Embedded Files
Preventive, anomaly prediction of crusher bridging
Application
Application
Rock bridging is a common problem in mining, especially in the operation of medium capacity, jaw crushers. Bridging refers to the blockage or jamming of the crusher due to the input material forming an arch or 'bridge' across the jaw of the crusher. The blockage prevents further material from entering, ultimately leading to a halt in the plant. This results in substantial downtime as the problem needs to be manually corrected, often by using a rock breaker. It is shown below that the highest contributor to downtime is rockbreaking at the ROM and at the jaw.
Early detection of such bridging events can significantly reduce the downtime and increase the efficiency of the operation. It enables timely intervention to dislodge the bridged material before it becomes a full-blown problem that halts production.
Minealytics utilises anomaly detection models for this purpose. Anomaly detection involves identifying unusual patterns or outliers in the data, which do not conform to expected behaviour. In this context, the anomalies represent potential bridging events. The key to these models is the utilisation of historical and real-time operational data to identify signs that indicate the potential for a bridge to occur.
In the case of the Minealytics Overload Detection model, it is trained on process data and the trained model is used to predict potential bridging events 5 seconds and 20 seconds ahead of their occurrence. The accuracy of over 80% on validation data means that the model can correctly predict the majority of the bridging incidents, providing valuable lead time for intervention.
The challenge lies in the fact that bridged events are not directly captured in the control system. The model uses indirect evidence, such as the rock breaker running signal, to identify these events. While this approach has been largely successful, it has led to some false positives where the model predicted a bridge where none existed. These 'false bridges‘ in the training dataset are a factor to be considered and addressed in data engineering and application integration phases. It has been shown the an overall 95% accuracy is achievable with minimum to no interruptions caused by the system by gradually increasing corrective response as the confidence increases.
This is where continuous monitoring, model refinement, and further integration with specialise sensor technologies like specialised cameras provides significant improvements. With more accurate and diverse data, the model's precision increases, reducing the number of false positives and further enhancing the efficiency of mining operations.
Alarm distribution highlighting ROM and Jaw bridging as the #1 and #2 highest occurring alarms at crushing plants
Rockbreaking and constrained control contributing the most to downtime
Overload Detection Model for Jaw Bridge Prediction
Overload Detection Model for Jaw Bridge Prediction
The Minealytics Overload Detection model uses a supervised classification model to detect the bridging. 23 input features have been used for the training and it has been shown that using only the crusher and grizzly currents result in a significant reduction >15% in accuracy.
In contrast, radar-based systems may struggle to differentiate between normal ore levels and potential bridging incidents, especially when the material characteristics are heterogeneous. They can also be affected by environmental conditions such as dust, humidity and temperature, requiring frequent calibration for optimal performance. Moreover, radar systems only provide distance or depth information, which may not be sufficient to analyse the nature or cause of a potential bridging scenario.
The flexibility and adaptability of camera-based systems make them highly compatible with data-driven, AI-based analytical tools. This is crucial for predictive maintenance and real-time decision-making, helping to transform mining operations into smarter, more autonomous entities. With lower costs for advanced camera hardware and the increasing efficacy of computer vision algorithms, camera-based systems offer a more versatile and cost-effective solution for modern mining operations looking to optimise their crusher circuit.
Grizzly current increases as crusher bridges and its current flatlines
Increasing grizzly cannot be observed but this is still picked up by the Overload detection model
False bridge prediction, Where crusher is clearly struggling while the grizzly receives oversize and its current shots up.
Training Metrics and Model Evaluation
Training Metrics and Model Evaluation
The best results have been achieved by the 21 input model with 5 second prediction horizon
95 % validation accuracy
By increasing the prediction horizon to 20 seconds 11% drop in validation accuracy has been seen:
84 % validation accuracy
By using only the crusher and grizzly currents 36% drop has been observer:
59 % validation accuracy
Typical model explanation during inference with the optimal model inputs on a 10 minute input window.
Realtime bridge detection marking the correct bridges over relevant crusher and grizzly current trends
Neural Preventive Anomaly Detection for Bridging and Blocked Chutes
Neural Preventive Anomaly Detection for Bridging and Blocked Chutes
A primary jaw crusher bridge prediction system employing a time series prediction neural network offers significant advantages in enhancing the efficiency and safety of crushing operations in mining. The system is trained on historical data to recognise patterns leading to material bridging—a common issue where materials form an arch over the crusher intake, hindering material flow and potentially causing operational disruptions and equipment damage.
The key advantage of this system lies in its ability to predict the likelihood of bridging events up to 20 seconds in advance using upstream process signals (ROM level, feeder pressure or current etc.), providing a crucial window for preemptive action. As the system forecasts an impending bridge, it progressively increases the response strength, such as by adjusting the apron feeder speed, to mitigate the risk. This dynamic throttling of the feeder speed based on real-time predictions helps maintain a consistent material flow, reducing the chances of operational halts and enhancing overall productivity.
Moreover, the adaptability of such a neural network model allows for its application in predicting bridging in other critical areas, such as chutes. This scalability means that the same predictive technology can be applied across different points in the material handling process, offering a comprehensive solution to a widespread challenge in mining operations.
The use of a supervised learning approach in training the neural network ensures that the model is well-acquainted with the specific characteristics and variations in the mining operation, leading to highly accurate predictions. This accuracy is vital for making informed decisions on controlling the feed rate and other operational parameters, ultimately leading to a more streamlined, efficient, and safer crushing process. The integration of such AI-driven predictive systems marks a significant step towards more autonomous, data-driven mining operations, where real-time data and predictive analytics drive decision-making and operational adjustments.
Technology
Technology
By training a machine learning model on labeled images of typical and atypical ore conditions, Minealytics can achieve a high level of granularity in real-time monitoring. For example, the algorithm could be trained to recognise and differentiate between ore, empty space, and potential bridges. When the system detects a bridge forming, it can alert operators to take corrective action immediately, thus preventing potential jams, equipment damage, or inefficiencies.
The benefit here is twofold: not only does semantic segmentation allow for real-time, automated bridge detection, but it also offers the potential for predictive analytics. By continuously analyzing the characteristics and formation patterns of the ore flow, a well-trained model can predict the likelihood of a bridge forming before it even happens. This predictive capability can be invaluable in shifting from reactive to proactive management, enhancing operational efficiency and equipment longevity. Given these advantages, semantic segmentation aligns well with the aim of transforming mining operations into more data-driven, autonomous systems.
Neural Modeling of Ore Characteristics
Neural Modeling of Ore Characteristics
The impact of ore hardness on the crushing process is a critical factor in mining operations, yet it poses a significant challenge in terms of measurement and modeling. Ore hardness directly influences the amount of energy required for effective crushing, the wear rate on crushing equipment, and the overall efficiency of the process. Traditional methods of estimating ore hardness through crusher parameters, such as electrical current draw, offer some insights but are often inadequate. These parameters can indicate changes in the crushing load, but the relationship between current draw and ore hardness is not straightforward and can be influenced by various other factors, making empirical modeling difficult and often unreliable.
This complexity has led to the exploration of advanced techniques like neural models for a more accurate assessment of ore hardness. Neural networks, with their ability to learn from large datasets and identify intricate patterns, can be trained on a range of data inputs, including crusher operation parameters, to develop a more precise understanding of the relationship between these variables and ore hardness. By continuously analysing operational data, these models can adapt to the subtle changes in ore characteristics, providing a dynamic and more accurate representation of ore hardness in real time. This capability is especially valuable in optimising the crushing process, as it allows for more informed and responsive adjustments to the crusher settings, improving energy efficiency, reducing equipment wear, and enhancing overall process control. In the realm of data-driven, AI-enhanced mining operations, the use of neural models for measuring ore hardness exemplifies the shift towards more sophisticated, adaptive, and efficient management of natural resources.
Page updated
Google Sites
Report abuse