Anomaly detection and diagnosis in manufacturing processes : a comprehensive framework

Abstract

We call this the two-step neural network algorithm, which is capable of training and predicting on multiple machines. The first step of this algorithm is an unsupervised autoencoder trained to capture the normal behavior of multiple machines. The output of this autoencoder and also the latent variables learned by the autoencoder are then used as input features to a new step-two neural network. This next neural network trains on the normal-abnormal state of previous wafers. The state information is trained in a semi-supervised manner by using both cpk performance data and raw sensor training data. Both fully-connected and one-dimensional convolutional network architectures have been experimented and compared. A comparison of performance is also presented between PCA and neural network-based anomaly detection.

Our results prove our intuition that cluster quality can be used for finding parameters for unsupervised anomaly detection performance. We also show a positive correlation between cluster quality and metrology performance. The results also show that all three methods have their own strengths and weaknesses with different algorithms performing the best in different aspects. Overall, the results show the efficacy of the framework at detecting anomalies and providing the diagnosis for the various anomalies within the data set.

The anomaly detection framework performance is demonstrated by experimenting on wafer trace data from the semiconductor manufacturing process. This data is generated by three different process nodes or types, with data from 13 different chambers.

Complex manufacturing processes require constant monitoring to mitigate potential mechanical failures and other inconsistencies to achieve higher yields and lower costs. These processes comprise many recipe steps performed on multiple tools. This paper presents a comprehensive framework called Anomaly Identification for Process Monitoring Online Framework (AIPMOF) for automated anomaly detection and monitoring of manufacturing processes. This is an adaptable multivariate online system that could detect anomalies on the fly and it also provides the associated diagnosis of potential root causes.

The framework has five components: feature transformation, anomaly detection and performance prediction, diagnosis, validation, and reporting. Each component is replaceable and adaptable with a different method. The feature transformation component converts raw time-series sensor data into a tensor of step-sensor properties. The online anomaly detection component is responsible for the actual machine learning modeling.

Our initial approach to anomaly detection is based on a moving window principal component analysis (PCA) algorithm, which does not require performance data with labels such as the metrology. This moving window approach allows real-time detection of parameter shifts, drifts, and mechanical faults. The diagnosis sub-component provides top contributing features for the diagnosis of faults and validates these top features with statistical significance tests of shifts.

The validation component provides both a supervised and unsupervised validation scheme for a comprehensive evaluation of performance. We present a cluster quality-based unsupervised metric that can be used for unsupervised validation of anomaly detection performance. This metric can also be used for unsupervised hyper-parameter training. We also present a supervised validation scheme that uses the Process Capability Index (Cpk) for automatically classifying batches of wafers as normal or abnormal based on the limited amount of metrology measurements.

The anomaly detection with the moving window PCA algorithm can only capture linear relationships. We propose an advanced algorithm based on autoencoders for nonlinear anomaly detection. This is also an unsupervised and real-time anomaly detection method. We call it the moving autoencoder anomaly detection algorithm (MOAAD). It uses a similar structure to the PCA algorithm. It replaces the linear normal model used in the PCA-based algorithm, with a non-linear autoencoder. Both of these approaches are capable of diagnosis based on top influential feature ranking and they can also tune their hyper-parameters using the cluster quality metric.

The anomaly detection with these two approaches have some limitations, including the fact that a separate model must be trained per machine, and it does not leverage performance information from metrology. Therefore, we also propose an alternative approach for anomaly detection based on an advanced neural network algorithm.

The final component in our framework is a diagnosis-reporting tool based on R Notebooks. This is a dynamic report generator that provides diagnosis, validation and other various statistics in a user-friendly manner. R Notebook combines both codes and results into a single document that allows users to generate their own further analyses if needed.