Three Common Myths About Predictive Maintenance in Mining

Although the potential is immense, studies show that most Predictive Maintenance programs fail to drive real value for mining companies.

In this blog, I will discuss three common myths about predictive maintenance and how mining companies can leverage AI sensor fusion to transform their predictive maintenance programs into value drivers.

Myth 1. The data collected by the mining companies can be easily used for Predictive Maintenance

Myth 2. Models that predict failures give (enough) value.

Myth 3. Deploying sensors is enough for Predictive Maintenance.

Myth 1. The Big Data miners collect can be easily leveraged for Predictive Maintenance.

Big Data vs. the Right Data

The first common misconception is that the big data that miners collect can be easily leveraged for predictive maintenance. Unfortunately, the reality is much more complex. 

The existing information is mainly based on PLC Tags that are stored historically. These are hundreds of gigabytes of data, not fully relevant to perform predictive maintenance properly. 

Let’s take a look at the automation pyramid below.

The Automation Pyramid

The automation pyramid shows all the organization’s data sources and information layers. The main parts of the data come from the lower layers of the pyramid.

First and foremost, these are the sensors installed on the machines that measure temperatures, vibration, power, and so on, in addition to other values calculated by the PLC systems, the machine’s computers, and the alarms. 

The upper layers of the pyramid contain additional data sources, like ERP and MES, which are the tags used to train predictive maintenance models. These tags tell the model whether the machine had malfunctioned, and if it did, what was the reason, how the machine functioned when it was new, or how it behaved just before the replacement. This is the information available in most mining companies today.

However, if we were to talk to field condition monitoring experts whose primary focus is increasing the machines’ reliability, they would say they use entirely different data. Yes, they will refer to the Historian’s data mentioned above, but they will treat it as a secondary data type where predictive maintenance is concerned.

Data Sources Used by the Mining Condition Monitoring Experts

These are the data sources they would look at first:

• Vibration – Data collected from the vibration sensors gives us an excellent diagnosis of the various components and the ability to know which rotating parts are not in optimal health. 
• Oil – The field ConMon teams will also focus on the oil analysis, which gives us knowledge about impurities in the oil, debris in the oil, and so on.
• Thermography – They will also examine the thermography reading obtained via temperature gauges, measuring overheating, overload, and various electrical problems. 
• And finally, they will mention the historian data, but it will be, as we said, quite secondary to all other data sources. 

 As you can see, there is a very large discrepancy between the information that experts in the field used for condition monitoring and the information collected by predictive maintenance solutions that are limited to the telemetry of the machines.

Performing condition monitoring only based on historical data is equivalent to asking a physician to diagnose a patient based on the data collected from the Apple watch. To properly diagnose a patient, the physician will order lab tests, CT/MRI scans, and many other check-ups to get a comprehensive picture of the patient’s health.

 To properly diagnose the machines’ health and execute predictive maintenance programs, in-depth data collected from various sources is mandatory.

Myth 2. Models that predict failures provide (enough) value

Most failures require different maintenance actions to prevent downtime. Prescribing a specific maintenance action is critical for value realization.

The mining companies’ data science teams are tasked with building models that predict machine failures. Even when designed successfully, they can’t provide the necessary value. Mining machinery is highly complex, and even when the models successfully predict malfunctions, more insights are required to manage them. 

After all, our goal is not only to predict a malfunction but to prevent the shutdown from happening and the subsequent shutdowns later on. Additionally, we are looking to prescribe a specific maintenance action to avoid this shutdown, deriving value from the Predictive Maintenance model that translates into fewer failures, fewer unexpected shutdowns, fewer work orders, and increased throughput and safety. 

Additionally, we are looking to prescribe a specific maintenance action to avoid this shutdown, deriving value from the Predictive Maintenance model that translates into fewer failures, fewer unexpected shutdowns, fewer work orders, and increased throughput and safety. 

The multitude of potential failure causes

Let's take a look at the possible failure causes. As mentioned above, the existing Predictive Maintenance solutions are based on machine reports. It can be a vibration-based report, oil-based report, thermography, or a model developed by the data science teams. A report will identify a fault in a machine, for example, bearing wear. The complexity of the problem is that many maintenance actions can be performed depending on the root cause. 

Here are the possible reasons for the wearing bear failure:

• Overload
• Oil contamination
• Imbalance in the conveyor belt strap
• Lubrication failure.

Each of these root causes will trigger a different maintenance action to treat the failure properly. As long as we do not have a reference to the specific maintenance action required, the likelihood of treating the cause of the problem itself is very low. 

When a machine malfunctions, the existing solutions don’t uncover the actual root cause of the failure, and the only thing that is treated is the symptom of the problem, perpetuating the ongoing loop of recurring machine failures.

In the illustrated example, if we address the symptom - replacing the bearing, the failure will happen again, and the cycle will continue, imposing two significant threats: 

1. Inefficient maintenance - flooding the maintenance teams with work orders and causing them to chase their tails
2. Potential safety hazard  - placing the maintenance team in the proximity of faulty machines in a critical health state.

Treating the underlying failure cause, not the symptom

Treating the underlying failure cause can prevent the malfunction from recurring, lowering the total number of malfunctions and thus increasing equipment reliability, the MTBF, and the safety of the staff. 

This can be achieved by a solution that fuses all the data sources and uncovers the root cause of the failure - in this case - lubrication failure and prescribes a specific maintenance action, e.g., “Check the ALS system and relubricate bearing.” By addressing the root cause and prescribing the appropriate maintenance action, mining companies can improve the value of their predictive maintenance programs and drive better outcomes.

Myth 3. Deploying sensors is enough for Predictive Maintenance.

We can’t stress the importance of sensors in the mining industry enough. Sensors play a crucial role in the mining sites’ day-to-day operations. Just like with data and models, the operational needs for sensors can be complex.

Not all sensors are created equal.

Sensors can be divided into two categories.

Type 1. Sensors that output raw data like vibration, temperature, torque, etc.

By deploying these sensors, the mining companies commit to allocating significant resources. Data science teams will need a lot of time and expertise to analyze the collected metrics and pinpoint the failures. Now imagine the amount of human power needed to perform this analysis for all the machines on the site – it could be impossible.

Type 2. Sensors that come with a built-in analytics layer that analyzes the raw data and processes it on some level.

The problem is that this analytics is specific to a certain type of sensor and is not enough.

Let’s look at an example of an engine of the reclaimer in which bearing wear was discovered. Potentially, there could be several root causes causing this failure. How can we identify the root of the failure and take the appropriate maintenance action? It could be an overload – a very common problem in mining equipment, caused by the stones passing through the reclaimer walls. To rule out this cause, we need to look at the vibration metrics, motor temperature measurements, and the current. We require data from several sensors to rule out the overload.

If we think that the root cause is oil contamination, we need to check vibration and oil analysis. To rule out the imbalance of the conveyor belt, we need additional indicators, such as belt centering, which can be obtained via photography. In this case, weightometer can indicate how many tons of material per hour pass through this belt. 

And finally, for lubrication, we need to check whether lubrication alerts were triggered and what was the density and secretions of the lubricant. We need to process the variety of data together to rule out irrelevant root causes and uncover what is causing this failure. 

If, after looking at all data sources, we see that the lubrication issues cause the failure, we understand that the automatic lubrication system (ALS) is not completing its cycle. In addition to replacing the bearing, the maintenance teams know that they also need to check the ALS system and ensure that the equipment is being lubricated properly.

Multi-sensor fusion is required to run effective Predictive Maintenance programs.

One thing is clear, whether the sensor only provides raw data, or comes with a built-in analytics layer, it is still not enough for running an effective predictive maintenance program. We need to look at the variety of sensors together. Just like doctors order various tests to make a diagnosis and prescribe a treatment for their patients when it comes to mining equipment, we need to analyze data from all available sources to determine the root cause of the failure and the maintenance actions needed to fix it.

Sensor fusion - the Razor Labs approach.

This is how the DataMind AI approach works when correctly performing predictive maintenance and condition monitoring. 

First, it is mandatory to perform a fusion of all types of relevant information, collecting big data and analyzing all the information that field maintenance experts use to perform condition monitoring – vibration, oi analysis, CCTV, and telemetry. This fusion can not only predict when the failures occur but indicate the exact root cause and prescriptive actions required to ensure that the failure doesn’t happen again.  

Second, it is required to ensure the coverage of the entire site’s critical equipment not to miss a single malfunction and ensure ongoing monitoring and identification of all the malfunctions.

Lastly, the involvement of domain experts is required to launch the correct maintenance processes and to ask the relevant questions about malfunctions indicated by the models. 

In this blog series, we have examined the Three Most Common Myths about the future of Predictive Maintenance in the Mining Industry. We believe in its relevance for all asset-intensive industries. It is not enough to collect data, generate AI models and deploy sensors. To run effective Predictive Maintenance programs, the companies must apply the sensor fusion approach, strive for full online machinery coverage, and involve the existing field maintenance experts to utilize the insights most efficiently.