Research On CNC Real Time Thermal Compensation System By GPHA Deep Neural Network Based On STM32-DSP

A novel thermal error modeling method is proposed based on Deep Network with stacked RBMs(DN) and GA-PSO-based Hybrid Algorithm(GPHA) in this paper. Thermal error becomes the largest contributor to manufacturing error of precision machining. Compensation is an effective way to reduce thermal error. The machine tool thermal error variation is largely influenced by specific operating parameters adopted, Four key temperature points of G200 drilling machine are firstly obtained to construct temperature features sets; then, we propose a method for temperature pattern recognition with deep learning method using Hinton and Salakhutdinov's Deep Auto Encoder(DAE), which are probabilistic generative neural networks composed of Restricted Boltzmann Machine(RBM). Before training RBM using defined energy function, a 3-layer deep network of RBM is adopted to extract the feature of input space of temperature data initially, and then Grid Search method is utilized to determine the optimized hidden layer neurons and learning parameter during training period. We conduct a visualization procedure to further symbolize feature map, using connection weights between hidden layer and visible layer of RBMs, thus temperature pattern recognition is fulfilled by specific feature map by visualization process. Once the classification is effected, the deformation is predicted by respective back propagation neural network(BPNN), BPNN is advantageous for its greatest capability in mapping complex relationship from input to output. But the local optimum makes it susceptible to initial weight and threshold. A simulated experiment shows that conventional optimization method such as GA, PSO are insufficient to solve this problem. To solve this problem, we present Tabu-Search Particle Swarm Optimization(TSPSO) by incorporating Tabu list and penalty function to reconstruct the fitness function. Based on the TBPSO and Genetic Algorithm, a novel Hybrid Algorithm(GPHA) is also presented. GPHA considers both local information and global information in searching solution space. emulation result shows that GPHA is more effective for ascertaining initial parameters of BPNN. Ultimately, a real-time compensation system is constructed to test the performance of this hybrid algorithm. The compensation system recognizes three temperature patterns and the experiment shows that the prediction is largely enhanced compared with single BPNN model; additionally, the proposed optimization method achieves higher prediction accuracy than other ones. Finally the real-time compensation results show that diameter error of work piece reduces from 26 um to 3.5um, which demonstrates the effectiveness of the proposed thermal compensation method...