Automatic Synthesis of Recurrent Neurons for Imitation Learning From CNC Machine Operators

Analyzing time series data in industrial settings demands domain knowledge and computer science expertise to develop effective algorithms. AutoML approaches aim to automate this process, reducing human bias and improving accuracy and cost-effectiveness. This article applies an evolutionary algorithm to synthesize recurrent neurons optimized for specific datasets. This adds another layer to the AutoML framework, targeting the internal structure of neurons. We developed an imitation learning control system for an industry CNC machine to enhance operators' productivity. We specifically examine two recorded operator actions: adjusting the engagement rates for linear feed rate and spindle velocity. We compare the performance of our evolved neurons with support vector machine and four well-established neural network models commonly used for time series data: simple recurrent neural networks, long-short-term-memory, independently recurrent neural networks, and transformers. The results demonstrate that the neurons evolved via the evolutionary approach exhibit lower syntactic complexity than LSTMs and achieve lower error rates than other networks. They yield error rates 270% lower for the first operation action, while the error rates are 20% lower for the second action. We also show that our evolutionary algorithm is capable of creating skip-connections and gating mechanisms adapted to the specific characteristics of our dataset.