Random noise and coherent noise interfere with seismic data collected in the field, resulting in the reduction of the signal-to-noise ratio, which affects the subsequent processing of seismic data, such as seismic migration and imaging. Therefore, it is necessary to develop an efficient and adaptive method to attenuate random and coherent noise in real seismic data. Conventional supervised learning algorithms need to manually generate a large number of labels to train the network, which is very difficult in the field of seismic exploration where the data volume is small. In addition, supervised learning-based noise attenuation methods are expensive in terms of computation and labor costs. To solve this problem, this paper constructs an adaptive deep learning framework based on unsupervised learning strategies to attenuate random and irregular (erratic) noise in multi-dimensional seismic data. This method uses the corresponding structure of encoding and decoding to compress and reconstruct data features. In order to improve the network's attention to important waveform features, this paper uses a soft attention mechanism to assign more weight to important waveform features in a weighted way. In this paper, the multi-dimensional noisy data is segmented into a large number of one-dimensional noisy signals, which are input into the network for iteration, so as to adaptively attenuate random and erratic noise in seismic data. This small-scale signal denoising method can effectively improve the noise attenuation performance of the network and help to avoid artifacts. In this paper, we use a more robust Huber loss function to attenuate random and erratic noise, which combines the root-mean-square error with l2 norm and the average absolute error loss with l1 norm. In addition, a Total Variation (TV) regularization term is added to the constructed network to capture the local smooth structure of the seismic data. By adjusting the weight of Huber loss function and TV regularization term, the network can obtain the best denoising performance. The method constructed in this paper can be directly used for attenuation of random and erratic noise of multi-dimensional seismic data, and ensure transverse continuity of seismic signals after reconstruction. We compare the proposed framework with classical seismic data denoising methods and noise attenuation methods based on unsupervised learning to analyze the advantages and disadvantages of each method. The test results of 2D and 3D synthetic data and actual seismic data show that the proposed method has better noise attenuation and useful signal protection capabilities.