Autonomous vehicles and mobile robots usually encounter challenges while perceiving the surrounding environment due to limited field of view. Image stitching technology stitches multiple images to construct a wider field of view, enabling these autonomous agents having a comprehensive environmental perception. While the performance of most image stitching approaches using visible images are affected by illumination changes and parallax problems, infrared cameras are robust and independent of the effect from environmental illumination. This paper proposes an unsupervised deep learning image stitching algorithm that fuses infrared images (ISA-FIR) to resolve parallax problems. The algorithm consists of two parts: a feature flow-based feature contrast algorithm that matches features extracted by a pyramid network and a multi-grid regression network in which infrared images that provide object depth information are fused to support the final stitching. In addition, most image stitching datasets nowadays contain images captured from one perspective, while real-world images collected by moving autonomous vehicles and robots usually contain significant parallax information that will affect the final stitching performance. To fill in the gap in datasets with parallax-tolerant images and validate our method, we construct a four-wheel robot and propose a novel dataset for image stitching. The dataset contains pairs of visible light images and infrared images of different scenes from different perspectives captured by the robot. Experimental results suggest that our method outperforms conventional and state-of-the-art image stitching approaches in all scenes, proving the advantages of our ISA-FIR in image stitching tasks.

Aiming at the limitation that current feature selection algorithms are only available for univariate vibration sequence, this paper proposes a multivariate vibration sequences feature selection algorithm named SVM-RFE-GA based on support vector machine recursive feature elimination algorithm (SVM-RFE) and genetic algorithm (GA). Taking a 220 kV high voltage shunt reactor as the research object, we build a mechanical fault simulation platform, set up 5 kinds of equipment states and collect multivariate vibration sequences of different equipment states at 24 sampling positions on its surface. We construct the feature pool from the time domain, frequency domain and time-frequency domain. For single vibration sequences, we rank the features and select features preliminarily by SVM-RFE. Then, the preliminarily select features are further optimized by GA algorithm to select the feature combination with the highest accuracy and the least number. The experimental result shows that the proposed method can select the common feature combination of multivariate vibration sequences, and the combination can ensure the highest fault diagnosis accuracy and the minimum number of features.

Multivariate time series anomaly detection plays a crucial role in ensuring the safety and stability of industrial systems. Currently, reconstruction-based methods rely on the powerful capability of deep learning models to extract temporal features, showing excellent performance in modeling normal patterns. However, they also tend to extract unnecessary anomalous features, leading to reconstruct of anomalies well. To address this issue, we propose a Transformer with Sinusoidal Prototype Guidance for Time Series Anomaly Detection (SPTAD). We introduce a sinusoidal prototype guidance module that uses structural priors of periodic subspaces to constrain the model to reconstruct within the normal pattern space by aggregating normal patterns based on prototype similarity. It enhances the model's representation capability in the latent space while avoiding interference from anomalous features. Moreover, SPTAD employs Transformer to capture long-term dependencies in time series, improving its ability to model global temporal features. Experiments on three real-world datasets demonstrate that SPTAD achieves an average F1 score of 95.59%, significantly outperforming six baseline methods.

Detecting anomalies in multivariate time series data collected from industrial systems is crucial for intelligent operations and maintenance. Most existing methods leverage deep neural networks to learn normal patterns for reconstructing input data, where anomalies exhibit higher reconstruction errors. However, due to the overly powerful feature extraction and generation capabilities of these models for time series, they tend to reconstruct anomalies well, leading to high false negative rates. To address this problem, we propose a novel time series Diffusion model with Self-Conditioning guidance for Anomaly Detection (DSCAD), which utilizes self-conditional control to effectively suppress the reconstruction of anomalies without affecting the original input. We introduce a self-conditioning guidance strategy that extracts coarse-grained features from intermediate results during the diffusion process as target vectors to guide the reconstruction toward generating expected normal data. Moreover, to capture long-term trends and periodic patterns in time series, we employ a Transformer as the denoising module during the reverse process. Furthermore, we introduce a novel detection criterion based on the target vectors to amplify normal-abnormal distinguishability of anomaly scores, thereby improving the detection performance. Extensive experiments conducted on five publicly available datasets demonstrate that DSCAD achieves an average F1 score of 95.23 %, outperforming other state-of-the-art methods.

Chatter is a common state in milling and turning, which will reduce the machining quality of parts. For taking adequate measures to avoid chatter, chatter detection is necessary. However, chatter recognition still has many difficulties for both turning and milling. Due to the material removal effect, the chatter frequency will change. The existing frequency band extraction methods, such as ensemble empirical mode decomposition (EEMD) and wavelet packet transform, can not fully reflect the chatter information. As a result, a novel online chatter recognition method is proposed based on spectrum characteristics, which can be applied to the chatter recognition for both turning and milling under the combination of multiple cutting parameters black simultaneously. Firstly, we carried out the hammering experiment to obtain the frequency response function and modal frequency of the machining system before and after machining. The signal distribution characteristics of turning and milling signals are analyzed. Four features sensitive to chatter are extracted, including spectral standard deviation, frequency spectral expectation (FSE), spectral skewness, and relative power spectral entropy (RPE). Taking these features as inputs, a chatter recognition model is established based on an extreme gradient boosting (XGBoost) algorithm. Finally, we conducted the turning and milling tests. The experimental results show that the proposed chatter recognition model can be effective under different cutting parameters for turning and milling. Besides, results also show the chatter recognition model has extraterritorial applicability.

Increasing the generalizability of intelligent diagnostic models amidst data distribution shifts is receiving growing attention. Nevertheless, current domain generalization methods primarily enhance fault diagnosis in variable working conditions or machines. Due to the lack of prior knowledge to determine which features are task-unrelated and which features are task-related, existing methods typically learn coupled features. Facing industrial diagnosis scenarios across bearings and artificial-to-real faults, coupled features induce false correlations that limit the model's generalizability. To address this challenge, this paper proposes a decoupled interpretable robust domain generalization network (DIRNet) to enhance model generalizability by interpretably transferring fault-related components. First, this paper constructs a neural basis function decoupling module to disentangle the signal into fault-related and fault-unrelated basis functions. Second, a dynamic Shapley pruning network is proposed to dynamically prune the fault-unrelated neural basis functions, achieving the generalization of fault-related basis functions. Third, we introduce a loss function that relies on interpretable basis function selection to enhance the expressive capability of the basis function decoupling module. Experiments on a self-collected industrial distributed fault bearing case and two laboratory cases are carried out. The results demonstrate that DIRNet can obtain generalizable fault-related components to effectively deal with industrial across bearings and artificial-to-real fault diagnosis scenarios compared to the previous methods.

Photoelectric smoke detectors are the most cost-effective devices for very early warning fire alarms. However, due to the different light intensity response values of different kinds of fire smoke and interference from interferential aerosols, they have a high false-alarm rate, which limits their popularity in Chinese homes. To address these issues, an embedded spatial–temporal convolutional neural network (EST-CNN) model is proposed for real fire smoke identification and aerosol (fire smoke and interferential aerosols) classification. The EST-CNN consists of three modules, including information fusion, scattering feature extraction, and aerosol classification. Moreover, a two-dimensional spatial–temporal scattering (2D-TS) matrix is designed to fuse the scattered light intensities in different channels and adjacent time slices, which is the output of the information fusion module and the input for the scattering feature extraction module. The EST-CNN is trained and tested with experimental data measured on an established fire test platform using the developed dual-wavelength dual-angle photoelectric smoke detector. The optimal network parameters were selected through extensive experiments, resulting in an average classification accuracy of 98.96% for different aerosols, with only 67 kB network parameters. The experimental results demonstrate the feasibility of installing the designed EST-CNN model directly in existing commercial photoelectric smoke detectors to realize aerosol classification.

Gas pipeline leakage can cause adverse social impacts such as waste of resources and environmental pollution. Existing leakage detection methods ignore the practical problem that aging or damaged sensors can affect detection accuracy. In this paper, we propose a method for gas pipeline leakage detection using multichannel acoustic signals. We consider the signals that acquired by damaged acoustic sensors as outliers. First, Local outlier factor is used to calculate the outlier degree of each channel signal, and channels with outliers are removed. Dynamic time warping barycenter averaging is then used to integrate the remaining channel signals into one composite signal. Next, we extract the features of the composite signal in the time domain, frequency domain, and time frequency domain. Finally, One-class support vector machine is used to determine whether a gas pipeline leakage occurs. In the experiment, we validated the effectiveness of the proposed method on multichannel acoustic signals, which were collected at compressor stations operating in PipeChina. Experimental results show that the proposed method has a high recognition accuracy. The dataset is composed of the acoustic signals produced by the real operation of the pipeline, indicating that the proposed method has a good practical application value.

We have obtained the polarization data cube of the VRO 42.05.01 supernova remnant at 1240 MHz using the Five-hundred-meter Aperture Spherical radio Telescope (FAST). Three-dimensional Faraday Synthesis is applied to the FAST data to derive the Faraday depth spectrum. The peak Faraday depth map shows a large area of enhanced foreground RM of 60 rad m-2 extending along the remnant's "wing" section, which coincides with a large-scale HI shell at -20 km/s. The two depolarization patches within the "wing" region with RM of 97 rad m-2 and 55 rad m-2 coincide with two HI structures in the HI shell. Faraday screen model fitting on the Canadian Galactic Plane Survey (CGPS) 1420 MHz full-scale polarization data reveals a distance of 0.7-0.8d_SNR in front of the SNR with enhanced regular magnetic field there. The highly piled-up magnetic field indicates that the HI shell at -20 km/s could originate from an old evolved SNR.

Source datasets are selected with high similarity to target datasets to improve the effect of transfer learning. The low similarity between the source and target datasets may lead to negative transfer. This paper proposes a similarity-based time series source dataset selection method for transfer learning. First, to reduce the complexity of the operation, we use Dynamic time-warped barycentric averaging to obtain the prototype signals of each dataset and convert the dataset similarity calculation into the similarity calculation between the prototype signals. The dynamic time warping (DTW) algorithm, commonly used to calculate time-series signal similarity, does not consider the effect of signal complexity and the amount of data the dataset contains on the signal similarity calculation. This article makes improvements to the above issues. Experiments are carried out on 20-time series datasets archived by the University of California Riverside. Experimental results show that the average transfer effect of source datasets selected by the proposed method is better than the commonly used dataset similarity measurement method based on DTW.

Gas-insulated switchgear (GIS) is one of the most important power devices in the power system. Once a fault occurs, it is difficult to determine and locate the fault due to its fully enclosed nature, which may lead to greater accidents and huge losses. Many GIS anomaly detection methods based on vibration and deep learning (DL) have been developed in recent years. However, these DL-based methods only utilize one single sampling point data on GIS, while not all faults are reflected at the points currently in use due to the signal attenuation. In this paper, an embedded position variational auto-encoder (EP-VAE) model for anomaly detection of GIS devices is presented, which can model multipoint vibration signals. We fuse the position information and the corresponding one dimension vibration signal to the two dimension grayscale images by Gram matrix. Then, the normal grayscale images are feed into the EP-VAE model, which is optimized by KL divergence and reconstruction error jointly. In the test phase, when one of these two losses is higher than the normal images, it is identified as an abnormal fault. We use the vibration data from the actual operation of GIS to carry out the anomaly detection experiments. The results show that our model EP-VAE has a detection accuracy of 93.12% for normal vibration grayscale images, and an abnormality detection accuracy of 87.94%.

Anomaly detection of gas-insulated switchgear is very important to ensure the reliability of power systems. We propose an anomaly detection algorithm for gas-insulated switchgear based on transfer learning. This algorithm includes the feature extraction and one-class classifier training stages. Firstly, we design a descriptiveness loss function based on cross entropy and a compactness loss function based on variance. they formed a composite loss function to optimize the feature extraction network. The feature extraction network is initialized by a pre-trained model and fine-tuned in batches using reference and target data. The target data is further processed into feature data using a fine-tuned feature extraction network, and input to the one-class classifier for training and testing. The experimental results on the gas-insulated switchgear anomaly detection dataset show that the proposed anomaly detection algorithm has a significant improvement over the current popular anomaly detection algorithms.

We propose the measurement method for aerosol Sauter mean diameter based on the optical scattering response of the combined volume-surface area of particles to reduce the error caused by inconsistent integration in existing measurements.

Convolution neural network (CNN)-based fault diagnosis methods have been widely adopted to obtain representative features and used to classify fault modes due to their prominent feature extraction capability. However, a large number of labeled samples are required to support the algorithm of CNNs, and, in the case of a limited amount of labeled samples, this may lead to overfitting. In this article, a novel ResNet-based method is developed to achieve fault diagnoses for machines with very few samples. To be specific, data transformation combinations (DTCs) are designed based on mutual information. It is worth noting that the selected DTC, which can complete the training process of the 1-D ResNet quickly without increasing the amount of training data, can be randomly used for any batch training data. Meanwhile, a self-supervised learning method called 1-D SimCLR is adopted to obtain an effective feature encoder, which can be optimized with very few unlabeled samples. Then, a fault diagnosis model named DTC-SimCLR is constructed by combining the selected data transformation combination, the obtained feature encoder and a fully-connected layer-based classifier. In DTC-SimCLR, the parameters of the feature encoder are fixed, and the classifier is trained with very few labeled samples. Two machine fault datasets from a cutting tooth and a bearing are conducted to evaluate the performance of DTC-SimCLR. Testing results show that DTC-SimCLR has superior performance and diagnostic accuracy with very few samples.

The severe doxorubicin (DOXO) side effect of cardiomyopathy limits it clinical application as an effective anticancer drug. Although Ca2+ overload was postulated as one of the mechanisms for this toxicity, its role was, however, disputable in terms of the contractile dysfunction. In this work, the dynamics of the intracellular Ca2+ signal were optically mapped in a Langendorff guinea pig heart. We found that DOXO treatment: (1) Delayed the activation of the Ca2+ signal. With the reference time set at the peak of the action potential (AP), the time lag between the peak of the Ca2+ signal and AP (Ca-AP-Lag) was significantly prolonged. (2) Slowed down the intracellular Ca2+ releasing and sequestering process. Both the maximum rising (MRV) and falling (MFV) velocity of the Ca2+ signal were decreased. (3) Shortened the duration of the Ca2+ signal in one cycle of Ca2+ oscillation. The duration of the Ca2+ signal at 50% amplitude (CaD50) was significantly shortened. These results suggested a reduced level of intracellular Ca2+ after DOXO treatment. Furthermore, we found that the effect of tachypacing was similar to that of DOXO, and, interestingly, DOXO exerted contradictory effects on the tachypaced hearts: it shortened the Ca-AP-Lag, accelerated the MRV and MFV, and prolonged the CaD50. We, therefore, concluded that DOXO had a different effect on intracellular Ca2+. It caused Ca2+ underload in hearts with sinus rhythm; this might relate to the contractile dysfunction in DOXO cardiomyopathy. It led to Ca2+ overload in the tachypaced hearts, which might contribute to the Ca2+-overload-related toxicity.

The leakage of dangerous gas at the gas station may cause explosions and other serious accidents, threatening the safety of workers. This paper designs and implements a gas leakage locating device based on ultrasonic technique. The device can communicate with an intelligent inspection robot which alarms the gas leakage in real time and tracks the leaking sources. This paper first builds a hardware platform, and then realizes a four-channel signal based leak location algorithm by combining decision tree and neural network algorithm of smoothing pre-treatment. Finally, each software module of the gas leak location device is constructed. Testing of the locating capability of the device was completed in practical working environment.

The abnormal operating conditions of pipeline equipment in natural gas compressor station, such as gas pipeline leakage or equipment failure, lead to negative social impacts. To solve the current problem of untimely and inaccurate manual maintenance, this study proposes a 24/7 real-time method for monitoring the operation of pipeline equipment in natural gas compressor station based on a spiral microphone array. First, the Linearly Constrained Minimum Variance (LCMV) beamformer is used to enhance the signal in the target angle direction and suppress the interference in other directions. Then we train the support vector machine (SVM) model based on the station’s Mel-Frequency Cepstral Coefficients (MFCC) feature dataset for diagnosis. The results show that the method is of great feasibility and reliability.

Exploring the law of unsafe behavior of operators is helpful to prevent accidents. At present, research on industry operation mainly focuses on what factors affect the operation safety and how these factors affect operation safety, and most of them are for a certain type of industry operation. Operation behavior of operators is analyzed for multiple types of work. By studying subject and object of behavior, the law of unsafe behavior of operators is obtained. Firstly, the salient object detection model is used to separate and extract behavior subject and object. Then, the optical flow is used to obtain motion feature of subject and object, and shape feature of subject is obtained based on a saliency graph. Through analysis of obtained features, it is found that for industry operation behavior with simple actions, gentle speed, and small range of motion, and the target speed variation features can be used to distinguish safe behavior and unsafe behavior. This study can help to find the unsafe behaviors of operators in time, prevent accidents, and provide a reference for identification of unsafe behaviors of operators.

High voltage shunt reactor is an important equipment of power transmission systems. The accurate assessment of their operating status and the timely and correct diagnosis of faults and defects concern the operation safety of the entire grid. Health assessment of high voltage shunt reactors based on vibration signal, which can be used to characterize the hidden troubles of it, is a topic widely studied in deep learning and fault diagnosis. A large number of samples are needed to train the deep learning model, but it is not easy to acquire enough fault samples in the actual scene. In this paper, we utilize a Deep Convolutional Generative Adversarial Networks (DCGAN) to generate synthetic fault samples and enlarge the fault dataset to train the Convolution Neural Network (CNN) fault detection model. Results reveal that the performance through the CNN model can be improved by 3% with the synthetic samples generated by DCGAN, which is better than that of traditional Synthetic Minority Oversampling Technique (SMOTE) algorithm.

GIS is one of most important power devices in power systems. However, much research focused on the partial discharge rather than the mechanical fault. In this paper, mechanical fault detection based on vibration signals were used. Firstly, vibration signals were collected on the surface of GIS equipment. Then vibration images were generated from the vibration signal sequences. A vibration images CNN (VI-CNN) mechanical fault detection model was used in this paper based on the vibration images. Finally, the VI-CNN model performed well during the test. The performance reveals that the method can be used in GIS monitoring system to detect the mechanical state of GIS.