4.0: Machine Learning Engineering Processes
Learning Objectives
After reading this chapter, you will be able to:
- Describe the MLE processes introduced in ASPICE 4.0
- Apply ML engineering practices in automotive systems
- Integrate ML components with traditional software
- Address safety considerations for ML-enabled systems
Chapter Overview
The Machine Learning Engineering (MLE) process group is new in ASPICE 4.0, addressing the unique challenges of developing ML-enabled systems in safety-critical domains.
Important Note: This book extends ASPICE 4.0 with additional processes (MLE.5 Deployment and MLE.06 Integration) to provide practical implementation guidance. When evaluating against the official ASPICE 4.0 standard, only MLE.1-4 are normative. These additional processes align with common industry practice for ML deployment in automotive systems.
Operational Design Domain (ODD): The set of operational conditions, environmental factors, and use cases for which the ML system is designed. ODD definition is critical for ML safety and directly constrains training and testing datasets.
The following diagram provides an overview of the MLE process group, showing the relationships between requirements, architecture, training, testing, and deployment processes.
Why MLE in ASPICE 4.0?
Industry Drivers
| Driver | Description |
|---|---|
| ADAS Growth | Advanced Driver Assistance Systems require ML |
| Sensor Fusion | Multi-sensor perception needs ML algorithms |
| Predictive Maintenance | ML for failure prediction |
| User Experience | Voice recognition, gesture control |
| Autonomous Vehicles | Path planning, decision making |
Traditional vs ML Development
The diagram below contrasts the traditional software development lifecycle with the ML development lifecycle, highlighting the additional data management, training, and validation stages unique to ML.
MLE Process Summary
Note: AI automation levels (L1-L3) are defined in Chapter 03.01 AI Automation Framework.
ASPICE 4.0 Standard Processes
| Process | Purpose | AI Automation Level |
|---|---|---|
| MLE.1 | Establish ML requirements | L1 |
| MLE.2 | Define model architecture | L1-L2 |
| MLE.3 | Train and learn models | L2-L3 |
| MLE.4 | Test and validate models | L2 |
Book Extension Processes
| Process | Purpose | AI Automation Level |
|---|---|---|
| MLE.5 | Deploy models to target | L2 |
| MLE.06 | Integration guidance | L1 |
MLE-Specific Challenges
Safety-Critical ML Concerns
| Concern | Description | Mitigation Approach |
|---|---|---|
| Black Box | Model internals not interpretable | Explainable AI (XAI) |
| Edge Cases | Rare scenarios not in training | Scenario coverage analysis |
| Distribution Shift | Field data differs from training | Continuous monitoring |
| Adversarial Inputs | Attacks on ML perception | Robustness testing |
| Determinism | Same input → different output | Fixed seeds, validation |
Data Quality Requirements
The following diagram illustrates the key data quality dimensions -- accuracy, completeness, consistency, timeliness, and representativeness -- that must be addressed for safety-critical ML systems.
Process Relationships
The diagram below shows how MLE processes integrate with the SWE process group, illustrating the handoff points between ML model development and traditional software engineering workflows.
Key Work Products
| WP ID | Work Product | Owner | Description |
|---|---|---|---|
| 17-62 | ML requirements specification | MLE | ML-specific requirements |
| 04-08 | ML architecture description | MLE | Model architecture |
| 11-06 | Trained model | MLE | Weights, parameters |
| 11-07 | Training dataset | MLE | Curated, labeled data |
| 08-62 | ML test specification | MLE | Validation strategy |
| 13-62 | ML test results | MLE | Validation results |
| 04-09 | ML deployment specification | MLE | Target integration |
Integration with Traditional Processes
MLE-SWE Integration Points
| Integration Point | MLE Responsibility | SWE Responsibility |
|---|---|---|
| Requirements | ML capabilities | API requirements |
| Architecture | Model structure | Integration architecture |
| Interface | Input/output format | Data preparation |
| Deployment | Optimized model | Runtime integration |
| Verification | Model validation | System testing |
Chapter Sections
ASPICE 4.0 Standard Processes
| Section | Topic | Focus |
|---|---|---|
| 08.01 | MLE.1 Requirements Analysis | ML-specific requirements |
| 08.02 | MLE.2 Model Architecture | Architecture patterns |
| 08.03 | MLE.3 Training & Learning | Training processes |
| 08.04 | MLE.4 Model Testing | Validation methods |
Book-Extended Processes
| Section | Topic | Focus |
|---|---|---|
| 08.05 | MLE.5 Model Deployment | Target deployment |
| 08.06 | MLE.06 Integration | Integration patterns |
Prerequisites
| Prerequisite | Covered In |
|---|---|
| SWE processes | Chapter 6 |
| System engineering | Chapter 5 |
| AI automation framework | Chapter 3 |
| Safety standards | Part IV (Implementation) |
Related Standards: For ML safety considerations, see ISO 21448 (SOTIF) which addresses safety of the intended functionality for perception-based systems.