Deep Learning-Based Error Mitigation for Assistive Exoskeleton With Computational-Resource-Limited Platform and Edge Tensor Processing Unit

Recently we introduced a new model-based fault injection method implemented as a highly customizable Simulink block called FIBlock. It supports the injection of typical faults of essential heterogeneous components of Cyber-Physical Systems (CPS), such as sensors, computing hardware, and network. The FIBlock allows to tune a fault type and configure multiple parameters to tune error magnitude, fault activation time, and fault exposure duration. The FIBlock is able to generate various types of highly adjustable CPS faults. We demonstrated the performance of the FIBlock on a Simulink case study representing a lower-limb EXOLEGS exoskeleton, an assistive device for the elderly in everyday life. In particular, we discovered the spatial and temporal thresholds for different fault types. Upon exceeding said thresholds, the Dynamic Movement Primitives-based control system could no longer adequately compensate errors. In this paper, we proposed a new Deep Learning-based approach for system failure prevention. We employed the Long Short-Term Memory (LSTM) network for error detection and mitigation. Error detection is achieved using the prediction approach. The LSTM models are mitigating the detected errors with computed predictions only when they were subject to the imminent failure (i.e., exceeded the aforementioned thresholds). To compare our approach with previous findings, we trained two LSTM models on angular position and angular velocity signals. For evaluation, we performed fault injection experiments with varying fault effect parameters. The ‘Sensor freeze’ fault was injected into the angular position sensor, and the ‘Stuck-at 0’ fault was injected into angular velocity sensor. The presented Deep Learning-based approach prevented system failure even when the injected faults were substantially exceeding thresholds. In addition, reasoning for data access point choice has been evaluated. We compared two options: (i) the input data for LSTM is provided from the sensor output and (ii) from the controller output. In the paper, the pros and cons for both options are presented. We deployed the trained LSTM models on an Edge Tensor Processing Unit. For that, the models have been quantized, i.e. all the 32-bit floating-point numbers (such as weights and activation outputs) were converted to the nearest 8-bit fixed-point numbers and converted to the TensorFlow Lite models. The Coral USB Accelerator was coupled with a Raspberry Pi 4B for signal processing. The result proves the feasibility of the proposed method. Because the LSTM models were converted to the 8bit integer TensorFlow Lite models, it allowed firm real-time error mitigation. Furthermore, the light weight of the system and minimal power consumption allows its integration into wearable robotic systems.