Seeing Inside the System: AI-Powered Fault Diagnosis for Automotive Software

by

Author: Denis Avetisyan

A new approach combines deep learning with explainable AI to pinpoint and understand errors in complex automotive systems, improving both performance and safety.

This review details a hybrid deep learning framework for intelligent fault detection and diagnosis in automotive software, validated using Hardware-in-the-Loop simulation and featuring enhanced interpretability through explainable AI techniques.

Despite advancements in data-driven machine learning for automotive systems, a critical gap remains in the interpretability of fault detection and diagnosis (FDD) models. This paper, ‘An explainable hybrid deep learning-enabled intelligent fault detection and diagnosis approach for automotive software systems validation’, addresses this challenge by introducing a novel hybrid 1D convolutional neural network-gated recurrent unit (CNN-GRU) model integrated with explainable AI (XAI) techniques. Validated through hardware-in-the-loop (HIL) simulations, the proposed approach not only enhances the accuracy of fault localization but also provides crucial insights into the reasoning behind its predictions. Can this increased transparency and adaptability unlock new levels of safety and efficiency in the development and validation of complex automotive software?

The Inevitable Cascade: Modern Vehicle Complexity and Fault Detection

The proliferation of advanced driver-assistance systems, electric powertrains, and interconnected electronic control units has dramatically increased the complexity of modern vehicles. This intricacy, while enhancing performance and safety features, simultaneously presents a significant challenge to effective fault detection. A seemingly minor malfunction in one system can cascade into broader operational failures, potentially compromising vehicle safety and passenger wellbeing. Consequently, identifying and isolating faults within these complex architectures is no longer simply a matter of preventative maintenance, but a critical safety concern demanding robust and increasingly sophisticated diagnostic approaches. The sheer number of sensors, actuators, and communication networks necessitates a paradigm shift in how automotive faults are detected, diagnosed, and ultimately, resolved.

Contemporary vehicles generate an immense and rapidly changing stream of data from numerous sensors and electronic control units – a far cry from the diagnostic approaches of even a decade ago. Traditional fault diagnosis, often relying on manual inspection or limited onboard diagnostics, simply cannot keep pace with this deluge of information. The sheer volume of data overwhelms technicians, while its velocity – the speed at which it’s generated – means that by the time a problem is identified through conventional methods, critical damage may have already occurred or safety been compromised. This lag between fault occurrence and detection hinders preventative maintenance, increases repair costs, and ultimately impacts vehicle reliability and passenger safety. Consequently, there’s a growing need for automated, real-time diagnostic solutions capable of sifting through this data storm and pinpointing anomalies before they escalate into major failures.

The proactive identification of vehicular faults represents a paradigm shift from reactive repair to preventative maintenance, dramatically reducing downtime and associated costs. Contemporary vehicles generate a constant stream of diagnostic data; analyzing this information before a component fails allows for scheduled interventions, minimizing unexpected disruptions and potentially averting safety-critical events. This approach extends beyond simple repairs; it enables optimized maintenance schedules tailored to individual vehicle usage and conditions, maximizing lifespan and resale value. Consequently, accurate and timely fault identification isn’t merely about fixing problems-it’s about enhancing vehicle reliability, lowering the total cost of ownership, and ensuring continuous, safe operation.

A Hybrid Architecture: Decoding Temporal Signals in Automotive Systems

The proposed Hybrid CNN-GRU model integrates a 1-dimensional Convolutional Neural Network (CNN) with a Gated Recurrent Unit (GRU) to leverage the benefits of both architectures. The CNN component is designed to automatically extract relevant spatial features directly from the raw sensor data. These extracted features are then fed into the GRU, a type of recurrent neural network, which is specifically structured to analyze and model the temporal dependencies within the sequential data. This combined approach allows the model to not only identify important characteristics in the data but also to understand how these characteristics evolve over time, improving overall performance in tasks requiring sequential data analysis.

The hybrid architecture utilizes a one-dimensional convolutional neural network (1dCNN) to process sensor data as a spatial signal, identifying relevant features within each data sample. This 1dCNN component learns localized patterns and representations directly from the raw sensor inputs. Simultaneously, a Gated Recurrent Unit (GRU) network is employed to analyze the sequential characteristics of the extracted features over time. The GRU component’s recurrent structure allows it to maintain an internal state representing the history of observed feature patterns, effectively modeling temporal dependencies and the evolution of these features across the time series data.

Evaluation of the proposed Hybrid CNN-GRU model on independent test datasets yielded an accuracy of 97.40% for Fault Localization (FLM) and 97.19% for Fault Type Classification (FTCM). These results represent a quantifiable improvement in performance compared to both traditional fault diagnosis methodologies and implementations utilizing single deep learning architectures. The observed accuracy gains indicate the model’s enhanced capability in identifying both the location and specific type of faults within the system under evaluation.

Validating Resilience: Real-Time Simulation and the Hardware-in-the-Loop Approach

The Hybrid CNN-GRU model’s validation process utilized a Hardware-in-the-Loop (HIL) simulation environment to replicate the operational conditions of an internal combustion engine. This HIL setup incorporated a high-fidelity model of an ASM Gasoline Engine, allowing for comprehensive testing of the fault detection and localization algorithms. The ASM engine model provided realistic data streams, including sensor readings and actuator signals, which served as input to the Hybrid CNN-GRU model. This approach facilitated the assessment of the model’s performance in a closed-loop system, mirroring the complexities of a real-world automotive application and enabling the evaluation of its response to various fault scenarios.

Real-time validation of the Hybrid CNN-GRU model was conducted to evaluate its performance characteristics under constraints comparable to those experienced during actual engine operation. This methodology involved subjecting the model to dynamically changing inputs and assessing its processing speed and responsiveness. By simulating time-critical conditions, such as rapid engine speed fluctuations or transient load demands, we were able to measure the model’s latency in fault detection and localization. This approach ensured the model’s viability for real-world deployment where timely and accurate diagnostics are essential, and helped identify potential bottlenecks in processing speed that might occur under operational stress.

Rigorous testing of the Hybrid CNN-GRU model within the Hardware-in-the-Loop simulation environment yielded high levels of diagnostic accuracy. The model demonstrated 97.40% accuracy in correctly localizing faults, identifying the specific component experiencing the issue. Furthermore, the model achieved 97.19% accuracy in classifying the type of fault present. Critically, this performance was achieved with minimal latency, indicating the model’s suitability for real-time applications and time-sensitive control systems. The testing encompassed both isolated single faults and the more complex scenario of concurrent, multiple faults occurring simultaneously.

Beyond Prediction: Illuminating the Logic of Intelligent Systems

Recognizing the limitations of “black box” machine learning models, the system incorporates Explainable AI (XAI) techniques to foster trust and understanding in its predictions. This integration moves beyond simply what the model predicts to illuminate why a particular decision was reached. By dissecting the complex calculations within the model, XAI reveals the relative importance of different input features driving each prediction. This transparency is crucial for stakeholders who require accountability and insight, enabling them to validate results, identify potential biases, and ultimately, confidently integrate the system’s outputs into critical decision-making processes. The inclusion of XAI doesn’t just enhance the system’s functionality; it fundamentally shifts the paradigm from opaque prediction to interpretable intelligence.

To move beyond ‘black box’ predictions, the system leverages several techniques designed to reveal the reasoning behind its outputs. Specifically, methods like Integrated Gradients, DeepLIFT, Gradient SHAP, and DeepLIFT SHAP were employed to decompose the model’s predictions, effectively tracing them back to the input features that exerted the most influence. These approaches don’t simply identify that a feature mattered, but quantify how much each feature contributed to the final prediction, providing a granular understanding of the model’s decision-making process. This decomposition allows for a more transparent and trustworthy system, enabling users to understand not just the outcome, but the rationale behind it, and fostering confidence in the model’s reliability.

Initial model training demanded 22,998.75 seconds, a considerable computational investment. However, a focused feature selection process dramatically reduced this timeframe to just 5,413 seconds, representing a substantial optimization of resources. This improvement is particularly noteworthy when contrasted with alternative models; recurrent neural networks (RNNs) required 373.92 seconds, long short-term memory networks (LSTMs) took 699.93 seconds, and gated recurrent units (GRUs) completed training in 1356.55 seconds. The accelerated training time not only enhances the practicality of the model but also allows for more rapid iteration and refinement of its performance.

The pursuit of automotive software validation, as detailed in this work, inherently acknowledges the transient nature of system stability. The presented hybrid deep learning approach, focusing on intelligent fault detection and diagnosis, attempts to cache this stability through predictive modeling and explainable AI. As Andrey Kolmogorov observed, “The most important discoveries often come from unexpected places.” This sentiment resonates with the innovative combination of deep learning and XAI techniques, seeking to illuminate the ‘unexpected’ faults within complex systems. The paper’s emphasis on feature importance, a critical aspect of explainability, allows for a deeper understanding of system behavior over time-a necessary step in gracefully managing inevitable decay. The study’s validation through Hardware-in-the-Loop simulations underscores a pragmatic approach to mitigating latency, the ‘tax’ every request must pay, within a dynamic environment.

What’s Next?

The pursuit of intelligent fault detection, as demonstrated by this work, is less about achieving a static endpoint and more about establishing a versioning system for inevitable decay. Automotive software, like all complex systems, will accrue anomalies; the question isn’t if failures occur, but how gracefully the system ages and how effectively it communicates its decline. This research, while promising, underscores a fundamental tension: the demand for increasing model complexity clashes with the need for transparent, interpretable diagnostics. The arrow of time always points toward refactoring, demanding continuous adaptation of these models to maintain both accuracy and understandability.

Future iterations must grapple with the limitations of current feature importance techniques. While they offer glimpses into model reasoning, they often present a fragmented view, a post-hoc reconstruction rather than inherent clarity. A fruitful avenue lies in integrating explainability into the deep learning architecture itself – building models that are intrinsically transparent, rather than relying on external probes. Furthermore, the reliance on HIL simulation, while valuable, represents a controlled environment. The true test will be deploying these techniques in the chaotic, unpredictable landscape of real-world operation.

Ultimately, the field must acknowledge that perfect fault prediction is a chimera. The more pressing challenge isn’t eliminating all errors, but building systems that can anticipate, isolate, and mitigate failures with minimal disruption. This requires a shift in perspective – from reactive troubleshooting to proactive resilience. The true metric of success won’t be the number of faults detected, but the time required to recover from them.

Original article: https://arxiv.org/pdf/2603.08165.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

See also:

2026-03-10 19:53