Efficient hardware implementation of interpretable machine learning based on deep neural network representations for sensor data processing

Abstract. With the rising number of machine learning and deep learning applications, the demand for implementation of those algorithms near the sensors has grown rapidly to allow efficient edge computing. Especially in sensor-based tasks like predictive maintenance and smart condition monitoring, the goal is to implement the algorithms near the data acquisition system to avoid unnecessary energy consumption caused by extensive transfer of raw data. Deep learning algorithms achieved good results in various fields of application and often allow the efficient implementation on dedicated hardware and common AI accelerators like graphic and neural processing units. However, they often need more interpretability to analyze upcoming results. For this purpose, this paper presents an approach to represent trained interpretable machine learning algorithms, consisting of a stack of feature extraction, feature selection, and classification/regression algorithms, as deep neural networks. This representation retains the interpretability but allows efficient implementation on hardware to process the acquired data directly on the sensor node. The representation is based on dissembling the inference of the trained interpretable algorithm into the basic mathematical operations to represent them with deep neural network layers. The technique to convert the trained interpretable machine learning algorithms is described in detail and applied to parts of an open-source machine learning toolbox. The accuracy, runtime, and memory requirements are investigated on four datasets, implemented on resource-limited edge hardware. The deep neural network representation reduced the runtime compared to a common Python implementation by up to 99.3 % while retaining the accuracy. Finally, a quantization method was successfully applied to interpretable machine learning algorithms, gained an additional reduction of 64.8 % in runtime, and reduced the memory requirement up to 75.6 % compared to the full precision implementation.