Inside Alstom's AI-Powered Train Inspection Workflow

Rethinking Rail Maintenance for the AI Era

Rail operators face an unrelenting challenge: maintain aging and complex rolling stock to the highest safety standards while controlling costs and minimizing downtime. Traditional inspection methods — manual visual checks, paper-based reporting, time-based maintenance schedules — served the industry for decades, but they are fundamentally limited. Human inspectors fatigue. Paper trails get lost. Time-based schedules replace components too early or too late.

When Alstom, one of the world's leading rail transport companies, set out to modernize its train inspection and maintenance operations, the requirements were clear: AI-powered defect detection that works at depot speed, digital workflows that capture evidence and generate audit trails, and a system that complements — rather than replaces — the expertise of seasoned maintenance engineers.

sensfix was selected to deliver on that vision. Here is how the deployment works, what it has achieved, and why the combination of visual and acoustic AI is proving transformative for rail maintenance.

Nine Proprietary Detection Models for Critical Components

At the heart of the Alstom deployment are nine proprietary computer vision models, each trained on thousands of annotated images of specific train components captured under real depot conditions. These are not generic object detection models adapted for rail. They are purpose-built classifiers and segmentation models designed to detect the precise failure modes that matter for each component category:

• Brake pad wear detection — measuring remaining pad thickness from visual inspection images, flagging components approaching minimum thresholds
• Pantograph analysis — detecting carbon strip wear, misalignment, and structural deformation in the current collection system
• Wheel profile assessment — identifying flange wear, flat spots, and tread anomalies that affect ride quality and safety
• Undercarriage inspection — scanning for corrosion, crack propagation, missing fasteners, and fluid leaks across the bogie assembly
• Additional specialized models covering door mechanisms, coupling systems, suspension components, HVAC units, and electrical cabinet conditions

Each model outputs not just a binary pass/fail classification but a severity score and confidence rating. This granularity allows maintenance planners to prioritize interventions based on actual condition rather than arbitrary schedules, shifting the operation toward true condition-based maintenance.

Audio AI for Compressor Monitoring

Not every defect is visible. Some of the most critical failure modes in rail equipment manifest first as changes in sound — a bearing beginning to fail, a compressor valve leaking, a motor winding developing an imbalance. These acoustic signatures are often subtle, buried beneath ambient depot noise, and easily missed by human ears.

sensfix deploys audio AI models specifically trained on industrial compressor acoustics as part of the Alstom solution. Maintenance technicians use standard recording equipment to capture short audio samples during routine checks. The AI analyzes these recordings in real time, comparing them against baseline acoustic profiles and flagging anomalies that indicate developing faults.

The combination of visual and acoustic AI delivers insights that neither modality can achieve alone. A compressor may look perfectly normal during a visual inspection while its acoustic signature reveals an internal fault that will lead to failure within weeks. Conversely, a visible crack may not yet affect acoustic performance but demands immediate attention. Multimodal analysis catches both.

TaskflowDigitizerAI: From Inspection to Action

Detection is only half the challenge. The other half is ensuring that findings are documented, communicated, and acted upon through a structured workflow. This is where TaskflowDigitizerAI transforms the inspection process.

Every inspection performed through the sensfix platform generates a digital workflow record that includes:

This last point — the audit trail — is critical for rail operators. Regulatory compliance in the rail industry demands comprehensive documentation of every inspection and maintenance action. Manual paper-based systems are error-prone, time-consuming to compile, and difficult to audit. TaskflowDigitizerAI generates compliance-ready documentation automatically, as a byproduct of the normal inspection workflow rather than as a separate administrative burden.

The Regulatory Compliance Imperative

Rail safety regulators in Europe, North America, and Asia require operators to demonstrate that rolling stock is maintained according to approved maintenance plans, that inspections are performed at prescribed intervals, and that all findings are documented and addressed. Failure to maintain adequate records can result in operating restrictions, fines, or — in the worst case — preventable accidents.

The sensfix platform addresses this imperative directly. Every data point — every image, every audio recording, every AI classification, every human decision — is stored with full traceability. Regulators and auditors can access a complete digital history of any component's inspection and maintenance lifecycle. This level of documentation would be prohibitively expensive to produce manually. With sensfix, it is a natural output of the AI-assisted inspection process.

Benchmark Results: 75% Inspection Time Reduction

The business case for AI-assisted inspection is ultimately measured in operational metrics. In a benchmark study conducted in collaboration with Rolls-Royce Power Systems, sensfix demonstrated a 75% reduction in inspection time compared to traditional manual methods.

This is not a marginal improvement. A 75% time reduction means that a depot performing four hours of inspection work per train can complete the same scope in one hour. Extrapolated across a fleet of hundreds of trains, each requiring periodic inspection, the impact on depot throughput, labor allocation, and vehicle availability is transformative.

Key metrics from the deployment and benchmarks include:

• 75% reduction in total inspection time per vehicle
• Significant improvement in defect detection rates compared to manual-only inspection
• Near-elimination of documentation errors and missing inspection records
• Faster turnaround from defect detection to maintenance action, reducing the window of risk

Why Multimodal AI Is the Future of Industrial Inspection

The Alstom deployment illustrates a broader principle that we believe will define the next decade of industrial maintenance: single-modality AI is insufficient for complex physical systems. Trains, turbines, manufacturing equipment — these are systems where failure modes manifest across multiple sensory domains simultaneously.

A purely visual system will miss acoustic precursors. A purely acoustic system will miss visible damage. A system that combines both, orchestrated by an intelligent rule engine that understands the relationships between visual and acoustic indicators, delivers a level of diagnostic capability that exceeds the sum of its parts.