Lightweight vs Foundation Models for Industrial Fault Detection: A Comparative Study for Edge Deployment on Industrial Sensor Signals

Abstract Predictive maintenance based on industrial sensor signals is an important application of machine learning. Performing inference close to the monitored machine can reduce communication requirements and response time, although the available computational resources depend on the sensing architecture. In parallel, large pre-trained foundation models for time series promise competitive accuracy with limited task-specific training. This thesis investigates whether such models provide a favourable accuracy-to-cost trade-off for industrial fault detection, compared with compact models designed for the task. The comparison includes a foundation-model surrogate, two lightweight signal-informed methods, and an unsupervised baseline. The same protocol is applied to four datasets. The methods are evaluated in terms of classification accuracy, latency, model size, and computational complexity. The unsupervised baseline assesses whether the labelled tasks can be explained by natural clustering alone. Following the recommendations of Roque et al. [1], all results are reported, including the datasets on which the proposed method does not achieve the highest accuracy. No method achieves the highest accuracy on every dataset. The lightweight methods, however, require substantially fewer resources in all experiments. The reservoir-based method also provides the highest accuracy on the two principal bearing-vibration benchmarks. The foundation-model surrogate is never the most accurate method and exhibits the largest variation across datasets. On CWRU, the reservoir-based model is more than three orders of magnitude smaller and approximately sixteen times faster than the foundation-model surrogate, while also achieving higher accuracy. A sensitivity analysis over thirty reservoir initialisations yields an accuracy of 98.98% ±0.21 percentage points on CWRU, indicating that the reported result is not attributable to a favourable random seed. Quantising the readout weights to 8-bit integers reduces their size fourfold, to 9.78kB, with a loss of 0.22 percentage points in accuracy. Additional feature and error analyses identify the conditions under which the lightweight methods perform less favourably. Overall, the results indicate that signal-informed models provide a more suitable accuracy-to-cost trade-off than the evaluated transformer-based surrogate for fault detection on resource-constrained edge platforms.