better output and metrices

README.md

@@ -14,10 +14,12 @@ AI-powered predictive maintenance for solar PV inverters using ML models and LLM
 
 ## Features
 
-- **Anomaly Detection**: Isolation Forest + LSTM Autoencoder
-- **Failure Forecasting**: XGBoost models for time-to-failure and failure probability
-- **AI Diagnosis**: Gemini LLM provides maintenance recommendations
-- **REST API**: FastAPI-based service with automatic documentation
+- **Advanced Anomaly Detection**: Isolation Forest + LSTM Autoencoder with reconstruction error analysis
+- **Comprehensive Health Monitoring**: Real-time health scoring, anomaly detection, and trend analysis
+- **Fault Classification**: Multi-class fault type prediction (Normal, Short Circuit, Degradation, Open Circuit, Shadowing)
+- **Predictive Maintenance**: XGBoost models for time-to-failure, remaining useful life, and failure probability
+- **AI Diagnosis**: Gemini LLM provides detailed maintenance recommendations with urgency levels
+- **REST API**: FastAPI-based service with automatic documentation and validation
 
 ## Quick Start
 
@@ -70,32 +72,67 @@ Analyzes sensor data and returns ML predictions with optional AI diagnosis.
 {
   "ml_output": {
     "asset_id": "PV_INVERTER_001",
+    "health_score": 0.892,
+    "anomaly_score": 0.0342,
+    "health_trend_200step": -0.0123,
     "failure_probability": 0.12,
     "expected_ttf_days": 450.5,
     "expected_rul_days": 9800.0,
+    "predicted_fault_type": "Normal",
+    "fault_confidence": 0.78,
     "confidence": 0.85
   },
   "agent_output": {
-    "diagnosis": "Minor voltage fluctuations detected...",
-    "urgency": "Low",
-    "recommended_action": "Continue monitoring...",
-    "justification": ["Voltage within range", "Current stable"]
+    "system_health": {
+      "status": "Good",
+      "health_score_explanation": "Health score of 0.892 indicates good system performance at 89% of optimal",
+      "anomaly_analysis": "Anomaly score of 0.0342 shows normal operation with minimal deviations",
+      "trend_assessment": "Slight negative trend (-0.0123) suggests gradual performance decline to monitor"
+    },
+    "failure_risk": {
+      "probability_assessment": "12% failure probability within 30 days represents low risk",
+      "time_to_failure_analysis": "450 days until predicted failure allows time for scheduled maintenance",
+      "remaining_life_summary": "9,800 days remaining useful life indicates good system longevity"
+    },
+    "fault_diagnosis": {
+      "primary_fault": "System classified as 'Normal' with 78% confidence, no major faults detected",
+      "confidence_interpretation": "85% overall confidence provides reliable analysis for decision making"
+    },
+    "maintenance_recommendations": {
+      "urgency_level": "Low",
+      "immediate_actions": [
+        "Continue routine performance monitoring",
+        "Schedule panel inspection within 3 months"
+      ],
+      "monitoring_schedule": "Next comprehensive inspection in 90 days",
+      "preventive_measures": "Implement quarterly maintenance checks"
+    },
+    "key_insights": [
+      "System health is good but showing early degradation signs",
+      "No immediate action required, continue standard monitoring",
+      "Schedule maintenance before 450 days to prevent issues"
+    ]
   }
 }
 ```
 
 ### Data Processing Pipeline
 
-1. **Input Validation**: Ensures voltage/current arrays match and contain sufficient data points
-2. **Data Preparation**: Pads input to 100 data points for consistent processing
+1. **Input Validation**: Ensures voltage/current/temperature arrays match and contain sufficient data points
+2. **Data Preparation**: Filters active periods and downsamples for efficient processing
 3. **Feature Engineering**: Creates 10 statistical features using rolling window analysis:
    - Voltage mean/standard deviation
-   - Power mean/standard deviation
+   - Power mean/standard deviation  
    - Power delta and slope
    - Normalized efficiency
    - Temperature mean/standard deviation/delta
-4. **ML Inference**: Processes features through anomaly detection and prediction models
-5. **Agent Analysis**: LLM analyzes ML results for human-readable diagnosis (if API key provided)
+4. **ML Inference**: 
+   - Anomaly detection via Isolation Forest + LSTM reconstruction error
+   - Health scoring and trend analysis
+   - Fault type classification (5 classes)
+   - Time-to-failure and remaining useful life prediction
+   - Failure probability calculation with weighted risk factors
+5. **Agent Analysis**: LLM analyzes all ML metrics for comprehensive maintenance diagnosis (if API key provided)
 
 ## Configuration
 

agent/prompts.yaml

@@ -1,19 +1,54 @@
 maintenance:
   system: |
-    You are a maintenance decision AI.
-    You must reason ONLY from the provided JSON.
-    Do NOT invent data.
+    You are an expert photovoltaic (PV) maintenance AI specializing in solar panel fault diagnosis and predictive maintenance.
+    
+    PARAMETER MEANINGS:
+    - health_score: Current system health (0-1, higher = healthier)
+    - anomaly_score: LSTM reconstruction error (lower = more normal)
+    - health_trend_200step: Health change over last 200 data points (positive = improving, negative = degrading)
+    - failure_probability: Risk of failure in next 30 days (0-1, higher = higher risk)
+    - expected_ttf_days: Days until predicted failure
+    - expected_rul_days: Remaining useful life in days
+    - predicted_fault_type: Specific fault classification (Normal, Short Circuit, Degradation, Open Circuit, Shadowing)
+    - fault_confidence: Confidence in fault type prediction (0-1)
+    - confidence: Overall prediction confidence (0-1)
+    
+    You must reason ONLY from the provided JSON data. Do NOT invent information.
   
   user_template: |
-    INPUT:
+    ANALYZE THIS PV SYSTEM DATA FOR MAINTENANCE INSIGHTS:
     {phase2_output}
     
+    Provide a concise maintenance analysis explaining what each parameter means for this PV system.
+    
     MANDATORY: Return output strictly in JSON format only. Do not include any markdown, code blocks, or extra text.
     
     OUTPUT FORMAT:
     {{
-      "diagnosis": "...",
-      "urgency": "Low | Medium | High",
-      "recommended_action": "...",
-      "justification": ["...", "..."]
+      "system_health": {{
+        "status": "Excellent | Good | Fair | Poor | Critical",
+        "health_score_explanation": "Brief explanation of health_score meaning",
+        "anomaly_analysis": "Brief explanation of anomaly_score implications",
+        "trend_assessment": "Brief explanation of health_trend_200step significance"
+      }},
+      "failure_risk": {{
+        "probability_assessment": "Brief explanation of failure_probability meaning",
+        "time_to_failure_analysis": "Brief explanation of expected_ttf_days implications",
+        "remaining_life_summary": "Brief explanation of expected_rul_days significance"
+      }},
+      "fault_diagnosis": {{
+        "primary_fault": "Brief explanation of predicted_fault_type and fault_confidence",
+        "confidence_interpretation": "Brief explanation of overall confidence level"
+      }},
+      "maintenance_recommendations": {{
+        "urgency_level": "Low | Medium | High | Critical",
+        "immediate_actions": ["2 specific, actionable steps"],
+        "monitoring_schedule": "Recommended next check timing",
+        "preventive_measures": "One key long-term strategy"
+      }},
+      "key_insights": [
+        "Most important finding",
+        "Second key insight", 
+        "Third insight with implication"
+      ]
     }}