def main():
"""
:param arg:
输入文件路径：
文件树结构（例如）：
——arg
——人工 （此目录下的wav文件均会被读取，并将标记为人工井内施工）
20180310253.wav
——2018.1.3
20180103111.wav
——机械施工
——放缆
:return:
"""
with open('.\\test_config.json', encoding='UTF-8') as file:
test_config = json.load(file, strict=False)
path = test_config['test_data_path']
seg_param = test_config['seg_param']
seg = seg_param['seg'] # 4s分段
nw = seg_param['nw'] # 帧长约23ms*4
n_mfcc = seg_param['n_mfcc'] # mfcc 维数
save_file = test_config['hmm_model']
data_dict = load_audio(path, seg, nw, n_mfcc)
# save_file = '..\\model'
hmms_model = hmms()
model_list = common.find_ext_files(save_file, ext='.npy')
for model_file in model_list:
hmms_model.load_one_model(model_file) # 加载模型
# hmms_model.load_model(save_file)
model_num = len(hmms_model.hmms)
for key in data_dict:
print('当前预测音频文件夹：%s' % key)
instance = data_dict[key]
audio_num = instance.get_num()
species_count = np.zeros(model_num)
for j in range(audio_num):
# print('\t音频名：%s' % instance.audio_name[j])
frame_data = instance.frame[j]
length = instance.frame_num[j]
predicts = hmms_model.batch_predict(frame_data, length=length)
count = np.bincount(predicts)
# for i in range(len(count)):
# print('\t\t%s:%d' % (hmms_model.model_name[i], count[i]), end='\t')
major = np.argmax(count)
if hmms_model.model_name[major] == '背景音':
count[major] = 0
major = np.argmax(count)
species_count[major] += 1
# print('\n\t\t预测结果：%s\n' % hmms_model.model_name[major])
print('当前种类识别分布：')
species_count /= audio_num
for i in range(model_num):
print('\t%s: %f' % (hmms_model.model_name[i], species_count[i]))
def load_audio(path, seg, nw, n_mfcc):
# seg = 4 # 2s分段
# nw = 1024 # 帧长约23ms*4
# n_mfcc = 32 # mfcc 维数
folder_list = common.list_files(path)
data_dict = {}
for folder in folder_list:
species = folder.split('\\')[-1]
wav_list = common.find_ext_files(folder, '.wav')
instance = dataObejct()
for wav in wav_list:
wavname = wav.split('\\')[-1]
wav_data, fs = common.read_wav_file(wav)
# 多通道仅取一通道
if wav_data.ndim > 1:
wav_data = wav_data[:, 0]
wav_data = wav_data.T
ref_value = 2**12 - 1
wav_data = wav_data / ref_value # wave幅值归一化
filter_data = common.butter_worth_filter(wav_data,
cutoff=1000,
fs=fs,
btype='high',
N=8)
seg_mfcc, frame_num = enframe_and_feature_extract(
filter_data, seg, nw, fs, n_mfcc)
# # 输出为[n_mfcc, n_sample]
# mfcc_data = librosa.feature.mfcc(y=filter_data, sr=fs, n_mfcc=n_mfcc, n_fft=nw, hop_length=inc)
instance.append(audio_name=wavname,
origin=filter_data,
frame=seg_mfcc,
frame_num=frame_num)
data_dict[species] = instance
return data_dict
def load_and_train(path, seg, nw, n_mfcc, save_path):
"""
:param path: 音频存储路径
:param nw: 帧长
:param n_mfcc: mfcc特征维数
:return:
hmm_models, hmm模型对象
test_list，list[instance('frame':分帧数据 ,'origin':原始音频数据（滤波）, 'frame_num':帧数), ]
"""
hmm_models = hmms(n_iter=1000)
list_folders = common.list_files(path)
test_list = []
for i in range(len(list_folders)):
name = list_folders[i].split('\\')[-1]
config_name = os.path.join(path, name + '.json')
with open(config_name, encoding='UTF-8') as file:
config = json.load(file)
n_components = config['n_components']
n_mixs = config['n_mixs']
audio_num_for_train = config['audio num for train']
# audio_num_for_train=10
list_wavs = common.find_ext_files(list_folders[i], ext='.wav')
print('%d = %s num：%d' % (i, list_folders[i], len(list_wavs)))
list_wavs = shuffle_list(list_wavs)
instance = dataObejct()
instance.set_name(name)
debug = 0
for wavname in list_wavs:
if debug >= audio_num_for_train:
break
debug += 1
wav_data, fs = common.read_wav_file(wavname)
# 多通道仅取一通道
if wav_data.ndim > 1:
wav_data = wav_data[:, 0]
wav_data = wav_data.T
ref_value = 2**12 - 1
wav_data = wav_data / ref_value # wave幅值归一化
filter_data = common.butter_worth_filter(wav_data,
cutoff=1000,
fs=fs,
btype='high',
N=8)
seg_mfcc, frame_num = enframe_and_feature_extract(
filter_data, seg, nw, fs, n_mfcc)
# # 输出为[n_mfcc, n_sample]
# mfcc_data = librosa.feature.mfcc(y=filter_data, sr=fs, n_mfcc=n_mfcc, n_fft=nw, hop_length=inc)
instance.append(audio_name=wavname,
origin=filter_data,
frame=seg_mfcc,
frame_num=frame_num)
split_ = int(instance.get_num() / 2)
shuffled_instance = instance.shuffle()
train_samples = shuffled_instance[split_:]
# 确定训练集最大帧数，留作样本平衡使用。
train_samples.recompute_total()
train_samples.set_name(name)
frames = np.empty((0, n_mfcc))
frames_num = []
npy_name = name + '_' + str(n_mixs) + '_' + str(n_components) + '.npy'
save_name = os.path.join(save_path, npy_name)
if not os.path.exists(save_name):
for j in range(len(train_samples.origin)):
frame_data = train_samples.frame[j]
frame_data = frame_data.reshape((-1, n_mfcc))
frames = np.vstack((frames, frame_data))
frames_num += train_samples.frame_num[j]
if sum(frames_num) != frames.shape[0]:
print('sum(frames_num) = ', sum(frames_num))
print('total frames = ', frames.shape[0])
raise ValueError('sum(frames_num) != frames.shape[0]')
hmm_models.train_one(frames, frames_num, n_components, n_mixs,
name)
else:
print('\t模型%s已存在，加载模型' % npy_name)
hmm_models.load_one_model(save_name)
test_samples = shuffled_instance[:split_]
test_samples.set_name(name)
test_list.append(test_samples)
hmm_models.save_model(save_path)
return hmm_models, test_list
def load_optical_data(path, seg, nw, n_mfcc):
"""
:param path: 音频存储路径
:param nw: 帧长
:param n_mfcc: mfcc特征维数
:return:
train_list，list[instance('frame':分帧数据 ,'origin':原始音频数据（滤波）, 'frame_num':帧数), ]
test_list，list[instance('frame':分帧数据 ,'origin':原始音频数据（滤波）, 'frame_num':帧数), ]
"""
list_folders = common.list_files(path)
train_list = []
test_list = []
for i in range(len(list_folders)):
name = list_folders[i].split('\\')[-1]
list_wavs = common.find_ext_files(list_folders[i], ext='.wav')
print('%d = %s num：%d' % (i, list_folders[i], len(list_wavs)))
list_wavs = shuffle_list(list_wavs)
instance = dataObejct()
instance.set_name(name)
debug = 0
for wavname in list_wavs:
if debug >= 200:
break
debug += 1
wav_data, fs = common.read_wav_file(wavname)
# 多通道仅取一通道
if wav_data.ndim > 1:
wav_data = wav_data[:, 0]
wav_data = wav_data.T
ref_value = 2**12 - 1
wav_data = wav_data / ref_value # wave幅值归一化
filter_data = common.butter_worth_filter(wav_data,
cutoff=1000,
fs=fs,
btype='high',
N=8)
seg_mfcc, frame_num = enframe_and_feature_extract(
filter_data, seg, nw, fs, n_mfcc)
# # 输出为[n_mfcc, n_sample]
# mfcc_data = librosa.feature.mfcc(y=filter_data, sr=fs, n_mfcc=n_mfcc, n_fft=nw, hop_length=inc)
instance.append(audio_name=wavname,
origin=filter_data,
frame=seg_mfcc,
frame_num=frame_num)
split_ = int(instance.get_num() / 2)
shuffled_instance = instance.shuffle()
train_samples = shuffled_instance[split_:]
# 确定训练集最大帧数，留作样本平衡使用。
train_samples.recompute_total()
train_samples.set_name(name)
test_samples = shuffled_instance[:split_]
test_samples.set_name(name)
train_list.append(train_samples)
test_list.append(test_samples)
return train_list, test_list
def Run():
with open('sys_config.json', encoding='UTF-8') as file:
data = json.load(file)
src_path = data['src_path']
dst_path = data['dst_path']
hmm_param = data['hmm_param']
hmm_model_path = data['hmm_model']
seg = hmm_param['seg'] # hmm分段
nw = hmm_param['nw'] # 帧长
n_mfcc = hmm_param['n_mfcc'] # mfcc维数
# hmm 加载初始化
print("loading hmm models")
hmms = hmm_optical_sensing.hmms()
model_list = common.find_ext_files(hmm_model_path, ext='.npy')
for model_file in model_list:
hmms.load_one_model(model_file) # 加载模型
print("hmm models is loaded.")
# 根据类别名创建文件夹
print("building folder for alarm audio")
for name in hmms.model_name:
dir = os.path.join(dst_path, name)
common.mkdir(dir)
print("folder build.")
# SQL server 初始化
sql_param = data['sql_param']
isconnect = int(sql_param['isconnect'])
if isconnect == 1:
host = sql_param['host']
user = sql_param['user']
pwd = sql_param['pwd']
db = sql_param['db']
table = sql_param['table']
ms = MsSQL(host=host, user=user, pwd=pwd, db=db)
print('isconnect==1! the alarm will be insert to the SQL server!')
elif isconnect == 0:
ms = None
table = 'empty'
print('isconnect==0! the alarm will not be insert to the SQL server!')
else:
raise ValueError('isconnect value error! it should be 0 or 1!')
# ms.InsertAlarm(' ', ' ', ' ', ' ')
print("Detecting", end='')
count = 0
while True:
count += 1
if count % 5 == 0:
# print('\b\b\b\b', end='')
# print(' ', end='')
# print('\b\b\b\b', end='')
print("\rDetecting", end='')
else:
print('.', end='')
sys.stdout.flush()
wav_list = common.find_ext_files(src_path, ext='.wav')
time.sleep(1)
for wav in wav_list:
filename = wav.split('\\')[-1]
try:
predict_result = hmms.predict_wav(wav,
seg=seg,
nw=nw,
n_mfcc=n_mfcc)
except Exception as e:
print('\rfile %s detection error %s' % (filename, e))
continue
result = hmms.model_name[predict_result]
audio_save_path = os.path.join(dst_path, result, filename)
shutil.move(wav, audio_save_path)
# print("Event Occur: %s---%s" % (filename, result))
print(("\rEvent Occur: %s---%s" % (filename, result)), end=' ')
sys.stdout.flush()
# 添加到数据库
if ms is not None:
date = filename.split('.')[0]
# ms.InsertAlarm(table, date=date, alarmtype=result, audio_save_path=audio_save_path)
ms.UpdateAlarm(table,
date=date,
alarmtype=result,
audio_save_path=audio_save_path)
print('写入数据库')
else:
print('')
# 确保当前list的wav文件已被处理，不被重复处理。
for wav in wav_list:
if os.path.exists(wav):
os.remove(wav)