Bearings and gears are crucial components in mechanical transmission systems, and their fault diagnosis is of great significance for ensuring the safe operation of equipment. To effectively extract the features of rotating machinery fault signals and solve the problem of strong dependence of classifiers on feature extraction, this paper proposes a fault diagnosis model based on dilated convolution and improved black winged kite optimized least squares support vector machine (BKA-LSSVM). Firstly, the one-dimensional vibration signal is transformed into a two-dimensional time-frequency image with high-resolution time-frequency representation using synchronous compression wavelet transform. Secondly, a multi-scale cascaded dilated convolution module is constructed, and the dilation rate adjustment mechanism is used to achieve hierarchical and multi granularity extraction of fault features, effectively capturing fault mode features at different scales. The results of the fully connected layer are used as inputs to the BKA-LSSVM classification layer, and a nonlinear growth model is introduced to dynamically adjust the disturbance coefficient. A random search mechanism is constructed to improve the BKA. Finally, the improved BKA is used to optimize the parameters of LSSVM to improve the classification accuracy of the model. Validation was conducted on two datasets, and the experimental results showed that the proposed model had an accuracy rate of over 87% when the sample size was 10, and an accuracy rate of over 95% when the signal-to-noise ratio was -4. This validates that the proposed model has stronger noise resistance and generalization performance compared to the comparison model.