A degradation assessment method for industrial robot in flexible manufacturing systems based on multiple physics-informed neural network

Abstract

With the expansion of industrial robot applications, degradation monitoring is valuable for manufacturing systems. The article develops a framework to predict the remaining useful life (RUL) of industrial robots, which is called Physical-Informed Channel-Attention-Inception-Unet (PICAIU), it can be used to evaluate the reducer degradation. In order to improve the training efficiency and prediction accuracy, it embeds knowledge of robot dynamics and servo controller into a data-driven neural network, adding channel-attention mechanism to the Inception-Unet to fuse motion and power feature, thereby obtaining accurate motor current estimates. Based on the aging experimental data of robot harmonic gear transmission, relevant life curves have been established, which can be used to evaluate other robots. We conducted a series of experiments to discuss the robustness and generalization of the proposed method. The ablation experiments showed that the introduction of physical information constraints improved training efficiency and convergence accuracy, effectively alleviating the sample size requirement of channel self-attention mechanism. The RUL prediction error was 26% lower than the best baseline model. Under the same batch size conditions, PICAIU only need 42% training rounds of the corresponding baseline model to let fitting error be lower than 10%, saving training time and sample costs significantly. Through cross validation with two robots, when generalizing from a rich dataset to poor dataset, the prediction error is only 6.4%, which has potential prospects in real industrial scenarios with relatively few samples.