To improve the efficiency and accuracy of the field diagnosis of insulation layer deterioration of the cable intermediate joint, a non-contact diagnosis method based on adaptive deep learning of surface temperature is proposed. First, infrared thermal imaging was performed on the insulating surface of the cable joint and cables at both ends. The surface temperatures of multiple symmetric areas on both sides of the center of the cable joint and cables at both ends were collected without contact. Subsequently, a deep learning network based on a two-hidden autoencoder extreme learning machine was constructed to mine the deep hidden features in the surface temperature data. The extracted deep hidden features were used as input to the random forest diagnosis model. A quantum rotation gate with a nonlinear dynamic adaptive rotation angle was further proposed to improve the update strategy of the quantum firework algorithm and optimize the parameters of the diagnostic model. Finally, by combining the infrared temperature of the joint surface and loss angle tangent value of the insulating medium, a dataset was constructed to train and test the diagnostic model in the field. The experimental results show that the improved quantum fireworks algorithm can better approximate the global optimal solution and has high convergence efficiency. The deep learning random forest diagnostic model exhibited strong feature extraction and classification capabilities, whereby the classification accuracy and stability of the diagnostic model were effectively improved after parameter optimization, and better diagnostic results were achieved under the condition of a small sample training set. Therefore, noncontact diagnosis of joint insulation deterioration is achievable.