Fast AutoAugment-Driven Lightweight Few-Shot Industrial Inspection Image Classification with Calibrated Confidence

In modern industrial production, quality inspection and control play a crucial role in enhancing production efficiency and ensuring product reliability. Traditional inspection methods primarily rely on manual labor, which not only results in low efficiency but also introduces human error, leading to unstable inspection outcomes. With the advancement of science and technology, automated image processing technology has emerged, revolutionizing industrial testing. By leveraging computer vision and image analysis algorithms, automated image processing technology can rapidly and accurately identify and analyze defects and anomalies in images, significantly improving inspection efficiency and accuracy. This technology not only enables real-time monitoring of product quality on production lines but also provides data support for process optimization, driving industrial production toward intelligent and automated development.

Industry 4.0 is revolutionizing the world of manufacturing by merging modern digital technologies with conventional production techniques. One of the most crucial advancements is the use of robotics and automated systems for quality checking, which improves accuracy, consistency, and real-time sensitivity. In this paper, deep robotics-based quality control systems are presented with an in-depth analysis focusing on the incorporation of artificial intelligence (AI), machine vision, and sensor networks. The transformation towards Industry 4.0 has heightened the demand for smart and automated quality inspection systems that can work with speed, accuracy, and flexibility. Traditional inspection methods tend to fail in meeting the increasing complexity and amount of contemporary manufacturing processes. Robotics and automation, enabled by machine learning, computer vision, and sensor fusion, offer a revolutionary solution to such issues. This paper contribution acknowledges the manner in which these technologies are transforming industrial quality assurance in the context of Industry 4.0.

To address the challenges of high noise levels in multi-source data, easy masking of key features, and difficulty in tracing anomalies in industrial quality inspection, a multi-level feature attention fusion quality anomaly detection and tracing algorithm (MFA-QADT) is proposed. This algorithm utilizes a three-module architecture: preprocessing, feature extraction, and detection and tracing. In the preprocessing stage, density peak clustering and dynamic normalization are used to filter noise and unify dimensions. In the feature extraction stage, atrous convolution and Transformer self-attention are combined with a quality weighting mechanism to enhance the representation of key features. In the detection and tracing stage, an improved VAE model is used to identify and locate anomalies. Experimental data is collected from 58,000 quality inspection data samples (50,000 normal and 8,000 abnormal samples) from an automotive chassis parts manufacturer. Four algorithms, including control charts, support vector machines (SVMs), CNN-VAEs, and LSTM-Attention, are compared. Results show that the MFA-QADT anomaly detection accuracy reached 98.2% (a 4.3% improvement over CNN-VAE), a recall rate of 97.5% (a 5.1% improvement over LSTM-Attention), and a traceability accuracy of 95.3%. The detection latency was 0.32 seconds, and even under 30% noise, it maintained an accuracy rate of 92.1%. Its robustness and real-time performance meet the requirements of industrial quality inspection.

Measuring the roundness of a circular workpiece is a crucial issue of quality control and inspection in industry. In this area, maximum inscribed circle (MIC) and Maximum circumscribing circle (MCC), Minimum zone circle (MZC) and Least Square Circle (LSC) are four commonly used methods. In particular, MIC, MCC, and MZC, which are non-linear constrained optimization problems, have not been thoroughly discussed lately. This study proposes a roundness measuring method that applies the Particle Swarm Optimization Algorithm (PSO) to compute MIC, MCC and MZC. To facilitate the PSO process, five different PSO methods are encoded using a radius (R) and circle center (x, y) and extensively evaluated using an experimental design, in which the impact of inertia weight, maximum velocity and the number of particles on the performance of the particle swarm optimizer is analyzed. The proposed method is verified with a set of testing images and benchmarked with the GA-based (Genetic Algorithm) method (Chen, 2000). The experimental results reveal that the PSO-based method effectively solved the MIC, MCC, and MZC problems and outperforms GA-based method in both accuracy and the efficiency. As a result, several industrial applications are presented to explore the effectiveness and efficiency of the proposed method.

The term Data Fusion refers to integrating knowledge from at least two independent sources of information such that the result is more than merely the sum of all inputs. In our project, the knowledge about a given specimen comprises its acquisitions from optical 3D scans and Computed Tomography with a special focus on limited-angle artifacts. In industrial quality inspection those imaging techniques are commonly used for non-destructive testing. Additional sources of information are digital descriptions for manufacturing, or tactile measurements of the specimen. Hence, we have several representations comprising the object as a whole, each with certain shortcomings and unique insights. We strive for combining all their strengths and compensating their weaknesses in order to create an enhanced representation of the acquired object. To achieve this, the identification of correspondences in the representations is the first task. We extract a subset with prominent exterior features from each input because all acquisitions include these features. To this end, regional queries from random seeds on an enclosing hull are employed. Subsequently, the relative orientation of the original data sets is calculated based on their subsets, as those comprise the - potentially defective - areas of overlap. We consider global features such as principal components and barycenters for the alignment, since in this specific case classical point-to-point comparisons are prone to error. Our alignment scheme outperforms traditional approaches and can even be enhanced by considering limited-angle artifacts in the reconstruction process of Computed Tomography. An analysis of local gradients in the resulting volumetric representation allows to distinguish between reliable observations and defects. Lastly, tactile measurements are extremely accurate but lack a suitable 3D representation. Thus, we also present an approach for converting them in a 3D surface suiting our work flow. As a result, the respective inputs are now aligned with each other, indicate the quality of the included information, and are in compatible format to be combined in a subsequent step. The data fusion result permits more accurate metrological tasks and increases the precision of detecting flaws in production or indications of wear-out. The final step of combining the data sets is briefly presented here along with the resulting augmented representation, but in its entirety and details subject to another PhD thesis within our joint project.

Sparse random neural networks for online anomaly detection on sensor nodes

In industrial quality inspection, large amounts of data on the desired product appearance are available at the time of training, while significantly few defective samples are available. In this study, we proposed new memory-augmented adversarial autoencoders to detect and localize defects in real-time using defect-free samples alone for model training. This research was conducted by reconstructing images using an adversarial autoencoder and detection results from the Fréchet Markov distance (FMD). A threshold was determined based on the statistical characteristics of defect-free samples in the training set. Innovatively, we introduced a memory module and redesigned the reconstruction loss function to avoid the situation where the reconstruction ability is too strong or too poor, which lead to missing detection of defects. Then we proposed FMD, which can accurately measure the distance between the distribution of test samples and positive samples. Moreover, the statistics-based threshold determination method is used to meet different industrial needs. The accuracy, robustness, and computational overhead of the proposed MAA were evaluated using three datasets obtained from the production line and two benchmark datasets. The results indicated the effectiveness and ability of the proposed method to adapt to the real-time nature of industrial production.

Synthetic data has emerged as a vital tool in computer vision research, yet procedural generation using 3D computer graphics remains underexplored compared to generative adversarial networks (GANs). Our method offers greater control over generated images, making it particularly valuable for domains like industrial quality inspection, where real data is often sparse. We present a method for generating physically based rendered images of an injection-molded cup, simulating two common defects - short shot and color streak. The approach automates defect generation with variable size and severity, along with pixel-perfect segmentation masks, significantly reducing labeling effort. Synthetic data was combined with a small set of real images to train semantic segmentation models and explore domain expansion, such as inspecting parts in novel colors not represented in real-world datasets. Experiments demonstrate that the method enhances defect detection and is especially effective for domain expansion tasks, such as inspecting parts in new colors. However, challenges persist in segmenting smaller defects, underscoring the need for balanced synthetic datasets and probably also for customized loss functions.

Accurate prediction of underground pillar stability statuses (i.e., stable, unstable or failed) is of great significance for mining safety, but the limited prior data for pillar failure makes it a critical challenge. The present study proposes a novel feature fusion learning (FFL) framework to address this issue by integrating feature fusion, machine learning (ML) models, and category voting. The FFL strategy expands limited datasets through cross-reconstruction of feature vectors, converting multi-class stability prediction into a binary classification task while preserving original feature attributes. Comparative analysis of the proposed FFL strategy with two widely used data augmentation techniques (i.e., SMOTE and GAN), combined with three ML models [Random Forests (RF), LightGBM and XGBoost] demonstrates the superiority of FFL. Results reveal that FFL achieves higher accuracy (0.89), precision (0.88), recall (0.86), and F1-score (0.87) when paired with the XGBoost model, outperforming conventional methods. SHAP analysis underscores pillar stress, uniaxial compressive strength (UCS), and width-to-height ratio as dominant predictors, while feature fusion enriches dimensionality without compromising interpretability. The framework exhibits robustness across datasets with varying feature vectors and sample sizes, validating its adaptability to real-world scenarios with sparse data. This work provides a reliable and interpretable tool for pillar stability assessment, enhancing the safety of deep mining operations.

An experimental investigation was conducted to explore the feasibility of replacing the current inspection method by a portable coordinate measuring machine (CMM) for daily quality inspection in automotive industry. The experiment focusses on inspecting stamped body parts due to the significant number of inspection points that could be time consuming for inspection activities. Two inspection methods were performed, i.e. the application of checking fixture (CF) which is the current method and the second method was using a portable CMM. Both inspections were performed onto the same loose freeform surface of an outer door panel. Factors related to inspection duration, economical and data reliability were all compared to understand the practicality of replacing the current inspection method with a portable CMM. It was found that a portable CMM provide a significant shorter inspection duration, more economical and better data reliability but at the same time is unable to match the performance of conventional inspection onto freeform surface. The conclusion was made that a portable CMM alone is not sufficient to replace the current inspection method thus require further study.

With the rapid development of artificial intelligence technology, data-driven methods show great potential in the simulation and fault diagnosis of complex industrial systems. However, the application of traditional intelligent methods in aeroengine health management still faces many challenges, including the limited number of initial fault samples and the insufficient ability to capture gas path performance degradation. To solve these problems, this paper proposed a new dual cross-modal feature fusion module and a new model-agnostic meta-learning module to construct an intelligent simulation and diagnosis framework for engine gas path. This framework can effectively integrate physical model and sensor measurement data, and realize high-precision simulation under gas path performance degradation and new fault diagnosis under small sample condition. Through the example verification and comparative analysis, the results show that under the condition of performance degradation, the simulation error of gas path parameters is only 0.074%; in the case of limited initial fault samples, the fault diagnosis accuracy is as high as 99.4%, which is significantly better than the existing intelligent methods. The research shows that the proposed intelligent simulation and diagnosis framework can provide strong support for the aeroengine intelligent health management system.

Convolutional neural network-based image super-resolution methods have achieved great success in recent years. However, the huge memory and computational costs make most of the existing methods difficult to be applied to resource-constrained scenarios such as edge devices. To tackle this problem, we propose a generic, lightweight and efficient feature fusion block to replace the commonly used 1*1 convolution. In addition, we propose the enhanced shallow residual blocks to improve the super-resolution performance. By combining these two novel blocks, we design an efficient similar feature fusion network for single image super-resolution, based on the observation that cross-layer features of the same channel usually have high similarities. The similar feature fusion block utilizes similarity as a guide for feature clustering, enabling efficient and high-performance cross-layer feature fusion. On the other hand, the enhanced shallow residual blocks are used as the base feature extraction model for the network to improve super-resolution performance in conjunction with the feature fusion module. In the enhanced shallow residual blocks, we combine convolution with identity connection to maintain the similarity of cross-layer features that are fed into the similar feature fusion block. The spatial attention mechanism is also introduced to reinforce the useful spatial features. Experimental results on the benchmark datasets show that the proposed method can achieve comparable results to state-of-the-art methods with a small number of parameters.

To enable the effective fusion of time-series features extracted from vacuum circuit breakers’ (VCBs) vibration signals and to predict mechanical performance remaining useful life (RUL) under small-sample conditions, this study proposes a RUL prognosis method based on multi-temporal feature fusion through extracting features reflecting different mechanical component operational signatures. Firstly, a short-time energy algorithm is applied to vibration signals to extract state features, specifically characterizing the electromagnetic triggering unit, transmission mechanism, and spring mechanism of the VCB operating system. Subsequently, a performance degradation model is established using functional principal component analysis, with Bayesian inference dynamically updating model parameters to enable multi-feature degradation prediction. Finally, a tensor fusion method integrates degradation prediction results from three features to construct a vibration-based fused degradation indicator, achieving the RUL prognosis of VCB mechanical performance. Experimental results demonstrate that the proposed fused indicator maintains high similarity to the offline-measured reference indicator, with mean absolute error, mean square error, and root mean square error all falling below 0.1. Additionally, dynamic prediction accuracy improves as the number of operations increases. This research provides reliable decision support for online performance and RUL prognosis of VCBs in power systems.

In the era of industrial transformation and intelligent manufacturing, ensuring the surface quality of metal materials is crucial for maintaining component performance and system reliability. Although deep learning-based visual inspection has shown promise in automating defect detection, existing methods heavily rely on large-scale labeled data, which are often scarce and costly to obtain in industrial settings. Furthermore, the high inter-class similarity and variability of metal surface textures pose additional challenges to robust defect classification under low-data regimes. To address these issues, we propose an attention-guided few-shot learning framework specifically tailored for metal surface defect classification. Our method comprises two key components. First, we design a Dual-Branch Attention Module to enhance feature extraction by explicitly modeling both channel-wise dependencies and spatial saliency. This module leverages lightweight convolutional operations to highlight discriminative regions and mitigate feature degradation due to limited training data. Second, we introduce a Cross-Set Guided Attention mechanism to improve semantic alignment between support and query samples. By employing scaled dot-product attention, the model dynamically adjusts feature representations based on cross-sample correlations, thereby enabling fine-grained discrimination of visually similar defect types. Extensive experiments conducted on benchmark metal defect datasets demonstrate that our framework significantly outperforms existing few-shot learning baselines in both classification accuracy and generalization capability. The proposed method provides a practical and efficient solution for real-time industrial quality inspection in data-scarce scenarios.

Scientific studies on fishes have considerable importance to the fishery industry and marine ecosystem. Fish images are typically collected by scuba divers or autonomous underwater vehicles and then annotated manually by marine biologists. In recent years, with the rise of deep learning, the field of image classification has been greatly improved. But, fish image classification which can be considered as fine-grained problem is more challenging than common image classification. Usually, fish images are low-quality and belong to small-scale data, while classic CNN models require a large quantity of high-quality data to gain excellent results. For few-shot fish images, it's hard to get rich diversity by rotation, slight transform, and upsampling. In this paper, our method could obtain few-shot classes which are translated from large-scale sample classes, while maintaining the intra-class similarity, diversity and inter-class identifiabilityso as to improve the classification effect of fish image.