How do you update and improve your predictive maintenance algorithm over time?

In my experience, the first challenge represents the structure and control of the data. The condition monitoring signals may come from different sources (sensors, SCADAS, manual inputs )and in different formats. However, signals must be processed in one place and under the same structure (data set). Labels—vital in predictive maintenance—need to come from a catalog of detectable failures across the company. This catalog allows synchronized data and label matching, which analysts often feed to the algorithm.

We should remember that predictive maintenance algorithms are not static; they evolve alongside technological advancements, data availability, and domain-specific knowledge. By embracing these strategies, organizations can achieve more reliable predictions and optimize maintenance schedules. Continuous improvement is essential to enhance the accuracy and effectiveness of these algorithms. Some strategies for continuous improvements of PdM algorithms could be Leverage Machine Learning (ML), Incremental learning, feedback loop, feature engineering, ensemble methods, anomaly detection, transfer learning, and regular maintenance of predictive models.

In condition monitoring, working with supervised systems by label assignment is normally the way to train our algorithm, which synchronizes labels with time-related data and the corresponding detection of failures or anomalies in the equipment or machines. Then, the algorithm should correlate the data behavior when a failure is labeled by doing a retrospective regression of the data.

The dashboard, with its alerts and reports, will therefore be absolutely key in order to keep in close touch with the performance of your algorithm. These features include automatic alerts for risk failures and generating reports for the long-term effectiveness of the said algorithms. This is in addition to the fact that the maintenance teams are going to be involved, going through the alerts and actions that need to be taken.

The maintenance department's feedback on diagnoses is the key to fine-tuning the algorithm. Maintenance staff can easily evaluate the effectiveness and relevance of failure detection and severity through maintenance interventions.

Many projects underestimate the challenge of closing feedback loops to a plant's operations, where maintenance professionals and operators have the insights, that allow to update and refine your algorithm. Put consideration into the interfaces you use to make your predictions available within operations. Do they create an experience, that makes it effortless for users to provide feedback? Additionally, if you architect your data infrastructure thoughtfully, you reduce the effort of integrating valuable data, that likely already exists in your enterprise software (i.e., maintenance work orders, warranty claims, spare parts orders, ...).

When data is included in addition to sensor and SCADA outputs, such as ambient temperature, process parameters, and types of operation, it can unveil some unexpected relations. For example, the correlation of the night shifts to an issue regarding the condition of the emulsion oil in the steel processing plant where I was working was due to different operators' behavior at night than during the day shifts.

The algorithm stakeholders are the executive board, the analyst, the maintenance department, and production. Analysts ensure the right diagnostics, often issuing the notification or maintenance order directly to the planners. The maintenance department confirms the algorithm's recommendations and diagnostics. A good practice on this topic is tracking and monitoring the rejection of recommendations. For this, feedback could improve the assertiveness of machine learning algorithms within predictive maintenance. The main focus of the production and executive boards is usually common: reducing downtime and increasing equipment availability.

This is important for the continuous adaptation and finetuning of predictive maintenance algorithms, given comprehensive data analysis, stakeholder feedback, and operational insights. Further, this continuous process would increase accuracy and relevance, pushing maintenance optimization and asset management strategies.