Background

This project was developed for the Finals of the 2025 National Smart Manufacturing Big Data Analysis Competition. The objective was to analyze servo control signals from CNC machines executing circular test paths. The core challenge was to distinguish between normal machine operation and machines with foundation leveling anomalies using internal sensor data.

Problem

Precision CNC machining relies on sub-micron accuracy, where even the slightest physical instability, such as a foundation leveling error, can cause systematic path deviations. These minute mechanical shifts are often masked by the high-frequency noise of servo motors, making them invisible to standard monitoring tools until defective parts are produced. The critical challenge was to isolate these specific, non-random systematic reversals from complex sensor data to prevent costly production offsets and ensure consistent machining quality.

Solution

We intentionally bypassed complex deep learning models to build a transparent and interpretable statistical pipeline. The process began by analyzing feature centroids using Cosine Similarity, which revealed that the foundation anomalies manifested as a systematic signal reversal rather than random noise. To capture these dynamics, we expanded the raw sensor streams into nine distinct time-series statistical features per channel, including spectral centroids and zero-crossing rates.

For the modeling phase, we implemented a two-stage approach designed for linear separability. First, we utilized Linear Discriminant Analysis (LDA) with regularization to project the high-dimensional feature set onto a single optimal axis, maximizing the distance between normal and abnormal states. We then fed this projected data into a Linear Support Vector Machine (SVM) to establish a robust decision boundary. This architecture not only ensured high classification accuracy but also allowed us to visualize exactly how the two states separated without the opacity of a neural network.

Fig1. Scatter plot showing the clear linear separation between State 1 and State 2

Results

Achieved 95% accuracy (0.92 F1 score) in classifying CNC servo health, successfully isolating foundation anomalies from normal operation on multi-axis signals.

Fig.2 K-Fold Cross-Validation Performance Summary (Internal Validation)

Delivered a fully reproducible inference pipeline by leveraging regularized LDA and linear SVM, significantly reducing computational overhead compared to deep learning baselines.

Certificate: 2025 National Smart Manufacturing Finals Award

2025 Intelligent Manufacturing & Big Data: CNC Anomaly Detection via Linear ML