import gradio as gr import numpy as np import librosa import librosa.display import matplotlib matplotlib.use("Agg") import matplotlib.pyplot as plt import pywt import io from PIL import Image from tensorflow.keras.models import load_model from sklearn.preprocessing import StandardScaler import json import random import plotly.express as px # ================================ # CẤU HÌNH # ================================ SAMPLE_RATE = 22050 MAX_DURATION = 5 TIME_STEPS = 20 USE_DENOISE = True model = load_model("Huan_luyen_6_huhong.h5") def load_scaler_from_json(filepath): with open(filepath, 'r') as f: data = json.load(f) scaler = StandardScaler() scaler.mean_ = np.array(data['mean_']) scaler.scale_ = np.array(data['scale_']) scaler.n_features_in_ = len(scaler.mean_) return scaler scaler = load_scaler_from_json("scaler.json") with open("label_map.json", "r") as f: label_map = json.load(f) index_to_label = {v: k for k, v in label_map.items()} # ================================ # HÀM TIỀN XỬ LÝ # ================================ def denoise_wavelet(signal, wavelet='db8', level=4): coeffs = pywt.wavedec(signal, wavelet, level=level) sigma = np.median(np.abs(coeffs[-1])) / 0.6745 uthresh = sigma * np.sqrt(2 * np.log(len(signal))) coeffs_denoised = [pywt.threshold(c, value=uthresh, mode='soft') for c in coeffs] return pywt.waverec(coeffs_denoised, wavelet) def create_sequences(mfcc, time_steps=20): return np.array([mfcc[i:i+time_steps] for i in range(len(mfcc) - time_steps)]) def cat_2s_ngau_nhien(y, sr, duration=2): if len(y) < duration * sr: return y start = random.randint(0, len(y) - duration * sr) return y[start:start + duration * sr] # ================================ # VẼ ẢNH (numpy array) # ================================ def fig_to_numpy(fig): buf = io.BytesIO() fig.savefig(buf, format="png", dpi=90, bbox_inches="tight") buf.seek(0) img = Image.open(buf) plt.close(fig) return np.array(img) def tao_anh_mel(file_path): y, sr = librosa.load(file_path, sr=None, mono=True) y = cat_2s_ngau_nhien(y, sr) S = librosa.feature.melspectrogram(y=y, sr=sr, n_mels=128) S_dB = librosa.power_to_db(S, ref=np.max) fig, ax = plt.subplots(figsize=(6, 3)) librosa.display.specshow(S_dB, sr=sr, x_axis='time', y_axis='mel', ax=ax, cmap='magma') ax.set_title("Phổ tần Mel", fontsize=10) return fig_to_numpy(fig) def tao_wavelet_transform(file_path): y, sr = librosa.load(file_path, sr=None, mono=True) y = cat_2s_ngau_nhien(y, sr) coef, _ = pywt.cwt(y, scales=np.arange(1, 128), wavelet='morl', sampling_period=1/sr) fig, ax = plt.subplots(figsize=(6, 3)) ax.imshow(np.abs(coef), extent=[0, len(y)/sr, 1, 128], cmap='plasma', aspect='auto', origin='lower') ax.set_title("Phổ sóng con (Wavelet)") ax.set_xlabel("Thời gian (s)") ax.set_ylabel("Tần số (scale)") return fig_to_numpy(fig) def tao_waveform_image(file_path): y, sr = librosa.load(file_path, sr=None, mono=True) y = cat_2s_ngau_nhien(y, sr) fig, ax = plt.subplots(figsize=(6, 2.5)) librosa.display.waveshow(y, sr=sr, ax=ax, color='steelblue') ax.set_title("Biểu đồ Sóng Âm (Waveform)") ax.set_xlabel("Thời gian (s)") ax.set_ylabel("Biên độ") return fig_to_numpy(fig) def tao_waveform_denoise(file_path): y, sr = librosa.load(file_path, sr=None, mono=True) y = cat_2s_ngau_nhien(y, sr) y_denoised = denoise_wavelet(y) fig, ax = plt.subplots(3, 2, figsize=(10, 8)) # 1. Waveform librosa.display.waveshow(y, sr=sr, ax=ax[0,0], color='red') ax[0,0].set_title("Waveform - Trước lọc") librosa.display.waveshow(y_denoised, sr=sr, ax=ax[0,1], color='green') ax[0,1].set_title("Waveform - Sau lọc") # 2. FFT freqs = np.fft.rfftfreq(len(y), 1/sr) fft_y = np.abs(np.fft.rfft(y)) fft_y_denoised = np.abs(np.fft.rfft(y_denoised)) ax[1,0].plot(freqs, fft_y, color='red') ax[1,0].set_xlim(0, 8000) ax[1,0].set_title("FFT - Trước lọc") ax[1,1].plot(freqs, fft_y_denoised, color='green') ax[1,1].set_xlim(0, 8000) ax[1,1].set_title("FFT - Sau lọc") # 3. Spectrogram D1 = librosa.amplitude_to_db(np.abs(librosa.stft(y)), ref=np.max) D2 = librosa.amplitude_to_db(np.abs(librosa.stft(y_denoised)), ref=np.max) img1 = librosa.display.specshow(D1, sr=sr, x_axis='time', y_axis='log', ax=ax[2,0], cmap="magma") ax[2,0].set_title("Spectrogram - Trước lọc") fig.colorbar(img1, ax=ax[2,0], format="%+2.0f dB") img2 = librosa.display.specshow(D2, sr=sr, x_axis='time', y_axis='log', ax=ax[2,1], cmap="magma") ax[2,1].set_title("Spectrogram - Sau lọc") fig.colorbar(img2, ax=ax[2,1], format="%+2.0f dB") plt.tight_layout() return fig_to_numpy(fig) # ================================ # VẼ BIỂU ĐỒ Top-3 (Plotly) # ================================ def ve_top3_chart(probs): labels = [index_to_label[i] for i in range(len(probs))] values = probs * 100 top_idx = np.argsort(values)[::-1][:3] fig = px.pie( values=[values[i] for i in top_idx], names=[labels[i] for i in top_idx], title="Top-3 dự đoán" ) return fig # ================================ # DỰ ĐOÁN # ================================ def bao_san_sang(file_path): if not file_path: return "", None, None, None, None return ( "✅ Âm thanh đã sẵn sàng. Nhấn kiểm tra ngay!", tao_anh_mel(file_path), tao_wavelet_transform(file_path), tao_waveform_image(file_path), tao_waveform_denoise(file_path) ) def du_doan(file_path): if not file_path: return "❌ Chưa có âm thanh.", None signal, sr = librosa.load(file_path, sr=SAMPLE_RATE, mono=True) signal, _ = librosa.effects.trim(signal) signal = librosa.util.fix_length(signal, size=SAMPLE_RATE * MAX_DURATION) if USE_DENOISE: signal = denoise_wavelet(signal) mfcc = librosa.feature.mfcc(y=signal, sr=sr, n_mfcc=13).T mfcc = scaler.transform(mfcc) X_input = create_sequences(mfcc, time_steps=TIME_STEPS) if len(X_input) == 0: return "⚠️ Âm thanh quá ngắn để phân tích.", None y_preds = model.predict(X_input, verbose=0) avg_probs = np.mean(y_preds, axis=0) pred_index = np.argmax(avg_probs) confidence = avg_probs[pred_index] * 100 pred_label = "HƯ HỎNG KHÁC" if confidence < 50 else index_to_label[pred_index] html = f"""
📋 Kết Quả:
✅ Tình trạng: {pred_label.upper()}
📊 Độ tin cậy: {confidence:.2f}%
Xác suất từng lớp:
""" for i, prob in enumerate(avg_probs): html += f"- {index_to_label[i]}: {prob*100:.1f}%
" html += "
" return html, ve_top3_chart(avg_probs) # ================================ # RESET # ================================ def reset_output(): return "", None, None, None, None, "", None # ================================ # GIAO DIỆN # ================================ with gr.Blocks(css=""" #check-btn { background: #007acc; color: white; height: 48px; font-size: 16px; font-weight: bold; border-radius: 10px; } """) as demo: gr.HTML("""
""") gr.Markdown("""

CHẨN ĐOÁN HƯ HỎNG TỪ ÂM THANH ĐỘNG CƠ

""") with gr.Row(): audio_file = gr.Audio(type="filepath", label="📂 Tải File Âm Thanh", interactive=True) audio_mic = gr.Audio(type="filepath", label="🎤 Ghi Âm", sources=["microphone"], interactive=True) thong_bao_ready = gr.HTML() btn_check = gr.Button("🔍 KIỂM TRA NGAY", elem_id="check-btn") output_html = gr.HTML() with gr.Accordion("📊 Phân tích Âm Thanh", open=False): mel_output = gr.Image(label="Mel Spectrogram", type="numpy") wavelet_output = gr.Image(label="Wavelet Transform", type="numpy") waveform_output = gr.Image(label="Waveform", type="numpy") waveform_denoise_output = gr.Image(label="So sánh", type="numpy") top3_chart = gr.Plot(label="Top 3 dự đoán") # --- Upload/ghi âm → chỉ báo sẵn sàng + vẽ ảnh audio_file.change( fn=bao_san_sang, inputs=audio_file, outputs=[thong_bao_ready, mel_output, wavelet_output, waveform_output, waveform_denoise_output] ) audio_mic.change( fn=bao_san_sang, inputs=audio_mic, outputs=[thong_bao_ready, mel_output, wavelet_output, waveform_output, waveform_denoise_output] ) # --- Nút kiểm tra → chỉ dự đoán btn_check.click( fn=lambda f1, f2: du_doan(f1 if f1 else f2), inputs=[audio_file, audio_mic], outputs=[output_html, top3_chart] ) # --- Reset khi clear audio_file.clear(fn=reset_output, outputs=[ thong_bao_ready, mel_output, wavelet_output, waveform_output, waveform_denoise_output, output_html, top3_chart ]) audio_mic.clear(fn=reset_output, outputs=[ thong_bao_ready, mel_output, wavelet_output, waveform_output, waveform_denoise_output, output_html, top3_chart ]) demo.launch()