@djm204/agent-skills 1.0.0

package/templates/business/predictive-maintenance/.cursor/rules/failure-prediction.mdc

@@ -0,0 +1,65 @@
+---
+description: Failure Prediction
+alwaysApply: false
+---
+
+# Failure Prediction
+
+Guidelines for building and operating failure prediction models and anomaly detection systems.
+
+## Prediction Approaches
+
+| Approach | Data Required | Best For |
+|----------|-------------|----------|
+| Physics-based | Engineering models | Well-understood failure modes |
+| Statistical (Weibull, Cox) | Historical failure data | Fleet-level life estimation |
+| Machine Learning | Sensor + failure history | Complex multi-sensor patterns |
+| Hybrid | All of the above | Critical assets, mixed failure modes |
+
+## Model Development Pipeline
+
+1. Define failure modes and target variable
+2. Collect and label historical data (run-to-failure + maintenance records)
+3. Engineer features (sensor stats, operating context, time-based)
+4. Train and validate (time-series cross-validation)
+5. Set classification thresholds (balancing precision vs recall)
+6. Deploy with monitoring (drift detection, performance tracking)
+7. Retrain on new failure data (continuous improvement loop)
+
+## Remaining Useful Life (RUL) Estimation
+
+- **Similarity-based**: Compare current degradation curve to historical failures
+- **Regression-based**: Predict time-to-failure from feature trajectory
+- **Threshold-based**: Extrapolate trend to known failure threshold
+
+## Anomaly Detection
+
+| Method | Use Case | Sensitivity |
+|--------|----------|-------------|
+| Statistical (Z-score, IQR) | Single-sensor deviations | Low-Medium |
+| Isolation Forest | Multi-sensor anomalies | Medium |
+| Autoencoders | Complex pattern deviations | Medium-High |
+| LSTM-based | Temporal sequence anomalies | High |
+
+## Model Performance Metrics
+
+| Metric | Target | Interpretation |
+|--------|--------|----------------|
+| Precision | > 85% | % of alerts that are true positives |
+| Recall | > 90% | % of actual failures caught |
+| Lead time | > 14 days | Average warning before failure |
+| False alarm rate | < 5% | Alerts with no confirmed issue |
+
+## Failure Mode Library
+
+Maintain a catalog per asset type:
+
+- **Failure mode**: What breaks (bearing wear, seal leak, winding insulation)
+- **Detection signature**: Which sensors and features indicate this mode
+- **Typical lead time**: How far in advance degradation is detectable
+- **Consequence**: Safety, production, and cost impact
+
+## Common Pitfalls
+
+- **Training on healthy data only**: Include run-to-failure data. If none exists, use accelerated testing or physics simulation.
+- **Ignoring operating context**: Normalize sensor readings against load/conditions before applying thresholds.

package/templates/business/predictive-maintenance/.cursor/rules/maintenance-scheduling.mdc

@@ -0,0 +1,61 @@
+---
+description: Maintenance Scheduling
+alwaysApply: false
+---
+
+# Maintenance Scheduling
+
+Guidelines for optimizing maintenance timing, balancing asset health against operational constraints.
+
+## Strategy Selection by Asset Criticality
+
+| Criticality | Strategy | Rationale |
+|------------|----------|-----------|
+| Safety-critical | Predictive + time-based floor | Cannot risk failure; dual approach |
+| Production-critical | Predictive primary | Maximize uptime, minimize unplanned stops |
+| Standard | Condition-based | Act when degradation detected |
+| Non-critical | Reactive | Run to failure; replacement cost < monitoring cost |
+
+## Scheduling Decision Tree
+
+```text
+Model predicts degradation?
+├── Yes → RUL > next planned shutdown?
+│   ├── Yes → Schedule for next planned shutdown
+│   └── No → RUL > minimum lead time for parts/labor?
+│       ├── Yes → Schedule dedicated maintenance window
+│       └── No → Emergency maintenance; escalate
+└── No → Continue monitoring at current frequency
+```
+
+## Maintenance Window Optimization
+
+Factors to balance when scheduling:
+
+- **Asset RUL estimate** with confidence interval
+- **Production schedule** (minimize impact on output)
+- **Parts availability** (lead time for critical spares)
+- **Labor availability** (skilled technician scheduling)
+- **Co-location opportunities** (batch nearby assets in same window)
+- **Weather and access constraints** (outdoor or remote assets)
+
+## Spare Parts Management
+
+| Part Criticality | Stocking Strategy | Reorder Trigger |
+|-----------------|-------------------|-----------------|
+| Critical (no alternative) | Safety stock on-site | Below minimum + lead time buffer |
+| Important (workaround exists) | Regional depot stock | Below reorder point |
+| Standard (commodity) | Supplier agreement | Just-in-time ordering |
+
+## Work Order Integration
+
+Every predictive maintenance recommendation generates a work order containing:
+
+- Asset ID, location, predicted failure mode, and confidence level
+- Recommended action, urgency, required parts/tools/skills
+- Estimated duration and RUL window (earliest/latest completion)
+
+## Common Pitfalls
+
+- **Over-maintaining**: Replacing at 50% remaining useful life "just to be safe" wastes remaining life and increases costs. Optimize the intervention point.
+- **Ignoring constraints**: Model recommends Tuesday; plant runs 24/7. Integrate recommendations with production scheduling.

package/templates/business/predictive-maintenance/.cursor/rules/overview.mdc

@@ -0,0 +1,55 @@
+---
+description: Predictive Maintenance
+alwaysApply: false
+---
+
+# Predictive Maintenance
+
+Principal-level guidelines for sensor-driven failure prediction and maintenance optimization.
+
+## Scope
+
+This ruleset applies to:
+
+- Sensor data ingestion and real-time processing
+- Failure prediction and anomaly detection
+- Maintenance scheduling optimization
+- Asset lifecycle management
+- Alert systems and severity classification
+- CMMS/EAM system integration
+
+## Core Philosophy
+
+**Every unplanned breakdown is a failure of the monitoring system.** The goal is to detect degradation gradients early enough to schedule maintenance at the optimal economic moment - not too early (wasting remaining useful life) and not too late (risking catastrophic failure).
+
+## Fundamental Principles
+
+### 1. Predict, Don't React
+
+Shift from time-based schedules to condition-based triggers. Maintain equipment based on what the data says, not what the calendar says.
+
+### 2. False Alarms Kill Trust
+
+A model that cries wolf will be ignored when the turbine actually fails. Optimize for precision alongside recall.
+
+### 3. Degradation Is Gradual, Failure Is Sudden
+
+Detect the degradation curve before it hits the cliff. Trend analysis matters more than point-in-time thresholds.
+
+### 4. Maintenance Is an Economic Decision
+
+Every intervention has a cost; every deferral has a risk. Optimize the tradeoff between maintenance expense and failure consequence.
+
+### 5. Context Over Raw Data
+
+Sensor readings without asset history, operating conditions, and failure mode context are meaningless noise.
+
+## Decision Framework
+
+When evaluating a maintenance signal:
+
+1. **Confidence**: What is the model's certainty? Is this a known failure pattern?
+2. **Consequence**: What happens if this asset fails? Safety, production, cost impact?
+3. **Remaining Useful Life**: How much operating time remains before probable failure?
+4. **Maintenance Window**: When is the next feasible maintenance opportunity?
+5. **Cost Tradeoff**: Is preventive action cheaper than the expected cost of failure?

package/templates/business/predictive-maintenance/.cursor/rules/sensor-analytics.mdc

@@ -0,0 +1,66 @@
+---
+description: Sensor Analytics
+alwaysApply: false
+---
+
+# Sensor Analytics
+
+Guidelines for ingesting, processing, and analyzing industrial sensor data streams.
+
+## Data Ingestion Pipeline
+
+```text
+Sensors → Edge Processing → Stream Ingestion → Feature Engineering → Storage
+                                                                        ↓
+                                                               Analysis & Models
+```
+
+## Sensor Types and Data Characteristics
+
+| Sensor Type | Measurement | Typical Frequency | Key Patterns |
+|------------|-------------|-------------------|-------------|
+| Vibration | Acceleration (g) | 1-10 kHz | Harmonics, bearing defects |
+| Temperature | Degrees C/F | 1-60 sec intervals | Thermal trends, hotspots |
+| Pressure | PSI/Bar | 1-10 sec intervals | Leak detection, blockages |
+| Current/Voltage | Amps/Volts | 1-60 sec intervals | Motor degradation |
+| Flow rate | Volume/time | 1-60 sec intervals | Blockages, pump wear |
+| Acoustic | Decibels, frequency | 1-10 kHz | Cavitation, looseness |
+| Oil analysis | Particle count, viscosity | Daily-Monthly | Wear debris, contamination |
+
+## Data Quality Requirements
+
+- **Completeness**: Flag gaps > 5 minutes; interpolate gaps < 1 minute
+- **Accuracy**: Validate against known reference points quarterly
+- **Timeliness**: Edge-to-analysis latency < 30 seconds for critical assets
+- **Consistency**: Normalize units across sensor vendors and asset types
+
+## Feature Engineering
+
+Key derived features for failure prediction:
+
+- **Rolling statistics**: Mean, std dev, min, max over 1h/24h/7d windows
+- **Rate of change**: First and second derivatives of key measurements
+- **Frequency domain**: FFT for vibration, acoustic data
+- **Cross-sensor correlations**: Temperature-vibration, pressure-flow relationships
+- **Operating context**: Load level, ambient conditions, time since last maintenance
+
+## Edge Processing Rules
+
+- Filter obvious noise (sensor reboot spikes, communication errors)
+- Apply calibration offsets at the edge before transmission
+- Aggregate high-frequency data to summary statistics for long-term storage
+- Forward raw data only when anomaly triggers require investigation
+
+## Common Pitfalls
+
+### Ignoring Sensor Drift
+
+Wrong: Assume sensor readings are accurate indefinitely.
+
+Right: Schedule sensor calibration checks. Track measurement drift and apply corrections.
+
+### Drowning in Data
+
+Wrong: Store every raw reading at maximum frequency forever.
+
+Right: Tier storage: raw data for 30 days, summary statistics for 2 years, anomaly events permanently.