether the system will fail at a certain time in the future according to the past and present operation state of the system. Online fault prediction is also known as system life prediction. Its purpose is to predict the remaining service life (rul) before fault occurrence and process collapse given the current machine state and past operating conditions.<br>Inspired by the deep structure network, aiming at the high-dimensional and nonlinear characteristics of complex industrial process data, which leads to the problems of large amount of calculation and long operation time, this paper studies the fault prediction method based on time convolution network. By improving the time convolution network and combining the attention mechanism, the accuracy of model prediction is improved, and the online fault prediction is completed . The main research is as follows<br>(1) A fault monitoring method based on random forest stack noise reduction self coding is implemented<br>In some fault cases, not all process variables are disturbed and may not contain meaningful information about the fault, so an important problem is to eliminate redundant features. In view of the nonlinear and high-dimensional characteristics of complex industrial processes, the random forest algorithm (RF) is used to filter out important characteristic variables as input data to achieve input data dimensionality reduction. Then, multiple de-noising auto encoders are stacked to construct a deep learning network structure; the original features are reconstructed by using stack denoising auto encoder (sdae) algorithm, and the square prediction error (SPE) is constructed according to the reconstruction error as the fault monitoring statistics to predict the fault.<br>(2) A fault prediction method based on sftcn is studied to predict the process fault state<br>A fault prediction method based on sftcn is studied to predict the process fault state. In order to solve the problem that the derivative of relu activation function in the residual module of time convolution network is zero in the negative interval of X, the negative gradient will be set to zero, and the neurons may not be activated, the swish activation function is introduced to design a swish and FRN based temporal convolutional network based on swish activation function and filter response normalization Network, SFTCN）。 Then the obtained SPE is composed of time series, and the state trend prediction of SPE is realized by using the prediction model of sftcn.<br>(3) A multivariable attention sftcn method is proposed to predict the fault state trend of complex industrial processes<br>Because the fault is affected by many variables, a new data set is formed by combining the original features, and a multivariable attention sftcn method is proposed. The attention mechanism is used to distribute the output of TCN to improve the prediction accuracy.<br>(4) Experimental study on real data<br>The proposed method is applied to an actual production data to test the rationality and effectiveness of the fault prediction method. The results show that the proposed method can effectively extract the deep features of variable data, significantly improve the accuracy of fault prediction, ensure the monitoring speed and superior prediction ability, and guide the operators to find the abnormal in time and effectively prevent accidents.<br>