In the railway system, the switch machine is a critical device to ensure the safe and smooth operation of trains. Fault diagnosis of the S700K switch machine is crucial for accident prevention and the maintenance of railway operations. To address the shortcomings of traditional diagnostic methods in terms of speed and accuracy, a diagnostic model integrating a deep residual shrinking network with a bidirectional long short-term memory network is proposed. First, the power curve of the switch machine is preprocessed. Next, DRSN is used to automatically learn features from the preprocessed data and compress the data length, improving the speed of diagnosis. Its attention mechanism and soft thresholding reduce the influence of noise features, and the DRSN structure helps to overcome network degradation and overfitting issues. Following that, the bidirectional structure of BiLSTM is utilized to capture complex relationships in the time-series data. Finally, a Softmax classifier is employed for fault classification. Simulation results show that the accuracy, precision, and recall rates of the DRSN-BiLSTM model all exceed 98.3%. Compared with models such as DRSN, deep neural network, and convolutional neural network, the diagnostic accuracy of this model is improved by at least 1.47%. Even when Gaussian white noise in the range of 15~40 dB is added, the accuracy remains above 92.7%, an improvement of at least 2% over other models. This model not only ensures the efficiency of the training process but also improves the accuracy of point machine fault diagnosis and demonstrates excellent robustness in noisy environments.