https://www.mdu.se/

This paper presents an interpretable machinelearning model for anomaly detection in door locks using torque data. The model aims to replace the human tactile sense in the quality control process, reducing repetitive tasks and improving reliability. The model achieved an accuracy of 96%, however, to gain social acceptance and operators' trust, interpretability of the model is crucial. The purpose of this study was to evaluate anapproach that can improve interpretability of anomalousclassifications obtained from an anomaly detection model. Weevaluate four instance-based counterfactual explanators, three of which, employ optimization techniques and one uses, a less complex, weighted nearest neighbor approach, which serve as ourbaseline. The former approaches, leverage a latent representation of the data, using a weighted principal component analysis, improving plausibility of the counter factual explanations andreduces computational cost. The explanations are presentedtogether with the 5-50-95th percentile range of the training data, acting as a frame of reference to improve interpretability. All approaches successfully presented valid and plausible counterfactual explanations. However, instance-based approachesemploying optimization techniques yielded explanations withgreater similarity to the observations and was therefore concluded to be preferable despite the higher execution times (4-16s) compared to the baseline approach (0.1s). The findings of this study hold significant value for the lock industry and can potentially be extended to other industrial settings using timeseries data, serving as a valuable point of departure for further research.

Historically, cylinder locks’ quality has been tested manually by human operators after full assembly. The frequency and the characteristics of the testing procedure for these locks wear the operators’ wrists and lead to varying results of the quality control. The consistency in the quality control is an important factor for the expected lifetime of the locks which is why the industry seeks an automated solution. This study evaluates how consistently the operators can classify a collection of locks, using their tactile sense, compared to a more objective approach, using torque measurements and Machine Learning (ML). These locks were deliberately chosen because they are prone to get inconsistent classifications, which means that there is no ground truth of how to classify them. The ML algorithms were therefore evaluated with two different labeling approaches, one based on the results from the operators, using their tactile sense to classify into ‘working’ or ‘faulty’ locks, and a second approach by letting an unsupervised learner create two clusters of the data which were then labeled by an expert using visual inspection of the torque diagrams. The results show that an ML-solution, trained with the second approach, can classify mechanical anomalies, based on torque data, more consistently compared to operators, using their tactile sense. These findings are a crucial milestone for the further development of a fully automated test procedure that has the potential to increase the reliability of the quality control and remove an injury-prone task from the operators.

Artificial intelligence in manufacturing systems is currently most used for quality control and predictive maintenance. In the lock industry, quality control of final assembled cylinder lock is still done by hand, wearing out the operators' wrists and introducing subjectivity which negatively affects reliability. Studies have shown that quality control can be automated using machine-learning to analyse torque measurements from the locks. The resulting performance of the approach depends on the dimensionality and size of the training dataset but unfortunately, the process of gathering data can be expensive so the amount collected data should therefore be minimized with respect to an acceptable performance measure. The dimensionality can be reduced with a method called Principal Component Analysis and the training dataset size can be estimated by repeated testing of the algorithms with smaller datasets of different sizes, which then can be used to extrapolate the expected performance for larger datasets. The purpose of this study is to evaluate the state-of-the-art methods to predict and minimize the needed sample size for commonly used machine-learning algorithms to reach an acceptable anomaly detection accuracy using torque measurements from locks. The results show that the learning curve with the best fit to the training data does not always give the best predictions. Instead, performance depends on the amount of data used to create the curve and the particular machine-learning algorithm used. Overall, the exponential and power-law functions gave the most reliable predictions and the use of principal component analysis greatly reduced the learning effort for the machine-learning algorithms. With torque measurements from 50-150 locks, we predicted a detection accuracy of over 95% while the current method of using the human tactile sense gives only 16% accuracy.