Since the industrial revolution, there has been a steady and continued effort to bring more automation and efficiency to the manufacturing process. Milling machines have long been a valuable tool to create precise parts quickly and effectively. CNC Milling machines have been the next evolution of automation in manufacturing and have allowed for complex parts to be produced quickly and with high precision.

Although CNC milling machines are able to semi-autonomously produce parts effectively, they still require significant human interaction to operate. This human interaction is especially true since many tasks are completed in open loop configuration with little to no feedback. To try to address this issue there has been significant effort in the literature to develop systems to provide feedback to the machine controller. This work is often focused on detecting anomalies such as chatter, broken tools and other conditions that will impact the surface finish or machine health.

A limitation of much of the current work is that it tends to be machine or material specific. These approaches developed in the lab often do not scale well to production as they require custom setups or complex machine dynamics to be studied. To overcome this problem, this thesis proposes a machine learning based solution that leverages deep learning to create a solution that can potentially be quickly and easily transferred to machines in production.

In this thesis we demonstrate that by using simple accelerometers mounted on the spindle of a CNC milling machine, we can create an LSTM-Autoencoder to detect anomalies such as chatter. This feat is accomplished by creating an artificial neural network that is trained on sensor data from stable cutting conditions. This network aims to reproduce the original signal with as little error as possible. If the network is provided data from a stable condition it will reconstruct the signal with little error, but if it is presented data from an anomaly condition it will reconstruct with significant error, which indicates that an anomaly is present.

In Chapter 4 we show that this approach can also be achieved by implementing what is known as transfer learning. In transfer learning we begin with a network that is trained on one source data set, and then transfer the knowledge to another target data set. We investigate under what conditions this is most feasible and demonstrate that we can train a network from a robust data set on one three-axis CNC machine and then transfer it to another three-axis CNC machine. We also demonstrate that this method works for both chatter detection and broken tool detection.

In Chapter 5, we introduce an incremental learning method based on ensemble learning. This approach takes the LSTM-Autoencoder trained previously as a strong learner and has weak learners continually learn as new data is made available. This approach is shown to have comparable results to a large network trained from scratch and improves the performance of a system trained with transfer learning.

Taking these transfer learning and incremental learning algorithms, we extend the approach to anomaly detection for five axis CNC milling machines in Chapter 6. This is accomplished by introducing a stacked ensemble learning approach by transferring the encoder from the three axis CNC anomaly detection algorithm and then combining it with an encoder and decoder that is trained from the target data. Incremental learning is then integrated by adding weak learners to this strong learner. These weak learners allow the network the ability to improve the performance of the system to be comparable to a network trained from scratch with a fraction of the data.

Lastly, we demonstrate how these approaches can be used for multi-class prediction in Chapter 6. In this chapter, we use the LSTM-Autoencoder to perform dimensionality reduction. We then use this dimensionality reduced output and apply a one-versus-all SVM classifier and Platt scaling to obtain a probabilistic prediction of the classes of interest. This approach allows us the ability to both detect and differentiate cutting with broken tools and chatter conditions.

The approaches presented in this thesis demonstrate that this proposed approach is capable of not only detecting chatter in a specific lab setting, but can potentially be used to detect multiple anomalies across a variety of machines and materials. This allows users to potentially scale these approaches to many machines quickly with minimal setup and minimal configuration.