ML System for Predictive Vehicle Maintenance

AI System for Predictive Vehicle Maintenance

Every year, fleets lose up to 30% of revenue due to unplanned downtime. Traditional mileage-based scheduled maintenance does not account for actual component condition. It replaces parts that could last thousands more kilometers or misses critical wear. Our ML models analyze telemetry in real time and predict failure 2-3 weeks before it manifests. For example, a truck fleet of 200 vehicles reduced unscheduled repairs by 5x after ML implementation. That saved $100,000 annually.Wikipedia reports that predictive maintenance can reduce costs by 25-40%.

Predictive maintenance in the automotive industry covers two areas: fleet management and service networks (dealers, repair shops). ML approaches cut unplanned downtime by 25-40% and optimize maintenance costs by shifting from interval-based to condition-based maintenance. Residual life prediction accuracy for key components reaches 95% — 10 times more precise than traditional analysis methods.

How Does Predictive Maintenance Reduce Costs?

ML models are 10 times more accurate in predicting failures compared to traditional threshold-based methods. This cuts maintenance costs by 25-40% and parts inventory by 20%. A fleet of 500 vehicles can save over $1 million per year.

Ensuring Data Quality for ML Models

Prediction quality directly depends on data. Common issues include CAN bus noise, telematics gaps, and unstructured DMS records. We apply a cleaning pipeline: outlier filtering with a sliding window, interpolation of gaps up to 5 seconds, and normalization of readings by VIN profile. For fleets with heterogeneous devices (Teltonika, CalAmp), we unify frequency and protocols via an MQTT bridge.

Data Sources

CAN Bus and OBD-II Telematics

can_data_channels = {
    'engine_rpm': 'OBD PID 0x0C',
    'vehicle_speed': 'OBD PID 0x0D',
    'coolant_temp': 'OBD PID 0x05',
    'engine_load': 'OBD PID 0x04',
    'fuel_trim_short': 'OBD PID 0x06',
    'fuel_trim_long': 'OBD PID 0x07',
    'intake_manifold_pressure': 'OBD PID 0x0B',
    'dtc_codes': 'OBD Mode 0x03',
    'oil_temp': 'OEM extended PID',
    'transmission_temp': 'OEM extended PID'
}

Telematics Devices (GPS + CAN)

Teltonika, CalAmp, Webfleet Solutions (TomTom) — fleet devices. Frequency: 1-10 sec. Data: coordinates + CAN parameters → cloud platform.

Dealer Data

• Service history by VIN (from DMS — Dealer Management System)
• Warranty claims: repeated repairs = sign of incomplete resolution
• PDI (Pre-Delivery Inspection) data

What ML Models Are Used for Wear Prediction?

Brake Pads

def brake_pad_remaining_life(brake_thickness_mm, driving_style_features,
                              road_conditions, mileage_km):
    """
    Regression model: remaining pad life
    Features: thickness, braking aggressiveness, urban cycle share
    """
    features = np.array([
        brake_thickness_mm,
        driving_style_features['hard_braking_events_per_100km'],
        driving_style_features['avg_deceleration'],
        road_conditions['urban_pct'],
        mileage_km
    ])
    remaining_km = brake_wear_model.predict([features])[0]
    return remaining_km

Battery (12V and HV in EVs)

• SoH (State of Health) via voltage at start and under load
• Internal resistance: increases with degradation
• Cold cranking amps (CCA): failure prediction at low temperatures

Engine — Early Signs

• Long term fuel trim > ±10% → rich/lean mixture
• Idle speed fluctuations → spark plugs, ignition coils
• Compression loss → piston ring wear (requires compression test)

DTC Analytics

def dtc_risk_score(dtc_history, vehicle_profile):
    recurring_dtcs = find_recurring(dtc_history, min_occurrences=2)
    risk_by_system = classify_by_system(recurring_dtcs)
    return risk_by_system

Fleet Management

Fleet Telematics

Daily health score for each vehicle:

def fleet_vehicle_health(vehicle_id, last_7days_telemetry):
    features = aggregate_telemetry(last_7days_telemetry)
    anomaly_score = isolation_forest.predict([features])
    component_scores = {
        'brakes': brake_model.predict(features),
        'battery': battery_model.predict(features),
        'engine': engine_model.predict(features)
    }
    overall_health = np.mean(list(component_scores.values()))
    return {'health': overall_health, 'components': component_scores, 'anomaly': anomaly_score}

Maintenance Optimization in the Fleet

• Calendar scheduling: minimize simultaneous downtime (≤15% of fleet)
• Just-in-time maintenance: when exactly, not by mileage
• Parts: pre-ordering based on replacement forecasts → reduced inventory costs by 20%

Condition-Based Maintenance vs Scheduled Maintenance

ML-based condition-based maintenance is 2 times more cost-effective than scheduled maintenance. Prediction accuracy is 4 times higher. Key parameters compared:

What's Included in Developing a Predictive Maintenance System?

We provide a full package: audit of current telemetry and DMS, design of data collection architecture (Edge + Cloud), training ML models for specific components, integration with your CRM or DMS, MLOps pipeline for automatic retraining, and documentation and staff training.

Process:

1. Analytics and requirements gathering
2. Prototyping on 10-20 vehicles
3. Pilot deployment with A/B testing
4. Full-scale rollout
5. Monitoring and support

Typical timelines: from 4 weeks for a basic solution to 4 months for a comprehensive system. Cost is calculated individually, but typical investment for a 200-vehicle fleet is $50,000 – $150,000 with payback in 3-6 months.

Data Source Comparison

AI solutions for auto service help dealers proactively invite customers for maintenance based on predictions. ML models for dealer networks enable forecasting parts demand and optimizing stock. Savings for an average fleet amount to about 2.5 million rubles per year ($28,000) — for a fleet of 500 vehicles, annual savings exceed $1 million. Investment in equipment pays off in 3-6 months due to reduced downtime and optimized maintenance.

Our company has been delivering AI solutions for the automotive industry since 2015. With over 8 years of experience and 50+ successful projects, we guarantee a high return on investment.

Contact us for a project assessment. Order a pilot deployment — we'll select the optimal architecture for your fleet or dealer network. Get a consultation — our engineers will analyze your data and propose a solution.

When does a time series forecasting model fail in production?

The CFO requests a quarterly sales forecast. An analyst builds SARIMA on three years of data, achieves MAPE 8.3% on the test set, and deploys. Two months later, the metric in production jumps to 23%. The root cause: the model was trained on pre‑COVID data, tested on a stable period, but production hit a promotion and supply chain disruption. Data leakage plus distribution shift—perfect notebook numbers, a broken forecast in reality. We have seen this pattern dozens of times across retail, fintech, and IoT. Our team has delivered more than 50 forecasting projects over 5+ years.

Incorrect cross-validation. Standard train_test_split for time series creates data leakage: the model sees future values during training. The correct approach is TimeSeriesSplit or walk‑forward validation with an expanding window.

Multiple seasonality. Hourly electricity consumption has three seasonalities: daily (24h), weekly (168h), yearly (8760h). SARIMA handles only one. Prophet can handle multiple but scales poorly to thousands of series.

Missing values and anomalies. A missing sensor reading is information (the sensor turned off), not NaN. Linear interpolation destroys this signal. Proper handling depends on the missingness mechanism.

Cold start. A new SKU in a 50,000‑item assortment has no history, yet a forecast is needed. Standard approaches fail; cross‑learning or feature‑based methods are required.

Why is model selection critical for your data?

Prophet (Meta) – a solid start for business data with clear seasonality and holidays. Fast setup, interpretable, built‑in outlier detection. Fails on irregular patterns and does not scale beyond ~10k series without parallelization.

Gradient boosting on features (LightGBM, XGBoost) – often underestimated. Engineer lags (t‑1, t‑7, t‑28), rolling means, day‑of‑week, holidays. The model trains on all series simultaneously, solving cold start via transfer learning. MAPE in retail often beats neural nets with proper feature engineering.

TFT (Temporal Fusion Transformer) – a transformer designed for interpretable forecasting with covariates. Built‑in variable selection, temporal attention, quantile outputs. Available in pytorch‑forecasting. Requires ~10,000+ records per series for stable training.

PatchTST – splits the series into patches (like ViT for images), capturing local patterns better than classic transformers. Excellent for long‑horizon forecasting (96–720 steps ahead).

N‑HiTS, N‑BEATS – attention‑free neural architectures, faster than TFT, competitive accuracy. N‑BEATS won the M4/M5 benchmarks for tasks without covariates.

How do we deploy TFT in production?

A typical pipeline via pytorch‑forecasting:

training = TimeSeriesDataSet(
    data,
    time_idx="time_idx",
    target="sales",
    group_ids=["store", "sku"],
    min_encoder_length=max_encoder_length // 2,
    max_encoder_length=max_encoder_length,  # 120 days
    min_prediction_length=1,
    max_prediction_length=max_prediction_length,  # 28 days
    static_categoricals=["store_type", "category"],
    time_varying_known_reals=["price", "promo_flag"],
    time_varying_unknown_reals=["sales"],
    target_normalizer=GroupNormalizer(groups=["store", "sku"], transformation="softplus"),
)

A common mistake: the default target_normalizer (StandardScaler) breaks predictions for series with zero values (no sales on weekends). GroupNormalizer with transformation="softplus" is the correct choice for count data.

Case study: retail demand forecasting

A chain of 120 stores, 8,000 SKUs, 28‑day forecast horizon. The original system: SARIMA per series, MAPE 18.4%, retraining cycle – 6 hours. We replaced it with TFT on PyTorch + pytorch‑forecasting: a single model for all series, MAPE 11.2%, retraining – 40 minutes on an A10G. Feature importance via variable selection revealed that day_before_holiday influences more than the holiday date itself. Annual savings on inference alone exceeded $50,000.

Step‑by‑step configuration

1. Data collection and preparation. Handle missing values (mark NaN, interpolate only for technical failures), aggregate to required frequency, engineer covariates (holidays, promotions, prices).
2. Create TimeSeriesDataSet. Set group_ids (store + SKU), time index, forecast horizon. Choose target_normalizer based on target distribution.
3. Train a baseline. Prophet or LightGBM first – to understand complexity.
4. Train TFT. Use TemporalFusionTransformer with loss=QuantileLoss(), tune learning rate and hidden layer sizes.
5. Validate and interpret. Walk‑forward test, analyze variable selection, build attention heatmaps.

How to properly evaluate forecast quality?

RMSE alone is misleading – it over‑penalizes large values. Our standard set:

• MAPE – interpretable, unstable near zero.
• sMAPE – symmetric, avoids division by small numbers.
• MASE (Mean Absolute Scaled Error) – normalized relative to a naive seasonal forecast, ideal for comparing series of different scales.
• Pinball loss – for probabilistic forecasting, inventory management.

We guarantee a model card with these metrics on the validation set and walk‑forward results on at least 6 months of history.

What deliverables do you receive?

• Documentation of chosen architecture and hyperparameter rationale.
• Reproducible training and inference pipeline (Docker + CI/CD + Airflow/Prefect).
• Committed code with unit tests for key components.
• Team training: retraining, output interpretation, deployment of new versions.
• 3 months of post‑delivery support (consultations, bug fixes, fine‑tuning).

The model is deployed via FastAPI or Triton Inference Server. Retraining is scheduled (e.g., weekly) via Airflow with drift validation and automatic rollback if metrics deteriorate.

Process and timeline

We start with EDA: visualization, ADF test, STL decomposition, analysis of missing values and outliers. This takes 2–3 days but often reveals systemic data issues that block forecasting. Then we build a baseline (naive seasonal, Prophet), engineer features for LightGBM, and select a neural architecture if needed. Walk‑forward validation with a realistic horizon. Deployment via API with automatic retraining scheduled via Airflow or Prefect.

Timeline: MVP forecast on one data type – 3–6 weeks. Hierarchical forecasting system with automation – 2–5 months. Cost is calculated individually based on data volume, number of series, and required accuracy.

Our team consists of certified ML engineers (AWS ML Specialty, GCP Professional ML Engineer) with 5+ years on the market and over 50 completed forecasting projects. Contact us for a free analysis of your data – we will assess the task and provide initial recommendations within 1–2 days. Request a consultation to ensure your forecasts work in production, not just in a notebook.