LSTM Network-Based Complex Industrial Processes Operating Performance Assessment Method And Application

Comprehensive,accurate and detailed cognition of complex industrial processes operating performance is a higher requirement for the development of artificial intelligence in modern industry.For most industrial processes with highly nonlinear and dynamic,it is difficult to establish a reliable operating performance assessment model.In this thesis,three operating performance assessment methods based on long shortterm memory(LSTM)network and non-optimal traceability methods are proposed.The details of the research are as follows.(1)To address the problem that complex industrial processes have complex characteristics such as nonlinearity,dynamic and large inertia,which make it difficult to accurately perceive the process operating performance,a supervised slow feature analysis-based LSTM network(SSFALSTM)is proposed for complex industries operating performance assessment.The method utilizes the LSTM network to enhance the model’s ability to perceive the dynamic behavior of process,and utilize slow feature constraints to filter out fast-changing short-term noise and disturbances,forcing the network to selectively learn slow-changing features that reflect the intrinsic dynamics of the process,removing redundant information while enhancing the interpretability of the network learning results.In addition,the comprehensive economic indexes information is utilized to force the network to focus on information that is relevant to the comprehensive economic indexes during the learning.Further,cascaded performance identification model constructs the complete operating performance assessment framework.For non-optimal operating performance,an autoencoder-LSTM(AE-LSTM)model based on sparse optimization is proposed to trace the non-optimal factors,locate the main causal variables,and guide the operators’ regulation strategies.Finally,the effectiveness of the proposed method is validated on the dense medium coal preparation process.(2)The shallow network has limited learning capability,which makes it difficult to mine deep operating performance feature in complex industrial process data,leaving the complex operating performance feature representation insufficiently accurate.Thus,the SSFALSTM is extended to deep learning in order to enable the network to perceive the dynamic information of the process more deeply.However,it is difficult to comprehensively perceive the process operating performance by relying only on a single dynamic feature,so an industrial processes operating performance assessment method based on static-dynamic feature fusion is proposed.The method stacks the SSFALSTM vertically to deeply mine the dynamic behavior of the process,while utilizing residual connections to map the static information of the data directly to the upper layer of the network to fuse with the dynamic information,forcing the network to focus on both static and dynamic features of the process at a time and efficiently use the implicit information of the data,so that the model can get more reliable assessment results.Finally,the effectiveness of the proposed method is validated on the dense medium coal preparation process.(3)The excellent performance of neural networks depends on the effective training of sufficient high-quality data,but actual industrial processes are affected by production conditions,environmental disturbances,and uncertain disturbances,and often cannot obtain adequate quality data in a limited period of time.To this end,this thesis proposes an operating performance assessment method based on graph convolution networkLSTM(GCN-LSTM).Firstly,the graph is constructed by mining the correlation between data using the self-attention mechanism and defining it as the priori knowledge,and then the GCN is utilized to combine data and knowledge for graph learning to discover the topological structure between data,while the LSTM network is utilized to sense the dynamic of the process and find the intrinsic operation mode and change rule of the process.Further,the topological information learned by the GCN is adaptively weighted fused with the dynamic information to obtain comprehensive features as subsequent inputs.Finally cascade the trained softmax classifier to identify the current process operating performance.Finally,the effectiveness of the proposed method is validated on the dense medium coal preparation process.