2. Anomaly Detection System for Automatic Defective Products’ Inspection
Related literature demonstrates that machine learning, data mining, mathematical methods, and deep learning approaches provide users with a convenient way to analyze data from manufacturing sites for anomaly detection (
Table 1).
Table 1 lists the current approaches for anomaly detection. Some anomaly detection-related research used machine-learning approaches. For example, Winters et al.
[4] investigated control charts and autoregressive models to detect anomalies for predictive maintenance; Erdmann
[5] used an unsupervised approach to recognize anomalous sensor data for predictive maintenance; Minarini
[6] proposed a framework for detecting anomalies in log-based predictive maintenance; Alaoui-Belghiti et al.
[7] proposed unsupervised online methods using optimal transport to recognize the anomalies for predictive maintenance; Farbiz et al.
[8] proposed cognitive analytics with unsupervised learning to predict the machine status for equipment health maintenance; Carrasco et al.
[9] proposed a framework for temporal unsupervised algorithm evaluation, which detects time-series analytics early. In addition, data mining approaches such as correlation analysis
[10] have been applied to equipment maintenance. For instance, Perini
[11] used the Markov Chain and autoencoder for off-road vehicle maintenance.
Table 1.
The main anomaly detection-related approaches in predictive maintenance.
Recently, anomaly detection using images has attracted many researchers
[19]. Therefore, this
res
tudy earch applied image data in data analytics for anomaly detection. First,
weresearchers used the Publish or Perish software to survey emerging studies’ time trends and disciplinary distribution and related literature were retrieved. Notably, this
studyresearch included keywords from top-cited articles to present the current technological trends: “deep learning,” “unsupervised learning”, “semi-unsupervised learning”, “supervised learning”, and “restructure”. These keywords are high-frequency for anomaly detection-related topics concerned with image issues. Then, the retrieved papers were sorted based on their average citations per year and total citations.
Table 2 lists the main methods identified.
Table 2.
The main anomaly detection approaches from 2017 to 2022 that used images.
Most findings have demonstrated that convolutional neural network (CNN) and image process approaches are the core approaches that drive research associated with anomaly detection, especially in the medical, manufacturing, and transportation fields. The result of
ourthe literature retrieval also showed that deep learning approaches, such as the ImageNet dataset, cancer diagnosis, and defective product detection, exhibited high accuracies in some cases. As a result, deep learning for anomaly detection has gained massive popularity in anomaly detection-related research. Of these deep learning approaches, CNN can learn directly from data. Previously, Haselmann et al.
[22] used a deep CNN to recognize an anomalous surface. In their study, the CNN framework extracted the features of the cell image through a self-learning capability. Similarly, while Xu et al.
[23] used a hierarchical CNN to detect anomalous chest X-ray images, Nguyen et al.
[25] used a deep CNN to restructure images and detect the anomalous region of magnetic resonance imaging scans. Khan et al.
[24] also used a deep CNN to recognize anomalous spectrograms. Accordingly, generative adversarial networks (GANs) are another deep-learning method that uses an ensemble neural network model (e.g., CNN) for automatically discovering and learning the regularities or patterns in input data. GANs are also usually used in data augmentation. In previous studies, while Deecke
[20] used the GAN in visual inspection for anomaly detection, Berg et al.
[35] used it to identify anomalous contaminated image data. Zhou et al.
[21] also used the GAN to identify anomalous retinal optical coherence tomography images. The above studies show that deep learning approaches achieve state-of-the-art performance in image anomaly detection using images. However, for the expected reason, the original image data have noise. For example, Chithirala et al.
[36] indicated that image data with noise may influence the deep learning analysis result. Therefore, the original image dataset must be preprocessed. A study reported that while the filters used in image processing can enhance the image or edge features
[37], they also effectively remove noise and blurred areas in images. Hence, reducing the noise in an image is the main preprocessing stage before being imported into the deep learning model for training
[38]. It has also been reported that while performing image denoising, sufficiently retaining the features in the image is necessary, which is a critical aspect during preprocessing.
Some studies have proposed image processing technologies: mathematical methods or neuron networks (NNs) to analyze image data for identifying an anomaly. For example, Zhang et al.
[28] proposed a three-stage tensor decomposition and divided it into three steps for detecting anomalies from hyperspectral images. Similarly, while Ayhan et al.
[29] proposed a two-step alignment approach for multispectral image anomaly detection, Cohen and Hoshen
[39] used correspondences based on a multiresolution feature pyramid in the sub-image detection. In addition, Mishra et al.
[40] applied the transformer method in image localization and anomaly detection. However, Mishra et al.
[41] proposed a deep reconstruction-based pyramidal approach to exact image features for anomaly detection. Likewise, although Zhuang et al.
[30] proposed a robust hyperspectral image denoiser to process its anomaly detection, Müller et al.
[27] used NN for feature extraction. Alternatively, another study employed the support vector machine and Gaussian mixture models to recognize anomalous images Müller et al.
[27]. Accordingly, while Cozzolino and Verdoliva proposed an autoencoder method for recognizing the anomalous-spliced image region, Li et al.
[33] proposed a Gaussian distribution model estimation method to identify the anomalous image descriptors in traffic videos. Furthermore, although Verdoj and Grangetto
[31] used the graph Fourier transform to enhance the efficacy of the Reed–Xiaoli detector for medical image anomaly detection, Wang et al.
[34] used the discrete probability model and deep autoregressive module for image anomaly detection. Lastly, Vojir et al.
[32] also proposed a reconstruction module to identify unknown objects for autonomous driving anomaly detection.