DTCEncoder: A Swiss Army Knife Architecture for DTC Exploration, Prediction, Search and Model Interpretation

Abstract

Diagnostic Trouble Code (DTC) events, produced in vehicles, assist in knowing the occurrence of faults in different modules and can be used for predictive maintenance by detecting patterns. While performing Exploratory Data Analysis (EDA) or correlating specific DTC events is an easy task, searching for patterns in long multivariate DTC sequences can be very challenging. Instead of performing analysis for individual DTCs, a self-supervised approach using a Long Short-Term Memory (LSTM) network was introduced recently to perform the next DTC prediction. Despite its merits, such an approach is not interpretable for engineers who need to understand the decisionmaking process of the model. In this paper, we introduce the DTCEncoder, a recurrent neural network that incorporates Gated Recurrent Units (GRU) and an attention mechanism to encode DTC sequences into low dimensional representations, and that serves as a unified approach to (i) efficiently represent multivariate event sequences and predict the next event, (ii) interpret what the model learns and what it uses for the next prediction, and (iii) perform efficient semantic search for individual DTCs and DTC sequences.

Explainable Predictive Maintenance: A Survey of Current Methods, Challenges and Opportunities

“…Articles Attention (VI-D1) [56], [58], [64], [78], [133] Fuzzy (VI-B1) [52], [53], [126], [141] Knowledge-based (VI-B2) [103], [121], [142] Sparse Networks (VI-A1) [46], [104], [112] Interpretable Filters (VI-B3) [45], [49], [ [145] Rule-based Interpretations (VI-B8) [146] fault trees for domestic heaters. C4.5 is used to learn the failure thresholds of the sensor data.…”

“…Xia et al[58] and Hafeez et al[64] tackled interpretable fault diagnosis in two separate ways. Xia et al looked at hierarchical attention by grouping the features by systems and subsystems.…”