# -*- coding: utf-8 -*- import numpy as np import traceback import pandas as pd from json import JSONDecodeError from scipy.stats import norm from scipy.stats import f from scipy.stats.distributions import chi2 import json import sys import requests import datetime import jenkspy import xlrd from smote import smote import config from recon import sequence,sequence_d def train(XMat, p): m = np.array(XMat).shape[1] # 取参数个数，即矩阵列数 average = np.mean(XMat, axis=0) # axis=0，分别对各列求均值 std = np.std(XMat, axis=0) # axis=0，计算各列的标准差 m, n = np.shape(XMat) # 取矩阵的行数m和列数n avgs = np.tile(average, (m, 1)) # 将average的数乘1倍，再以此为基，复制到每一行，变成m行的矩阵 stds = np.tile(std, (m, 1)) # 将std的数乘1倍，再以此为基，复制到每一行，变成m行的矩阵 data_adjust = np.divide(XMat - avgs, stds) # 计算XMat-avgs，即每个元素与列平均值之间的差，再用此差除以每列数据的标准差 covX = np.cov(data_adjust.T) # 计算协方差矩阵（对称阵） # corr=np.corrcoef(data_adjust.T) featValue, featVec = np.linalg.eig(covX) # 求解协方差矩阵的特征值和特征向量 # ############可能协方差为0（一列数据相同，在训练的时候不要选择一样的数据，就不能除。############################ featValue = np.real(featValue) # 返回复杂参数的实部 featVec = np.real(featVec) # 返回复杂参数的实部 index = np.argsort(-featValue) # 按照featValue进行从大到小排序 -featValue转置后把元素变成其相反数 featValue = featValue[index] # 按照排序进行重建 featVec = featVec[:, index] featValue_sum = np.divide(featValue, np.sum(featValue)) # 特征值分别除以特征值之和 per = 0 # 特征值百分比 k = 0 # 主元个数 for precent in featValue_sum: per += precent k = k + 1 if per > p: break # 记录主元数小与设定值p的总值和个数 finalData = [] if k > n: # 如果k比总数大，就返回 k必须小于特征数 print "k must lower than feature number" return else: # 注意特征向量是列向量，而numpy的二维矩阵(数组)a[m][n]中，a[1]表示第1行值 selectVec = np.matrix(featVec[:, :k]) # 所以这里需要进行转置 取第k列的特征向量，将ndarray对象转为matrix矩阵 finalData = np.dot(data_adjust, selectVec).dot(selectVec.T) # 将data_adjust，selectVec，selectVec.T三者相乘 reconData = np.add(np.multiply(finalData, stds), avgs) # 重构值 将finalData和stds相乘，然后与avgs相加 Train_X_min = np.min(XMat, axis=0) # 训练值最小值 Train_X_max = np.max(XMat, axis=0) # 训练值最大值 Train_X_mean = np.mean(XMat, axis=0) # 训练值平均值 Train_X_std = np.std(XMat, axis=0) # 训练值方差 Train_X_bais = XMat - reconData # 训练值偏差 Train_X_bais_max = np.max(np.abs(Train_X_bais), axis=0) # 训练值偏差最大值 axis=0 对各列求 Train_X_bais_min = np.min(np.abs(Train_X_bais), axis=0) # 训练值偏差最小值 Train_X_bais_mean = np.mean(np.abs(Train_X_bais), axis=0) # 训练值偏差平均值 Train_X_bais_std_upperB95 = np.array(np.abs(1.96 * np.std(Train_X_bais, axis=0) + Train_X_bais_mean))[ 0] # 训练值偏差标准差 Train_X_bais_std_upperB99 = np.array(np.abs(2.58 * np.std(Train_X_bais, axis=0) + Train_X_bais_mean))[0] Train_X_bais_std_lowerB95 = np.array(np.abs(1.96 * np.std(Train_X_bais, axis=0) - Train_X_bais_mean))[ 0] # 训练值偏差标准差 Train_X_bais_std_lowerB99 = np.array(np.abs(2.58 * np.std(Train_X_bais, axis=0) - Train_X_bais_mean))[0] QCUL_95_line = [] # 限值 QCUL_99_line = [] for index1 in range(len(Train_X_bais_std_upperB95)): QCUL_95_line.append(max(Train_X_bais_std_upperB95[index1], Train_X_bais_std_lowerB95[index1])) QCUL_99_line.append(max(Train_X_bais_std_upperB99[index1], Train_X_bais_std_lowerB99[index1])) QCUL_95_line = np.array(QCUL_95_line) QCUL_99_line = np.array(QCUL_99_line) ################################################################################# # 计算阈值----------------QUCL--------------------################################################ theta1 = np.sum(featValue[k:]) theta2 = np.sum(np.power(featValue[k:], 2)) theta3 = np.sum(np.power(featValue[k:], 3)) h0 = 1 - 2 * theta1 * theta3 / (3 * np.power(theta2, 2)) ca_95 = norm.ppf(0.95, loc=0, scale=1) QCUL_95 = theta1 * np.power( h0 * ca_95 * np.sqrt(2 * theta2) / theta1 + 1 + theta2 * h0 * (h0 - 1) / np.power(theta1, 2), 1 / h0) # 置信域为百分之95 # QCUL_95_line = Train_X_bais_std*2.58 # +Train_X_mean#反归一化阈值 ca_99 = norm.ppf(0.99, loc=0, scale=1) QCUL_99 = theta1 * np.power( (h0 * ca_99 * np.sqrt(2 * theta2) / theta1 + 1 + theta2 * h0 * (h0 - 1) / np.power(theta1, 2)), 1 / h0) # 置信域为百分之99 # QCUL_99_line = Train_X_bais_std*1.96 # + Train_X_mean # 反归一化阈值 # 计算阈值----------------T2UCL--------------------########################################### f_95 = f.ppf(0.95, k, m - k) T2CUL_95 = k * (m - 1) * (m + 1) * f_95 / (m * (m - k)) # 置信域为百分之95 T2CUL_95_line = np.sqrt(T2CUL_95) * Train_X_std / np.sqrt(m) # +Train_X_mean#反归一化阈值 f_99 = f.ppf(0.99, k, m - k) T2CUL_99 = k * (m - 1) * (m + 1) * f_99 / (m * (m - k)) # 置信域为百分之99 T2CUL_99_line = np.sqrt(T2CUL_99) * Train_X_std / np.sqrt(m) # +Train_X_mean#反归一化阈值 # 计算阈值----------------综合--------------------################################################# gfi_95 = (k / pow(T2CUL_95, 2) + theta2 / pow(QCUL_95, 2)) / (k / T2CUL_95 + theta1 / QCUL_95) hfi_95 = pow((k / T2CUL_95 + theta1 / QCUL_95), 2) / (k / pow(T2CUL_95, 2) + theta2 / pow(QCUL_95, 2)) Kesi_95 = gfi_95 * chi2.ppf(0.95, hfi_95) # 卡方分布 Kesi_95_line = np.sqrt(Kesi_95) * Train_X_std / np.sqrt(m) # 反归一化阈值 gfi_99 = (k / pow(T2CUL_99, 2) + theta2 / pow(QCUL_99, 2)) / (k / T2CUL_99 + theta1 / QCUL_99) hfi_99 = pow((k / T2CUL_99 + theta1 / QCUL_99), 2) / (k / pow(T2CUL_99, 2) + theta2 / pow(QCUL_99, 2)) Kesi_99 = gfi_99 * chi2.ppf(0.99, hfi_99) # 卡方分布 Kesi_99_line = np.sqrt(Kesi_99) * Train_X_std / np.sqrt(m) # 反归一化阈值 # 计算对应的指标矩阵 numbel_variable = featValue.shape[0] selectVec = featVec[:, 0:k] featValue_sort = featValue # [index] # 排序后的特征值 C_ = np.eye(numbel_variable) - np.dot(selectVec, selectVec.T) X_SPE = C_.T # 99 II99 = featValue_sort.copy() II99[:k] = II99[:k] * T2CUL_99 II99[K:] = QCUL_99 # 95 II95 = featValue_sort.copy() II95[:k] = II95[:k] * T2CUL_95 II95[K:] = QCUL_95 DIAG_Fai99 = np.linalg.inv(np.diag(II99)) DIAG_Fai95 = np.linalg.inv(np.diag(II95)) D_Fai = featVec.copy() DIAG_T2 = np.linalg.inv(np.diag(featValue_sort[:k])) D_T2 = selectVec.copy() m_spe = X_SPE @ X_SPE.T m_fai_99 = D_Fai @ DIAG_Fai99 @ D_Fai.T m_fai_95 = D_Fai @ DIAG_Fai95 @ D_Fai.T m_T2 = D_T2 @ DIAG_T2 @ D_T2.T spe_recon = get_m(XMat.shape[1],m_spe) fai_99_recon = get_m(XMat.shape[1],m_fai_99) fai_95_recon = get_m(XMat.shape[1],m_fai_95) T2_recon = get_m(XMat.shape[1],m_T2) # cos检验值 R = per # 相关性 items = [('Train_X_min', np.around(Train_X_min, decimals=3).tolist()), ('Train_X_max', np.around(Train_X_max, decimals=3).tolist()), ('Train_X_std', np.around(Train_X_std, decimals=3).tolist()), ('Train_X_mean', np.around(Train_X_mean, decimals=3).tolist()), ('Train_X_bais_max', np.around(Train_X_bais_max, decimals=3).tolist()), ('Train_X_bais_min', np.around(Train_X_bais_min, decimals=3).tolist()), ('Train_X_bais_mean', np.around(Train_X_bais_mean, decimals=3).tolist()), ('QCUL_95', np.around(QCUL_95, decimals=10).tolist()), ('QCUL_99', np.around(QCUL_99, decimals=10).tolist()), ('QCUL_95_line', np.around(QCUL_95_line, decimals=3).tolist()), ('QCUL_99_line', np.around(QCUL_99_line, decimals=3).tolist()), ('T2CUL_95', np.around(T2CUL_95, decimals=3).tolist()), ('T2CUL_99', np.around(T2CUL_99, decimals=3).tolist()), ('T2CUL_95_line', np.around(T2CUL_95_line, decimals=3).tolist()), ('T2CUL_99_line', np.around(T2CUL_99_line, decimals=3).tolist()), ('Kesi_95', np.around(Kesi_95, decimals=3).tolist()), ('Kesi_99', np.around(Kesi_99, decimals=3).tolist()), ('Kesi_95_line', np.around(Kesi_95_line, decimals=3).tolist()), ('Kesi_99_line', np.around(Kesi_99_line, decimals=3).tolist()), ('speRecon', np.around(spe_recon, decimals=3).tolist()), ('T2Recon', np.around(T2_recon, decimals=3).tolist()), ('Fai95Recon', np.around(fai_95_recon, decimals=3).tolist()), ('Fai99Recon', np.around(fai_99_recon, decimals=3).tolist()), ('mSPE', np.around(m_spe, decimals=3).tolist()), ('mFai99', np.around(m_fai_99, decimals=3).tolist()), ('mFai95', np.around(m_fai_95, decimals=3).tolist()), ('mT2', np.around(m_T2, decimals=3).tolist()), ('COV', np.around(covX, decimals=3).tolist()), ('K', k), ('R', np.around(R, decimals=3).tolist()), ("featValue", np.around(featValue, decimals=3).tolist()), ("featVec", np.around(featVec, decimals=3).tolist()), ("selectVec", np.around(selectVec, decimals=3).tolist())] # model_info=json.dumps(dict(items)) res_items = [('Model_info', dict(items)), ('Model_type', 'PCA')] result = dict(res_items) # json.dumps(result) return json.dumps(result) def get_m(dimension, m): ke_si_matrix = np.zeros(shape=(dimension, dimension)) for i in range(dimension): for j in range(dimension): ke_si_matrix[i, j] = m[i, j] / m[i, i] return ke_si_matrix def clearmain(info): try: Train_Data = info["Train_Data"] condition=info["conditon"].replace("=","==").replace(">=",">").replace("<=","<") times = Train_Data["time"].split(';') points = Train_Data["points"].split(',') interval = Train_Data["interval"] if interval == 10000: DCount = 60 elif interval == 100000: DCount = 6 elif interval == 300000: DCount = 5 else: DCount = 4 dead = Train_Data["dead"].split(',') limit = Train_Data["limit"].split(',') uplower = Train_Data["uplow"].split(';') percent = info["Hyper_para"]["percent"] count=0 ItemsInfo, SamplingTimePeriods = [], [] Constraint = "" for i in range(len(points)): iteminfo = {} iteminfo["ItemName"] = points[i] # 加点 if (dead[i] == "1"): # 判断是否参与死区清洗 iteminfo["ClearDeadZone"] = "true" else: iteminfo["ClearDeadZone"] = "false" if (limit[i] == "1"): # 参与上下限清洗 limits = uplower[i].split(',') if (isnumber(limits) == True): # 输入上下限正确 isnumber 是否为数字 count += 1 Constraint += "[" + points[i] + "]>" + limits[0] + " and " + "[" + points[i] + "]<" + limits[1] + " and " ItemsInfo.append(iteminfo) if(count!=0): Constraint = Constraint[:len(Constraint) - 4:] else: Constraint="1==1"#没有上下限清洗 Constraint+=" and ("+condition+")" for i in range(len(times)): Eachsampletime = {} timess = times[i].split(',') Eachsampletime["StartingTime"] = timess[0] Eachsampletime["TerminalTime"] = timess[1] SamplingTimePeriods.append(Eachsampletime) Constraint = Constraint.replace("\n", " ") url = f"http://{config._CLEAN_IP}/exawebapi/exatime/GetCleaningData?ItemsInfo=%s&SamplingTimePeriods=%s&Constraint=%s&SamplingPeriod=%s&DCount=%d" % ( ItemsInfo, SamplingTimePeriods, Constraint, interval, DCount) response = requests.get(url) content = json.loads(response.text) data = np.array([item for item in content["ClearData"]]).T try: smote_data = info["smote"] # smote_data = False except KeyError: smote_data = False if smote_data: try: smote_index = [points.index(item["pointId"]) for item in info["smote_config"] if item["LAY_CHECKED"]] smote_num = [int(item["number"]) for item in info["smote_config"] if item["LAY_CHECKED"]] max_value = [float(item["max"]) for item in info["smote_config"] if item["LAY_CHECKED"]] min_value = [float(item["min"]) for item in info["smote_config"] if item["LAY_CHECKED"]] except KeyError: pass else: if len(smote_num) != 0: data, *_ = smote(data, smote_index, smote_num, max_value, min_value) result = train(data, percent) result = result.replace("NaN", "-1") result=json.loads(result) result["BeforeCleanSamNum"]=content["BeforeCleanSamNum"] result["AfterCleanSamNum"]=content["AfterCleanSamNum"] result["CleanOrNot"] = True return json.dumps(result) except Exception as e: result = [{"CleanOrNot": False, "msg": traceback.format_exc()}] return json.dumps(result, ensure_ascii=False) def test_offline(model, LockVariable, Data_origin): test_data = (Data_origin - model["Train_X_mean"]) / model["Train_X_std"] SPE_list = [] FAI_list = [] m_fai = np.array(model["mFai99"]) m_spe = np.array(model["mSPE"]) ke_si_matrix = np.array(model["Fai99Recon"]) line = model["Kesi_99"] paraState = np.zeros(Data_origin.shape) # 故障方向矩阵 f_matrix = np.zeros(Data_origin.shape) # 故障幅值矩阵 for i in range(test_data.shape[0]): data = test_data[i, :] SPE_list.append(data @ m_spe @ data.T) FAI_list.append(data @ m_fai @ data.T) if data @ m_fai @ data.T > line: t_c, f = sequence(data, line, m, ke_si_matrix) # 计算故障方向和幅值 paraState[i] = t_c f_matrix[i] = f final_data = test_data - f_matrix recon_data = np.add(np.multiply(final_data, model["Train_X_std"]), model["Train_X_mean"]) # 重构值 error_data = Data_origin - recon_data items = {"reconData": recon_data.tolist(), "errorData": error_data.tolist(), "R": 0, "SPE": SPE_list, "FAI": FAI_list, "paraState": paraState.tolist()} return json.dumps(items)