AI-based vehicle health management system

Project background

Modern vehicles are costly to fix and maintain. For example, warranty claims, maintenance, and vehicle downtime costs affect the car fleet and influence customer satisfaction. To identify and settle these costs, vehicle health management systems, or VHM systems, use sensors and software to monitor vehicle performance, detect and predict problems, and recommend needed maintenance.  

Today, most VHM systems rely on diagnostic trouble codes. These codes identify an issue only after it has already occurred. Besides, they provide a piece of limited information only. It adds extra work for technicians to identify the cause of the problem. 

Modern vehicles use telematics to send diagnostic trouble codes and additional information to the cloud. Still, telematics data is too general and limited. As an effective alternative to the current solutions, our client created an AI-powered solution to monitor the health of vehicles effectively, processing thousands of different signals in real-time.

One of the project’s main goals was to create a possibility for users to compare and evaluate different anomaly detection models on unsupervised data.

Business needs

Modern vehicles are complex machines that are costly to fix and maintain. For instance, warranty claims, recalls, maintenance, and vehicle downtime costs reduce commercial fleet and customer satisfaction. To face and address them, vehicle health management systems, or VHM systems, use sensors and software to monitor vehicle performance, detect and predict problems, and recommend needed maintenance.

– One of the project’s goals was to create a possibility for users to compare and evaluate different anomaly detection models on unsupervised data.  

– The critical goal was to enable faster selection of unsupervised anomaly detection model for every new client dataset having no or very few labeled anomalies. 

– Another goal is to develop a robust data augmentation framework adjustable for different client datasets to help recognize unknown types of anomalies.   

– The client needed to strengthen the in-house development team with skilled ML developers and Data Scientists. 

– Our client was looking for a development team with good communication skills ready to start fast. 

Product features

1. Select the anomaly. Semi-automatic selection of unsupervised anomaly detection model for specific client datasets
2. Anomaly streaming. Offline and online streaming anomaly detection
3. Detection and compression. Changepoint detection and data compression
4. Data augmentation. Automatic data augmentation
5. Samples generating. Generating synthetic abnormal samples for each client dataset having no ground truth anomalous data available

Solution

Using Deep Learning technologies, the client’s solution can determine the vehicle’s health conditions and then flag any deviations from the expected behavior.

This VHM solution can also predict when a vehicle’s health condition could lead to failure and help identify the cause of the problem.

– Unlike other similar VHM solutions that can only operate in the cloud, this one can operate onboard vehicles in real-time, assisting fleet owners to predict vehicle health issues.

– CHI Software Data Scientist is currently working on implementing and testing Deep Learning models for unsupervised anomaly detection and auto-ml solutions based on synthetically generated anomalous data along with research on related state-of-the-art models and novel ideas in this field.

– The tasks of the CHI Software’s Machine Learning Engineer include obtaining models from Data Scientists, running them on Big Data, and parallelizing calculations on the cloud.

Our technology stack

• Python
• Microsoft Azure
• Databricks
• PySpark
• PyTorch
• Keras
• Horovod
• matplotlib
• Pandas
• NumPy
• sklearn
• seaborn
• plotly
• scipy

Client values

1. Building precise and powerful anomaly detection tools for offline and online streaming anomaly detection in time-series automotive data.
2. Developing a robust framework for data augmentation applicable to various kinds of data.
3. Recognition of previously unseen anomalies.