Deep learning-based anomaly detection using one-dimensional convolutional neural networks (1D CNN) in machine centers (MCT) and computer numerical control (CNC) machines

Introduction

Managing failures in manufacturing processes is necessary in the manufacturing industry. Manufacturing machines can cause a loss in the manufacturing process, sometimes because of their high complexity or default. An anomaly during production is not beneficial to a company because it affects its overall success. In the case of a manufacturing machine’s failure, an alternative solution should be used, such as a spare system in an emergency. This will be very costly and will require regular maintenance, even if they remain unused. Therefore, manufacturing industries seek cost-effective and efficient ways to handle machine failures. Predictive maintenance is one of the options used for this purpose; it provides a promising path to machine failures and avoids anomalies in the manufacturing process (Barlow & Hunter, 1960; Hoppenstedt et al., 2018; Saxena et al., 2008).

MCT and CNC machines are used in manufacturing processes to manufacture different metals. These machines are handled by coding on a PC via a programming application. Various kinds of commands are generated through coding to perform various operations. Multiple data types are produced from these machines, such as temperature, vibration, power, spindle, and noise. Many researchers have proved that from all kinds of data, vibration, spindle, energy, and noise data are the main reasons for machine failure (Jiang et al., 2019). Figure 1 shows a picture of an MCT machine (Kim et al., 2022).

Moreover, the trend toward machine learning produces an interrogation of the possibility of predicting manufacturing machine failure, thus detecting anomalies in the manufacturing process. Although the manufacturing industry has witnessed some technical developments to address these defects and failures, their practical application remains a challenging task for many reasons (Hoppenstedt et al., 2018; Kammerer et al., 2019). Based on these considerations, our study presents a real-world case conducted in cooperation with a manufacturing company. The company uses computer numerical control (CNC) and machine centers (MCT) to drill metals. Several defects during these manufacturing processes can be observed, reducing the devices’ efficiency. Thus, for a company, defect detection and prevention remain of utmost importance as they will increase the productivity and efficiency of its machines. However, defect detection in these machines remains a complex task that requires considerable time and effort. Thus far, many manufacturing companies have employed selected choice experts to detect anomalies based on their long-term experience. However, such experts are generally expensive and sometimes ineffective. Moreover, these experts took too much time to observe several machines. Based on sensor data collected from MCT machines, we aim to minimize manual decisions regarding defect detection and improve productivity (Feng et al., 2019; Sadaf, Athar & Azam, 2016; Schlechtendahl et al., 2015; Vodenčarević et al., 2011).

The company’s real-world data are related to an MCT machine that monitors records during the drilling of the metals. The process must be continuously monitored to reduce costly downtime. Continuous monitoring generates large amounts of sensor data related to several MCT machine components.

This study aimed to assess and predict the early failure of MCT machines. Training our proposed one-dimensional convolutional neural network (1D-CNN) base model that can use sensor-based data collected from a manufacturing machine capable of predicting defects will significantly contribute to our research. Our experimental results demonstrate the 1D CNN’s superior performance in anomaly detection for MCT machines. The model’s ability to accurately identify anomalies, even with limited training data, underscores its practical value in real-world manufacturing settings. This innovation enhances the fault diagnosis process and contributes to the broader adoption of deep learning techniques in the industrial sector. In the past, researchers have proposed different techniques to predict early defects in these machines. Still, there are some drawbacks, such as some of them using only vibration data and some of them using only spindle values in their research. Also, some of the researchers did not perform feature engineering techniques. Moreover, there is a need for comprehensive evaluation metrics that go beyond accuracy to include precision, recall, F1-score, and statistical significance tests. Considering these factors, we have tried to fill the research gap using a combination of three types of datasets obtained from the CNC and MCT machine, and we also performed features engineering techniques to enhance the performance of our proposed model. This 1D-CNN-based deep-learning (Du, 2018; Lei, Pan & Huang, 2019) model assists small manufacturing companies in early predicting anomalies in CNC and MCT machines. Furthermore, we have conducted paired t-tests to statistically evaluate the performance differences between the proposed 1D CNN model and other traditional machine learning and deep learning models. This evaluation not only determines our model’s superior performance in terms of precision, recall, and F1 score but also confirms the statistical significance of the observed improvements, thereby reinforcing the reliability and robustness of our approach.

A summary of our work is provided below:

• Applying a state-of-the-art 1D CNN deep learning model on a custom sensor dataset from a private manufacturing company’s CNC and MCT machine.

• We are pioneers in using a combination of spindle datasets, power, and vibration data, as it is easy for small companies to record spindle values in MCT and CNC machines.

• After applying different preprocessing techniques, we extracted seven features, a unique combination that improved the accuracy of our proposed model. No one else used this combination in research on manufacturing machines.

• To address the class imbalance problem, we utilized the class weight technique, which helped balance both effective and defective classes.

• We trained a state-of-the-art 1D CNN deep learning model and compared its performance with eight other machine learning classifiers, including Random Forest (RF), XGBoost, multi-layer perceptron, logistic regression (LR), K-nearest neighbor (KNN), Support vector machine (SVM), decision tree, and naïve Bayes.

• Furthermore, the proposed model’s performance was compared with other deep learning models, such as long short-term memory (LSTM) and the hybrid CNN and LSTM (CNN + LSTM) model.

• Finally, we have conducted paired t-tests to statistically evaluate the performance differences between the proposed 1D CNN model and all other models.

The remainder of this paper is organized as follows:

Related work is discussed in ‘Related Work’, and ‘Materials & Methods’ describes the methods and materials used. ‘Proposed Algorithm’ briefly explains the proposed algorithm. The results are given in ‘Results’, and finally, ‘Conclusions and Future Work’ provides conclusions and future directions.

Related Work

Various machine and deep learning algorithms have been employed with notable success in anomaly detection within manufacturing processes. However, the 1D-CNN utilized in our research presents distinct advantages over these methods, particularly in CNC and MCT machines and other manufacturing applications.

One-dimensional convolutional neural networks (1D CNNs) are particularly well-suited for analyzing sequential data, such as the time-series data generated by sensors in MCT and CNC machines. Unlike traditional machine learning methods, which often rely on handcrafted features, 1D CNNs automatically learn the temporal dependencies and patterns directly from the raw data. This capability is crucial for capturing the dynamic behavior of manufacturing processes, where anomalies may manifest as subtle temporal variations in the sensor readings (Hussain, Ali & Kim, 2022a; Yao et al., 2017).

Machine-learning algorithms detect anomalies using different datasets produced by manufacturing machines. Recently, Pittino et al. (2020) investigated anomaly detection methods in in-production manufacturing machines. They demonstrated that the nature of the available data, featuring any anomaly, was vital for algorithmic choice. In this study, automatic anomaly detection methods using statistical machine learning and control charts are developed and deployed effectively in an in-production environment. They used support vector machine-based fault diagnostics for the induction motors. Their study was based on vibration, current signals, and critical fault detection features. These features were fed into the SVM as inputs to diagnose faults. The obtained model was tested under several conditions and proven highly effective (Pittino et al., 2020). Chen et al. (2019) proposed a method for predicting transient stability in power systems using the XGBoost machine learning model. By simulating the New England 10-machine 39-bus power system with PSASP, the study generates time-domain data to train the model. The approach leverages feature selection techniques to enhance model performance and emphasizes interpretability through decision rules and feature importance scores. Their proposed XGBoost model outperforms other machine learning classifiers, including RF, DT, SVM, and neural network (NN), in accuracy and efficiency (Chen et al., 2019). Using NASA hypertext transfer protocol (HTTP) log datasets, Henriques et al. (2020) combined XGBoost and K-means algorithms to detect anomalies in the dataset. The built model can be used for forensic and compliance auditing analyses of security management (Henriques et al., 2020). Gao, Ma & Yang (2002) used classifiers to assign anomaly labels to a set of k-length windows obtained from a training dataset. The classifier can detect anomalous windows that are either ignored or assigned a defect label (Gao, Ma & Yang, 2002). For a sequence of sensor data collected from an operational aircraft system (Budalakoti et al., 2006; Budalakoti, Srivastava & Otey, 2008; Sriastava, 2005) Anomaly detection techniques were applied to handle sequence-based problem formulation and identify faulty functional runs. Selvaraj & Min (2023) retrofitted a CNC machine to monitor energy consumption and detect defects in real-time. Using data from a power meter, they extracted different features across time, frequency, and time-frequency domains. They used five machine learning classifiers for anomaly detection: DTs, RFs, k-NN, SVM, and Bagging classifiers. Their proposed model achieved high accuracy and was deployed using AWS for real-time monitoring (Selvaraj & Min, 2023).

Deep learning technology has risen quickly and achieved significant success in several domains, including object detection, intelligent robots, saliency detection, sound event detection for innovative city safety, UAV blade problem diagnostics, and parking garage sound event detection. Researchers have used deep learning methods to detect product defects and have improved productivity and quality (Yang et al., 2020). Intelligent manufacturing, a cornerstone of Industry 4.0, heavily relies on machine vision-based surface defect detection to ensure product quality. CNNs play a pivotal role in this context by providing robust and automated solutions for detecting surface defects across diverse industries, thus supporting the goals of Industry 4.0 (Khanam et al., 2024). Tao et al. research makes two contributions: The first is a multitask 1D-CNN system for wire defect detection, which combines object detection and object categorization networks. Second, they suggested that deep learning should be applied in this area first. The real-world dataset of an industrial production plant shows how well the proposed approach identifies defects (Tao et al., 2018). Kim et al. (2022) proposed an abnormal data classification method for intelligent devices in CNC machines. They proposed a hybrid method that combined singular spectrum analysis (SSA) techniques, which were very efficient in decreasing the noise in time-series data, with convolutional autoencoder (CAE), which performed very well in their time-series data (Kim et al., 2022). Jiang et al. (2022) presented a novel approach to counter-error prediction using deep learning and reinforcement. Gubernatorov et al. proposed a method to detect faults in CNC machine gearboxes using vibration data. They developed a CNN-based VGG model that distinguishes between healthy (OK) and faulty (NOK) processes during high-speed machine operations. Their proposed research integrates their model into an edge-to-cloud architecture for automated fault detection and data annotation (Gubernatorov & Gavrilenkov, 2024). Chengyang, Sitong & Wansheng (2021) proposed a method in which the axial and radial thermal errors of horizontal and vertical spindles are modeled using a deep-learning convolutional neural network (CNN). Their model blends thermocouple data with the thermal image to accurately reflect the temperature field of the spindle (Chengyang, Sitong & Wansheng, 2021). Zhao et al. (2021) proposed an improved target defect detection algorithm for steel surfaces. They improved the target detection accuracy on the NUE-DET dataset by using the R-CNN algorithm (Zhao et al., 2021).

Anomaly detection to natural language processing deep learning has been effectively used in time series modeling and prediction problems in various application domains. Recurrent neural networks (RNNs) were developed to address the limitations of conventional feed-forward neural networks for modeling sequences. However, LSTM networks have recently become popular for sequence and sequence-to-sequence learning because of the well-known vanishing gradients problem of RNN (Brunelli et al., 2019). They used the internal data of CNC machines to track errors in the feeding axis and modeled it as a nonlinear auto-regressor LSTM (Jiang et al., 2022). Masci et al. proposed a multiscale pooling technique that can detect flaws on steel surfaces and capture photos of various sizes as inputs. Their proposed method outperformed many traditional methods available in the market (Masci et al., 2013). Sharma, Chawla & Ram (2020) developed an automatic CNC program for operating multiple types of holes in CNC machines using support vector machines and restricted Boltzmann machines (RBM) with a deep belief network (DBN). Li, Bedi & Melek (2023) proposed an LSTM Autoencoder-based method to detect anomalies in CNC machines. They utilized transfer learning for this purpose. Their initial trained model could distinguish between stable and unstable cutting conditions based on vibration data obtained from the accelerometer. They introduced additional layers and re-trained their model to enhance the detection accuracy. Experimental results validate the ability of their proposed model to detect anomalies, with performance metrics showing promise for industrial application in the automated manufacturing industry (Li, Bedi & Melek, 2023).

Table 1 provides a comprehensive comparative analysis of several research studies, summarizing the machine used, dataset, model, and contribution of each study.

Materials & Methods

We followed different steps in this study. Each is explained below:

Class weight technique

Different techniques are used for data balancing, including the class weight. Using this technique, we develop a weighting algorithm to calculate the loss function. Depending on the imbalanced dataset, weights were assigned to both the majority and minority classes. A threshold value should be defined to maintain balance among different classes. This threshold value helps to increase or decrease class weights, so ultimately, it will prevent the bias of the model toward any particular class (Hussain, Ali & Kim, 2022a). The class weight is expressed in Eq. (1): (1)${\displaystyle w^i=\frac{n\text{\_}instances}{n\text{\_}classes\times n\text{\_}instancesi}}$

where,

wi = wight of each class and I = Each Class

n_instances = Total number of instances in the whole dataset

n_classes = Total number of classes

n_instancesi = Total number of rows in each class.

Table 2:

Overview of data distribution.

Class	Total data	No. of values for power	No. of values for spindle	No. of values for vibration
Effective	9,811	9,811	9,811	9,811
Defective	16,315	16,315	16,315	16,315
Total	26,126	26,126	26,126	26,126

DOI: 10.7717/peerjcs.2389/table-2

Table 3 shows an implementation of the class weight technique in our study. We calculated the class weight using Eq. (1) and assigned 1.33 weights for the minority class i.e., effective class. This gave more weight to the minority class than the majority class, i.e., defective. Resultantly, it helped to address the data imbalance issue in our dataset and improved the performance of our model.

Table 3:

Class weights.

Class	Weight
Defective (0)	0.66
Effective (1)	1.33

DOI: 10.7717/peerjcs.2389/table-3

Complete architecture of proposed methodology

A complete picture of our work is given in Fig. 3.

Proposed Algorithm

We used three different types of sensor data to develop our prediction algorithm. Different types of data are power, spindle, and vibration. In addition, we applied the other preprocessing techniques discussed in ‘Proposed Algorithm’ to make our data ready for use in our proposed algorithm. Using this algorithm, we predicted the anomaly of the CNC and MCT machines. The function and Training of our proposed model are explained by the pseudo-code given in Table 4.

Data format and measurements

We created a matrix of 24 ×30 = 720 feature vectors, where 24 denotes the total number of attributes, and 30 is the number of rows in each window. Figure 4 explains this process, later used for model training and testing while predicting the results.

To avoid robustness and overfitting of results, the entire dataset was divided into 80% for training, 10% for validation, and 10% for testing. This division was implemented for all eight machine learning and three deep learning models. This data division simplifies the analysis and ensures that our model’s performance metrics are more straightforwardly interpretable.

Results

In this section, we discuss our study’s experimental results. We compared the performance of our proposed 1D CNN model with eight machine learning models—random forest, XGB, MLP, LR, SVM, KNN, DT, and naïve Bayes—and one deep learning model, the LSTM. Additionally, we evaluated a hybrid model combining 1D CNN and LSTM (1D CNN+LSTM).

Conclusions and Future Work

While several researchers have focused on manufacturing, research on MCT and CNC machines remains sparse. This study addresses this gap by applying a 1D-CNN deep learning model for anomaly detection in MCT/CNC machines. Our results demonstrate that the 1D-CNN model effectively predicts MCT machines’ operational status and defects, outperforming eight traditional machine learning models and one deep learning model. Our model achieved the highest accuracy and outperformed a hybrid 1D CNN and LSTM model, even with a small dataset.

Although our study was limited to data collected for a single day, the 1D CNN’s strong performance suggests its potential for effective anomaly detection with limited data. Future work will use data from multiple days and varying operational conditions to further validate our findings to ensure the model’s robustness and adaptability. Exploring more advanced deep learning techniques may enhance MCT machine productivity.

Supplemental Information