Research On Diagnosis Technology Of Indicator Diagram Based On Deep Learning

The working environment of pumping well is very complex in the production process.It is not only affected by the comprehensive factors of pumping unit,rod and pump,but also by various physical and environmental factors in the well,which makes the fault types of pumping units more and difficult to detect.It is the main method of fault diagnosis for pumping unit to judge the fault and locate the cause of the fault quickly by the method of diagram analysis.However,the traditional classification and recognition methods of indicator diagram use the artificial experience features,and need strong prior knowledge.Based on the strong feature extraction and learning ability of deep learning,this paper proposes the idea of self-learning indicator feature classification recognition.Through learning the local characteristics of the indicator map,the paper designs the adaptive network to realize the classification and recognition of the indicator diagram,and makes the model lightweight improved,and the robustness of the model is enhanced by the way of enhanced learning,which makes the method actually be on the embedded equipment Migration of applications becomes possible.The specific work and innovation of this paper include:Firstly,the traditional complex features of indicator diagram are analyzed and summarized,and the corresponding feature extraction and feature fusion methods are designed.Opencv is used to extract complex features of power graph.In addition to the image recognition,the most effective moment feature is identified,and the engineering parameters of displacement load are used to make up for the global limitation of image moment.The feature fusion method is used to process the features,which improves the accuracy of the indicator recognition.The effectiveness of the method is verified by classification and comparison.Secondly,in order to solve the problem of small sample in deeplearning of indicator recognition,a self encoder is designed to enhance the sample,and the data set of the indicator map is constructed.Based on the data set,the paper compares and analyzes the influence of various model parameters on the model recognition rate and convergence by using the classical network structure thinking,and designs the network model suitable for the classification and identification of indicator map.The algorithm of self encoder coding layer is also proposed to improve the convergence performance of the model.Finally,aiming at the practical application of deep learning method ofindicator recognition in embedded equipment at edge end,the lightweight and reinforcement learning scheme of the model is proposed.The paper uses mobileNet network to decompose the model in volume layer,reduce the parameter and calculation amount,and design the intensive learning scheme of embedded equipment with actor critical reinforcement learning strategy as the core.The new knowledge is studied to fine tune the model with the training knowledge reserved.The performance comparison of parameter adjustment is given to adapt to different actual needs.