Industrial anomalous sound detection (ASD) is typically trained exclusively on normal samples, making self-supervised learning a promising paradigm for label-scarce settings. However, existing discriminative ASD methods still exhibit three practical limitations. First, most methods rely on a fixed Mel-scale front end, even though fault-related frequency bands vary across machine types, potentially obscuring informative anomalous components in the mid- and high-frequency ranges. Second, spectrogram generation can attenuate long-range temporal dependencies present in raw audio, thereby reducing sensitivity to temporally evolving anomalies. Third, widely used backbone networks often fail to achieve an effective balance between representational capacity and computational efficiency, thereby limiting deployment in resource-constrained industrial environments. To address these limitations, AFSTgram-DFE is proposed as a self-supervised ASD framework that integrates adaptive filtering with dynamic feature extraction. In the spectrogram branch, the filtering strategy is adaptively selected between uniform and Mel-scale filter banks according to the fault-sensitive frequency characteristics of each machine type. In the temporal branch, an attention mechanism and an autocorrelation-based module are introduced to recover temporal cues lost during spectrogram generation. In addition, a parameter-free feature enhancement module and a dynamic feature extraction network with complementary large and small convolutional kernels are employed to enhance discriminative representation learning while preserving lightweight inference. Finally, anomaly discrimination is performed using an ArcFace-based classifier in conjunction with a Gaussian mixture model. Experiments on the DCASE 2020 Challenge Task 2 dataset show that AFSTgram-DFE achieves an average AUC of 95.10% and an average pAUC of 90.25%, outperforming existing mainstream methods.