PdM for Mining Fleet

A mining fleet is the most critical and expensive asset of a mining operation. Operating and maintaining that equipment at high performance is expensive, and unplanned maintenance events have a direct impact on the production outcome and costs. The failure of a single component could affect the entire system, and one day of unexpected downtime could cost many thousands of dollars [22]. Mining operations worldwide typically adopt preventative maintenance based on the equipment's manufacturer schedules. These estimations can be drastically affected by the actual use of the equipment.

According to the literature review described in [22], multiple approaches can be used to perform predictive analytics in a fleet.

• • Summary statistics - Evaluates the behavior of equipment to learn relevant statistics. For example, counting faults and determining equipment efficiency.
• • Hypothesis testing - Comprises the investigation of causal relationships.
• • Clustering - Applied to find novel concepts and allows heterogeneous groups of items to be grouped by similarity. For instance, fault categories can be grouped according to sensor data.
• • Classification - Can be used to categorize a variety of faults.
• • Anomaly detection - Can be applied to find abnormal signals due to a fault or malfunction.
• • Frequent pattern mining - Can identify correlations between variables and could be used to search faults that often occur together.
• • Process mining - Used to reconstruct sequences of activities using event data and may be applied for deviation detection.
• • Sensor selection - Approach used to improve the accuracy of the chosen models.

The analytical evaluation of the fleet can be categorized into two major types: onboard analytics and fleet-wide analytics. In the onboard analytics, data is processed locally in the equipment, which means that only a single equipment data is used in the predictive models. The benefits of this approach include (1) fog computing (a computational resource that enables real-time analytics and supports low latency and lightweight computing) and (2) the capacity of supporting simple models to analyze individual vehicle's sensors, such as expert systems. On the other hand, fleet-wide analytics combine data from the entire fleet, performing cross-fit analytics. The benefits include (1) higher accuracy in identifying faults, (2) capacity of accurately identifying no-fault events, (3) uncovering new types of faults, by applying similarity approach, (4) capacity of searching for root causes of the faults and (5) great performance for monitoring the normal status of many similar vehicles (a most common scenario) since fault data is often scarce. Finally, it is known that systems that integrate both fleet-wide and onboard analytics produce better PdM results [22].

In terms of advanced analytics techniques for PdM, a variety of approaches are found in the literature. In underground mining, predictive failure using data from Load Haul Dump (LHD) is a rich research field. A typical LHD has more than 150 sensors.

Using data from an underground mine in Canada, [23] studied the application of genetic algorithms for PdM of mining machinery. Assuming that failures of mining equipment caused by various agents follow the biological evolution theory, the team created software that analyzed data of an LHD and reached satisfactory results. Other successful applications involve supervised and unsupervised ML algorithms. According to Ref. [24], a supervised learning approach using Support Vector Machine and NN can be used to classify engine condition of an LHD in three categories: normal, suboptimal, or imminent failure. The model’s inputs considered for training the algorithm included engine data (torque, coolant temperature, oil pressure, turbo boost pressure, and speed), wheel-based speed, and historical based speed. Another ML case relied on unsupervised ML using One-Class Support Vector Machine to predict conditions leading to engine failure.

In the thesis published in Ref. [11], the Sequential Pattern Mining technique was used for analyzing big data collected from a North America mining company. The dataset presents records from eleven trucks for nine months. Sequential Pattern Mining is a data mining technique that searches for patterns that occur consecutively in a database or patterns that have an association with time or other values. The approach has demonstrated great applicability to uncover important patterns in sequential data. In the studied case, three failure codes were selected and, first, several patterns between the two same codes were identified. Second, a variety of patterns were uncovered in the last three and five shifts that anticipate the breakdown. Despite some improvement opportunities reported by the author, the prediction rate was more than 90% in the last five shift events.

The mining company Barrick Gold decided to improve the asset health monitoring system of their truck haul fleet in order to increase maintenance efficiency and reduce costs. The operation where the initiative was developed - Pueblo Viejo - had 34 haul trucks, and the project involved real-time gathering information using the available systems and sensors at minimum cost. Instead of keeping with the traditional reactive approach, the maintenance team is now one step ahead of the failure. As a result, the company saved $500,000 and reduced the total number of failures from engine, brake, and suspension faults by 30% [17].

In a pilot study with a leading mining operator, the industrial AI software company Uptake processed 365 million data points from a fleet of 44 Caterpillar 797Fs to uncover insights that resulted in a 3% downtime reduction and a 5% improvement in haul truck utilization. The most difficult challenge reported was related to adding context to the machine fault codes before training the ML model so that the false failures could be distinguished from true failures that require action [24].

A predictive analytics solution provided by the PdM software company Dingo provides predictive models based on AI and ML techniques. The Anomaly

Detection and Remaining Useful Life models are developed by collecting and combining failure data from the actual equipment. One of Dingo's mining clients was experiencing a degrading final drive life on Caterpillar 789 C&D trucks, which resulted in significant costs and downtime impacts to the operation. Before the implementation of the predictive analytics solution, the average life of final drives had decreased by more than 30% (from 19,092 to 13,229 hours). After understanding the most common failure modes (broken/ worn teeth in the central gear and worn copper washers) and using the only condition monitoring data available for these components (oil analysis), a team of data scientists identified the most correlated oil analysis indicators to failure and developed proprietary ML models to predict future failures.