Isolation Forests and Deep Autoencoders for Industrial Screw Tightening Anomaly Detection

Abstract

Within the context of Industry 4.0, quality assessment procedures using data-driven techniques are becoming more critical due to the generation of massive amounts of production data. In this paper, we address the detection of abnormal screw tightening processes, which is a key industrial task. Since labeling is costly, requiring a manual effort, we focus on unsupervised detection approaches. In particular, we assume a computationally light low-dimensional problem formulation based on angle–torque pairs. Our work is focused on two unsupervised machine learning (ML) algorithms: isolation forest (IForest) and a deep learning autoencoder (AE). Several computational experiments were held by assuming distinct datasets and a realistic rolling window evaluation procedure. First, we compared the two ML algorithms with two other methods, a local outlier factor method and a supervised Random Forest, on older data related with two production days collected in November 2020. Since competitive results were obtained, during a second stage, we further compared the AE and IForest methods by adopting a more recent and larger dataset (from February to March 2021, totaling 26.9 million observations and related to three distinct assembled products). Both anomaly detection methods obtained an excellent quality class discrimination (higher than 90%) under a realistic rolling window with several training and testing updates. Turning to the computational effort, the AE is much lighter than the IForest for training (around 2.7 times faster) and inference (requiring 3.0 times less computation). This AE property is valuable within this industrial domain since it tends to generate big data. Finally, using the anomaly detection estimates, we developed an interactive visualization tool that provides explainable artificial intelligence (XAI) knowledge for the human operators, helping them to better identify the angle–torque regions associated with screw tightening failures.

An Empirical Study on Anomaly Detection Algorithms for Extremely Imbalanced Datasets

“…The discrimination performance is given by the AUC = 1 0 ROCdK. It should be noted that the AUC measure is a popular measure that contains two main advantages [34,35]: quality values are not affected by the imbalanced rate of the target class; and quality values are easy to interpret (50% -performance of a random classifier; 70% -good; 80% -very good; 90% -excellent; and 100% -ideal classifier). After executing the 5-fold cross-validation, the AUC results are aggregated by computing the median value (which is less sensitive to outliers when compared withe the average value).…”

“…Let (L I , L 1 , ..., L H , L O ) denote the structure of a dense (fully connected) Deep FeedForward Network (DFFN) with the layer node sizes, where L I and L O represent the input and output layer sizes and H is the number of hidden layers. The proposed AE is based on an architecture that previously obtained high quality anomaly detection results in a industrial anomaly detection task [26,34,35]. It assumes L I = L O , a symmetrical encoder and decoder structure (e.g., L 1 = L O−1 ) and the popular ReLu activation function is used by all hidden neural units with the exception of the output layer, which assumes a linear activation function.…”

AutoOC: Automated multi-objective design of deep autoencoders and one-class classifiers using grammatical evolution

“…Besides the ANN structure, deep learning architectures include a large number of additional hyperparameters. In order to reduce the search space, using modeling knowledge from previous OCC works (e.g., [39,13]) we fixed some choices, such as the usage of the MAE measure as the loss function for both AE and VAE and usage of Batch Normalization layers for AE. We also restricted the search space for some hyperparameters.…”

“…Also known as unary classification, OCC can be viewed as a subclass of unsupervised learning, where the Machine Learning (ML) model only learns using training examples from a single class [8,9]. This type of learning is valuable in diverse real-world scenarios where labeled data is non-existent, infeasible, or difficult (e.g., requiring a costly and slow manual class assignment), such as fraud detection [10], cybersecurity [11], predictive maintenance [12] or industrial quality assessment [13].…”

Categorical Attribute traNsformation Environment (CANE): A python module for categorical to numeric data preprocessing

“…CANE was used in several scientific studies as a means of reducing the number of inputs (after the categorical to numeric transform) to feed predictive ML models. Diverse real-world applications were addressed, including mobile performance marketing [5,9], Industry 4.0 anomaly detection [4,10] and quality prediction [8,11]. As a consequence of the adoption of CANE python package, the aforementioned studies have observed a reduced computational effort when preprocessing categorical data.…”