When a machine experiences failure, reactive maintenance has to be performed — which results in unscheduled downtime and production delays. The best antidote to prevent such inconveniences is predictive maintenance.
Predictive maintenance is the proactive technique of using extensive historical and real-time operational data to forecast when a machine will fail. It makes use of condition-monitoring tools to detect deterioration in machine operations and performance. Also, it works as an extension of condition monitoring and uses sensor data to measure the parameters of machines to determine their health and safety.
Condition-monitoring tools combined with artificial intelligence and machine learning techniques forecast expected machine failure. This information can be used to proactively perform maintenance activities without disrupting normal machine operations. At the same time, predictive maintenance helps to prevent unexpected machine failure and excessive maintenance.
Predictive maintenance helps in:
reducing maintenance costs
maintenance scheduling and planning
IIoT sensors connected to machines collect operational data and store them in a database. Predictive maintenance employs different systems and methodologies to predict machine failure with the help of the data repository. The following topics cover some of the prominent techniques used in predictive maintenance as well as their characteristics.
The knowledge-based approach in predictive maintenance relies on the knowledge and expertise of maintenance professionals to predict equipment failure — historical operational data and fault data are extrapolated to predict expected machine failure. This approach does not consider mathematical and physical models for forecasting machine failures.
The major downside is the low accuracy of the model. The inherent nature of this approach makes it difficult to apply it to complex systems with many moving parts.
The knowledge-based approach can be further broken into three:
Rule-based: The machine failure is predicted on conditions like ‘if-then’ which relies on an iterative process to check the rules. The rules are developed by experts with knowledge and expertise in the machines and processes.
Case-based: Each case is analyzed by a team of experts to predict the lifecycle of the machine and its parts. This approach is majorly subjective and relies heavily on the expertise of maintenance teams.
Fuzzy-logic based: This approach is very similar to rule-based systems. The only difference is the addition of fuzzy logic to the mix. It employs Boolean logic to deal with partial truth values.
The model-based predictive maintenance approach relies on the physical characteristics of the system. This knowledge of physics is used to determine the remaining useful life of machines and their parts. The physics of the system is simulated with computer algorithms with real-time data to forecast machine failures.
The model-based prediction approach can be further subdivided depending on how the prediction algorithm is structured. The subdivisions are the hidden Markov model, the mathematical model, the probability distribution model, and the filter model.
This approach has far higher accuracy than the knowledge-based approach, but building the physics model is complex and can be a daunting task for large systems with multiple machines and parts.
Data-driven prediction model
Data-driven prediction models completely rely on machine learning algorithms to make fault predictions. The IIoT sensors connected to the machines collect data continuously. The data is further vetted and cleaned before processing — to predict faults, machine learning algorithms are applied to this data.
To be used, this model is determined by the parameters of equipment that require prediction. The major benefit of using a data-driven model is its ability to identify relations and insights hidden in the vast amount of data. Another advantage is the ability of machine learning algorithms to process with ease multivariate, high-dimensional data.
Two types of machine learning models can be used for prediction, supervised learning algorithms and unsupervised learning algorithms. Artificial Neural Network (ANN), which is a supervised machine learning algorithm, is the most widely used algorithm for predictive maintenance.
Support Vector Machine (SVM) algorithms have been proven to have high accuracy compared to ANN. In recent times, deep learning models have also been used for predictive maintenance.
Due to the difficulty in data acquisition and relying on expertise, one single approach is not feasible for all facilities — a hybrid approach can be taken to alleviate the limitations of a single model.
The models discussed in the previous sections can be used in different permutations and combinations to develop a hybrid model. The first principles of different models can be combined in the hybrid model to make better predictions. Combining data-driven models and physics-based models has proven to have high efficiency and accuracy compared to other combinations.
Different combinations of models work for different scenarios. If you plan on implementing a multi-model approach to predictive maintenance for your production process, you need to evaluate your requirements. The combination should be feasible for your requirements and constraints.
Maintenance activities have become more efficient and advanced with the help of the Industrial Internet of Things (IIoT) and machine learning algorithms. Predictive maintenance helps to reduce the number of maintenance activities and maintenance costs. An increase in machine uptime and productivity are also benefits of implementing predictive maintenance in your production process.
Predictive maintenance can be approached and executed in many different ways. The efficiency rates and accuracy vary depending on the industry and the equipment parameters. The use of machine learning in predictive maintenance is unavoidable.
Each implementation of predictive maintenance will have different sets of constraints and requirements. The methodology selected should accommodate the specific needs of the plant to generate useful predictions.
Bryan Christiansen founded Limble CMMS in 2015 where he serves as CEO and directs creation of software to help managers organize, automate and streamline maintenance operations.