Abstract: Addressing the issues of low recognition accuracy, weak anti-interference capability, and insufficient operational stability in the fault detection of belt conveyor idlers, an intelligent diagnosis method based on feature signal fusion (TFM) and multi-input one-dimensional convolutional neural network (MI-1DCNN) is proposed. This method first collects the operating audio of the idlers through a microphone and employs db4 wavelet basis and unbiased risk estimation threshold for noise reduction, aiming to enhance the signal-to-noise ratio and suppress background interference. On this basis, the time-domain features, frequency-domain features, and Mel Frequency Cepstral Coefficients (MFCC) along with their first-order and second-order differential coefficients are extracted and normalized to construct a TFM feature set. This feature set is then input into the MI-1DCNN model equipped with multi-scale convolutional kernels. The precise identification of the idler state is achieved through the multi-path feature fusion mechanism and Softmax classifier. Experiments conducted based on data collected from a coal mine site demonstrate that the proposed method achieves an average recognition accuracy of 98.65% for faulty idlers, representing an improvement of 1.50% and 1.03% compared to the improved wavelet threshold-BP-RBF network and MFCC-KNN-SVM methods, respectively. In industrial tests, the method still maintains a recognition accuracy of 98.4%, exhibiting strong robustness and potential for application.