基于主成分分析-粒子群优化算法-支持向量机的混合气体分类识别方法研究

Study on recognition and classification of gas mixtures based on principal component analysis-particle swarm optimization and support vector machine

目的解决多传感器设备在混合气体识别率和预测精度方面存在的问题,提高对混合气体的检测准确度,保障接触危险气体的作业人员的安全。方法采用了美国加州大学尔湾分校(UCI)的“动态气体混合物中的气体传感器阵列”公共数据集,使用MATLAB 2021b软件进行仿真测试。数据集的传感器阵列包括16个传感器。使用70个一氧化碳气体样本、80个乙烯气体样本、70个空气样本、80个一氧化碳与乙烯混合气体样本,共300个气体样本进行训练。提出了一种基于机器学习和传感器阵列技术的可燃混合气体分类方法,该方法首先利用主成分分析(principal component analysis,PCA)来降低输入数据的维数,再采用粒子群优化(particle swarm optimization,PSO)算法对支持向量机(support vector machine,SVM)的超参数进行优化,然后与未经优化的SVM法与未经优化的BP神经网络法所得出的气体识别分类结果进行对比。结果数据集经过PCA处理后将原始数据从16维降为4维,累计贡献率达到99%以上。该算法在空气、一氧化碳、混合气体和乙烯4种气体的定性识别中,准确率达到100%(50/50),分别比未经优化的SVM法准确率(90%,45/50)和BP网络法准确率(98%,49/50)高10%和2%。结论基于主成分分析-粒子群优化算法-支持向量机方法能够准确地对混合气体进行识别和分类,提高传感器报警速度与准确度,及时发现生产中可能存在的危险。但对组分更复杂的气体的识别效率和准确性,仍须进一步研究。

Objective Address the challenges posed by multi-sensor devices in terms of recognition rate and prediction accuracy of gas mixtures to enhance the detection accuracy of flammable gas,thereby ensuring the safety of personnel exposed to hazardous gas.Methods The public dataset“Gas Sensor Arrays for Dynamic Gas Mixtures”from the University of California,Irvine(UCI)was employed for simulation tests using MATLAB 2021b.The sensor array in the dataset comprised 16 sensors.A total of 300 gas samples were used for training,including 70 carbon monoxide samples,80 ethylene samples,70 air samples,and 80 carbon monoxide-ethylene mixture samples.A flammable gas mixture classification method was proposed based on machine learning and sensor array technology.Principal Component Analysis(PCA)was first utilized to reduce the dimensionality of the input data,then Particle Swarm Optimization(PSO)was employed to optimize the hyperparameters of the Support Vector Machine(SVM).Finally,the gas recognition and classification results were compared with those obtained using unoptimized SVM and unoptimized BP neural network methods.Results After PCA processing,the dimensionality of the original data was reduced from 16 to 4,with a cumulative contribution rate exceeding 99%.This algorithm achieved 100%accuracy(50/50)in the qualitative identification of air,carbon monoxide,mixed gas,and ethylene,which was 10%and 2%higher than the accuracy of unoptimized SVM(90%,45/50)and BP neural network(98%,49/50),respectively.Conclusions The proposed method based on PCA,PSO,and SVM can accurately identify and classify gas mixtures,improving the speed and accuracy of sensor alarms and facilitating the timely detection of potential hazards in production.However,further research is needed to study the efficiency and accuracy of identifying gases with more complex compositions.