Method And Implementation Of Equipment Running State Prediction Based On Temporal Convolutional Network

In modern manufacturing production systems,effective monitoring and forecasting of equipment operating status is conducive to troubleshooting hidden troubles in advance,and ensuring that equipment operates safely at low cost.However,with the rapid development of intelligent manufacturing,the widespread deployment of sensors has collected a large amount of time series data that records equipment operating status information.The large scale,complex type,and uneven value density of these data make the data processing and model training process complicated and time-consuming,it is difficult to meet the requirements for real-time prediction of operating status.Moreover,the problem of data imbalance also severely restricts the improvement of model generalization and prediction accuracy.In this context,this paper proposes a method for predicting the operating state of equipment based on the Att-MDTCN model.The model is based on a Temporal Convolutional Network and combines the Attention Mechanism of feature dimensions and a weighted loss function to improve model performance.The model mainly includes the following parts: 1)Aiming at the problem that the original time convolutional network cannot receive and process multi-variable data,the model designs a stacked convolution kernel by splicing convolution kernels,and realizes the processing of multivariate data on the basis of ensuring that it still has the characteristics of dilated causal convolution,and fully obtains the timing characteristics of the data;2)Aiming at the problem that the activation function used by the original Temporal Convolutional Network will deactivate the neuron,replace the original activation function with the Re LU activation function with parameters,so that the network increases the ability to derive the input that is less than zero;3)Considering that the importance of different variable features may have an impact on the prediction results,the Attention Mechanism is introduced.Different from the traditional Attention Mechanism selection related time steps,this model applies the Attention Mechanism to the feature dimension,assigns different weight parameters to the feature through the attention mechanism system,and improves the accuracy of the model in the feature dimension;4)Aiming at the problem of data set imbalance,a weighted loss function is designed.During the training process,when a minority sample is incorrectly predicted,a larger penalty is given,thereby amplifying the loss of the model.Improve the problem of poor model generalization performance and low prediction accuracy caused by unbalanced data sets.Finally,through many verification experiments,the rationality and effectiveness of each module design of the Att-MDTCN model proposed in this paper is proved,and the final model prediction accuracy rate reaches 84.58%.And in a series of comparative experiments with RNN,LSTM and GRU,it is proved that the Att-MDTCN model proposed in this paper can double the prediction speed and still maintain a better prediction accuracy.