This paper presents the development of an easy-to-deploy and smart monitoring IoT system that utilizes vibration measurement devices to assess real-time condition of bulldozers, power shovels and backhoes, in non-stationary operations in the mining industry. According to operating experience data and the type of mining machine, total loss failure rates per machine fleet can reach up to 30%. Vibration analysis techniques are commonly used for condition monitoring and early detection of unforeseen failures to generate predictive maintenance plans for heavy machinery. However, this maintenance strategy is intensively used only for stationary machines and/or mobile machinery in stationary operations. Today, there is a lack of proper solutions to detect and prevent critical failures for non-stationary machinery. This paper shows a cost-effective solution proposal for implementing a vibration sensor network with wireless communication and machine learning data-driven capabilities for condition monitoring of non-stationary heavy machinery in mining operations. During the machine operation, 3-axis accelerations were measured using two sensors deployed across the machine. The machine accelerations (amplitudes and frequencies) are measured in two different frequency spectrums to improve each sensing location's time resolution. Multiple machine learning algorithms use this machine data to assess conditions according to manufacturer recommendations and operational benchmarks Proposed data-driven machine learning models classify the machine condition in states according to the ISO 2372 standards for vibration severity: Good, Acceptable, Unsatisfactory, or Unacceptable. After performing field tests with bulldozers and backhoes from different manufacturers, the machine learning algorithms are able to classify machine health status with an accuracy between 85% - 95%. Moreover, the system allows early detection of “Unacceptable” states between 120 to 170 hours prior to critical failure. These results demonstrate that the proposed system will collect relevant data to generate predictive maintenance plans and avoid unplanned downtimes.

Vibratory screens are used in mining to classify mineral and send it to different pathways, normally using conveyor belts. Vibration analysis techniques are commonly used for condition monitoring and early detection of unforeseen failures to generate predictive maintenance. This paper proposes a novel solution to implement wireless sensors forming an instrumentation dedicated network combined with data-driven machine learning for monitoring vibrating screens. The system is optimized explicitly for vibratory equipment, which sets it apart from general-purpose condition monitoring systems. Embedded sensors are battery-powered and robust to withstand constant vibratory movement. The data used for training the machine learning models are gathered from a lab setup and discrete element simulations. The test bench consisted of a lab-scale vibratory screener, in which 3-axis accelerations, cumulative damage and wear are measured using sensors embedded in the rubber screens. Proposed data-driven machine learning models classify each screen condition in states according to the ISO 2372 standards for vibration severity. The system can identify random failures (based on test bench measured data) as progressive degradation failures over time (based on discrete element methods simulation results). The accuracy of the classification algorithms consistently ranges from 95% to 98%. Moreover, the system allows the early detection of unacceptable states up to 168 hours before the screen’s end-of-life predation by an expert. The system is characterized for (i) avoiding unplanned downtime and consequently (ii) increase operational availability. The system is intended to notify users when an abnormal operation is detected and impending failure events in the early stage.

Mining mobile machinery in non-stationary operations faces high levels of wear and unpredictable stress, posing significant challenges for predictive maintenance (PdM). This paper introduces a hierarchical inference network for PdM consisting on edge sensor devices, gateways, and cloud services for real-time condition monitoring. The system dynamically can adjusts inference locations – on-device, on-gateway, or on-cloud – based on trade-offs between real-time demands and conditions such as accuracy, latency, and battery range. The edge-based architecture enables rapid decision-making directly on-sensor or on-gateway, achieving classification accuracies above 90% while reducing latency up to 30% and power consumption on sensor nodes by approximately 45% regarding the cloud inference mode. This is critical to ensure machinery uptime in remote, rugged environments. The use of Tiny-Machine-Learning (TinyML) optimization approaches allow optimal accuracy and model compression for efficient deployment of deep learning models on IoT edge devices with limited hardware resources. The ESN-PdM hierarchical framework offers a scalable and adaptive solution for reliable condition monitoring and anomaly detection, contributing to advancing technology in PdM frameworks for real-world industrial applications.

Millions of workers are required to wear reusable respirators in several industries worldwide. Reusable respirators include filters that protect workers against harmful dust, smoke, gases, and vapors. These hazards may cause cancer, lung impairment, and diseases. Respiratory protection is prone to failure or misuse, such as wearing respirators with filters out of service life and employees wearing respirators loosely. Currently, there are no commercial systems capable of reliably alerting of misuse of respiratory protective equipment during the workday shifts or provide early information about dangerous clogging levels of filters. This paper proposes a low energy and non-obtrusive functional building block with embedded electronics that enable breathing monitoring inside an industrial reusable respirator. The embedded electronic device collects multidimensional data from an integrated pressure, temperature, and relative humidity sensor inside a reusable industrial respirator in real time and sends it wirelessly to an external platform for further processing. Here, the calculation of instantaneous breathing rate and estimation of the filter’s respirator fitting and clogging level is performed. The device was tested with ten healthy subjects in laboratory trials. The subjects were asked to wear industrial reusable respirator with the embedded electronic device attached inside. The signals measured with the system were compared with airflow signals measured with calibrated transducers for validation purposes. The correlation between the estimated breathing rates using pressure, temperature, and relative humidity with the reference signal (airflow) is 0.987, 0.988 and 0.989 respectively, showing that instantaneous breathing rate can be calculated accurately using the information from the embedded device. Moreover, respirator fitting (well-fitted or loose condition) and filter’s clogging levels (≤60%, 80% and 100% clogging) also can be estimated using features extracted from absolute pressure measurements combined to statistical analysis ANOVA models. These experimental outputs represent promising results for further development of data-driven prediction models using machine learning techniques to determine filters end-of-service life. Furthermore, the proposed system would collect relevant data for real-time monitoring of workers’ breathing conditions and respirator usage, helping to improve occupational safety and health in the workplace.

Forklifts are mobile heavy machines that are used to transport, lift, or lower objects without the high physical effort of the operator. They work in different types of industries such as logistics, retail, food, mining, and construction, among others. Qualified personnel usually operate the forklifts to handle heavy loads in an environment surrounded by other workers. This creates a high risk of accidents due to the lack of visibility with a loaded forklift, the random movement of the workers around the area and possible risk maneuvers sometimes required in a normal day of operation. For example, in Chile 2000 accidents occur per year due to one of the mentioned situations. For this reason, the detection of risk maneuvers to prevent accidents is essential. This article shows a cost-effective solution proposal to implement an inertial sensor network with a dedicated wireless communication and automatic deep-learning algorithms to detect forklift risk events. A test bench was designed where a crane forklift equipped with four inertial sensors performed normal and risky maneuvers, according to the Occupational Safety and Health Administration (OSHA) 3949. During the forklift operation, the sensors measured the accelerations and angular velocities in three axes. Videos of the operation were also taken as reference. In this paper, we developed convolutional neural networks (CNN) and long-term memory (LSTM) algorithms to infer a risky maneuver from the inertial sensors data and compared it to the outcome of a video-based model trained on data labeled by a risk-prevention engineer. After field testing with the forklift, the inertial data-based algorithms had an average F1 of 0.93 versus video analysis which had an average F1 of 0.95. However, models based on inertial data take a quarter of the time to make the inference compared to video-based models.

This paper proposes a transfer learning-based approach to enhance video-driven safety risk detection in industrial environments, addressing the critical challenge of limited generalization across diverse operational scenarios. Conventional deep learning models trained on specific operational contexts often fail when applied to new environments with different lighting, camera angles, or machinery configurations, exhibiting a significant drop in performance (e.g., F1-score declining below 0.85). To overcome this issue, an incremental feature transfer learning strategy is introduced, enabling efficient adaptation of risk detection models using only small amounts of data from new scenarios. This approach leverages prior knowledge from pre-trained models to reduce the reliance on large-labeled datasets, particularly valuable in industrial settings where rare but critical safety risk events are difficult to capture. Additionally, training efficiency is improved compared with a classic approach, supporting deployment on resource-constrained edge devices. The strategy involves incremental retraining using video segments with average durations ranging from 2.5 to 25 min (corresponding to 5–50% of new scenario data), approximately, enabling scalable generalization across multiple forklift-related risk activities. Interpretability is enhanced through SHAP-based analysis, which reveals a redistribution of feature relevance toward critical components, thereby improving model transparency and reducing annotation demands. Experimental results confirm that the transfer learning strategy significantly improves detection accuracy, robustness, and adaptability, making it a practical and scalable solution for safety monitoring in dynamic industrial environments.