Our customer, a large manufacturing company, experienced the periodic malfunctioning of the specific components of their hydraulic system. This created a major problem for the company, disrupting the whole hydraulic system and, as a result, bringing extra maintenance costs and affecting the operations of the entire facility.
As the project progressed, we developed four machine learning models for the customer that could predict the degradation level of the key components of the hydraulic system with over 90% accuracy. Specifically, these included the ML models for predicting the cooler condition; valve condition; internal pump leakage; and hydraulic accumulator.
The Intelliarts team achieved average accuracy on test sets as high as 98% for each of the models. To accomplish this, we processed the records from IIoT sensors in the dataset and conducted a detailed data analysis. We also prevented the problem of model overfitting caused by a huge amount of features in the system, which could affect the model’s ability to generalize. The ML algorithms we chose included XGBoost classifier, StratifiedKFold, and RandomForestClassifier as the best fit for the prediction models.
After deploying the models, the Intelliarts team also developed API endpoints, triggered after each load cycle of the system. This should help the engineers identify when the model performance drops, and the model requires retraining or tuning. Finally, together with the management of the company, we conducted staff training sessions on how to use the ML solution in their day-to-day procedures to its fullest potential.
At the end of the project, the customer was completely satisfied with the results. We achieved 98% accuracy for each ML model, despite the original expectation of 90%, and provided a workable ML solution for predicting component failures and allowing proper maintenance of the hydraulic system. Our team continues to monitor the performance of the models and is ready to retrain them on demand. In the long run, the company expects improved productivity and safety, cost reduction, and a low probability of unexpected downtime.