Condition Monitoring of Machine Tool Ball Screw Feed Drives Through Signal Analysis and Artificial Intelligence

This thesis is set in the context of the large volume of work directed to improving the overall equipment effectiveness (OEE) of manufacturing machines. Of the three OEE factors, performance receives much research attention since it provides simple metrics of parts per hour produced. However, availability and quality, which are the other two factors, can play an equally important role. In the past, high availability has been achieved by time-based or preventive maintenance techniques, which can be expensive and wasteful due to needless repairs and replacement of useful parts. This research aims to develop a cost-effective strategy for machine tool maintenance that improves availability and accuracy by adopting a condition-based or predictive maintenance approach.

The approaches under investigation use both machine learning and deep learning techniques to analyse continuous time-series signals to assess a machine tool's condition. For this research, the focus is on applying the techniques to the ball screw assembly of axis feed drives. This is one of the most common machine tool parts whose degradation can affect its availability and positional accuracy. This research data is obtained from experiments on a gantry-type machine tool with two ball screws, where one is good, and the other is worn.

In the machine learning approach, wavelet and fast Fourier transforms are employed for data processing on time-series vibration readings before extracting useful features for model training. These extracted features consistently show better accuracy across several machine learning algorithms than those obtained via classical methods. Deep learning is then investigated as an alternative method of analysing time-series data. The chosen approach utilises a pre-trained deep learning neural network based on convolution, which had been successfully used to learn from image files. The novelty in this research arises from the use of convolution-based deep learning on time series data. It does this by the conversion of the vibration signals to image files. The method of converting time-series data streams to images relevant for this analysis has been established and verified.

Test results show that the wavelet and fast Fourier transform (FFT) features used in the machine learning approach can outperform the statistical features in classifying the condition of the ball screw. With at least a 98 % accuracy across the examined machine learning networks compared to a range of 87 % (support vector machine) to 96 % (k nearest neighbour). On the other hand, the deep learning technique can achieve at least 98 % or 100 % accuracy when trained with raw and processed data, respectively. The deep learning approach has the advantage of requiring less data processing and better accuracy than the machine learning approach.

This research project will contribute to the manufacturing industry by improving the overall equipment effectiveness at a low cost. Furthermore, it can lead to real-time online condition monitoring with less overhead since there is no need for a data processing stage. This research's natural progression would be applying this approach to other parts of a machine tool or equipment. Furthermore, investigating and identifying specific faults and their progression would lead to a more sophisticated system for widespread deployment.