Learning machine monitoring models from sparse and noisy sensor data annotations

Present-day requirements on efficiency and quality of manufacturing processes necessitate constant monitoring of machine tools and machining processes. Although automated, sensor-based machine monitoring techniques are described in literature, real-world production shops still exhibit a high degree of human intervention, which tends to be both expensive and error-prone. This is due to three challenges that such machine monitoring systems are confronted with which this thesis will address:
First of all, long-term deployable systems require robust predictive models.
The models need to generalize across user-initiated adjustments of process parameters and changes of workpiece types, such that trained models still match the distribution of newly incoming test data by independence of covariate shift among training and testing data distributions. The variance in sensor data is often more influenced by such parameter adjustments and workpiece changes than by actual anomalies. This dominance of covariate shift over class-discriminative information in sensor data is challenging. Secondly, most performant predictive models are (semi-)supervised, requiring large sets of labeled sensor data. Annotation of anomalous data is expensive and comes with a severe risk of machine damages when deliberately provoking anomalies. Finally, high-performant models rely on high memory resources, long training and model execution times or specific hardware for training (e.g., GPUs). These requirements conflict with the desire of companies for industrially robust and compact embedded sensor systems and short model execution times (allowing timely alerts of and quick responses to potentially critical anomalies). Evaluation directly on embedded sensor systems allows for an increased data security, compactness allows to retrospectively equip machines with these sensor systems.
The first part of the thesis is concerned with defining features tailor-made for specific machine monitoring tasks that generalize across covariate shift. To that end, domain expertise about machine and process characteristics is included in custom-built preprocessing models for segmentation of sensor data and tracking of discrete frequency components. The second part of the thesis focuses on low-cost annotation and detection of “in-the-wild” recorded anomalies. A prototypical evaluation system was developed specifically for harsh industrial environments and deployed there. The system enables data recording, on-system evaluation and reporting of potential anomalies, both supporting operators in decision processes and allowing for annotation of sensor data by operators’ feedback to anomaly propositions. Evaluations of this prototypical system and the resulting data suggested that involved anomaly detection models were overly simple, so more sophisticated unsupervised neural anomaly detection models were tested.
In addition, two semi-supervised extensions trained with expert labels and automatically generated, therefore weak labels were compared. Both unsupervised and semi-supervised neural anomaly detectors prove to be well-suited, generalizing across several weeks of data exhibiting covariate shift. All presented methods respect constraints imposed by embedded systems used for machine monitoring and the need of timely responses to anomalies.