<p>As the manufacturing industry evolves towards high precision, efficiency, and intelligence, the precise control of the machining process has become increasingly crucial. Tool wear is not only directly related to the accuracy of machined parts but also closely linked to machining efficiency. However, tool wear is influenced by complex interactions among multiple factors. The monitoring signals show strong nonlinearity and non-stationarity in the spatial–temporal dimension, complicating the accurate extraction of relevant features during tool wear prediction and subsequently affecting prediction accuracy. Therefore, this paper proposes a tool wear prediction method based on a multi-scale dilated convolutional long short-term memory network with channel-spatial attention (CSA-MSDCLSTM). First, the monitoring signals are selectively intercepted to eliminate noise data. Second, a predictive model utilizes multi-scale dilated convolutions to extract features from signals across various scales. By incorporating channel-spatial attention, the network can automatically focus on essential features. The double-layer LSTM further enhances the extraction of spatial–temporal features, resulting in accurate predictions. Finally, data is input into the constructed model to experiment with tool wear prediction. This paper selects the PHM 2010 dataset from the milling experiment. The experimental results demonstrate that this method effectively extracts multi-dimensional spatial–temporal features from the monitoring signals, accurately predicts the tool wear state, and offers a more reliable and precise approach for prediction. This contributes to enhancing the accuracy of machined products and improving machining efficiency.</p>