Improvement and Application of Deep Belief Network Based on Sparrow Search Algorithm

Deep Belief Network (DBN) is one of the most popular network structures in the field of Deep Learning (DL). It consists of multiple Restricted Boltzmann Machines (RBMs) and an output layer, which gives it a strong feature extraction and non-linear mapping capability through layer-by-layer feature extraction and backward fine-tuning of the BP network. However, it is mainly affected by the number of neuron nodes in the hidden layer. Too few neurons will cause the DBN model to suffer from underfitting problems, and conversely too many neurons will also cause the DBN model to suffer from overfitting, so the performance of the DBN model is difficult to be satisfied directly, leading to some limitations in its application. Thus, this paper proposes an improved DBN model based on the sparrow search algorithm, which can overcome the problem of difficult neuron selection and find the optimal number of neural nodes per layer of the neural network through the sparrow search algorithm, greatly reducing the human debugging time and achieving excellent diagnosis performance. A comparison of the Western Reserve dataset shows that the method proposed in this paper has higher accuracy and more stable model training than the traditional DBN model.