An explainable hybrid deep learning-enabled intelligent fault detection and diagnosis approach for automotive software systems validation

This paper proposes a novel explainable hybrid deep learning framework combining 1D-CNN and GRU architectures with interpretability techniques like IGs and SHAP to enhance fault detection, diagnosis, and root cause analysis in automotive software system validation while overcoming the limitations of traditional black-box models.

The Gist

Imagine you are a master mechanic trying to fix a brand-new, self-driving car. But there's a catch: you can't open the hood, and the car's computer is speaking in a language you don't understand. Every time the car acts up, it just flashes a generic "Error" light. You know something is wrong, but you have no idea what is wrong, where it is, or why it happened.

This is exactly the problem engineers face with modern automotive software. They have massive amounts of data from test drives, but the "smart" AI models they use to find faults are Black Boxes. They give an answer ("The engine is failing!"), but they refuse to explain how they reached that conclusion. This makes it hard to trust the AI or fix the root cause.

This paper introduces a new solution: a "White Box" Detective for car software. Here is how it works, broken down into simple concepts:

1. The Hybrid Detective (The Brain)

The researchers built a new type of AI brain called a Hybrid 1dCNN-GRU. Think of this as a detective with two superpowers working together:

• The Spotter (1dCNN): Imagine a security guard scanning a crowd. This part of the AI looks at the data and instantly spots small, local patterns—like a sudden spike in temperature or a weird noise in the engine. It's great at finding "what" is happening right now.
• The Storyteller (GRU): Imagine a historian who remembers the last 10 years of events. This part of the AI looks at the sequence of events. It understands that a slight vibration now might be the result of a loose bolt from five minutes ago. It connects the dots over time.

By combining these two, the AI doesn't just see a snapshot; it understands the whole story of the car's behavior, even when multiple things go wrong at the same time (concurrent faults).

2. The "Why" Machine (Explainable AI)

The real magic of this paper is the Explainable AI (XAI) part. Usually, deep learning models are like a magician pulling a rabbit out of a hat—you see the rabbit, but you don't know the trick.

The researchers added a "Magic Reveal" layer. After the AI makes a diagnosis, it uses special techniques (like DeepLIFT and SHAP) to point to the exact ingredients that caused the prediction.

• Without XAI: "The car is broken."
• With XAI: "The car is broken because the fuel pressure dropped and the turbo speed spiked, which usually happens when the throttle valve is stuck."

It's like the detective not only catching the criminal but also showing you the fingerprint, the motive, and the timeline of the crime.

3. The Training Ground (HIL Simulation)

You can't test a new car by crashing it every day. So, the researchers used a Hardware-in-the-Loop (HIL) simulator.

• Imagine a flight simulator for pilots, but for cars.
• They built a super-realistic virtual car in a computer.
• They connected a real car computer (the "brain") to this virtual car.
• They then "injected" faults into the system—like pretending a sensor is broken or a wire is cut—while a human driver (or an automated one) drove the virtual car on highways and in cities.

This created a massive library of "what-if" scenarios to train their AI detective.

4. The Results: Smarter and Faster

The researchers tested their new "White Box" detective against older, simpler models (like standard RNNs or LSTMs).

• Accuracy: The new model was a superstar, getting it right 97% of the time, even when multiple things went wrong at once. The older models struggled, getting it right only about 40–70% of the time.
• Efficiency: Because the AI could now "see" which features mattered most (thanks to the XAI), they could throw away the useless data. They reduced the number of variables the model had to check from 24 down to 10.
  • Analogy: It's like cleaning your house. Instead of cleaning every single drawer in the house, the AI told you, "You only need to clean the kitchen and the bedroom." This made the training process 4 times faster without losing accuracy.

Why Does This Matter?

In the world of self-driving cars and safety-critical software, you can't just guess. If an AI says a car is safe, you need to know why.

• Trust: Engineers can trust the AI because they can see the logic.
• Speed: They can fix problems faster because they know exactly which part is failing.
• Safety: It helps prevent accidents by catching complex, simultaneous faults that older systems would miss.

In a nutshell: This paper teaches us how to build a car-fault-detecting AI that is not only incredibly smart but also honest about how it thinks, making our future roads safer and our engineers' jobs easier.

Technical Summary

Here is a detailed technical summary of the paper "An explainable hybrid deep learning-enabled intelligent fault detection and diagnosis approach for automotive software systems validation."

1. Problem Statement

The validation of Automotive Software Systems (ASSs) during the development phase (specifically Hardware-in-the-Loop or HIL testing) generates massive amounts of real-time data. Current challenges include:

• Complexity of Faults: Faults can occur individually or concurrently (simultaneously), and data is often imbalanced (healthy data vastly outweighs faulty data).
• Black-Box Limitations: While Deep Learning (DL) models offer high accuracy, their "black-box" nature prevents engineers from understanding why a fault was predicted. This lack of interpretability hinders root cause analysis (RCA), model adaptation, and trust in safety-critical applications.
• Inefficiency: Traditional knowledge-based analysis struggles with the volume of data, while existing data-driven models often lack the transparency required for ISO 26262 compliance and optimization.

2. Methodology

The authors propose a three-phase framework: Data Collection & Preprocessing, Hybrid DL Model Development, and Explainable AI (XAI) Integration.

A. Data Collection & Preprocessing

• Source: Data was collected from a high-fidelity HIL simulation using dSPACE SCALEXIO and MicroAutoBox II, simulating a turbocharged 6-cylinder gasoline engine.
• Scenarios: Virtual test drives included highway, city, and lane-change scenarios with both manual and automatic driving behaviors.
• Fault Injection: Real-time fault injection was applied to sensors and actuators to create datasets for:
  • Fault Identification (FTCM): Detecting fault types (Noise, Gain, Offset) and their combinations.
  • Fault Localization (FLM): Identifying fault locations (e.g., pedal sensor, steering angle, engine speed) and their combinations.
• Preprocessing: Steps included denoising, outlier removal, scaling/normalization, and handling class imbalance using Random Under Sampling (RUS), Class Weights (CWs), and SMOTE (Synthetic Minority Over-sampling Technique). Data was segmented into fixed-size windows for time-series analysis.

B. Hybrid DL Model Architecture

The core innovation is a Hybrid 1dCNN-GRU architecture designed to leverage the strengths of both Convolutional Neural Networks (CNN) and Gated Recurrent Units (GRU):

• 1dCNN Layers: Used for spatial feature extraction. They apply convolutional filters to capture local patterns, trends, and irregularities in the time-series data.
• GRU Layers: Used for temporal dependency modeling. They process the features extracted by the CNN to understand long-term sequences and time-dependent patterns, avoiding the vanishing gradient problem common in standard RNNs.
• Fully Connected (FC) Layers: Map the learned representations to output classes (7 classes for both FTCM and FLM).
• Optimization: Hyperparameters (window size, layer counts, learning rate) were tuned using Grid Search, Random Search, and Bayesian Optimization.

C. Explainable AI (XAI) Integration

To transform the model from a black box to a "white box," four XAI techniques were integrated to analyze feature importance and interactions:

1. Integrated Gradients (IGs): Tracks gradients from a baseline to the input to assign feature importance.
2. DeepLIFT: Decomposes output changes relative to a reference point to trace contributions back to inputs.
3. Gradient SHAP: Approximates Shapley values using gradients to evaluate feature importance while accounting for prediction uncertainty.
4. DeepLIFT SHAP: A hybrid approach combining DeepLIFT and SHAP to capture non-linear interactions and diverse baselines.

These techniques were used to compute:

• Global Feature Importance (GFI): Identifying top features across the entire dataset.
• Per-Class Feature Importance (PCFI): Understanding feature contributions specific to each fault class.
• Feature Interactions (FIs): Analyzing how pairs of features jointly influence predictions (crucial for concurrent faults).

3. Key Contributions

1. Novel Hybrid Architecture: Development of a 1dCNN-GRU model specifically for ASSs validation that handles both single and concurrent faults with imbalanced data.
2. Explainable White-Box Model: Integration of XAI techniques to provide interpretable predictions, enabling engineers to understand the logic behind fault detection and localization.
3. Model Optimization via XAI: Demonstration that using XAI to identify significant variables allows for model retraining with fewer features, significantly reducing computational cost without sacrificing much accuracy.
4. Comparative XAI Analysis: A thorough investigation of four XAI methods (IGs, DeepLIFT, Gradient SHAP, DeepLIFT SHAP) regarding their computational costs and performance in the automotive domain.
5. Industrial Validation: The approach was validated on a realistic HIL dataset involving user behavior and high-fidelity vehicle models, bridging the gap between academic research and industrial application.

4. Results

• Performance:
  • The proposed FLM (Fault Localization Model) achieved 97.40% accuracy, precision, recall, and F1-score.
  • The proposed FTCM (Fault Type Classification Model) achieved 97.19% across all metrics.
  • These results significantly outperformed baseline models (RNN, LSTM, and standalone GRU), which ranged from 43% to 85% accuracy.
• Efficiency vs. Accuracy Trade-off:
  • While the hybrid model had a higher training time initially (~23,000s for FLM), the integration of XAI allowed for feature reduction (from 24 features to 10).
  • Retraining the model with only the top 10 features reduced training time by ~76% (from 23,000s to ~5,400s) while maintaining high accuracy (95.6%).
• XAI Insights:
  • DeepLIFT and Gradient SHAP were found to be the most computationally efficient XAI methods.
  • DeepLIFT SHAP provided the most balanced and consistent feature rankings but required the highest computation time.
  • Feature interaction analysis successfully identified non-linear dependencies (e.g., between turbo speed and air mass flow), which is critical for diagnosing concurrent faults.

5. Significance

This research addresses a critical gap in automotive software validation by making advanced deep learning models interpretable.

• Safety & Trust: By providing a "white-box" view, safety engineers can trust the model's decisions and perform Root Cause Analysis (RCA) effectively, which is essential for ISO 26262 compliance.
• Operational Efficiency: The ability to reduce model complexity through XAI-driven feature selection lowers the computational burden, making real-time deployment on embedded systems more feasible.
• Concurrent Fault Handling: Unlike many previous studies that focus only on single faults, this approach successfully diagnoses simultaneous faults, a common and difficult scenario in real-world vehicle operations.
• Industrial Relevance: The use of HIL simulation with realistic user behavior validates the approach as a viable solution for the automotive industry, moving beyond theoretical datasets to practical engineering applications.