For the dispersion and fuzziness of abnormal characteristics of hydraulic pumps in large-scale die forging press, an abnormal monitoring method based on multi-signal fusion and convolutional neural networks was proposed. Firstly, a data processing method based on multi-signal fusion was proposed, and the time series signals for different types of sensors were converted into multi-channel feature maps by calculating the autocorrelation coefficients (ACF) of time series data from each channel, which effectively combined the features between each fault. Then, based on the feature maps, a convolutional neural network was constructed and improved to learn the health status of the device, which was used to predict the correlation of future key indicator parameters. Finally, the abnormal score was calculated by using the Fast Fourier Transform (FFT), and an adaptive error measurement method was proposed to detect the abnormal data of hydraulic pumps. Experimental analysis results show that the proposed method can effectively achieve the abnormal monitoring of the hydraulic pumps in large-scale die forging press.