Due to the characteristics of early minor faults in industrial processes, such as small data amplitudes and strong feature coupling, the detection performance of traditional autoencoder models for these faults is poor. In response, a method for early fault detection in industrial processes based on adversarial quadratic autoencoders and ensemble learning is proposed. Initially, a quadratic neuron is introduced into the hidden layer of a conventional autoencoder model to enhance its expressive power. Subsequently, an adversarial quadratic autoencoder is introduced, incorporating a GAN network during training to enforce feature learning to adhere to specific probability distributions. Then, employing the concept of ensemble learning, normal operational data is sampled, and each subset is used to train an adversarial quadratic autoencoder. Subsequently, two matrices, SPE and T2 statistical quantities, are generated for each subset model. A fusion strategy utilizing singular value decomposition within a single-step sliding window is employed to utilize the maximum singular value within each window as a detection statistic. The proposed method is validated using a numerical example and the TE process, demonstrating its effectiveness in early detection of minor faults in industrial processes.