def torch4FROLS_pipeline(data_type, configs, n_trial):
"""基于 Kalman 滤波器的各个算法的 pipeline
Args:
data_type: 数据类型
configs (dict): 配置字典
n_trial: 试验次数
"""
config = configs[data_type]
WGCI100 = []
for trial in range(n_trial):
term_selector = Selector(
f"{config['data_root']}{data_type}{config['term_path']}{trial+1}.mat"
)
# get data
normalized_signals, Kalman_H, candidate_terms, Kalman_S_No = term_selector.make_selection(
)
print(f'data_type: {data_type}, trial: ### {trial+1}')
kf = torch4FROLS(normalized_signals,
Kalman_H=Kalman_H,
n_epoch=config['n_epoch'])
y_coef = kf.estimate_coef()
# 总体误差
whole_y_error = np.var(kf.y_error, 0)
# 子模型情况
terms_mat = sio.loadmat(
f"{config['data_root']}{data_type}{config['term_path']}_WGCI{trial+1}.mat"
)
sub1 = []
for ch in range(5):
data_set = {
'normalized_signals': terms_mat['normalized_signals'],
'Hv': terms_mat['Hv'],
'Kalman_H': terms_mat['Kalman_H'][0, ch],
'terms_chosen': terms_mat['terms_chosen'][0, ch]
}
term_selector = Selector(data_set)
# get data
normalized_signals = term_selector.make_selection()[0]
# 构造 Kalman Filter
kf = torch4FROLS(normalized_signals,
Kalman_H=Kalman_H,
n_epoch=config['n_epoch'])
# 估计系数
y_coef = kf.estimate_coef()
# 误差
sub_y_error = np.var(kf.y_error, 0)
sub1.append(np.log(sub_y_error / whole_y_error))
WGCI100.append(np.asarray(sub1).T)
mean_WGCI = np.mean(WGCI100, 0)
var_WGCI = np.var(WGCI100, 0)
print(f"mean_WGCI = {mean_WGCI}, var_WGCI = {var_WGCI}")
fname1 = f"{config['data_root']}{data_type}_kalman4FROLS100_{config['est_fname']}WGCI100.txt"
save_3Darray(
fname1,
np.array(
[mean_WGCI * (mean_WGCI > 0.01), var_WGCI * (var_WGCI > 0.01)]))
def torch4FROLS_pipeline(data_type, configs, n_trial):
"""基于 Kalman 滤波器的各个算法的 pipeline
Args:
data_type: 数据类型
configs (dict): 配置字典
n_trial: 试验次数
"""
config = configs[data_type]
fname = f"{config['data_root']}{data_type}_torch4FROLS100_{config['est_fname']}"
term_selector = Selector(
f"{config['data_root']}{data_type}{config['term_path']}")
terms_set = term_selector.make_terms()
# get data
normalized_signals, Kalman_H, candidate_terms, Kalman_S_No = term_selector.make_selection(
)
y_coef = 0
for trial in range(n_trial):
print(f'data_type: {data_type}, trial: ### {trial+1}')
kf = torch4FROLS(normalized_signals,
Kalman_H=Kalman_H,
n_epoch=config['n_epoch'])
y_coef += kf.estimate_coef()
est_model = make_func4K4FROLS(y_coef / n_trial,
candidate_terms,
Kalman_S_No,
fname=fname)
# 输出结果
print(est_model)
def kalman4ARX_pipeline(data_type, configs, n_trial):
"""基于 Kalman 滤波器的各个算法的 pipeline
types = ['linear', 'longlag_linear']
Args:
data_type: 数据类型
configs (dict): 配置字典
n_trial: 试验次数
"""
config = configs[data_type]
term_selector = Selector(
f"{config['data_root']}{data_type}{config['term_path']}")
terms_set = term_selector.make_terms()
# get data
normalized_signals = term_selector.make_selection()[0]
fname = f"{config['data_root']}{data_type}_kalman4ARX100_{config['est_fname']}"
A_coef0 = 0
for trial in range(n_trial):
print(f'data_type: {data_type}, trial: ### {trial+1}')
# 构造 Kalman Filter
kf = Kalman4ARX(normalized_signals, config['max_lag'], uc=config['uc'])
# 估计系数
y_coef, A_coef = kf.estimate_coef(config['threshold'])
A_coef0 += A_coef
# 计算模型表达式并保存
est_model = make_linear_func(A_coef0 / n_trial, fname=fname)
# 输出结果
print(est_model)
def frokf(noise_var,
ndim,
dtype,
terms,
length,
root='../data/',
trials=100,
uc=0.01,
ntest=50):
assert dtype in ['linear', 'nonlinear'], 'type not support!'
ax = []
for trial in range(1, trials + 1):
# for trial in [trials]:
terms_path = root + f'{dtype}_terms{ndim}D_{noise_var:2.2f}trial{trial}.mat'
term = Selector(terms_path)
_ = term.make_terms()
normalized_signals, Kalman_H, candidate_terms, Kalman_S_No = term.make_selection(
)
# Kalman_S_No = np.sort(Kalman_S_No)
y_coef = 0
# 对FROKF多次实验取平均值
for _ in trange(ntest):
kf = Kalman4FROLS(normalized_signals, Kalman_H=Kalman_H, uc=uc)
y_coef += kf.estimate_coef()
y_coef /= ntest
terms_set = corr_term(y_coef, candidate_terms, Kalman_S_No)
flatten_coef, t = [], 0
for i in range(ndim):
tmp = []
for k in terms[t:t + length[i]]:
tmp.append(terms_set[i][k] if k in terms_set[i] else np.nan)
flatten_coef.extend(tmp)
t += length[i]
ax.append(flatten_coef)
return np.stack(ax)
def test_torch4FROLS(self):
terms_path = 'test_data/linear_terms.mat'
term = Selector(terms_path)
normalized_signals, Kalman_H, _, _ = term.make_selection()
kf = torch4FROLS(normalized_signals, Kalman_H, n_epoch=100)
y_coef = kf.estimate_coef()
self.assertTrue(isinstance(kf, torch4FROLS))
self.assertTrue(isinstance(y_coef, np.ndarray))
def kalman_pipeline(configs):
"""基于 Kalman 滤波器的各个算法的 pipeline
Args:
configs (dict): 配置字典
"""
for data_type in configs.keys():
config = configs[data_type]
term_selector = Selector(
f"{config['data_root']}{data_type}{config['term_path']}")
terms_set = term_selector.make_terms()
# get data
normalized_signals, Kalman_H, candidate_terms, Kalman_S_No = term_selector.make_selection(
)
# 构造 Kalman Filter
for algo_type in config['algorithm']:
print(f"{data_type} {algo_type}")
fname = f"{config['data_root']}{data_type}_{algo_type}_{config['est_fname']}"
if algo_type == 'Kalman4ARX':
kf = Kalman4ARX(normalized_signals,
config['max_lag'],
uc=config['uc'])
# 估计系数
y_coef, A_coef = kf.estimate_coef(config['threshold'])
# 计算模型表达式并保存
est_model = make_linear_func(A_coef, fname=fname)
# 保存平均结果
elif algo_type == 'Kalman4FROLS':
kf = Kalman4FROLS(normalized_signals,
Kalman_H=Kalman_H,
uc=config['uc'])
y_coef = kf.estimate_coef()
est_model = make_func4K4FROLS(y_coef,
candidate_terms,
Kalman_S_No,
fname=fname)
# 保存平均结果
elif algo_type == 'torch4FROLS':
kf = torch4FROLS(normalized_signals,
Kalman_H=Kalman_H,
n_epoch=config['n_epoch'])
y_coef = kf.estimate_coef()
est_model = make_func4K4FROLS(y_coef,
candidate_terms,
Kalman_S_No,
fname=fname)
# 保存平均结果
else:
print('!Not Defined!')
print(f"\n{data_type}_{algo_type} est model saved!\n")
# 输出结果
print(est_model)
def torch4FROLS_pipeline(data_type, configs, n_trial, id_correct, n_correct):
"""基于 Kalman 滤波器的各个算法的 pipeline
Args:
data_type: 数据类型
configs (dict): 配置字典
n_trial: 试验次数
"""
config = configs[data_type]
y_coef100 = np.zeros((100, 5, 5))
y_coef9 = np.zeros((100, 9))
for trial in range(n_trial):
fname = f"{config['data_root']}{data_type}_torch4FROLS100_{config['est_fname']}{trial+1}.txt"
term_selector = Selector(
f"{config['data_root']}{data_type}{config['term_path']}{trial+1}.mat"
)
terms_set = term_selector.make_terms()
# get data
normalized_signals, Kalman_H, candidate_terms, Kalman_S_No = term_selector.make_selection(
)
print(f'data_type: {data_type}, trial: ### {trial+1}')
kf = torch4FROLS(normalized_signals,
Kalman_H=Kalman_H,
n_epoch=config['n_epoch'])
y_coef = kf.estimate_coef()
print(kf.y_error, kf.y_error.shape)
y_coef100[trial] = y_coef
est_model = make_func4K4FROLS(y_coef,
candidate_terms,
Kalman_S_No,
fname=fname)
coef9 = []
Kalman_S_No_order = np.sort(Kalman_S_No)
for row in range(5):
for t in range(n_correct[row]):
idx = np.argwhere(
Kalman_S_No_order[row, :] == id_correct[data_type][row][t])
value = y_coef[row, idx]
coef9.append(value[0, 0])
y_coef9[trial] = np.array(coef9)
# 输出结果
# print(est_model)
fname1 = f"{config['data_root']}{data_type}_torch4FROLS100_{config['est_fname']}log.txt"
save_3Darray(fname1, y_coef100)
mean_y = np.mean(y_coef9, 0)
var_y = np.var(y_coef9, 0)
print(mean_y, var_y, sep='\n')
fname1 = f"{config['data_root']}{data_type}_torch4FROLS100_{config['est_fname']}log100.txt"
save_2Darray(fname1, np.array([mean_y, var_y]).T)
def kalman4ARX_pipeline(data_type, configs, n_trial):
"""基于 Kalman 滤波器的各个算法的 pipeline
types = ['linear', 'longlag_linear']
Args:
data_type: 数据类型
configs (dict): 配置字典
n_trial: 试验次数
"""
config = configs[data_type]
if data_type == 'linear':
y_coef100 = np.zeros((100, 5, 25))
else:
y_coef100 = np.zeros((100, 5, 50))
for trial in range(n_trial):
term_selector = Selector(
f"{config['data_root']}{data_type}{config['term_path']}{trial+1}.mat"
)
terms_set = term_selector.make_terms()
# get data
normalized_signals = term_selector.make_selection()[0]
fname = f"{config['data_root']}{data_type}_kalman4ARX100_{config['est_fname']}{trial+1}.txt"
print(f'data_type: {data_type}, trial: ### {trial+1}')
# 构造 Kalman Filter
kf = Kalman4ARX(normalized_signals, config['max_lag'], uc=config['uc'])
# 估计系数
y_coef, A_coef = kf.estimate_coef(config['threshold'])
y_coef100[trial] = y_coef
# 计算模型表达式并保存
est_model = make_linear_func(A_coef, fname=fname)
# 输出结果
# print(est_model)
fname1 = f"{config['data_root']}{data_type}_kalman4ARX100_{config['est_fname']}log.txt"
save_3Darray(fname1, y_coef100)
mean_y = np.mean(y_coef100, 0)
var_y = np.var(y_coef100, 0)
print(mean_y, var_y, sep='\n')
fname1 = f"{config['data_root']}{data_type}_kalman4ARX100_{config['est_fname']}log100.txt"
save_3Darray(fname1, np.array([mean_y, var_y]))
print(est_model)
# !Kalman4FROLS 测试
# !Selector 测试
# !线性模型
terms_path = './data/nor_linear_terms.mat'
term = Selector(terms_path)
terms_repr = term.make_terms()
# *保存候选项集合
fname = './data/linear_candidate_terms.txt'
np.savetxt(fname, terms_repr, fmt='%s')
# *selection
print(term)
normalized_signals, Kalman_H, candidate_terms, Kalman_S_No = term.make_selection(
)
# *构造 Kalman Filter
kf = Kalman4FROLS(normalized_signals, Kalman_H=Kalman_H, uc=0.01)
y_coef = kf.estimate_coef()
print(y_coef)
# *估计模型生成
est_model = make_func4K4FROLS(y_coef,
candidate_terms,
Kalman_S_No,
fname='./data/K4FROLS_est_model.txt')
print(est_model)
# !非线性模型
terms_path = './data/nor_nonlinear_terms.mat'
def kalman4ARX_pipeline(data_type, configs, n_trial):
"""基于 Kalman 滤波器的各个算法的 pipeline
types = ['linear', 'longlag_linear']
Args:
data_type: 数据类型
configs (dict): 配置字典
n_trial: 试验次数
"""
config = configs[data_type]
WGCI100 = []
for trial in range(n_trial):
# 计算总体情况
term_selector = Selector(
f"{config['data_root']}{data_type}{config['term_path']}{trial+1}.mat"
)
# get data
normalized_signals = term_selector.make_selection()[0]
print(f'data_type: {data_type}, trial: ### {trial+1}')
# 构造 Kalman Filter
kf = Kalman4ARX(normalized_signals, config['max_lag'], uc=config['uc'])
# 估计系数
y_coef, A_coef = kf.estimate_coef(config['threshold'])
# 总体误差
whole_y_error = np.var(kf.y_error, 0)
# 子模型情况
terms_mat = sio.loadmat(
f"{config['data_root']}{data_type}{config['term_path']}_WGCI{trial+1}.mat"
)
sub1 = []
for ch in range(5):
data_set = {
'normalized_signals': terms_mat['normalized_signals'],
'Hv': terms_mat['Hv'],
'Kalman_H': terms_mat['Kalman_H'][0, ch],
'terms_chosen': terms_mat['terms_chosen'][0, ch]
}
term_selector = Selector(data_set)
# get data
normalized_signals = term_selector.make_selection()[0]
# 构造 Kalman Filter
kf = Kalman4ARX(normalized_signals,
config['max_lag'],
uc=config['uc'])
# 估计系数
y_coef, A_coef = kf.estimate_coef(config['threshold'])
# 误差
sub_y_error = np.var(kf.y_error, 0)
sub1.append(np.log(sub_y_error / whole_y_error))
WGCI100.append(np.asarray(sub1).T)
mean_WGCI = np.mean(WGCI100, 0)
var_WGCI = np.var(WGCI100, 0)
print(f"mean_WGCI = {mean_WGCI}, var_WGCI = {var_WGCI}")
fname1 = f"{config['data_root']}{data_type}_kalman4ARX100_{config['est_fname']}WGCI100.txt"
save_3Darray(
fname1,
np.array(
[mean_WGCI * (mean_WGCI > 0.01), var_WGCI * (var_WGCI > 0.01)]))