Cement concrete pavements constitute critical transportation infrastructure, while slab corner voids represent a frequent distress that severely impairs pavement service performance. The Heavy Weight Deflectometer (HWD) serves as a core non-destructive testing instrument for evaluating pavement structural performance; however, temperature gradients significantly alter its deflection responses, which constitutes a key obstacle to accurate void identification. Existing studies on pavement mechanical responses based on HWD/FWD predominantly focus on variations in deflection magnitudes, with limited investigations into deflection ratio characteristics. Furthermore, there is a shortage of experimental data under controlled conditions with predefined void patterns and definite temperature gradients. This study selected the cargo apron of a newly built 4E-class civil-military integrated airport as the test site, established six test sections, embedded sensors, and conducted on-site HWD tests with preset artificial voids. The objective was to explore the HWD mechanical responses and deflection ratio characteristics of cement pavements under three temperature gradients:-12.5, 0, and 12.5°C/m. The results demonstrate that temperature gradients exert no obvious influence on mid-slab deflections but significantly affect slab corner deflections and deflection ratios. Negative temperature gradients cause upward warping of slab corners and increased deflections, whereas positive gradients induce downward warping and reduced deflections. Additionally, deflection ratios of voided slab corners are notably higher than those of non-voided ones. This study fills the data gap in HWD tests under conditions of known voids and temperature gradients, reveals their coupled effects on pavement deflection ratios, and identifies limitations of the single deflection ratio method in void identification. The findings provide essential experimental support for optimizing cement pavement void detection technologies and establishing multi-factor coupled void identification models.