Weihua Gui’s research while affiliated with Peng Cheng Laboratory and other places

Online detection of CO2 emissions can provide important data support for low-carbon operation of process industries. However, the cumbersome calibration process often limits the applications of gas sensors in field measurement. This article develops a total calibration-free gas detection method based on natural logarithm and linear convolution algorithms in TDLAS-WMS (ln-LC-WMS) and investigates its optimization in applications. Logarithm processing effectively eliminates the influence of light intensity on the harmonic, while linear convolution accurately quantifies the amplitude gain, which overcomes the problem of inaccurate filtering gain in the traditional demodulation method. Thereby, the harmonic is obtained with an accurate absolute amplitude that is positively correlated with gas concentration, and the gas concentration can be obtained without calibration. Meanwhile, to deal with the detection error caused by the mismatch between the preset and actual concentrations in the concentration inversion, an iterative process is generated to calculate the accurate gas concentration. To test the feasibility of the proposed method, three CO2 absorption lines near 2004 nm are applied for experimental verification. The experimental results indicate that to overcome the inaccurate preset concentration, within CO2 concentration of 5%, two iterations are sufficient to obtain the accurate result. Continuous measurements of a fixed CO2 concentration at different temperatures show high precision and stability, demonstrating that the proposed method is reliable and can obtain the gas concentration without additional calibration. The proposed method is expected to simplify the detection equipment and avoid calibration operations, which has good industrial application prospects.

Generative Adversarial Networks (GANs) face challenges in effectively utilizing deep features of varying resolutions and adequately fusing information characterized by typicality and multiplicity, particularly when confronted with extremely limited training samples. A multi-scale and local feature fusion based generative adversarial network (FPN LoFGAN) is proposed to address these issues. Initially, time-domain waveforms of vibration signals are transformed into time-frequency maps using the Continuous Wavelet Transform (CWT). Subsequently, feature maps are extracted through channel attention mechanisms and fused with multi-scale features via a Feature Pyramid Network (FPN). Finally, adversarial samples are generated using a multi-head attention-embedded generator, while a self-attention embedded discriminator ensures robust adversarial learning. The proposed method, applied to fault analysis and diagnosis on various planetary gearbox datasets, outperforms state-of-the-art methods, achieving superior performance regarding SSIM and FID metrics.

Key performance indicator (KPI) reflects the quality and efficiency of manufacturing operations, and KPI forecasting enables proper operations or controls in many industrial processes. However, existing KPI forecasting methods are inadequate for managing the advance prediction of KPI across multiple cycles effectively, which impedes precise and timely control in complex industrial processes. Therefore, we propose an adaptive encoder–decoder framework with partial teacher forcing strategy (PTF-ED) to enable flexible multihorizon KPI forecasting. First, we employ an encoder that processes the input time series and an attention layer to generate the context vectors. Then, we divide the measured KPIs into delayed and current time series, and design a delayed decoder and a current decoder in series to make the delayed and current time series correspond to the input time series. Especially, we propose a partial teacher forcing strategy to utilize the measured KPI efficiently and tackle the challenge of exposure bias in the current time series between training and inference phases. Moreover, we introduce a weighted multihorizon forecasting constraint in the model training loss to constrain the input–output correspondence across different sample intervals. The effectiveness of the proposed model has been validated through both a numerical simulation study and a case study in a real-world zinc flotation process.

Machine learning has been widely applied in industrial intelligent manufacturing. However, significant domain differences in data across factories make it difficult for models trained on a single factory dataset to achieve cross-factory reuse. Unsupervised Domain Adaptation is a method to address this issue, but its basic assumption is the source domain data is available. With increasing attention to data and internet security in the modern manufacturing industry, privacy protection of source data makes it unavailable. To address this challenge, we propose a contrastive learning-based secure unsupervised domain adaptation framework, which does not require source domain data and can achieve high-precision domain alignment by relying on the source domain well-trained model and the target domain unlabeled data. We conduct sufficient experimental studies on a digital recognition benchmark transfer task and a real industrial case, demonstrating that the proposed method outperforms state-of-the-art methods in terms of performance. It is worth mentioning that the proposed method can eliminate the dependence on source domain data, effectively ensuring cross-factory data privacy protection and providing new possibilities for intelligent networked collaborative manufacturing.

In fringe projection profilometry 3D measurement systems, the measurement of surfaces with high variability in reflectivity poses a challenge due to the limited dynamic range of cameras. The main solution involves using multiple exposures to modulate fringe intensity; however, it is inefficient. In this study, we introduce an attention-guided end-to-end phase calculation network to accelerate the multi-exposure structured light process for high dynamic range (HDR) measurements. We use attention modules to guide feature selection, enhancing relevant features and suppressing irrelevant features. Using the 12-step phase-shifting profilometry (PSP) as ground truth, our method accurately extracts the sine and cosine components of the fundamental frequency from a single pattern to retrieve the absolute phases. Tested on our metallic dataset requiring HDR imaging, our method achieves an absolute phase error of 0.084, close to that of the six-step PSP method (0.069), while using only 16.7% of the time. On the ceramic dataset, our method achieves 0.021 phase error, close to that of the four-step PSP (0.012). In quantitative measurements, our method achieves an accuracy of approximately 40 \upmu\text{m} μ m on standard spheres and plates. Overall, our method preserves the accuracy of multi-exposure PSP methods while significantly accelerating the 3D reconstruction process.