Machines are ubiquitous in every industry, be it in manufacturing or aerospace, logistics or automobiles that deal with machines in one shape or the other. Reliability of the machines on the production floor of a manufacturing company or wind turbine in a power plant or automobile for a logistics company are critical to the business operations of the company. For example, a critical machine failure on a production floor could not only impact its manufacturing process but could equally disrupt the entire supply chain network of its business. The business competitiveness, overall productivity and reliability of a manufacturing-based company is directly linked to its operational efficiency on its production floor. This is true with any other industries that relies on machines for its day-to-day operation.
In a manufacturing plant or on road, machine failures are not uncommon, but it does not happen without warning and does not happen daily. Machines in typical production floor or on road runs for weeks, months even years before succumbing to the effects of wear and tear. An inadvertent failure of a single machine component in a manufacturing process flow could adversely impact the manufacturing process flow, requiring unplanned workforce mobilization to attend to emergency maintenance needs. A logistics company could see severe impact to its operation if 10% of its fleet go down in the same week or in a month. This is where predictive maintenance could be key to business planning starting with budgeting to workforce scheduling, inventory optimization to cost control and overall supply chain improvement. Predictive maintenance in machine heavy industries hold the key to its business stability and growth.
Traditional Approach and Challenges
Traditional approach to avoiding fatal machine failures have been scheduled maintenance at certain intervals like we do with scheduled oil change in our cars after 3000 or 5000 miles of run. But this strategy could be a wasteful and costly option, considering we could be doing the maintenance or changing the oil prematurely or for that matter replace a costly equipment in a manufacturing plant prematurely. This would incur unnecessary loss in production time and company must bear the replacement cost of the old equipment. The traditional approach has been dependent upon the type or model of the equipment in use and usually follows the Manufacturer’s guidelines for scheduled maintenance. This approach does not rely on any operational data such as data collected from machine sensors which says a lot about the machine’s operational readiness.
In the recent times with the advent of Industrial IOT and sensor-based manufacturing processes, manufacturing plants could collect large amount of operational data such as temperature, pressure, humidity, and other process related parameters over days and months. The volume of this data is large and often collected in real or near real time. The new approach to predictive maintenance is data driven and thus is more accurate and can predict future failures with fact-based data points. Unlike traditional approach, the data driven predictive approach is not limited to alerting impending failure, but can recommend the type of maintenance required after a failure. With the new predictive data driven approach, businesses could improve the overall plant operational efficiency and maintain a reliable supply chain system with greater customer satisfaction and increased profitability.
Predictive Maintenance Use Cases
Predictive maintenance use cases can range from Aerospace to Utilities, Manufacturing to Transportation and Logistics to name a few. The most significant predictive maintenance use cases can be in aircraft industries, where failures in aircraft components have disastrous impact on lives and the aircraft itself. High volume machine data collected from Aircraft engines and other critical machine components can be used to schedule predictive maintenance in advance and avoid any fatal failures in flight.
Similarly predictive maintenance of Wind Turbines could prevent power generation interruption and save costly maintenance schedules otherwise. Manufacturing industries are the most common use cases for predictive maintenances where costly maintenances and unscheduled components failures could be avoided with data driven incident predictions and warnings. Similarly in Logistics and Transportation predictive maintenance could help optimize the maintenance cycle and prevent costly failures.
Key components of a predictive Maintenance Solution
The success of predictive maintenance models relies on 3 important premises.
1. Right data available for Model Training
Garbage In is Garbage out. The types of data that fits right for the use case are dependent of key factors like “What are we trying to predict” and the type of failures that can occur. Apart from this the “failure process” like slow degradation or sudden fatal failures will decide the type of data that needs to be collected. The complexity of failure and how the machines/systems correlate to each other are also key considerations in data selection. That is where subject matter expertise (SMEs) and domain knowledge is key in identifying and qualifying the data sources and their attributes.
Machines typically fail after years of service. For training an effective predictive model, one would need historical data over a long period of time to understand the degradation pattern. This is a key requirement in ML model training where the model accuracy depends on historical labelled data for model accuracy.
2. Accurately Framing the problem with clear path for action
In predictive maintenance the problem statement could be many folds. Questions such as “Will this equipment fail in next 7 days” or “how much remaining useful life (RUL) is left with this machine” are some examples on how to frame the problem statement. There should also be clear path to action for the predicted events such as workforce mobilization or inventory and budget planning that should follow.
Types of models trained or deployed should align with the overall business goal and priorities attached to achieving machine reliability and profitability.
3. Appropriate evaluation after prediction
One of the key decisions in evaluating Predictive models’ performance is choosing the right set of evaluation metrics after Model training. The choice of metrics you choose squarely depends on the type of parameters you want to improve such as better inventory optimization or better resource allocation or cost efficiency. For example, in classification models, choosing the model with highest accuracy may not be the right choice rather one should look at the False positive and False negative buckets to improve the right parameters in the business planning and processes.
A few of the common but important factors that could decide the evaluation approaches and metrics are factors like lower labor cost, lower spare part inventory, more throughput capacity, less energy consumption, better safety, and improved product quality. The Model evaluation metrics should align to one or more of these parameters based on business priority and the kind of reliability it is trying to achieve.
We’ll speak more about the modelling techniques and an overview of the process in the next post of the predictive maintenance use cases part 2.