The invention discloses a carrier roller fault diagnosis method and system based on machine learning and a storage medium. The carrier roller fault diagnosis method based on machine learning comprisesthe following steps of: S1, collecting audio data of a carrier roller; S2, extracting features of the audio data, wherein the features of the audio data specifically comprise one or any combination of sharpness, noise annoyance and speech interference level; S3, inputting the features of the audio data into a trained CART model, and identifying a running state of the carrier roller by using the CART model; S4, if the carrier roller operates abnormally, executing alarm, monitoring or control operation, and if the carrier roller operates normally, completing carrier roller fault diagnosis at the current moment, and executing a step S5; and S5, updating the moment, repeatedly executing the steps S1 to S4, and carrying out carrier roller fault diagnosis at the next moment. According to the carrier roller fault diagnosis method and the system, the carrier roller fault can be diagnosed in real time, and the method is easy to implement, low in cost and low in algorithm complexity.