Predictive Maintenance Method And Embedded Implementation Based On Recurrent Neural Network

Predictive Maintenance Technology(Pd M)is one of the pivotal technologies to achieve industrial modernization.It is broadly employed in various fields such as aerospace,rail transit,and mechanical equipment,etc.In the political strategies context of "Made in China 2025" and "Industry 4.0",the industrial equipment of factories matured to be more intricate,and the working environment tends to be worse.Pd M refers to a technology that anticipates the remaining useful life(RUL)of industrial equipment through sensor monitoring data of equipment and system.Through Pd M,maintenance remedies could be employed in advance for aging industrial equipment to warrant the safe operation of the factory.Embedded devices have the attributes of small size and high reliability,which made them become broadly employed in industrial control conditions.Applying Pd M technology on embedded devices can efficiently lessen the burden of cloud storage the pressure of data transmission.Maintenance expenses can be also abated.Accordingly,the study of predictive maintenance technology and its embedded implementation approach is of transcendent significance to the advancement of the industrial field.Deep learning has compelling feature extraction and data-fitting capabilities.The existing Pd M technology generally constitutes the mapping relationship between sensor data and RUL through a deep learning model.Yet,the RUL prediction approach based on deep learning has stringent demands for computing resources and storage space.If the model needs to be deployed in an embedded environment with less computing resources and less storage space,the model structure should be simplified and the model size should be reduced as much as possible without losing the prediction precision.To accomplish predictive maintenance tasks at a lower charge and higher real-time performance,this paper promotes a deep learning model for RUL prediction and proposes two compression methods based on the properties of the model.After experimental execution,the compressed model was successfully deployed to the embedded device.The main work results of this paper are as follows:1.Proposed an RUL prediction method based on a multi-layer recurrent neural network.The method extracts the time-series traits in the monitoring data through the parallel multi-layer GRU structure.Within the supervised training,the method can acquire the mapping relationship between sequence data and RUL.Aiming at the flaw of insufficient long-term memory of this model,the paper consolidates the attention mechanism with the multi-layer GRU model.The attention mechanism can designate corresponding weights by considering the importance of different data.Regarding,this paper applies the C-MAPSS data set of the NASA Failure Prediction Research Center for empirical attestation.Firstly,the characteristic parameters are selected based on the theory of information entropy.Then,the normalized data is constructed into time-series data employing the window sliding method.Finally,the model training method is formulated according to the evaluation criteria.Compared with the experimental issues with other means,it is convinced that the multi-layer GRU model has a more eminent prediction accuracy for RUL and is more fancied in model size.2.Investigated the model compression optimization method from the model pruning perspective and parameter quantization viewpoint.For the first part,the paper recommends a more advanced classification threshold pruning method based on the redundancy of the weight parameter distribution in the above model.In detail,adaptive dynamic parameter tailoring was employed on each layer model according to the distinct value of each layer of the model.In this condition,the model accuracy could be saved through retraining.On the other part,the research generated a dynamic quantification approach.According to the distribution range of collected data in the model operation,this method could formulate corresponding quantization schemes for different data.By employing 16 bit and 8bit fixed-point numbers,which are more space-saving,the model supersedes the 32-bit floating-point numbers in the primary model.The effectiveness of these two compression strategies has been proved by experiments.3.Completed the deployment of the multi-layer GRU model in embedded devices.Firstly,the model was compressed grounded on the proposed pruning and quantization methods.After compression,the model parameters were lessened by about 23 times.Next,the researchers constituted the compressed model in C language and transferred it to the microcontroller LPC1768.Finally,the model was validated on the C-MAPSS dataset.The experimental effects prove that the methods addressed in this paper could employ smaller storage space while ensuring that the two accuracy indicators of RMSE and Score are still at a high level.On resource-constrained embedded devices,these methods can accomplish high-precision RUL prediction,thereby achieving predictive maintenance tasks.