Among distributed optical fiber sensing technologies, optical frequency domain reflectometry (OFDR) has garnered widespread attention in aerospace, medical interventional devices, civil engineering, and other fields owing to its high spatial resolution and dynamic range. However, affected by the non-ideal phase tuning of the practical tunable laser light source, the interference signal between the local oscillating light and the scattering light at a specific position in OFDR will expand from a single frequency to a spectral band, thereby resulting in degradation of spatial resolution and demodulation accuracy. While the existing research on non-ideal phase tuning has focused on how much the broadening process can be compensated through software and hardware optimization, there is a lack of analysis on how the non-ideal tuning leads to degradation. The above motivates this paper to analyze the influence of different non-ideal tuning forms on the OFDR demodulation results, to point out the deficiencies of the traditional auxiliary interferometer compensation method, and to preliminarily verify the potential of deep learning in non-ideal tuning compensation.Taking the possible time-varying characteristics of the tuning rate and the random jitter of the initial phase into account, this work investigates the non-ideal phase forms of the tunable laser source from 30 sets of measured instantaneous frequency-time curves of the tunable laser source. The non-ideality of phase tuning caused by the jitters of the tuning rate/initial phase are analyzed by leveraging polynomial least squares fitting and zero-mean Gaussian distribution, respectively. From this, parameters can be introduced to simulate and semi-quantitatively analyze the impact of the non-ideal tuning on the strain-sensing performance of OFDR. In addition, taking advantage of the convenience of numerical simulation in generating and processing pseudo-random numbers, the convolutional neural network structure based on the Unet encoding and decoding mechanism is further trained to compensate for the non-ideal phase tuning.The results indicate that the phase tuning non-ideality resulting from the time-varying tuning rate can be approximated by a polynomial fitting curve. It affects localization more significantly in the case of smaller deviations and exhibits a dual effect on both localization and demodulation in the case of larger deviations. On the other hand, the tuning non-ideality caused by random phase jitter mainly impacts the strain demodulation accuracy and has a negligible influence on localization.This paper semi-quantitatively analyzes the impact on a typical OFDR strain sensing system, considering the time-varying nature of the tuning rate during the phase tuning process of a tunable laser and the random jitter of the initial phase. The results also reveal that the conventional compensation method utilizing auxiliary interferometer interpolation resampling can effectively mitigate the spreading due to the polynomial non-ideal form of tuning. However, it cannot completely eliminate its effect on the localization outcomes, nor can it eliminate the accuracy degradation caused by the random phase jitter of the light source. Based on this, the secondary compensation of the interpolated results through a convolutional neural network model brings the distributed strain measurement results closer to the given ideal values, demonstrating the feasibility of deep learning for the secondary compensation of residual phase non-ideal tuning forms.