工程科学与技术 (Jan 2025)
Deep Learning-based 3D Visualization Health Condition Monitoring Method for Anchor Digging Machines
Abstract
ObjectiveIn coal mining, anchor digging machines operate in harsh and complex working conditions. They are subjected to high-intensity loads, resulting in high-frequency failures, and the problem of difficult health management is becoming more prominent. Therefore, it is crucial to monitor the operation status of anchor digging machines and ensure their healthy operation. This study integrated deep learning and 3D visualization technology, designed a complete health monitoring system framework, and proposed a health monitoring method for anchor digging machines.MethodsBased on deep learning, a data-driven remaining useful life prediction (RUL) method was proposed for key components of anchor digging machines. A dual time-length transformer (DT) RUL prediction model incorporating a stacked denoising auto-encoder (SDAE) was constructed (SDAE-DT Net), and particle swarm optimization (PSO) was used for model hyper-parameter optimization. The model possessed the structure of dual time-length encoding, which meant that long-time sequences' features were retained and allowed efficient processing of tightly connected short-time series data. The model was improved by SDAE, which could accurately predict the RUL when there was a lot of noise interference in the dataset. Experimental validation was conducted utilizing the actual production dataset sourced from the coal mine and the Intelligent Maintenance System (IMS) dataset. The results showed that the SDAE-DT Net model had the highest accuracy and the best prediction effect. On this basis, a 3D visualization health condition monitoring method of the anchor digging machine with data interaction was proposed. The 3D model of the anchor digging machine and coal mining geological model were constructed using 3D visualization modeling technology. Finally, combined with examples, the 3D visualization health condition monitoring system of anchor digging machine was developed, which realized the data mapping between the integrated coal mining working face and the 3D visualization model, and verified the correctness and feasibility of the method proposed in this study.Results and DiscussionsFor the performance validation experiments of the SDAE-DT Net model, the sensitivity analysis experiments for PSO hyperparameter optimization showed that the best results were achieved when the population size and inertia factor were 40 and 0.5, respectively. The optimal hyperparameters were: the number of iterations was 239, the number of coder/decoder layers was [4, 4], the number of training samples was 153, and the number of hidden neurons was 107. At this time, the PSO-SDAE-DT Net model reached the optimal value of each evaluation index in the training set as 0.157, 0.899, 0.192, and 0.087. The ablation experiments explored the effects of the improvements on the model through DT and SDAE. The results showed that the SDAE-DT Net model was significantly more stable in the training process than the other three sets of experiments due to its ability to deep feature capture and noise suppression, with a loss of 0.087. Comparison experiments of the prediction results with commonly used models, such as BiGRU, LSTM, and BiLSTM, similarly showed the superiority of the proposed method. To compare with multiple existing methods, experiments were conducted using the IMS dataset. The results showed that the SDAE-DT Net model has an RMSE value of 0.056 and the best prediction error of 48.36 min, which was the best performance among the models compared. The prediction time of the proposed model was 32 seconds, which was greater than the SVM model's 28 seconds, but the RMSE value was less than 1.214, so both models have their advantages. As a result, the proposed model has the smallest prediction error and higher prediction accuracy. The developed three-dimensional visualization health monitoring system of the anchor digging machine can display real-time environmental data and the operating status of the digging work. In the actual production, the real-time monitoring operation status of the anchor digging machine was selected from the system at a certain moment and compared the normal value of the operation indexes with the actual value. The results showed that all indicators were in the normal range. To verify the real-time RUL prediction effect of the SDAE-DT Net model, the vibration signals of the cut-off boom bearing were collected online for prediction. The results showed that the RMSE value during real-time prediction was 0.098 and the time required was 36 seconds. The predicted RUL prediction value of the 85th sample was 917 min and the true value was 1009 min, with a prediction error of 2.94%. The experimental results were close to the results of the historical data.ConclusionsThe results demonstrate that the designed health condition monitoring framework for anchor digging machines can achieve real-time and accurate health condition monitoring. The constructed SDAE-DT Net model effectively combines the features of different time-length sequences in the data, which improves the RUL prediction accuracy in the presence of noise interference. Based on cloud data storage and processing, the developed three-dimensional visualized health condition monitoring system for anchor digging machines achieves data interaction and real-time monitoring. The proposed method realizes the monitoring of operating status and coal mining anomalies and can provide a theoretical basis for the efficient operation and maintenance of coal mining equipment.
