ИИ на базе LSTM снизил погрешность взвешивания медной руды на 64,6%

Aiming to improve measurement accuracy and support sustainable mining, researchers have developed and validated an intelligent error compensator based on long short-term memory (LSTM) recurrent neural networks for dynamic weighing systems on copper concentrate belt conveyors. They published their findings in Inventions. Copper Weighing Accuracy Imperative Peru's copper mining sector saw strong growth in the 2010s, yet a sharp price decline starting in 2018 severely impacted profit margins. This shift made traditional 30,000-ton bulk-shipment cargoes financially unviable, as loads often fell short. The industry responded by moving to containerized exports, a method requiring rigorous identification and weighing accuracy, adhering to protocols such as DIN EN ISO 6343. However, conventional conveyor belt scales, prone to issues including mechanical vibrations, belt tension fluctuations, and material impact forces, were simply inadequate for these new precision demands. Even small errors in weighing high-value copper concentrate resulted in substantial financial losses from over- or under-loading. This critical situation spurred an urgent search for intelligent, robust weighing solutions to ensure profitability, regulatory compliance, and sustainable mining operations. Download a free PDF copy of this page by clicking here. LSTM Compensator Development Process To overcome the deficiencies of existing weighing systems, this study developed an innovative intelligent error compensator, leveraging long short-term memory (LSTM) recurrent neural networks. This system acts as a predictive model, differing from conventional reactive filters, by anticipating weighing inaccuracies based on historical operational data. The research began with extensive data collection from an active industrial copper concentrate container filling circuit. Real-time inputs were gathered from multiple sensors: load cells for weight, speed sensors for belt movement, inclinometers for angle, current sensors for motor activity, and vibration sensors for mechanical disruptions. The experimental setup leveraged a conveyor belt inclined between 20 and 30 ° and equipped with a bridge-type scale for dynamic weighing; data collection was conducted at a nominal inclination of 25 °. Raw sensor data underwent rigorous preprocessing, including normalization and time window generation, which were crucial for LSTM network preparation. The central element was the LSTM architecture, crafted to model the intricate temporal dependencies in dynamic weighing errors. At each time step, the network's input layer received a five-dimensional vector of normalized features: load-cell voltage, belt speed, inclination angle, motor current, and vibration. Hidden LSTM layers, with their internal ‘gates,’ managed information flow, allowing the model to discern both short- and long-term patterns in error dynamics. The final LSTM layer's output was then passed to a dense layer with a linear activation function, yielding the compensated error prediction. Hyperparameter optimization, including the number of LSTM layers, units per layer, and optimal time window length, was conducted via a systematic grid search to ensure peak predictive accuracy. The model was trained on a comprehensive dataset, with performance rigorously assessed using MAPE, RMSE, and R2. This approach aimed to transcend predefined physical principles, learning directly from multisensor time series for proactive, adaptive error correction in copper concentrate weighing. Compensator Performance and Impact The LSTM-based compensator delivered impressive results, significantly sharpening dynamic weighing accuracy for copper concentrate belt conveyors. Through methodical fine-tuning, an optimal configuration emerged: a single LSTM layer with 20 units, a learning rate of 0.001, and a 20-step processing window, balancing modeling power with stable convergence. Independent testing showed dramatic improvements. The mean absolute percentage error (MAPE) plummeted from 8.5% in uncompensated systems to just 3.01%, a remarkable 64.6% reduction. Similarly, the root mean square error (RMSE) dropped from 12.3 to 4.2 tons, marking a 65.9% enhancement. A strong R2 of 0.95 affirmed the model’s robust explanatory power. These figures easily outperform conventional weighing methods, which typically have an error of 5–10%, and surpass basic filtering techniques, particularly given that validation occurred in a demanding, real-world industrial setting. Feature analysis gratifyingly confirmed the model’s physical common sense, with load-cell voltage as the dominant predictor (42%). This interpretability bolsters confidence, easing its acceptance by operators and regulators. Economically, the implications are immense: for a typical 30,000-ton copper concentrate shipment, the MAPE reduction could prevent an estimated $13 million in losses per vessel. This directly optimizes payload, avoids penalties, and drives sustainable mining practices. Enhancing Mining Weighing Precision This research successfully developed and validated an intelligent error compensator based on LSTM neural networks for dynamic weighing on copper concentrate belt conveyors. The approach addresses conventional system limitations by adeptly capturing nonlinear dynamics and complex temporal dependencies. An optimal LSTM configuration (one layer, 20 units, 0.001 learning rate) achieved a remarkable 64.6% accuracy improvement, reducing MAPE from 8.5% to 3.01%. This precision gain, attributed to the LSTM’s capacity for modeling nonlinear temporal relationships, promises substantial economic benefits for mining, potentially saving millions per shipment through reduced material losses and optimized logistics. The model's interpretability, confirmed by load-cell voltage as the primary error predictor, enhances its industrial credibility. While rigorously validated offline, its integration into live production environments is poised to establish new benchmarks for operational excellence and sustainability. Journal Reference Chambi N., Sanga C., et al. (2026). Intelligent Error Compensation in Copper Concentrate Belt Conveyors Using LSTM Recurrent Neural Networks for Sustainable Mining Operations. Inventions. 11(4). DOI: 10.3390/inventions11040085. https://www.mdpi.com/2411-5134/11/4/85.