Python环境下基于机器学习（多层感知机，决策树，随机森林，高斯过程，AdaBoost，朴素贝叶斯）的压缩机故障识别（出口阀泄漏，止逆阀泄露，轴承损伤，惯性轮损伤，活塞损伤，皮带损伤等）。

空压机是一种经典的动力设备,也被誉为企业产品生产的"生命气源",，广泛应用于制药工业、爆破采煤、矿上通风、风动实验等众多领域。空压机的工作机理是通过利用旋转电机的机械能对气体进行挤压，从而使得气体能够产生巨大的能量，利用充满能量的气体进行一些爆破、通风等作业，从而满足现实中的使用需求。

本项目使用声信号检测空压机的 7 种故障，分别为出口阀泄漏Leakage Outlet Valve（LOV）,入口阀泄露Leakage Inlet Valve(LIV)，止逆阀泄露Non-Return Valve(NRV)，轴承损伤Bearing，惯性轮损伤Flywheel，活塞损伤Piston，皮带损伤Riderbelt和健康状态。经过特征提取后，利用各种机器学习算法对空压机故障进行分类，试验台如下：

所测得的声信号以Flywheel为例

接下来开始步入正题，由于代码包含众多模块，为了避免篇幅过长，只写主函数。提取的特征可以选择维数是否约简，分类器也可选择是否进行参数优化（粒子群算法，pip install pyswarm），代码必要的标注都有了，也比较容易看懂。

import time
import pandas as pd
import numpy as np
import matplotlib.pyplot as pltMaxExpNo=10
counter=-1
#标签
labels=['Bearing','Flywheel','Healthy','LIV','LOV','NRV','Piston','Riderbelt']
import PrimaryStatFeatures  #特征提取模块
import FFT_Module  #FFT变换
#所要提取的特征
data_columns_PrimaryStatFeatures=['Mean','Min','Max','StdDv','RMS','Skewness','Kurtosis','CrestFactor','ShapeFactor']
data_columns_Target=['Fault']
Faults={labels[0]:int(0),labels[1]:int(1),labels[2]:int(2),labels[3]:int(3),labels[4]:int(4),labels[5]:int(5),labels[6]:int(6),labels[7]:int(7)}
for label in labels:for ExpNo in range(1,MaxExpNo+1):counter+=1file='Data\\'+label+'\\preprocess_Reading'+str(ExpNo)+'.txt'X=np.loadtxt(file,delimiter=',')if (counter%10==0): print('Lading files: ',str(counter/(len(labels)*MaxExpNo)*100),'% completed')StatFeatures=PrimaryStatFeatures.PrimaryFeatureExtractor(X)FFT_Features,data_columns_FFT_Features=FFT_Module.FFT_BasedFeatures(X)data_columns=data_columns_PrimaryStatFeatures+data_columns_FFT_Features+data_columns_Targetif (label==labels[0] and ExpNo==1): data=pd.DataFrame(columns=data_columns)StatFeatures[0].extend(FFT_Features)StatFeatures[0].extend([Faults[label]])df_temp=pd.DataFrame(StatFeatures,index=[counter],columns=data_columns)data=data.append(df_temp)input_data=data.drop(columns=['Fault'])
#数据标准化处理
#参考: http://benalexkeen.com/feature-scaling-with-scikit-learn/
from sklearn import preprocessing
normalization_status='RobustScaler'   
''' Choices:1. Normalization2. StandardScaler3. MinMaxScaler4. RobustScaler5. Normalizer6. WithoutNormalization   '''
input_data_columns=data_columns_PrimaryStatFeatures+data_columns_FFT_Featuresif (normalization_status=='Normalization'):data_array=preprocessing.normalize(input_data,norm='l2',axis=0)input_data=pd.DataFrame(data_array,columns=input_data_columns)
elif (normalization_status=='StandardScaler'):scaler = preprocessing.StandardScaler()scaled_df = scaler.fit_transform(input_data)input_data = pd.DataFrame(scaled_df, columns=input_data_columns)
elif (normalization_status=='MinMaxScaler'):scaler = preprocessing.MinMaxScaler()scaled_df = scaler.fit_transform(input_data)input_data = pd.DataFrame(scaled_df, columns=input_data_columns)
elif (normalization_status=='RobustScaler'):scaler = preprocessing.RobustScaler()scaled_df = scaler.fit_transform(input_data)input_data = pd.DataFrame(scaled_df, columns=input_data_columns)
elif (normalization_status=='Normalizer'):scaler = preprocessing.Normalizer()scaled_df = scaler.fit_transform(input_data)input_data = pd.DataFrame(scaled_df, columns=input_data_columns)
elif (normalization_status=='WithoutNormalization'):print ('No normalization is required')target_data=pd.DataFrame(data['Fault'],columns=['Fault'],dtype=int)DimReductionStatus=False
if (DimReductionStatus==True):for nComponents in range(1,110):#降维#主成分分析from sklearn import decompositionpca = decomposition.PCA(n_components=nComponents)pca.fit(input_data)input_data_reduced = pca.transform(input_data)#训练集和测试集划分from sklearn.model_selection import train_test_splitx_train,x_test,y_train,y_test=train_test_split(input_data_reduced,target_data,test_size=0.3,random_state=42,stratify=target_data)#使用 KNN（K 最近邻）训练import KNN_Classifiertest_accuracy_max=KNN_Classifier.KNNClassifier(x_train,x_test,y_train,y_test)plt.figure(10)plt.scatter(nComponents,test_accuracy_max)plt.xlabel('Number of utilized components based on PCA')plt.ylabel('Test Accuracy')#使用 SVC（支持向量分类器）进行训练import SVC_Classifiertest_accuracy_max=SVC_Classifier.SVCClassifier(x_train,x_test,y_train,y_test)plt.figure(11)plt.scatter(nComponents,test_accuracy_max)
else:from sklearn.model_selection import train_test_splitx_train,x_test,y_train,y_test=train_test_split(input_data,target_data,test_size=0.3,random_state=42,stratify=target_data)#使用决策树进行训练
import DT_Classifier
DT_Classifier.DTClassifier(x_train,x_test,y_train,y_test)from sklearn.neural_network import MLPClassifier
from sklearn.neighbors import KNeighborsClassifier
from sklearn.svm import SVC
from sklearn.gaussian_process import GaussianProcessClassifier
from sklearn.gaussian_process.kernels import RBF
from sklearn.tree import DecisionTreeClassifier
from sklearn.ensemble import RandomForestClassifier, AdaBoostClassifier
from sklearn.naive_bayes import GaussianNB
from sklearn.discriminant_analysis import QuadraticDiscriminantAnalysis#写入最佳分类器参数
file1 = open('Optimized Parameters for Classifiers.txt','w')
print('\nThis file contains the output optimized parameters for different classifiers\n\n\n',file=file1)#分类器参数优化
#*************************************
SVMOptStatus=True
KNNOptStatus=True
MLPOptStatus=True
DTOptStatus=True
import CLFOptimizer
#pip install pyswarm
#优化 SVC 参数
if (SVMOptStatus==True):SVM_kernels = ['linear','poly','rbf','sigmoid']for KernelType in SVM_kernels:print('\n\nstage: Optimizing SVC:',KernelType)SVCParams_opt,SVCAccuracy_opt=CLFOptimizer.SVCOPT(KernelType,x_train,x_test,y_train,y_test)print('\nClassifier: SVC-',KernelType,', Gamma=',SVCParams_opt[0],', PenaltyPrameter=',SVCParams_opt[1],', Test Acuuracy=',SVCAccuracy_opt,file=file1)#优化 KNN 参数
if (KNNOptStatus==True):print('\n\nstage: Optimizing KNN')KNNParams_opt,KNNAccuracy_opt=CLFOptimizer.KNNOPT(x_train,x_test,y_train,y_test)print('\nClassifier: KNN, n_neighbors=',KNNParams_opt,', Test Acuuracy=',KNNAccuracy_opt,file=file1)#优化 MLP 分类器
if (MLPOptStatus==True):print('\n\nstage: Optimizing MLP')MLPParams_opt,MLPAccuracy_opt=CLFOptimizer.MLPOPT(x_train,x_test,y_train,y_test)print('\nClassifier: MLP, hidden_layer_sizes=(',MLPParams_opt[0],',',MLPParams_opt[1],',',MLPParams_opt[2],'), Test Acuuracy=',MLPAccuracy_opt,file=file1)#优化决策树分类器
if (DTOptStatus==True):print('\n\nstage: Optimizing Decision Tree')DTParams_opt,DTAccuracy_opt=CLFOptimizer.DTOPT(x_train,x_test,y_train,y_test)print('\nClassifier: Decision Tree, max_depth=',DTParams_opt[0],', min_samples_split=',DTParams_opt[1],', min_samples_leaf=',DTParams_opt[2],', Test Acuuracy=',DTAccuracy_opt,file=file1)file1.close()#生成分类器名称及其配置
classifiers=[]
CLFnames=[]CLFnames= CLFnames + ["SVC-linear","K-Nearest Neighbors","Multi-Layer Perceptron","Decision Tree", "Random Forest", "Gaussian Process", "AdaBoost","Naive Bayes", "QDA"]#classifiers=classifiers+ [
#    SVC(kernel='linear',gamma=1.785,C=3.463),
#    KNeighborsClassifier(n_neighbors=3),
#    MLPClassifier(hidden_layer_sizes=(28,34,80,),alpha=1),
#    DecisionTreeClassifier(),
#    RandomForestClassifier(),
#    GaussianProcessClassifier(1.0 * RBF(1.0)),
#    AdaBoostClassifier(),
#    GaussianNB(),
#    QuadraticDiscriminantAnalysis()]classifiers=classifiers+ [SVC(),KNeighborsClassifier(),MLPClassifier(),DecisionTreeClassifier(),RandomForestClassifier(),GaussianProcessClassifier(),AdaBoostClassifier(DecisionTreeClassifier(),n_estimators=1000,learning_rate=1),GaussianNB(),QuadraticDiscriminantAnalysis()]#将分类结果写入文件
f = open('ClassificationResults.txt','w')
print('\nThis file contains an overall comparison of different classifiers performance\n\n\n',file=f)
import ClassificationModule
ClassificationModule.Classifiers(CLFnames,classifiers,x_train,x_test,y_train,y_test)

分类器优化参数结果

This file contains the output optimized parameters for different classifiers

Classifier: SVC- linear , Gamma= 1.61614372336136 , PenaltyPrameter= 6.455613795161285 , Test Acuuracy= 1.0

Classifier: SVC- poly , Gamma= 0.02472261261696812 , PenaltyPrameter= 8.084311259878596 , Test Acuuracy= 0.875

Classifier: SVC- rbf , Gamma= 0.003211151769260453 , PenaltyPrameter= 3.2794598503610994 , Test Acuuracy= 0.9583333333333334

Classifier: SVC- sigmoid , Gamma= 0.001 , PenaltyPrameter= 9.44412249946835 , Test Acuuracy= 1.0

Classifier: KNN, n_neighbors= 3 , Test Acuuracy= 0.875

Classifier: MLP, hidden_layer_sizes=( 19.0 , 97.0 , 14.0 ), Test Acuuracy= 1.0

Classifier: Decision Tree, max_depth= 2121.0 , min_samples_split= 6.0 , min_samples_leaf= 4.0 , Test Acuuracy= 1.0

Training accuracy for SVC-linear is: 0.9642857142857143 and Prediction accuracy is: 0.875

Training accuracy for K-Nearest Neighbors is: 0.8928571428571429 and Prediction accuracy is: 0.7916666666666666

分类结果

Training accuracy for Multi-Layer Perceptron is: 1.0 and Prediction accuracy is: 0.9583333333333334

Training accuracy for Decision Tree is: 1.0 and Prediction accuracy is: 0.7916666666666666

Training accuracy for Random Forest is: 1.0 and Prediction accuracy is: 0.9583333333333334

Training accuracy for Gaussian Process is: 1.0 and Prediction accuracy is: 0.8333333333333334

Training accuracy for AdaBoost is: 1.0 and Prediction accuracy is: 0.9583333333333334

Training accuracy for Naive Bayes is: 1.0 and Prediction accuracy is: 0.9166666666666666

Training accuracy for QDA is: 1.0 and Prediction accuracy is: 0.3333333333333333

完整代码可通过知乎学术咨询获得：

Python环境下基于机器学习的空压机故障识别

工学博士，担任《Mechanical System and Signal Processing》审稿专家，担任《中国电机工程学报》优秀审稿专家，《控制与决策》，《系统工程与电子技术》，《电力系统保护与控制》，《宇航学报》等EI期刊审稿专家。

擅长领域：现代信号处理，机器学习，深度学习，数字孪生，时间序列分析，设备缺陷检测、设备异常检测、设备智能故障诊断与健康管理PHM等。