Predictive Maintenance in Serial Production Using Deep Learning: Analysis of Multi-Layered Telemetry Data Streams and Maintenance History

Abstract

This article addresses the problem of predicting machine failures in a company conducting serial, multi-
variant production, comprising over one hundred workstations, including welding, spot welding,
gluing, and milling. We formulate predictive maintenance (PdM) as a binary early-warning task: at
time t, the model predicts whether a failure episode - defined as a computerized maintenance management
system (CMMS) corrective intervention causing unplanned downtime of at least 10 minutes -
will start within the next H hours (H = 4 h by default; H E {1, 4, 8, 24} h in the sensitivity analysis),
using the preceding W hours of telemetry history. The study integrated historical maintenance data,
including service requests, Mean Time Between Failures (MTBF), Mean Time To Repair (MTTR),
and failure cause codes, with current machine operation telemetry, such as vibration, temperature, current,
load, cycle times, and programmable logic controller (PLC) alarms. A complete pipeline was developed,
encompassing synchronization of multi-source data streams, feature engineering, and methods
for addressing class imbalance, such as weighted loss functions and focal loss. The analysis considered
recurrent neural network architectures - Long Short-Term Memory (LSTM) and Gated Recurrent
Unit (GRU) - as well as a one-dimensional convolutional neural network (1D-CNN). Model performance
was assessed using the Area Under the Receiver Operating Characteristic Curve (AUROC),
the Area Under the Precision-Recall Curve (AUPRC), and the F1-score, with decision-threshold adjustment
reflecting downtime costs. On the chronologically separated test set, the best-performing
LSTM model achieved AUROC = 0.924, AUPRC = 0.847, and F1-score = 0.823, indicating strong
potential for predictive maintenance in serial production.