eDatX
A flexible, intelligent and scalable solution
for edge-to-cloud vehicle data aggregation
tQCS is a sole solution and service partner for Excelfore eDatX in Asia-pacific region
including China, Korea, Japan and Vietnam.
Edge AI Anomaly Detection
In the automotive and IoT sectors, the way massive streams of data are processed is becoming a central focus. The industry is rapidly transitioning toward Software-Defined Vehicles (SDVs) and highly connected car ecosystems, where vehicles operate as intelligent, data-driven platforms. Each modern car can generate dozens of terabytes of sensor and operational data per day, ranging from high-resolution camera feeds to real-time telemetry from hundreds of ECUs. However, transmitting all this data to the cloud is not practical—network bandwidth is limited, cloud storage and processing costs are escalating, and the sheer volume makes it difficult to identify critical signals in the noise. As automotive and IoT systems become more connected and autonomous, the need to process, filter, and act on data closer to where it’s generated—at the edge—has never been more urgent.
To address this, Excelfore introduces Edge AI Anomaly Detection—an intelligent, efficient way to process and filter data directly at the source. By leveraging advanced AI algorithms on the device or within the vehicle, only relevant and potentially anomalous data is flagged and transmitted to backend servers. This edge-first approach dramatically reduces bandwidth consumption, lowers operational costs, and accelerates response times, enabling OEMs and IoT providers to make smarter, faster decisions without being overwhelmed by data noise. By doing so, Edge AI, as an add-on service alongside eDatX Data Aggregation within the device or vehicle, enhancing the data collection process and making eDatX significantly more efficient.
Real-Time Anomaly Detection
Ensemble-based unsupervised learning utilizes multiple decision trees to detect patterns and anomalies in high-dimensional time-series data such as CAN bus signals.
No labeled data required, making it ideal for real-time and complex automotive and IoT environments where labeling is challenging.
The anomaly detection algorithm is appropriately chosen for hardware resources, e.g., Random Cut Forest (RCF), LSTM, and Transformer
Lightweight and Efficient
Optimized for resource-constrained automotive/IoT embedded systems, including NXP i.MX series.
Inference time: 20 ms across all tested platforms, enabling real-time anomaly detection.
NPU utilization: exceptionally low, ranging from 0.025% to 1.51% (minimum) and 0.25% to 6.04% (realistic), leaving ample capacity for parallel workloads and additional AI models. Refer to C. Edge AI Performance
Framework Agnostic
Not tied to specific AI frameworks such as TensorFlow or PyTorch, allowing customers to develop models with their preferred toolchain.
Runs on a lightweight runtime in Embedded Linux, ensuring broad hardware and software compatibility.
Customer-Defined Models
SDKs and REST APIs provide full lifecycle management of AI models, including updates, diagnostics, and configuration.
Seamlessly integrate with customer workflows, enabling automated deployment and monitoring.
OTA-based model updates by Customer
Enables secure, remote model deployment without physical access to the vehicle.
Supports version control, rollback, and targeted deployment for specific vehicle groups.
OEMs can deploy their own AI models without being locked into Excelfore’s solution.
Integration with the Excelfore Platform
eDatX data orchestration: Securely transmits anomalous or raw data to the cloud for model retraining and continuous improvement with in-vehicle processing, such as granularity/buffer control based on rules updated in real-time by OTA.
Closed-loop OTA: Safely redeploys updated models to edge devices, supporting ongoing performance optimization.
In-vehicle middleware: Aggregates data across CAN, LIN, and Ethernet networks to deliver standardized input streams to AI models.
CASE STUDY
An automotive OEM tested Excelfore’s Edge AI anomaly detection during the development of an electric vehicle’s e-Axle CAN network.
The system runs a demonstration against two datasets: one composed mainly of non-anomalous CAN data, and another containing a mix of regular and anomalous messages from an e-Axle under real-world operating conditions.
Using a trained dataset and application-specific tuned RCF, LSTM, and Transformer-based models, Edge AI evaluates each CAN message in real time. For standard CAN traffic, the anomaly detection model consistently maintains a false positive rate below 1% with the current X1 version, as the accuracy in a non-anomalous scenario is evaluated using statistical metrics such as FPR and the anomalous percentage captured. However, when analyzing the e-Axle dataset, a higher proportion of anomalies is correctly flagged due to the presence of inherent irregularities, such as unexpected torque sensor values or transient voltage spikes.
Each CAN message is scored and categorized instantly:
CAN ID: 0771 – Anomaly Score: 0.7574 – Anomaly Detected (criteria 0.65)
CAN ID: 0032 – Anomaly Score: 0.5172 – Normal
CAN ID: 0048 – Anomaly Score: 0.4931 – Normal
By filtering data directly at the edge and transmitting only messages above the 0.65 anomaly threshold, the system reduces data transmission volume by up to 80%, lowering cloud storage and processing costs. The reduced data load also minimizes device-level preprocessing requirements, allowing more efficient use of CPU cycles and memory. As a result, the anomaly detection pipeline operates with sub-millisecond latency, enabling real-time identification of high-risk events such as component degradation or early-stage failures.
This demonstration clearly illustrates how Excelfore’s Edge AI can:
Drastically cut bandwidth and storage requirements by transmitting only the top 20% of data that matters.
Enable faster and more targeted cloud analytics, since irrelevant “noise” is filtered at the source.
Improve device efficiency, as less onboard computation is needed for non-critical data.
Accelerate fault detection, helping OEMs respond before minor irregularities become major failures.