CNH Industrial Pilot

Starting with the XAI Insight Services, are used early to perform tests on the models and the training data to understand the challenges related to the use case realization. The XAI Execution Services allow to execution of the XAI algorithms experimented earlier in a robust production-ready environment, once the models have been finalized. Moving on to the XAI Pipeline Lifecycle Management Services, those are being used in all aspects of the process, in order to securely create, schedule and monitor the execution of XAI pipelines. The Scalable Storage Services are used both as the source of input data to be processed and to store the refined data undergoing the process of data ingestion. Those are further utilized to keep the trained models and the associated prediction. Moving to the Data Collection & Governance Services and, in particular, to the Data Handler category, the File Data Harvester component finds particular use to ingest the raw machinery data in the proper format, following the CNH data model and ensuring type consistency. Finally, the platform management services are used to guarantee the security and authorization of the users in the platform, allowing access to data and pipelines only to CNH and the supportive partners of choice.

It is important to highlight that, although some of the components haven’t been marked as used, they are exploited indirectly through other components of the XMANAI platform. The above description only refers to the components used directly in the realization of the two use cases.

Another crucial aspect of the use case is the choice of the model to be used in order to accomplish the described tasks with the support of the XMANAI platform. The model must be well-performant with multiple irregular time series of different lengths. It also has to keep into account the possibility of adding new sensors in the future without compromising accuracy or forcing to homogenize data every time. For all these reasons an isolation forest approach was identified as the most suitable to use. This algorithm selects a feature and then randomly selects a split between the minimum and maximum values of the selected parameter. The core idea is that many splits are required in order to isolate a “normal” point while a small number of splits are required to isolate an anomaly. The sequence of splits that brings to an individual data point is called a path. Depending on the path length, an anomaly score is computed and interpreted as follows:

• A score close to 0.5 indicates a normal point.
• A score close to -0.5 indicates an anomaly.

With this basic concept in mind, two approaches were attempted:

1. The first approach consists of fitting an Isolation Forest for each available sensor, then checking the results of the models on the day of the occurrence of the failure. Finally, build plots to visualize and interpret the results. This method allows each individual model to be fitted on the specific sensor, increasing accuracy, and making each prediction independent of the number of samples available for other sensors. This way, however, prevents the user to catch the correlation between sensors (since each model only refers to an individual sensor). For this reason, a second approach was proposed in order to mitigate this issue.
2. The second approach relies on fitting an Isolation Forest for each group of sensors and checking the results of the models on the day of occurrence of the failure. Finally, some summary plots can be built by exploiting the visualization capabilities of the SHAP library to interpret the results. The crucial aspect of this method is the selection of the groups; these should be large enough to allow the model to correlate together as many sensors as possible, while avoiding grouping sensors with too different time series shape so that only a few samples have to be discarded/extended to harmonize all sensors in the group.

The combination of the two approaches described, allows the overall process to be precise enough thanks to the individual models tailored for each sensor while providing good correlation information thanks to the categorization of the sensors. The visualizations produced by each of the two approaches are presented in the section below.