Acoustic signals of industrial equipment play a critical role in intelligent manufacturing and predictive maintenance. Their non-intrusive sensing characteristics enable real-time capture of micro-level changes in internal components, providing a basis for early fault warning. Current industrial audio anomaly detection methods can be divided into supervised learning and unsupervised learning, each with its own advantages, disadvantages, and limitations. Supervised learning relies on a large amount of labeled data, whereas unsupervised learning only requires modeling of normal samples. In this paper, we propose a Multi-feature Adversarial Convolutional Autoencoder framework (MACAE), which includes a multi-feature fusion module and an adversarially constrained convolutional autoencoder architecture. By jointly employing log-Mel spectrogram and Mel-Frequency Cepstral Coefficient (MFCC) features, and incorporating adversarial training, the model’s discrimination capability and robustness against complex fault patterns are significantly enhanced. Experiments on the public MIMII dataset and the augmented OWED dataset demonstrate that MACAE markedly improves anomaly detection performance in complex noise environments. For valve-type equipment, the anomaly detection AUC values reach 87.0% and 84.3%, and for rail-type equipment, AUC values reach 94.3% and 86.0%, respectively—substantially outperforming the baseline method that uses a single Mel-spectrogram autoencoder, while maintaining stable performance across different signal-to-noise ratio scenarios.