5 April 2023 in IEEE Access
In the routine inspection process of railway catenary systems, the primary task is to find out the locations of various components accurately. The complex composition of the components in the catenary system and their large dimensional differences make the inspection of small components considerably difficult. Aiming at the problem of the difficulty in locating small components, a new locating method, named asymmetrically effective decoupled head-you only look once (AED-YOLO), for locating small components of the catenary has been proposed in this study. In this method, firstly, a small object detection layer has been added to improve the detection accuracy of the small-sized components such as fastener nuts and bracing wire. Secondly, to reduce missed and false detection errors of small components, the improved bidirectional feature pyramid network with high-order spatial interactions and recursive gated convolution has been used to fuse the features of different scales to further enhance the ability to detect small objects. Finally, an asymmetrically effective decoupled head has been proposed using different decoupled branches to decouple the classification and localization processes, thus further reducing the error in small-sized object classification and location. Experiments performed on the railway catenary dataset collected on-site show that the proposed localization method can efficiently improve detection accuracy, achieving a mean average precision of 93.5%. Thus, compared to the other methods, the proposed method can accurately locate small-sized components.
18 July 2024 in IEEE Transactions on Geoscience and Remote Sensing
Given the limited computing, storage, and communication transmission abilities of downhole edge devices, it is awkward to transmit massive imaging data collected by an edge device to the ground server. The predicament results in long latency times and severely degrades the efficiency of downhole state monitoring. To reduce the cost of transmitting logging image data, a low-cost quantized compressed sensing (CS) and transmission framework is proposed in this work. The method exploits the inherent sparsity of the logging image data to compress and transmit the image as a 1-D vector. By significantly reducing the amount of data that needs to be transmitted, quantized CS and transmission (QCST) achieves efficient communication of information. Compared to existing compression reconstruction methods, a hardware-deployed implementation of a neural network (NN) is presented for image compression. Furthermore, a data quantization method is designed that convert the floating-point to binary data for reduced transmission costs. Finally, an image reconstruction network is adopted to improve the quality of compressed logging images using the powerful nonlinear modeling capabilities of NN. Comprehensive experiments demonstrate that the use of CS for compression and quantized transmission, combined with NN for reconstruction, could achieve a peak-signal-to-noise ratio (PSNR) of about 20 dB at extremely low sampling rates of 0.01 (1%) on a real-field dataset. It also outperforms existing methods in terms of maximizing compression gain and reconstruction quality, providing a new solution for logging data compression and transmission. Code available at https://github.com/Qiang-Feng98/QCST .
12 August 2024 in Measurement Science and Technology
Metal couplers are susceptible to unpredictable failure and fracture under long-term high-load conditions in heavy-haul railway transportation. The current mainstream manual inspection method has the disadvantages of high subjectivity and high a priori knowledge requirements, thus not meeting the rapid analysis requirements of production companies. Therefore, in this study, an automated failure analysis method is proposed for heavy-haul coupler fractures. First, a novel image segmentation method (PermuteNet) combining a visual multilayer perceptron and a convolutional neural network is designed to segment different failure patterns of fracture surfaces. The proposed method uses two newly proposed modules—permute attention module and context attention module—to improve the network’s ability to perceive weakly differentiated objects, thereby improving the recognition ability of the model for different failure patterns. In addition, a deep supervisory function is adopted to accelerate the convergence speed of the network. Finally, the proposed image segmentation method is deployed on a computer in conjunction with a developed client application to implement a single-click detection function for coupler fracture pattern analysis. Experiments are performed using the heavy-haul coupler fracture dataset established using on-site data; the proposed segmentation method achieves a mean intersection over union of 77.8%, which is considerably higher than that of other existing methods. By using the client software, the single-click detection function of the fracture area is realized. Thus, the proposed method provides a more convenient and accurate fracture identification solution for factory inspectors and has broad application prospects.
8 May 2025 in IEEE Access
Speech Emotion Recognition (SER) technology analyses speech characteristics in human-computer interactions to understand user intent and improve interaction experience. It is widely used in the field of intelligent interaction. The significant challenge is to recognize the speech emotion of the speaker faster and more precisely. To address this issue, we propose a lightweight forward-backward independent temporal-aware causal network termed I-TCN to construct bi-directional efficient representations of causality in speech time domain sequences. Specifically, a forward-aware component constructed with dilated causal convolutions and skip connections is deployed to perform forward sequences feature modeling, which captures forward semantic information and causal relationships in the speech signal. It facilitates the prediction of emotional changes in the future. Furthermore, the backward temporal-aware module uses dilated causal convolutions to learn backward information and weighted fusion of multi-level temporal features to enhance the perception of backward emotion changes. Finally, different levels of forward-backward features are fused to refine historical-future emotion change trends and better perceive the details of emotion changes. Experimental results on six different linguistic datasets (EMODB: 95.52%; EMOVO: 92.00%; RAVDESS: 93.75%; SAVEE: 88.54%; CASIA: 94.50%; IEMOCAP: 71.47%) show that the emotion recognition capability of the proposed method is extremely competitive with state-of-the-art technologies. Meanwhile, the numerical results show that the proposed method has a good application prospect with a small number of parameters (0.21M) and low computational cost (80.72 MFOPs).
17 March 2026 in IEEE Access
Speech Emotion Recognition (SER) has a wide range of applications, as it analyzes acoustic features in speech signals to infer the speaker’s emotional state and enhance interaction experiences. For speech sequences with varying lengths and complex emotional characteristics, relying solely on fixed-scale feature information and temporal causal characteristics makes it challenging to accurately identify the speaker's emotional state. To address this issue, a multi-scale temporal causal network (MSTCN) is proposed to effectively capture multi-scale feature representations and long-range temporal dependencies in speech signals. Specifically, the multi-scale semantic feature extraction block performs semantic feature extraction at different receptive field scales for speech sequence features, capturing multi-scale semantic information within the speech sequence to establish a local-to-global feature learning pathway. Furthermore, the temporal-sensing block employs dilated causal convolutions to achieve bidirectional sequence feature modeling, capturing bidirectional semantic information and time-causality details within the speech signal to represent spatial-temporal characteristics of emotional changes. Simultaneously, a one-dimensional efficient attention (ECA) mechanism is introduced to reduce noise interference, enhancing the capture of effective speech sequence feature details. Finally, by integrating multi-layer forward-backward features, the proposed MSTCN precisely characterizes emotional change trends, achieving high-precision perception of emotional details. Experimental results on six language datasets (EMODB: 96.64%; EMOVO: 93.53%; RAVDESS: 93.13%; SAVEE: 91.25%; CASIA: 95.17%; IEMOCAP: 71.33%) demonstrate that the proposed method exhibits highly competitive emotion recognition capabilities compared to state-of-the-art techniques, indicating promising practical applications.