Smartphone-Based Vibration Monitoring for Drill Bit Wear Classification Using Lightweight CNN Fusion Strategies

Abstract This study investigates the feasibility of using a consumer-grade smartphone as a low-cost, fixtureless sensing node to monitor and classify drilling bit wear. Tri-axial vibration signals were acquired using the embedded MEMS accelerometer of a smartphone placed freely on the machine table, representing a practical worst-case deployment scenario without rigid sensor mounting. This work evaluates whether severely bandwidth-limited, sub-Nyquist vibration measurements can still retain diagnostic information for drilling-state identification and progressive drill wear. Raw vibration recordings were segmented, transformed into compact two-dimensional pseudo-RGB representations, and classified using lightweight convolutional neural network architectures. We systematically compared three multi-axis fusion strategies---Data-Level, Feature-Level, and Decision-Level fusion--- using LeNet-5 and Mini-MobileNetV2 models, with Support Vector Machine and Random Forest classifiers as traditional baselines. The results show that Mini-MobileNetV2 with Data-Level fusion achieved the best drilling-phase detection performance, whereas Feature-Level fusion provided the strongest tool-wear classification performance. Robustness analyses further demonstrated that Feature-Level and Decision-Level fusion were more resilient to simulated sensor saturation and angular misalignment than Data-Level fusion. Overall, the findings support the potential of smartphone-based vibration sensing as an accessible proof-of-concept approach for rapid, non-invasive tool condition monitoring, while highlighting the need for broader validation using additional machine-tool platforms, smartphone devices, and reference-grade sensing systems.