app.py

import gradio as gr
import numpy as np
import torch
import librosa
import soundfile as sf
from transformers import AutoFeatureExtractor, AutoModelForAudioClassification
import matplotlib.pyplot as plt
from matplotlib.colors import Normalize
import tempfile
import os

# Constants
SAMPLING_RATE = 16000
MODEL_NAME = "MIT/ast-finetuned-audioset-10-10-0.4593"
DEFAULT_THRESHOLD = 0.7

# Load model and feature extractor
feature_extractor = AutoFeatureExtractor.from_pretrained(MODEL_NAME)
model = AutoModelForAudioClassification.from_pretrained(MODEL_NAME)

# Equipment knowledge base
EQUIPMENT_RECOMMENDATIONS = {
    "bearing": {
        "high_frequency": "Recommend bearing replacement. High-frequency noise indicates wear or lubrication issues.",
        "low_frequency": "Check for improper installation or contamination in bearings.",
        "irregular": "Possible bearing cage damage. Schedule vibration analysis."
    },
    "pump": {
        "cavitation": "Pump cavitation detected. Check suction conditions and NPSH.",
        "impeller": "Impeller damage likely. Inspect and balance if needed.",
        "misalignment": "Misalignment detected. Perform laser shaft alignment."
    },
    "motor": {
        "electrical": "Electrical fault suspected. Check windings and connections.",
        "mechanical": "Mechanical imbalance detected. Perform dynamic balancing.",
        "bearing": "Motor bearing wear detected. Schedule replacement."
    },
    "compressor": {
        "valve": "Compressor valve leakage suspected. Perform valve test.",
        "pulsation": "Pulsation issues detected. Check dampeners and piping.",
        "surge": "Compressor surge condition. Review control settings."
    }
}

def analyze_frequency_patterns(audio, sr):
    """Analyze frequency patterns to identify potential issues"""
    patterns = []
    
    # Spectral analysis
    spectral_centroid = librosa.feature.spectral_centroid(y=audio, sr=sr)[0]
    spectral_rolloff = librosa.feature.spectral_rolloff(y=audio, sr=sr)[0]
    
    mean_centroid = np.mean(spectral_centroid)
    mean_rolloff = np.mean(spectral_rolloff)
    
    if mean_centroid > 3000:  # High frequency components
        patterns.append("high_frequency")
    elif mean_centroid < 1000:  # Low frequency components
        patterns.append("low_frequency")
        
    if mean_rolloff > 8000:  # Rich in harmonics
        patterns.append("harmonic_rich")
        
    return patterns

def generate_recommendation(prediction, confidence, audio, sr):
    """Generate maintenance recommendations based on analysis"""
    if prediction == "Normal":
        return "No immediate action required. Equipment operating within normal parameters."
    
    patterns = analyze_frequency_patterns(audio, sr)
    
    # Simple equipment type classifier based on frequency profile
    spectral_flatness = librosa.feature.spectral_flatness(y=audio)[0]
    mean_flatness = np.mean(spectral_flatness)
    
    if mean_flatness < 0.2:
        equipment_type = "bearing"
    elif 0.2 <= mean_flatness < 0.6:
        equipment_type = "pump"
    else:
        equipment_type = "motor" if np.mean(audio) < 0.1 else "compressor"
    
    # Generate specific recommendations
    recommendations = ["🔧 Maintenance Recommendations:"]
    recommendations.append(f"Detected issues in {equipment_type} with {confidence:.1%} confidence")
    
    for pattern in patterns:
        if pattern in EQUIPMENT_RECOMMENDATIONS.get(equipment_type, {}):
            recommendations.append(f"→ {EQUIPMENT_RECOMMENDATIONS[equipment_type][pattern]}")
    
    # General recommendations
    if prediction == "Anomaly":
        recommendations.append("\n🛠️ Suggested Actions:")
        recommendations.append("1. Isolate equipment if possible")
        recommendations.append("2. Perform visual inspection")
        recommendations.append("3. Schedule detailed diagnostics")
        recommendations.append(f"4. Review last maintenance records ({equipment_type})")
        
        if confidence > 0.8:
            recommendations.append("\n🚨 Urgent: High confidence abnormality detected. Recommend immediate inspection!")
    
    return "\n".join(recommendations)

def analyze_audio(audio_input, threshold=DEFAULT_THRESHOLD):
    """Process audio and detect anomalies"""
    try:
        # Handle file upload
        if isinstance(audio_input, str):
            audio, sr = sf.read(audio_input)
        else:  # Gradio file object
            with tempfile.NamedTemporaryFile(suffix='.wav', delete=False) as tmp:
                tmp.write(audio_input.read())
                tmp_path = tmp.name
            audio, sr = sf.read(tmp_path)
            os.unlink(tmp_path)
        
        # Convert to mono and resample if needed
        if len(audio.shape) > 1:
            audio = np.mean(audio, axis=1)
        if sr != SAMPLING_RATE:
            audio = librosa.resample(audio, orig_sr=sr, target_sr=SAMPLING_RATE)
        
        # Feature extraction and prediction
        inputs = feature_extractor(audio, sampling_rate=SAMPLING_RATE, return_tensors="pt")
        with torch.no_grad():
            outputs = model(**inputs)
            probs = torch.softmax(outputs.logits, dim=-1)
        
        # Get results
        predicted_class = "Normal" if probs[0][0] > threshold else "Anomaly"
        confidence = probs[0][0].item() if predicted_class == "Normal" else 1 - probs[0][0].item()
        
        # Generate spectrogram
        spectrogram = librosa.feature.melspectrogram(y=audio, sr=SAMPLING_RATE, n_mels=64, fmax=8000)
        db_spec = librosa.power_to_db(spectrogram, ref=np.max)
        
        fig, ax = plt.subplots(figsize=(10, 4))
        librosa.display.specshow(db_spec, x_axis='time', y_axis='mel', sr=SAMPLING_RATE, fmax=8000, ax=ax)
        plt.colorbar(format='%+2.0f dB')
        plt.title('Mel Spectrogram with Anomaly Detection')
        
        # Mark anomalies on plot
        if predicted_class == "Anomaly":
            plt.text(0.5, 0.9, 'ANOMALY DETECTED', color='red', 
                    ha='center', va='center', transform=ax.transAxes,
                    fontsize=14, bbox=dict(facecolor='white', alpha=0.8))
        
        spec_path = os.path.join(tempfile.gettempdir(), 'spec.png')
        plt.savefig(spec_path, bbox_inches='tight')
        plt.close()
        
        # Generate detailed recommendations
        recommendations = generate_recommendation(predicted_class, confidence, audio, SAMPLING_RATE)
        
        return (
            predicted_class,
            f"{confidence:.1%}",
            spec_path,
            recommendations
        )
        
    except Exception as e:
        return f"Error: {str(e)}", "", None, ""

# Gradio Interface
with gr.Blocks(title="Industrial Diagnostic Assistant 👨‍🔧", theme=gr.themes.Soft()) as demo:
    gr.Markdown("""
    # 🏭 Industrial Equipment Diagnostic Assistant
    ## Acoustic Anomaly Detection & Maintenance Recommendation System
    """)
    
    with gr.Row():
        with gr.Column():
            audio_input = gr.Audio(
                label="Upload Equipment Recording (.wav)",
                type="filepath",
                source="upload"
            )
            threshold = gr.Slider(
                minimum=0.5, maximum=0.95, step=0.05, value=DEFAULT_THRESHOLD,
                label="Detection Sensitivity", interactive=True
            )
            analyze_btn = gr.Button("🔍 Analyze & Diagnose", variant="primary")
            
        with gr.Column():
            result_label = gr.Label(label="Diagnosis Result")
            confidence = gr.Textbox(label="Confidence Score")
            spectrogram = gr.Image(label="Acoustic Analysis")
            recommendations = gr.Textbox(
                label="Maintenance Recommendations", 
                lines=10,
                interactive=False
            )
    
    analyze_btn.click(
        fn=analyze_audio,
        inputs=[audio_input, threshold],
        outputs=[result_label, confidence, spectrogram, recommendations]
    )
    
    gr.Markdown("""
    ### System Capabilities:
    - Automatic anomaly detection in industrial equipment sounds
    - Frequency pattern analysis to identify failure modes
    - Equipment-specific maintenance recommendations
    - Confidence-based urgency classification
    
    **Tip:** For best results, use 5-10 second recordings of steady operation
    """)

if __name__ == "__main__":
    demo.launch()