Objective <p>Due to the complexity of the internal characteristics of nuclear radiation detectors and conditioning circuits, nuclear pulse&#xa0;signals usually have a high-order nature. If nuclear pulse signals have broken tails, their digitally shaped results become distorted,&#xa0;affecting subsequent analysis. Accurate identification and repair of such distorted nuclear pulse signals are crucial for improving the&#xa0;performance of nuclear radiation detection systems.</p> Methods <p>This paper proposes an adaptive nuclear pulse processing method based on a dual-model architecture. The method consists of two stages. In the first stage, the CNN + LSTM model is used to identify the key parameters of the nuclear pulse (e.g., amplitude, decay time constant, shaping parameter) in real time, to determine whether the current measurement environment is drifting or distorted. In the second stage, based on the identification results, the corresponding repair model is dynamically selected or switched according to the different time constants to correct the distorted pulse and restore its true amplitude directly.</p> Results <p>Compared with the traditional fixed-repair algorithm, this method demonstrates higher accuracy in both two stages: In the first stage, the relative error of the CNN + LSTM model in recognizing the pulse amplitude and time constant is within 0.2%; in the second stage, the relative error of the pulse amplitude is only 0.22% after repairing the distortion pulse with the bidirectional LSTM. The simulation results show that even if the front-end circuit test data drifts, the adaptive nuclear pulse processing method proposed in this paper is still able to effectively deal with the higher-order nuclear pulse signal distortion and maintain high accuracy in recognizing amplitude values.</p> Conclusion <p>This method significantly improves energy spectrum accuracy, counting rate, and overall system reliability by accurately extracting pulse amplitudes under non-ideal conditions.</p>