Abstract: This study aimed to develop a stable online monitoring system for cleaning loss in wheat combine harvesters and to improve the recognition of grain impact signals under strong machine vibration and mixed separated materials. A double-layer differential piezoelectric sensing structure and a deep learning model based on continuous wavelet transform and squeeze-and-excitation attention were combined to suppress vibration interference, enhance discriminative time-frequency features, and provide a practical technical approach for accurate field monitoring of wheat cleaning loss. A double-layer piezoelectric sensor was designed with upper and lower stainless-steel sensing plates equipped with lead zirconate titanate piezoceramic elements. The upper plate collected mixed impact and vibration signals, whereas the lower plate mainly collected background vibration. The two-channel signals were subtracted in the host computer, and butyl-rubber damping components were used to reduce vibration transmission. Impact signals from wheat grains, straw, incompletely threshed wheat ears and background noise were collected on a bench test rig. The signals were resampled, transformed into time-frequency images by continuous wavelet transform, and classified using a squeeze-and-excitation convolutional neural network.The Huaimai 40 wheat variety was used for dataset construction. A total of 2,400 valid samples were obtained, including 1,200 grain impact samples and 1,200 negative samples composed of straw impact, wheat-ear impact and background-noise samples. The dataset was divided into training, validation and test sets at a ratio of 6:2:2. Each signal segment was normalized to 2,000 data points and converted into a 64 × 64-pixel color time-frequency image using a Morlet wavelet. The proposed network used three multi-scale convolutional blocks with convolution kernels of 3 × 3, 5 × 5 and 7 × 7, and squeeze-and-excitation modules were introduced to recalibrate channel features and enhance key frequency-band responses. On the test set, the proposed model achieved an accuracy of 94.2%, a precision of 95.7%, a recall of 92.5% and an F1-score of 94.1%. The confusion matrix showed that 222 of 240 grain samples and 230 of 240 non-grain samples were correctly identified. Ablation results showed that removing continuous wavelet transform reduced the accuracy by 2.1 percentage points, removing the squeeze-and-excitation module reduced it by 3.0 percentage points, and replacing continuous wavelet transform with short-time Fourier transform reduced it by 6.9 percentage points. Compared with support vector machine, random forest, multilayer perceptron and variable-scale one-dimensional convolutional neural network models, the proposed model obtained the best overall performance. The model contained about 1.15 million parameters, and the additional parameters introduced by the attention module accounted for less than 0.4% of the total. The average end-to-end processing time from raw signal interception to model inference was 78.6 milliseconds per frame, which was shorter than the 204.8-millisecond classification and interface-refresh cycle of the host-computer system. Bench tests further showed that the best sensor installation parameters were an inclination angle of 45° and a mounting height of 20 cm above the tail sieve. Detection accuracy decreased slightly as grain flow rate increased, mainly because continuous impacts and signal overlap became more likely at higher flow rates. Field validation was conducted in Xianggong Town, Linyi City, Shandong Province, using a Kubota 4LZ-4J combine harvester. Nine repeated runs were completed in the morning, at noon and in the afternoon on the same day to cover different crop moisture conditions. The average error rate between system outputs and manually measured cleaning-loss data was 6.57%.The developed double-layer differential sensor effectively reduced vibration interference, and the continuous wavelet transform combined with squeeze-and-excitation attention improved the extraction and recognition of wheat grain impact features. The system achieved high classification accuracy, acceptable computational complexity and near-real-time response in both bench and field tests. The proposed monitoring system provided accurate, stable and practical support for online detection of cleaning loss in wheat combine harvesters and could serve as a technical basis for intelligent harvesting-quality evaluation.