Abstract: The Crown pear is the primary fresh pear variety for consumption in China. Consumers prefer this pear because it has high juice content, high sweetness, and rich nutritional value. Firmness represents one of the most important quality indicators for pears. It directly shows how ripe the pear is and how well the pear can be stored after harvest. For these reasons, it is very important to measure pear firmness accurately. However, traditional methods for measuring fruit firmness mainly depend on destructive tests. These include the Magness-Taylor puncture test and compression tests. These destructive testing methods lead to large amounts of food waste. Because of this waste problem, these methods cannot be used on a large scale. This research aims to examine the connection between fruit firmness and vibration frequency features in Crown pears. It also works to solve two issues with current studies: regular prediction models have low accuracy, while deep learning models are difficult to understand. In order to achieve this, the study improved the standard convolutional neural network model. It created a new recognition model that combines an improved convolutional neural network with Deep Shapley Additive Explanations. The experiment used 508 Crown pears as test samples. The samples were split into five different groups. Each group was tested every thirteen days. Vibration data from the Crown pears was gathered using a Laser Doppler Vibrometer (LDV). The research applied an improved one-dimensional convolutional neural network method to build a firmness prediction model. This model was based on characteristic frequencies. It also used the Deep Shapley Additive Explanations (Deep SHAP) structure to help explain how the prediction model functions. The research compared this new model with several standard prediction models. These standard models were: partial least squares regression, support vector regression, extreme gradient boosting, and adaptive gradient boosting. The results show that the improved one-dimensional convolutional neural network prediction model can achieve high-precision prediction of Crown pear hardness using feature frequencies the R_P² was 0.945, the RMSE_P was 0.594N/mm, and the RPD_P was 4.272, and outperforms all traditional models across all metrics, with the most notable performance: its R² value is closest to the ideal value of 1, the RMSE is the lowest, at only 68.5% of the next-best XGBoost model, and the RPD value is significantly higher than other models, reflecting superior generalization ability.The results made it clear that the improved one-dimensional convolutional neural network prediction model could use characteristic frequencies to predict Crown pear firmness with great accuracy. The explanations from the Deep SHAP structure showed one key discovery. Vibration frequency features between 300 Hertz and 700 Hertz have a strong relationship with pear firmness. The interpretable deep learning model constructed in this study based on LDV vibration data achieves high-precision prediction while improving the interpretability and practicality of the model. It reveals the correlation between characteristic frequency and firmness, providing an efficient and reliable new method for non-destructive testing of Crown pear firmness, and also providing a theoretical basis for research in the field of fruit quality assessment.