Wind turbine (WT) fault diagnosis using supervisory control and data acquisition (SCADA) data is challenged by severe class imbalance, which deteriorates the recognition of minority class. To address this issue, a spatiotemporal feature extraction framework from integrated multivariate time series (IMTS) is proposed, with coordinated designs in data representation, model architecture, and loss regularization. First, a dual-label IMTS is constructed to fuse multi-source SCADA data via sliding windows. Second, a spatiotemporal feature learning architecture is developed by integrating a multi-scale convolutional neural network (MCNN) and a stacked long short-term memory (LSTM) network and designing the MCNN-LSTM model. Finally, an improved weighted cross-entropy loss is proposed, which incorporates kernel density estimation-based sample probabilities as regularization, thereby directing model attention toward minority and hard-to-classify samples. Experimental results on a real SCADA dataset demonstrate that the proposed MCNN-LSTM model achieves a Macro-recall score of 0.943 and a G-mean score of 0.942, outperforming comparison models in comprehensive performance.