Effective　anomaly　data　detection　and　anomaly　type　identification　can　help　to　improve　the　data　quality　of　wireless　sensor　networks.　The　classification-based　anomaly　detection　algorithm　has　difficulty　in　extracting　the　classification　characteristics　of　sensor　data,and　cannot　further　distinguish　the　types　of　anomaly　data.　The　anomaly　detection　method　based　on　spatiotemporal　characteristics　has　the　problem　of　over-reliance　on　the　hypothetical　distribution　of　data.　In　this　paper,an　anomaly　detection　algorithm　combining　spatiotemporal　characteristics　of　data　streams　and　multi-classification　model　is　proposed.　Firstly,based　on　the　Markov　chain,the　spatiotemporal　characteristics　of　the　sensor　data　stream　are　extracted.　Then,the　extracted　spatiotemporal　characteristics　are　used　as　the　input　characteristics　of　the　multi-class　convolutional　neural　network　model　to　detect　the　anomaly　of　the　data　stream　and　identify　the　anomaly　type.The　results　show　that　the　algorithm　exhibits　higher　rate　of　detection　accuracy　and　lower　rate　of　false　negatives　and　false　positives　on　different　data　sets,which　can　effectively　detect　the　anomaly　data　and　complete　the　type　identification.　©　2019,　The　Editorial　Office　of　Chinese　Journal　of　Sensors　and　Actuators.　All　right　reserved.