Deep learning has recently resulted in remarkable performance improvements in machine fault diagnosis using only raw input vibration signals with no signal preprocessing. However, research on machine fault diagnosis using deep learning has primarily focused on model architectures, even though optimizers and their hyperparameters used for training can have a significant impact on model performance. This paper presents extensive benchmarking results on the tuning of optimizer hyperparameters using various combinations of datasets, convolutional neural network (CNN) models, and optimizers with varying batch sizes. First, we set the hyperparameter search space and then trained the models using hyperparameters sampled from a quasi-random distribution. Subsequently, we refined the search space based on the results of the first step and finally evaluated model performances using noise-free and noisy data. The results showed that the learning rate and momentum factor, which determine training speed, substantially affected the model’s accuracy. We also discovered that the impacts of batch size and model training speed on model performance were highly correlated; large batch sizes led to higher performances at higher learning rates or momentum factors. Conversely, model performances tended to be high for small batch sizes at lower learning rates or momentum factors. In addition, regarding the growing attention to on-device artificial intelligence (AI) solutions, we assessed the accuracy and computational efficiency of candidate models. A CNN with training interference (TICNN) was the most efficient model in terms of computational efficiency and robustness against noise among the benchmarked candidate models.
A reverse vending machine motivates citizens to bring recyclable waste by rewarding them, which is a viable solution to increase the recycling rate. Reverse vending machines generally use near-infrared sensors, barcode sensors, or cameras to classify recycling resources. However, sensor-based reverse vending machines suffer from a high configuration cost and the limited scope of target objects, and conventional single image-based reverse vending machines usually make erroneous predictions about intentional fraud objects. This paper proposes a dual image-based convolutional neural network ensemble model to address these problems. For this purpose, we first created a prototype reverse vending machine and constructed an image dataset containing two cross-sections of objects, top and front view. Then, we chose convolutional neural network models widely used in image classification as the candidates for building an accurate and lightweight ensemble model. Considering the size and classification performance of candidates, we constructed the best-fit ensemble combination and evaluated its classification performance. The final ensemble model showed a classification accuracy higher than 95% for all target classes, including fraud objects. This result proves that our approach achieves better robustness against intentional fraud objects than single image-based models and thus can broaden the scope for target resources. The measurement results on lightweight embedded platforms also demonstrated that our model provides a short inference time that is enough to facilitate the real-time execution of reverse vending machines based on low-cost edge artificial intelligence devices.
In modern complex mechanical systems, machine faults typically occur in multiple components simultaneously, and the domain of collected sensor data changes continuously due to variations in operating conditions. Deep learning-based fault diagnosis approaches have recently been enhanced to address these real-world industrial challenges. Comprehensive labeled data covering compound fault scenarios and multi-domain conditions are crucial for exploring these issues. However, existing multi-domain datasets focus on a limited range of operating conditions, such as motor rotating speeds and loads. This limits their applicability to real-world industrial scenarios. To bridge this gap, we present a novel multi-domain dataset that incorporates these basic conditions and extends to various bearing types and compound machine faults. The deep groove ball bearing, the cylindrical roller bearing, and the tapered roller bearing were utilized to provide data that reflect diverse mechanical interactions between the shaft and the bearing. Vibration data were collected using a USB digital accelerometer at two sampling rates and six rotating speeds, encompassing three single bearing faults, seven single rotating component faults, and 21 compound faults of the bearing and rotating component. Additionally, the dataset provides spectrograms of vibration data using short-time Fourier transform (STFT) for data-driven analysis with a 2-D input. This dataset encompasses more complex compound fault and domain shift problems than those presented in conventional public vibration datasets, thereby aiding researchers in studying intelligent fault diagnosis methods based on deep learning.
26 August 2024 in IEEE Transactions on Instrumentation and Measurement
Machine fault diagnosis is important in modern industries, such as smart manufacturing, because machine failure can increase maintenance costs and safety risks. Although deep learning-based fault diagnosis approaches have succeeded remarkably in recent years, they often encounter challenges in accurately diagnosing machine faults in noisy environments using resource-constrained edge devices in real time. We propose FRFconv-TDSNet, a lightweight, noise-robust convolutional neural network (CNN) that prioritizes global feature extraction and integration between deep learning and vibration statistics. The key components of FRFconv-TDSNet are a full-receptive-field (FRF) CNN and a time domain statistics (TDS) extractor. A fundamental element of the FRF CNN is the FRF convolution block, which has a kernel size that is the same as the length of the input data; therefore, the FRF CNN can extract high-level features throughout the input data range. We ensure that the model is highly robust to noise while suppressing the increase in the computational costs of FRF convolution using small-channel depthwise separable convolutions and a downsampling block. The TDS values capture the necessary characteristics indicating machine conditions, such as the signal magnitude. We further improve the model’s noise robustness by concatenating the four TDS values calculated from raw vibration signals with the FRF CNN output. The evaluation results demonstrate the superiority of our model over ten comparison models in terms of accuracy for heavily noisy data and inference time on an edge device. Furthermore, visual interpretation of the model shows how FRF convolution and TDS integration contribute to the robustness of our model to data noise.
The characteristics of an earthquake can be derived by estimating the source geometries of the earthquake using parameter inversion that minimizes the L2 norm of residuals between the measured and the synthetic displacement calculated from a dislocation model. Estimating source geometries in a dislocation model has been regarded as solving a nonlinear inverse problem. To avoid local minima and describe uncertainties, the Monte-Carlo restarts are often used to solve the problem, assuming the initial parameter search space provided by seismological studies. Since search space size significantly affects the accuracy and execution time of this procedure, faulty initial search space from seismological studies may adversely affect the accuracy of the results and the computation time. Besides, many source parameters describing physical faults lead to bad data visualization. In this paper, we propose a new machine learning-based search space reduction algorithm to overcome these challenges. This paper assumes a rectangular dislocation model, i.e., the Okada model, to calculate the surface deformation mathematically. As for the geodetic measurement of three-dimensional (3D) surface deformation, we used the stacking interferometric synthetic aperture radar (InSAR) and the multiple-aperture SAR interferometry (MAI). We define a wide initial search space and perform the Monte-Carlo restarts to collect the data points with root-mean-square error (RMSE) between measured and modeled displacement. Then, the principal component analysis (PCA) and the -means clustering are used to project data points with low RMSE in the 2D latent space preserving the variance of original data as much as possible and extract clusters of data with similar locations and RMSE to each other. Finally, we reduce the parameter search space using the cluster with the lowest mean RMSE. The evaluation results illustrate that our approach achieves 55.1~98.1% reductions in search space size and 60~80.5% reductions in 95% confidence interval size for all source parameters compared with the conventional method. It was also observed that the reduced search space significantly saves the computational burden of solving the nonlinear least square problem.
26 May 2025 in IEEE Transactions on Instrumentation and Measurement
Failure detection in rotating machines is critical for reducing maintenance costs, economic losses, and safety-related accidents. Deep learning has become a robust and powerful solution for machine fault diagnosis, but it faces two major challenges: scarcity of labeled faulty data and variability of machine operating conditions. To address these challenges, we propose a novel hybrid machine anomaly detection methodology that integrates an innovative pretext task-based self-supervised learning framework with vibration statistics. First, we designed a pretext task to classify the rotating speeds and trained an auxiliary classifier on it. This auxiliary classifier effectively extracts fault-discriminative features related to rotating speed, an operating condition closely associated with various machine faults. Moreover, the proposed pretext task is applicable to the fault diagnosis of general rotating machines because the rotating speed can be easily obtained using a Hall sensor. Second, a hybrid feature was generated by combining the auxiliary classifier outputs with six statistical features derived from raw vibration signals. These statistical features augment the fault-discriminative ability of the auxiliary classifier outputs by incorporating the magnitude and asymmetry of the signal, which the auxiliary classifier alone cannot capture. The evaluation results demonstrated that our approach significantly outperformed 11 state-of-the-art anomaly detection methods on three different vibration datasets encompassing multiple operating conditions and a real-world run-to-failure scenario. In addition, our method is capable of performing real-time inference on an edge device platform, demonstrating its applicability to practical industrial setups. Furthermore, feature visualization analysis highlighted how the proposed hybrid feature enables a more distinct separation between faulty and healthy data samples than that achieved with statistical features or auxiliary classifier outputs alone.