Abstract:

Deep learning is one of the emerging techniques that shows good failure modes classification prediction results due to its flexibility in recognizing patterns from raw sensor data. However, it requires complex hyperparameter optimization, high training time, and high computational hardware resources for neural network architecture. On the other hand, classical machine learning algorithms rely heavily on domain knowledge and manual feature engineering which is not always available in the industry. Therefore, we present an alternative method that learns characteristics of multivariate raw time series data to perform failure mode classification. The method is based on the deep forest algorithm, which is composed of two main processes: multi-grained scanning and cascade forest. The multi-grained scanning process windows the data and screens the times series to generate feature vectors automatically based on class probability distribution and hence recognize patterns from data. The cascade forest uses the output of the multi-grained scanning process and creates layers of random forests to make predictions. Each layer will perform fault classification, and the number of layers will increase until the accuracy of the classification does not improve. This layer-by-layer process is similar to deep learning, where the algorithm architecture is composed of different hidden layers. The presented methodology directly works with raw data in three domains: time, frequency, and time & frequency domain. Also, the method is validated using data provided by the Prognosis and Health Management (PHM) data challenge 2022 competition for hydraulic rock drill multiple failure mode classifications. The results show that the presented methodology is faster, less complex, and more accurate than deep learning algorithms.

This article is available online here and to CALCE Consortium Members for personal review.