Due to the differences in data distribution caused by different locations of multiple measuring points， the fault diagnosis of the harmonic reducer is often ineffective. A fault diagnosis method for the harmonic reducer， based on a multiple feature spaces adaptation network （MFSAN）， is proposed. Firstly， the vibration signal of the harmonic reducer is transformed using continuous wavelet transform to construct a time-frequency diagram that characterizes its operational state. Secondly， the data measured by sensors at different positions are divided into multiple source domain and target domain data， which are mapped to different feature spaces to obtain feature representations for each measuring point position. Then， the adaptive network is used to automatically transfer the knowledge learned from the source domain to the target domain features and automatically align the feature distribution of a specific domain to learn multiple domain-invariant representations. Finally， a domain-specific decision boundary is used to align the output of the classifier， effectively solving the data distribution differences caused by sensor location. Experimental results of harmonic reducer diagnosis of an industrial robot show that the identification accuracy of this method is 99.72%， which is higher than that of other comparison methods. The effectiveness and feasibility of this method are thus verified.