Tianjin Vocational Institute

With the development of edge computing, AI algorithms are deployed on edge computing servers to reduce latency and protect privacy. The advancement of AI technology and the enhancement of computing capabilities of edge computing servers enable their application in more industries. Currently, in the aspect of factory safety production control, there exist problems such as insufficient intelligence and waste of human resources. To address these issues, this article designs a factory safety production control system based on AI technology. The system consists of four layers: perception layer, data layer, AI algorithm application layer, and user display management layer. It supports safety helmet detection, safety belt detection, human body detection, hot work operation detection, and left-behind object detection, featuring high levels of intelligence, security, and ease of use. It realizes the automatic identification and alarm of various safety hazards in industrial production environments. The project is targeted at a scale of approximately 1,000 cameras in the enterprise's production base, and an AI server is added in the monitoring room to perform intelligent recognition functions related to safety.

Active suicidal ideation and passive suicidal ideation are distinct but related constructs, each carrying significant clinical implications. However, their real-time predictors remain underexplored in intensive longitudinal studies. This ecological momentary assessment (EMA) study investigated the associations of thwarted belongingness (TB), perceived burdensomeness (PB), and hopelessness with changes in both types of ideation at 2-4 h intervals, and whether a history of suicide attempts moderates these relationships. Thirty-one college students (23 females; mean age = 19.83 ± 1.86 years) with current suicidal ideation participated, 17 of whom (54.84%) had a history of suicide attempts. Participants completed EMA assessments five times daily for one week. Data were analyzed using mixed linear models. TB (β = 0.05, p = 0.007) and hopelessness (β = 0.09, p = 0.006) predicted next-moment active ideation, but these associations became insignificant after all covariates were considered. PB was a robust predictor (β = 0.04, p = 0.032) of subsequent passive ideation, even after accounting for TB, hopelessness, and prior ideation. Notably, hopelessness predicted subsequent active ideation (β = 0.12, p = 0.034) only in participants with a history of suicide attempts. These findings underscore the distinct predictive roles of PB, TB, and hopelessness, highlighting the importance of considering suicide attempt history in risk assessment for suicidal ideation.

The combination of acoustics measurement and robotics technology has opened a new door in the field of fault monitoring. The use of robots can flexibly accommodate changing sound field environments in industries. The necessity of finding the best acquisition location was demonstrated through analysis of sound field radiation and comparison of multiple locations signals. Multiple parameters were fused to assist robot motion to further determine the best distance, height, and angle. The localization accuracy and resolution at the best location found are superior to those at other locations, which are within 1% and 8%, respectively. The addition of the sound field imaging function could visual interpretation of the spatial location of the fault source. This research has extensive application scenarios and demands in modern factories.

Auto-PLD is the automatic design algorithm framework of machine learning pipeline for the whole-process data analysis scenario. First, define a machine learning pipeline with five stages and can support the processing of continuous and discrete type features separately. Then, the automated pipeline design problem is decomposed into two sub problems: structure search and hyper parameter optimization, and an algorithm combining reinforcement learning and Bayesian optimization is proposed to alternately optimize the two subproblems. Finally, in order to improve the efficiency of automatic pipeline design, two parallelized pipeline construction methods are further proposed. Experimental results show that Auto-PLD outperforms auto-sklearn with most datasets. Moreover, with the increase of computing nodes, the parallelized Auto-PLD can further improve the pipeline building performance. The Auto-PLD-random method performed at 1, 4, and 8 hours respectively: 19, 22, 23, 27, 26, and 25. Automated Pipeline Design with Q-learning (Auto-PLD-Q) methods performed at 1, 4, and 8 hours: 18, 20, 26; 23, 27, and 27. The Auto-PLD-DeepC method performed at 1, 4, and 8 hours, respectively: 21,23,27; 22, 26, and 26. Auto-PLD-PG methods performed at 1, 4, and 8 hours, respectively: 19, 21, 27; 26, 26, and 25. Auto-LLE is an automated machine learning algorithm framework for lifelong learning scenarios. For the classification task based on concept drift and data imbalance, an algorithm based on weighted ensemble learning of adaptive model was proposed. Concept types are divided into “long-term concepts” and “short-term concepts,” and different types of concepts are handled separately using incremental learners and adaptive weight update methods, and automatically capture the concept drift and improve the model prediction performance. Based on Auto-PLD and Auto-LLE, we design and implement a system that supports both automated pipeline design and automated lifelong learning. In system design, high system accessibility and scalability are obtained by designing easy high-level programming interface and pluggable module integration. In task type, the common data analysis tasks such as classification, regression, and clustering are supported.

This paper proposes an algorithm of traffic sign recognition based on double attention fusion. By using multiscale feature extraction module, this algorithm integrates spatial and channel attention mechanism to enhance key features, and cooperates with light weight classifier to reduce computation cost. Experiments were carried out on a German Traffic Sign Recognition Benchmark (GTSRB) and a Belgian Traffic Sign Dataset (BTS), using template matching algorithms such as SIFT, YOLO and SSD to compare. It is shown that the precision of GTSRB data set is 98.7%, 3.2% higher than ResNet, and F1 value reaches 98.5%; for complex scene of BTS data set, the recall rate reaches 94.5%, mean average accuracy (mAP) reaches 92.3%, so it is obviously superior to traditional algorithms and deep learning algorithm. The proposed algorithm keeps high recognition precision in rainy days, nights and other special scenes. The reasoning speed reaches 45 frames/second to meet the real time demand of new energy vehicles. This algorithm can effectively improve traffic sign recognition performance, provide reliable technology support for new energy vehicle intelligent driving.

This paper presents an enhanced Multi-Layer Perceptron (MLP) model with Dropout regularization for credit scoring on imbalanced financial datasets. The proposed model introduces three key innovations: an adaptive Dropout rate mechanism that adjusts regularization based on layer-wise feature importance, a balanced mini-batch sampling strategy for handling class imbalance ($12.3 \%$ default rate), and an optimized network architecture with skip connections. Experimental results on a comprehensive dataset of $\mathbf{8 5, 0 0 0}$ records demonstrate superior performance with an AUC-ROC of 0.892 and F1-score of 0.836, representing a 4.7% improvement over baseline methods. The model achieves $\mathbf{9 2 \%}$ precision in high-risk identification while reducing computational requirements by $35 \%$. The architecture incorporates SHAP values and LIME for model interpretability, revealing payment history and debt utilization ratio as the strongest predictors. Our comprehensive evaluation across different economic cycles validates the model’s effectiveness in real-world financial applications, particularly in identifying minority class samples while maintaining interpretability and computational efficiency. The implementation demonstrates remarkable resilience to input noise and feature corruption, with performance variations remaining within acceptable bounds under stress testing.

The effectiveness of a training support program of individually tailored physiotherapeutic scoliosis-specific exercises (PSSEs) delivered via a digital care system is unclear. To determine the effectiveness of a digital care program in which patients receive PSSE supervision and guidance via the Healbone Intelligent Rehabilitation System and educational videos compared with a conventional intervention involving 3 treatment sessions conducted by qualified physiotherapists at outpatient clinics or specialized orthopedic institutions and unsupervised home-based PSSE training. This randomized clinical trial recruited patients aged 9 to 17 years from Peking Union Medical College Hospital in Beijing, China, who had adolescent idiopathic scoliosis with a primary curve Cobb angle of 10° or greater and skeletal immaturity (as determined via the Risser grading system). The trial was conducted from June 1, 2023, to August 10, 2024. Patients were randomly assigned to a digital care (DC) group, in which each patient received fully remote and home-based PSSE training supported by a digital care system, or a usual care group, in which each patient completed 3 treatment sessions conducted by qualified physiotherapists at outpatient clinics or specialized orthopedic institutions and home-based PSSE training without supervision. The primary outcome was a change in the Cobb angle of the major curve between baseline and month 6 of the intervention. Primary and secondary outcomes were evaluated using an independent samples t test. The proportion of patients who exhibited disease progression or improvement in the Cobb angle between the 2 groups was compared using a χ2 test. Of 591 patients assessed for eligibility, 436 did not meet the eligibility criteria, 12 declined to participate, and 15 withdrew before randomization. The remaining 128 patients were randomized to 2 parallel groups, with 64 patients in each group (mean [SD] age, 11.1 (2.2) years; 97 female [75.8%]), all of whom completed the baseline assessment and 6-month follow-up. The mean difference in the improvement of the Cobb angle between the DC group and the usual care group after a 6-month intervention was -4.23° (95% CI, -6.08° to -2.39°) in the intention-to-treat analysis and -4.01° (95% CI, -5.68° to -2.35°) in the per-protocol analysis. The posttreatment Cobb angle was adjusted by using analysis of covariance, and the mean of the difference in the adjusted posttreatment Cobb angle between 2 groups was -4.24° (95% CI, -6.09° to -2.38°). In this randomized clinical trial, the effectiveness of individually tailored PSSE training support delivered via a digital care system was superior to that of a conventional PSSE training model in improving the Cobb angle. This mode of exercise may be an effective and convenient alternative for individuals with adolescent idiopathic scoliosis. Chinese Clinical Trial Registry Identifier: ChiCTR2300076563.

Every year, the global production of marine debris reaches a staggering 400 million tons, and the amount of garbage continues to increase. Among this, about 14 million tons will eventually flow into the oceans, posing a significant threat to marine ecosystems. Apart from endangering marine ecology, this waste will ultimately come back to affect humans. To address the increasingly severe issue of marine debris, our team independently developed a surface garbage cleaning robot, DaYu No. 1, and designed a path planning algorithm specifically for DaYu No. 1, called DyNav. DaYu No. 1 aims to increase garbage cleaning efficiency and adaptability. The traditional A<inline-formula> <tex-math notation="LaTeX">$^{\ast }$ </tex-math></inline-formula> algorithm performs well in static environments, but the ocean is a dynamic environment where marine debris constantly shifts due to ocean currents or tides, making it extremely challenging for robots to collect marine debris. Therefore, our proposed DyNav algorithm can adaptively adjust paths to cope with environmental changes. We proposed DyNav, a new path planning algorithm that takes ocean currents into account, and the DaYu No. 1 robot with YOLOv7 and IMA algorithms built in. These will make cleaning up marine debris much more efficient and flexible. Compared to traditional path planning algorithms, DyNav demonstrates superior performance. We have validated the effectiveness of the DyNav algorithm through practical experiments. The results show that in dynamic environments, the robot can clean surface debris more quickly and flexibly while avoiding obstacles. This not only contributes to environmental protection, but also improves cleaning efficiency, offering an innovative solution for water environment management.

In order to achieve the goal of carbon neutrality, especially to guide and encourage young people to participate in this process, with the goal of achieving zero carbon emissions on campus. A multi-energy complementary green energy system composed of photoelectricity, wind energy, geothermal energy, and energy storage is comprehensively utilized to construct a campus comprehensive energy system, and intelligent energy scheduling is achieved through the operation of big data. MQTT, CoAP IoT protocol, and various industrial field control bus technologies such as ModBus, ProfiBus DP, CAN, MBus are adopted to achieve data acquisition, transmission, and control. A comprehensive management platform is established to explore the design of a new energy system for achieving zero carbon goals in university campuses.