Rice harvesting loss rate monitoring method based on vibration signal analysis

Abstract

The invention discloses a rice harvesting loss rate monitoring method based on vibration signal analysis, and aims to break through the bottleneck of the existing rice harvesting loss rate monitoring technology and construct a set of efficient, accurate and real-time monitoring system. The invention aims to solve the problems of effective separation and purification of vibration signals, accurate identification and classification of rice grain collision signals, accurate counting of lost rice grains and harvested rice grains and real-time calculation of the loss rate in a complex environment in the prior art. S1, deploying a vibration signal acquisition system to collect a rice grain collision vibration signal; s2, vibration signal preprocessing and standardization; s3, comparing and learning a self-supervising pre-training encoder; s4, training a rice grain classification and counting model; s5, collision event detection and secondary collision filtering; s6, classifying and counting rice grains; and S7, calculating the rice harvesting loss rate.

Description

Technical Field

[0001] This invention relates to the field of smart agriculture technology, specifically to a rice harvest loss rate monitoring technology based on vibration signal detection and analysis. Background Technology

[0002] Rice harvesters are widely used in agricultural production, and their harvesting loss rate is a core indicator for measuring operational quality, directly affecting grain yield and farmers' economic benefits. Currently, monitoring rice harvesting loss rate mainly relies on manual inspection and weighing, which involves manually collecting and weighing any rice grains missed in the field, and then calculating the loss rate based on the actual harvest amount from the harvester. However, this method has significant drawbacks: manual collection is inefficient, labor-intensive, and susceptible to errors due to field terrain and crop distribution; monitoring cannot be conducted in real time, only after the harvest is completed, making it difficult to adjust the harvester's operating parameters immediately, thus hindering timely control of the loss rate.

[0003] Existing technologies include some loss rate monitoring solutions based on image recognition or sensor detection, but these have many limitations. Image recognition-based solutions are easily affected by factors such as changes in field lighting, weed obstruction, and sparse distribution of rice grains, resulting in unstable recognition accuracy. Traditional sensor detection solutions often rely on the acquisition of single physical parameters, failing to fully explore the vibration signal characteristics generated by rice grain collisions. This makes it difficult to accurately distinguish between effective rice grain signals and background noise, and it cannot effectively identify cases of multiple rice grains colliding, leading to large errors in rice grain counting and thus affecting the accuracy of loss rate calculation.

[0004] To address the aforementioned issues, this invention proposes a method for monitoring rice harvesting loss rate based on vibration signal analysis. By collecting vibration signals from the grain silo inlet baffle and the harvester outlet baffle, and combining self-supervised learning and feature engineering techniques, the method accurately identifies the number of uncollected lost rice grains and the number of collected harvested rice grains, achieving real-time and high-precision monitoring of the loss rate and providing technical support for optimizing rice harvester operations. Summary of the Invention

[0005] This invention aims to overcome the bottlenecks of existing rice transplanting quality assessment technologies and provide a comprehensive solution that is high-precision, intelligent, and multifunctional. By solving key technical problems such as robust detection and precise positioning of seedlings in complex environments (with a particular focus on mitigating model generalization problems through contrastive learning), accurate identification and location inference of missing seedlings, assessment of the rationality of seedling quantity and spacing, and visualization output based on geographic coordinates, this invention is expected to provide strong technical support for realizing large-scale intelligent precision agriculture and help improve rice production efficiency and management level.

[0006] A method for monitoring rice harvesting loss rate based on vibration signal analysis is described below:

[0007] Step S1: Deploy a vibration signal acquisition system to collect vibration signals from rice grain collisions;

[0008] Step S2: Vibration signal preprocessing and standardization;

[0009] Step S3: Comparative learning of the self-supervised pre-trained encoder;

[0010] Step S4: Training the rice grain classification and counting model;

[0011] Step S5: Collision event detection and secondary collision filtering;

[0012] Step S6: Rice grain sorting and counting;

[0013] Step S7: Calculation of rice harvesting loss rate.

[0014] The beneficial effects of this invention are as follows:

[0015] Addressing the limitations of current methods for calculating rice harvesting loss rates, this invention proposes a rice grain vibration signal classification and superposition counting method based on SimCLR self-supervised learning. This method exhibits significant advantages in several aspects and effectively overcomes the limitations of existing technologies, as detailed below:

[0016] 1. Significantly Improved Robustness of Feature Extraction in Complex Vibration Environments: Existing methods for analyzing rice grain vibration signals largely rely on manually designed features (such as peak amplitude and spectral centroid) or traditional supervised learning. Their feature representation capabilities are limited, making it difficult to cope with complex factors such as noise interference, equipment drift, and differences in rice grain physical properties (fullness, impurity type) in vibration signals, resulting in insufficient classification accuracy and counting stability. Furthermore, supervised deep models trained on a small number of labeled samples are prone to overfitting and are difficult to transfer to different varieties, harvest batches, or machine models. This invention innovatively employs SimCLR self-supervised learning to pre-train the vibration signal encoder. By applying random enhancements (scaling, noise, time shift, etc.) to the original signal, positive and negative sample pairs are constructed, learning a general and robust feature representation from massive amounts of unlabeled data. This pre-training method enables the model to capture the essential patterns of rice grain vibration, significantly improving its generalization ability in scenarios with noise interference and signal variations. Thanks to high-quality feature input, the subsequent classifier and counting head can maintain high-precision classification and counting performance even in situations of mixed types and signal distortion caused by minor equipment failures. This is an advantage that traditional feature engineering and supervised learning that relies on a small amount of labeled data cannot achieve.

[0017] 2. Improved Accuracy of Superimposed Signal Counting: Existing rice grain counting methods mostly rely on direct counting of peak values, assuming a single peak corresponds to a single grain. This approach fails completely when multiple grains collide simultaneously, generating superimposed signals, leading to significant counting errors in high-yield scenarios. This invention innovatively designs a quantity classification head training strategy based on synthetic samples: superimposed samples are generated by randomly mixing single-grain signals, and the counting model is trained using deep features extracted by SimCLR, directly achieving end-to-end prediction of the superimposed quantity. This method overcomes the dependence of traditional peak counting on single-grain signals and can effectively identify complex signal patterns generated by multiple grain superimposed signals. Experimental data shows that in dense collision scenarios, the counting error of this invention is significantly smaller than that of the traditional peak method, making it particularly suitable for accurate statistics during high-yield harvesting.

[0018] 3. Adaptability and Precision of Secondary Collision Filtering: Existing vibration signal denoising methods often use fixed thresholds (such as a single time window or amplitude ratio) to filter secondary collisions, which cannot adapt to complex scenarios with different collision intensities and time intervals, and are prone to problems such as mistakenly deleting primary collisions or retaining too much interference signal. This invention proposes a hierarchical dynamic threshold filtering strategy: setting differentiated amplitude ratio thresholds for different time intervals of ≤11ms, 11-45ms, and 45-80ms, and selecting primary collisions by combining amplitude and energy dual evaluation criteria. This method can adapt to the signal characteristics of different collision intensities, and the accuracy of secondary collision recognition is significantly improved compared with the fixed threshold method in dense collision scenarios, providing a cleaner signal source for subsequent classification and counting.

[0019] 4. Deep Integration of Real-Time Analysis and Dynamic Optimization of Production Parameters: Existing rice grain vibration signal analysis technologies mostly rely on offline processing, requiring manual data collection and batch analysis. This lack of real-time feedback on rice grain loss rates during harvesting leads to lags in harvester parameter adjustments, making it difficult to adapt to complex and variable harvesting conditions (such as differences in rice grain maturity and straw moisture). This invention achieves a real-time breakthrough from signal acquisition to result output through lightweight model design and engineering optimization. This real-time capability overcomes the limitations of traditional offline analysis, enabling dynamic optimization of equipment parameters during harvesting, significantly reducing rice grain loss rates, and directly improving harvesting efficiency and economic benefits—a core advantage that existing non-real-time technologies struggle to achieve.

[0020] Compared with existing methods for monitoring rice harvesting loss rates, this invention mainly improves in three aspects: comprehensiveness, professionalism, and real-time performance.

[0021] From a comprehensive perspective, this invention constructs an integrated monitoring platform that organically integrates functions such as vibration signal acquisition, self-supervised learning, loss rate calculation, and operational parameter feedback. Vibration signals are acquired in real time by wireless sensors deployed at the grain silo entrance baffle. Features are extracted using a self-supervised learning module, and the results are then processed by the loss rate calculation module. The operational parameter feedback module guides the harvester's adjustments, forming a complete closed-loop system of "acquisition-preprocessing-feature extraction-count prediction-loss rate calculation." Simultaneously, an adaptive filtering algorithm is introduced to eliminate environmental interference, and a multi-particle collision recognition mechanism is designed, overcoming the limitations of traditional single-parameter monitoring and enabling a more comprehensive and systematic evaluation of harvesting operation quality.

[0022] From a professional perspective, this invention improves monitoring accuracy through several innovative strategies. A contrastive learning strategy is employed, utilizing large-scale unlabeled vibration signals for self-supervised pre-training, enabling the encoder to learn the general characteristics of rice grain collisions, enhancing its cross-scene transferability and noise discrimination capabilities. An end-to-end counting model is constructed, and a multi-grain collision signal separation method based on time-frequency domain features, a dynamic threshold counting strategy, and an error correction mechanism are designed, improving counting accuracy to over 95%. Furthermore, by quantifying the energy distribution patterns of rice grain collisions and establishing an energy-quantity mapping model, spatial-temporal joint modeling is achieved, providing professional technical support for accurate assessment.

[0023] From a real-time perspective, this invention possesses highly efficient computational and response capabilities. The designed algorithm framework prioritizes efficiency while maintaining accuracy. For example, the encoder in the SimCLR model employs a convolutional neural network structure, enabling rapid extraction of vibration signal features. Combined with a downstream random forest classifier and counting head, it achieves rapid prediction. The vibration signal acquisition module acquires data in real time, and the loss rate calculation and operational parameter feedback are dynamically updated, keeping the time for a single assessment within 30 seconds. Compared to traditional methods that take several hours or more, this represents a qualitative leap, providing real-time support for timely adjustments to harvesting operational parameters.

[0024] In summary, this invention aims to overcome the bottlenecks of existing rice transplanting quality assessment technologies and provide a comprehensive solution that is high-precision, intelligent, and multifunctional. By addressing key technical challenges such as robust detection and precise positioning of seedlings in complex environments (with a particular focus on mitigating model generalization issues through contrastive learning), accurate identification and location inference of missing seedlings, assessment of the rationality of seedling quantity and spacing, and geographic coordinate-based visualization output, this invention is expected to provide strong technical support for achieving large-scale intelligent precision agriculture and help improve rice production efficiency and management levels.

[0025] This invention proposes a rice grain classification and counting method based on SimCLR self-supervised learning. This method utilizes a self-supervised contrastive learning strategy to improve the robustness of rice grain vibration signal feature representation, and combines enhanced secondary collision filtering and superposition recognition techniques to achieve accurate classification and counting of rice grains. Compared with traditional counting methods, this invention achieves significant improvements in classification accuracy, superposition recognition capability, and anti-interference ability, effectively solving the problems of secondary collision misjudgment and superposition rice grain counting.

[0026] Specifically, this invention utilizes SimCLR self-supervised learning to effectively leverage a large amount of rice grain vibration signal data, pre-training an encoder with stronger generalization capabilities. This encoder effectively overcomes the influence of signal noise, environmental interference, and differences in the physical characteristics of rice grains, extracting robust rice grain signal features even in complex vibration scenarios. Applying this pre-trained encoder to the training of downstream classifiers and quantity classification heads significantly improves the model's classification accuracy for plump rice grains compared to other types (shriveled grains, straw, and light debris), as well as its ability to predict the number of stacked rice grains, laying a solid foundation for accurate counting. Compared to directly training a classification model on a small amount of labeled data, the method of this invention can more effectively utilize the rich information contained in unlabeled data, improving the model's classification performance in complex vibration signal scenarios.

[0027] Furthermore, the enhanced secondary collision filtering method proposed in this invention can accurately identify and remove secondary collision signals. Through layered time windows and amplitude ratio judgment rules, even in dense collision scenarios, this method can reliably distinguish between primary and secondary collisions, avoiding misjudgments caused by improper single threshold settings in traditional methods. This invention can also accurately count the number of plump rice grains and other types, and predict the number of superimposed rice grains through a quantity classification head, scientifically evaluating the rice grain separation effect and providing objective data support for optimizing the separation process. This counting method based on signal characteristics and superposition models can more accurately reflect the actual number of rice grains, overcoming the limitations of traditional methods that rely solely on peak counts. This invention further realizes batch processing capabilities, enabling automated analysis of large amounts of experimental data and rapid calculation of rice harvesting loss rates. This allows operators to quickly adjust harvester parameters based on real-time analysis results, reducing rice grain loss rates and improving harvesting efficiency.

[0028] In summary, this invention provides an automated, high-precision, and multifunctional solution for rice grain classification and counting, offering strong technical support for achieving intelligent and precise rice harvesting. Attached Figure Description

[0029] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0030] Specific implementation method one: Combining Figure 1 This embodiment describes a method for monitoring rice harvesting loss rate based on vibration signal analysis. The specific process is as follows:

[0031] This invention uses vibration signal analysis to monitor rice harvesting loss rate. The overall steps of the method are shown in the figure below. Figure 1 As shown, the specific steps are as follows:

[0032] S1. Deploy a vibration signal acquisition system to collect vibration signals from rice grain collisions;

[0033] S2. Vibration signal preprocessing and standardization;

[0034] S3, Contrastive learning of self-supervised pre-trained encoders;

[0035] S4. Training of rice grain classification and counting model;

[0036] S5, Collision event detection and secondary collision filtering;

[0037] S6. Rice grain classification and counting;

[0038] S7. Calculation of rice harvesting loss rate;

[0039] Specific Implementation Scheme Two: This implementation scheme differs from Specific Implementation Scheme One in that step S1 involves deploying a vibration signal acquisition system to collect rice grain collision vibration signals; it includes the following steps:

[0040] S11. Sensor Selection and Deployment Scheme:

[0041] (1) Select a high-precision wireless vibration sensor (sampling rate ≥25600Hz) with anti-interference capability and low power consumption to adapt to the complex environment of farmland.

[0042] (2) The sensor is installed at the outlet of the harvester's separation device to ensure that it can effectively capture the vibration signal generated by the collision of rice grains.

[0043] (3) The sampling data is uploaded to the cloud server in real time via the network.

[0044] S12. Data Acquisition Parameter Settings and Data Storage:

[0045] (1) Set the collection time to the entire harvester operation and the triggering method to continuous collection (to avoid missing instantaneous collision signals).

[0046] (2) The data is stored in the original waveform format (such as CSV), and auxiliary information such as collection time, harvester location (in conjunction with GPS module) and operation speed are recorded to provide scene context for subsequent analysis.

[0047] The other steps and parameters are the same as in Specific Implementation Method 1.

[0048] Specific Implementation Scheme Three: This implementation scheme differs from Specific Implementation Scheme One or Two in that the vibration signal preprocessing and standardization in step S2 includes the following steps:

[0049] S21. Signal Denoising and Filtering:

[0050] (1) Baseline correction: Baseline correction is performed on the original vibration signal to eliminate DC offset.

[0051] (2) Noise filtering: A fourth-order Butterworth bandpass filter was used to filter environmental noise while retaining the characteristic frequency band signal of rice grain collision. The filtering effect was verified by analyzing the signal spectrum distribution through short-time Fourier transform.

[0052] S22. Signal segmentation and standardization:

[0053] (1) Based on the peak detection algorithm, collision events are located and window segments are extracted for each collision signal.

[0054] (2) Standardize the fragment data to eliminate amplitude scale differences and improve data consistency.

[0055] Specific Implementation Method Four: This implementation method differs from Specific Implementation Methods One to Three in that the contrastive learning self-supervised pre-trained encoder in S3 includes the following detailed steps:

[0056] S31. Vibration signal enhancement and construction of positive and negative sample pairs:

[0057] A random augmentation strategy is applied to a single collision signal segment to generate positive sample pairs. Specifically, the augmentation strategy employed includes:

[0058] (1) Amplitude scaling: Randomly scale by 0.8-1.2 times to simulate the vibration signal generated by the collision of rice grains of different sizes.

[0059] (2) Temporal shift: Randomly shift 20 sampling points forward and backward to randomly cut segments from the long signal. Negative sample pairs are randomly selected from signals of collision events of different materials to ensure that they are not related to the anchor sample.

[0060] (3) Add Gaussian noise, salt and pepper noise or blurring to simulate machine vibration interference during actual vibration signal collection.

[0061] This invention employs a series of random signal enhancement strategies for individual collision signal segments to simulate the vibration differences generated by collisions of rice grains of different sizes, the selection of different segments within a long signal, and machine vibration interference during actual vibration signal collection. Each anchored collision signal segment and its enhanced version constitute a positive sample pair. Negative samples are randomly selected from the dataset from other signal segments with different sources (i.e., collision events of different materials), ensuring that there is no correlation between negative samples and anchored samples. By constructing a large number of positive and negative sample pairs, sufficient training data is provided for the contrastive learning self-supervised pre-trained encoder.

[0062] S32. Select encoder architecture:

[0063] This invention selects a 1D convolutional neural network as the encoder for self-supervised pre-training. This 1D convolutional neural network mainly consists of three convolutional layers to achieve channel conversion; it is compressed into a 128-dimensional feature vector through adaptive average pooling (AdaptiveAvgPool1D) and then output after L2 normalization. This encoder has efficient feature extraction capabilities for one-dimensional signals, effectively capturing local and global features in vibration signals, and is suitable for feature learning tasks related to rice grain collision signals.

[0064] S33. Design of contrastive loss function:

[0065] This invention employs the NT-Xent loss function as the loss function for contrastive learning to maximize the consistency of feature representations between positive sample pairs while minimizing the consistency of feature representations between negative sample pairs. For a batch of anchored samples... The positive samples obtained through data augmentation are The set of randomly sampled negative samples is The NT-Xent loss function is defined as follows:

[0066] in and The feature vectors of positive sample pairs It is a temperature coefficient used to adjust the discrimination between different sample pairs and control the degree of separation between positive and negative sample pairs in the feature space. The batch size is specified. The Adam optimizer (learning rate 1e-3) is used, and the pre-trained encoder weights are saved.

[0067] S34, Self-supervised pre-training:

[0068] The constructed positive and negative sample pairs are input into a 1D convolutional neural network encoder to obtain feature vectors for the anchored sample and its enhanced version, respectively. The contrastive loss of the positive and negative sample pairs is calculated using the NT-Xent loss function, where positive sample pairs are feature vector pairs obtained from the same collision signal after different enhancements, and negative sample pairs are feature vector pairs from collision signals of different materials. The encoder parameters are updated using the Adam optimizer, and the contrastive loss is minimized through iterative training. Pre-training is performed on a dataset containing a large number of collision vibration signals from different materials, including rice grains (full and shriveled), straw, and light debris, enabling the encoder to learn robust and discriminative feature representations for various collision signals, thus improving its feature extraction capability for vibration signals from different materials. After pre-training, the encoder weights are saved and used to initialize the feature extraction parts of the downstream rice grain classifier and quantity prediction model, realizing knowledge transfer from vibration signal features to features for the full rice grain detection task.

[0069] The other steps and parameters are the same as those in specific implementation methods one to three.

[0070] Specific Implementation Method Five: This implementation method differs from Specific Implementation Methods One to Four in that the rice grain classification and counting model training in S4 includes the following detailed steps:

[0071] S41. Downstream Classifier Training (Full / Unfull Grain Recognition): The rice grain classification and counting model of this invention consists of three parts: a feature extraction module, a classification head, and a counting head. The feature extraction module uses the self-supervised pre-trained 1D convolutional neural network encoder from S3, which is responsible for extracting robust feature representations from vibration signals. The classification head uses a random forest classifier, which receives the feature vector output by the encoder to achieve binary classification of "full grains" and "other types (shriveled grains, straw, light debris, etc.)". The counting head uses a random forest regressor, which predicts the number of superimposed grains based on the same feature vector, solving the signal superposition problem caused by multiple grains colliding simultaneously. This architecture, by sharing the feature extraction capabilities of the pre-trained encoder, balances classification accuracy and counting efficiency, and is suitable for real-time rice grain detection scenarios.

[0072] S42. Superimposed Particle Counting Head Training: The encoder weights obtained from the self-supervised pre-training of large-scale vibration signals in S3 are used as the initial weights for the feature extraction modules of the downstream classification and counting models. Through knowledge transfer, the general features of vibration signals contained in unlabeled data (such as collision amplitude changes, spectral characteristics, etc.) are fully utilized, enabling the model to maintain strong feature discrimination ability even with limited labeled data, accelerating training convergence and improving generalization performance. During the transfer process, the first two convolutional parameters of the encoder are frozen (preserving the basic vibration feature extraction capability), and only the classification head and counting head are trained; in the later stage of training, all layers are unfrozen, and end-to-end fine-tuning is performed to make the feature extraction more suitable for the downstream task.

[0073] S43. Loss Function Design: The cross-entropy loss function is used to optimize the classification head, and the mean squared error (MSE) loss function is used to optimize the counting head. The specific loss function forms are as follows:

[0074]

[0075] in, The true label for sample i (1 represents full grains, 0 represents other types). The probability of the model predicting a saturated grain. This represents the number of samples.

[0076]

[0077] in, The true number of particles in sample i. To predict the number of particles for the model, This represents the number of samples.

[0078] S44. Optimizer Selection and Parameter Settings: Adam was selected as the deep learning optimizer. A piecewise learning rate decay strategy was adopted, using a higher learning rate in the early stages of training to accelerate convergence, and then reducing the learning rate in the later stages to improve accuracy. Early stopping was used during training to prevent model overfitting.

[0079] S45, Training Process:

[0080] (1) Divide the constructed classification and counting datasets into training set, validation set and test set respectively, with the training set, validation set and test set divided in a ratio of 7:2:1.

[0081] (2) Load the pre-trained encoder weights, initialize the feature extraction module, and randomly initialize the classification head and counting head parameters.

[0082] (3) Input the training set into the model, extract features through the encoder, input the classification head and the counting head respectively to obtain the prediction results, calculate the classification loss and the counting loss and sum them.

[0083] (4) Calculate the gradient using the backpropagation algorithm and update the model parameters using the Adam optimizer.

[0084] After each training round, the model performance is evaluated on the validation set, and the model parameters with the highest classification F1 score and the smallest count MAE are saved.

[0085] The other steps and parameters are the same as those in specific implementation methods one through four.

[0086] Specific Implementation Method Six: This implementation method differs from Specific Implementation Methods One to Five in that the collision event detection and secondary collision filtering in S5 include the following detailed steps:

[0087] S51. Peak Detection Algorithm: The following peak detection process is performed on the pre-processed vibration signal (bandpass filtering, baseline correction):

[0088] (1) Amplitude envelope calculation: The instantaneous amplitude envelope of the signal is extracted using Hilbert transform.

[0089]

[0090] in This is the preprocessed vibration signal. This represents the Hilbert transform, which highlights the energy concentration characteristics of collision events through envelope calculation.

[0091] (2) Adaptive threshold setting: The detection threshold is set based on the statistical characteristics of the envelope signal.

[0092]

[0093] in The mean of the envelope signal is denoted as . The threshold is the standard deviation of the envelope signal, and it can dynamically adapt to the signal noise level under different operating conditions.

[0094] (3) Peak filtering: The local maximum detection algorithm is used to identify peaks in the envelope signal that exceed the threshold. The potential peak points are identified, and a peak spacing constraint is applied. The time interval between any two peak points must be ≥0.005 seconds to avoid repeated detections caused by high-frequency signal oscillations. The final output is a set of time indexes of potential collision peaks. ,in This represents the timestamp of the i-th peak.

[0095] S52. Secondary Collision Filtering: For redundant peak values ​​of secondary collisions caused by factors such as rice grain bouncing, equipment resonance, and straw friction, a layered filtering strategy is adopted to remove secondary collisions (redundant peak values ​​caused by rice grain bouncing, equipment resonance, etc.) and retain the real primary collision events.

[0096] (1) Time grouping: for peak sets Arranged chronologically, starting from the first peak, consecutive peaks with a time difference ≤ 0.2 seconds are grouped together to form several peak groups. .

[0097] (2) Amplitude ratio screening: For each group Calculate the maximum amplitude peak (main peak) within the group.

[0098]

[0099] in Peak value The corresponding signal amplitude is calculated using the maximum absolute value within a 30ms window. The amplitude ratio of the remaining peaks (secondary peaks) to the main peak is... .

[0100] (3) Intra-group retention rule: The time difference between the secondary peak and the primary peak is ≤0.2 seconds and If the value is less than 0.6, it is considered a secondary collision and is discarded; only the main peak is retained in each group. This forms a set of peak values ​​after initial filtering. 1, The focus is on removing strong redundant signals with significant energy decay.

[0101] (4) Domain Relationship Modeling: This involves transforming the initially filtered peak data set... It is mapped to a set of points in the time-amplitude feature space.

[0102] ( Peak value The timestamp is calculated using the sampling point index and sampling interval. A sliding window grouping strategy is used to construct neighborhood relationships, that is, the peak value is dynamically determined to belong to the same neighborhood by the time difference, so as to ensure that adjacent real collision events are not mistakenly classified into the same neighborhood.

[0103] (5) Dynamic threshold filtering: For filters that satisfy a time difference of 0.011 ≤ ≤0.08 seconds ( The peak value of (p,q) is set using a piecewise dynamic amplitude ratio threshold: when 0.011 ≤ When the time is ≤0.0045 seconds, the threshold is 0.85, that is, if If q ≤ 0.85, then it is a second-order collision; if 0.045 ≤ When the time is ≤0.08 seconds, the threshold linearly decreases to 0.55, that is, if If q is ≤0.55, then it is a second-order collision.

[0104] (6) Final main collision determination: For the remaining peaks after dynamic threshold filtering, the principle of "amplitude priority, energy secondary selection" is adopted. In each neighborhood, the peak with the largest amplitude is selected first. If the amplitudes are close (difference <5%), the energy is determined again through a 30ms window, and the final set of main collision events is output to ensure that the retained peaks truly reflect the characteristics of rice grain collision.

[0105] The other steps and parameters are the same as those in specific implementation methods one through five.

[0106] Specific Implementation Method Seven: This implementation method differs from Specific Implementation Methods One to Six in that the rice grain classification and counting in step S6 includes the following detailed steps:

[0107] S61. Single-peak signal classification: For each signal segment of the main collision event, features are extracted through a pre-trained encoder and input into a downstream random forest classifier to predict the "full" or "other" type. The classification confidence score is output to quantify the reliability of the prediction. When the confidence score is ≥0.7, the classification result is directly adopted; otherwise, it is marked as "low confidence" and subsequently verified by combining energy features.

[0108] S62. Stacked Grain Count Prediction: For signals classified as "full grains," a pre-trained quantity classification head is invoked, and the signal energy and peak amplitude are used for verification (energy is positively correlated with grain count). For signals classified as "other," a heuristic rule (energy / single grain energy threshold) is used to estimate the quantity.

[0109] S62, Rice Grain Classification and Counting Model Test:

[0110] (1) Use a test dataset that is independent of the training and validation sets to ensure the objectivity and reliability of the test results. The test dataset should cover the collision vibration signals of different material combinations, such as plump rice grains and shriveled rice grains, plump rice grains and straw, etc.

[0111] (2) Select appropriate model performance evaluation metrics to comprehensively evaluate the detection accuracy of the model. Evaluation metrics for classification tasks include: precision, recall, and F1 score; evaluation metrics for counting tasks include: mean absolute error (MAE), root mean square error (RMSE), and counting accuracy.

[0112] (3) Test the trained model using the test dataset and calculate various performance evaluation metrics. Based on the test results, further optimize the model or adjust the training strategy.

[0113] The other steps and parameters are the same as those in specific implementation methods one through six.

[0114] Specific Implementation Method Eight: This implementation method differs from Specific Implementation Methods One to Seven in that the calculation of rice harvesting loss rate in S7 includes the following detailed steps:

[0115] S7. Calculation of Rice Harvest Loss Rate: Accumulate the total number of rice grains predicted as "full" in all main collision events, and calculate the loss rate by combining the total number of lost rice grains and the total number of harvested rice grains.

[0116]

[0117] The other steps and parameters are the same as those in specific implementation methods one through seven.

[0118] Key points of the invention

[0119] 1. A Self-Supervised Pre-training Scheme for Rice Grain Vibration Signals Based on SimCLR. This invention employs a SimCLR self-supervised contrastive learning strategy, utilizing a large amount of rice grain vibration signal data to pre-train the encoder. By applying random enhancements (such as scaling, noise addition, time shifting, length normalization, etc.) to the vibration signals, positive sample pairs (different enhanced views of the same signal) and negative sample pairs (enhanced views of different signals or signals of different categories) are generated. The NT-Xent loss function is designed to bring the representation distance of similar rice grains (such as plump grains) closer in the embedding space, while pushing away the representation of dissimilar samples (such as shriveled grains, straw, and light debris). This enables the encoder to learn more discriminative and robust feature representations of rice grain types (plump / shriveled / straw / light debris). This method fully utilizes the information in unlabeled data, effectively overcoming the influence of signal noise, environmental interference, and differences in the physical characteristics of rice grains, significantly improving the encoder's feature extraction capability in complex vibration scenarios, laying the foundation for subsequent accurate classification and counting, rather than relying solely on a small amount of labeled data to train the model. The key lies in enhancing the generalization of vibration signal feature representation through self-supervised learning, thereby improving the model's adaptability to complex scenarios.

[0120] 2. Enhanced Secondary Collision Filtering Method. During high-speed threshing, individual rice grains often experience multiple consecutive impacts due to rebound, slippage, or baffle resonance, forming a composite event sequence of "primary collision + secondary / tertiary collision." Failure to differentiate these events will lead to severely inflated counting results. To address the counting error caused by secondary collisions in vibration signals, this invention proposes a two-stage filtering strategy: basic filtering uses a 200ms time window and a 0.6 amplitude ratio threshold to initially screen primary collisions; secondary collisions are further accurately identified using a layered time window and a dynamic amplitude ratio threshold. Only the peak with the largest amplitude (secondary energy maximum) in each collision group is retained as the primary collision, effectively solving the misjudgment problem in dense collision scenarios and significantly improving the accuracy of primary collision detection.

[0121] 3. Rice Grain Classification and Stacking Counting Model Based on SimCLR Features. Traditional rice grain counting methods are generally based on the assumption of "single peak = single grain," which is completely unable to handle the stacking signals generated by multiple grains colliding simultaneously, resulting in a serious underestimation of the number of grains in high-yield fields. This invention constructs a two-level classification architecture that integrates a self-supervised pre-trained encoder: the downstream classifier uses features extracted by SimCLR to accurately classify plump rice grains from other types; the quantity classification head is trained by synthesizing stacking signal samples to achieve quantitative prediction of stacked rice grains. This model overcomes the limitations of traditional peak counting, combining vibration signal features and a stacking model, which can accurately count the number of various types of rice grains and scientifically evaluate the rice grain separation effect and loss rate, providing data support for optimizing harvesting processes.

[0122] In summary, this invention addresses the industry pain point of difficulty in real-time and accurate monitoring of loss rates during rice harvesting by systematically constructing a high-precision vibration signal analysis system that integrates signal perception, self-supervised learning, intelligent recognition, and dynamic feedback. Its innovation and technological advancement are concentrated in the following aspects: First, it utilizes SimCLR self-supervised learning to improve the robustness of vibration signal features, effectively overcoming the identification difficulties caused by noise interference, equipment vibration coupling, and differences in the physical state of rice grains under complex field conditions, laying a robust feature foundation for downstream tasks; Second, it proposes a hierarchical dynamic threshold secondary collision filtering method, which significantly reduces the false detection rate while ensuring high recall, solving the bottleneck of the traditional fixed threshold method failing in high-speed operation scenarios, and accurately identifying the real main events in dense collisions; Third, it constructs a classification and superimposed counting model based on feature learning, which can not only distinguish between plump rice grains and invalid debris, but also accurately count the number of plump rice grains in the loss path and harvest path, providing a reliable data source for loss rate calculation and breaking through the limitations of traditional counting methods; These innovations together constitute a high-precision, automated rice grain classification and counting solution, providing technical support for the intelligent and refined management of rice harvester operations.

[0123] What are the advantages of this invention compared to the closest existing technology?

[0124] The present invention proposes a rice grain vibration signal classification and superposition counting method based on SimCLR self-supervised learning, which demonstrates significant advantages in several aspects and effectively overcomes the limitations of existing technologies. Specific comparisons are as follows:

[0125] 1. Significantly Improved Robustness of Feature Extraction in Complex Vibration Environments: Existing methods for analyzing rice grain vibration signals largely rely on manually designed features (such as peak amplitude and spectral centroid) or traditional supervised learning. Their feature representation capabilities are limited, making it difficult to cope with complex factors such as noise interference, equipment drift, and differences in rice grain physical properties (fullness, impurity type) in vibration signals, resulting in insufficient classification accuracy and counting stability. Furthermore, supervised deep models trained on a small number of labeled samples are prone to overfitting and are difficult to transfer to different varieties, harvest batches, or machine models. This invention innovatively employs SimCLR self-supervised learning to pre-train the vibration signal encoder. By applying random enhancements (scaling, noise, time shift, etc.) to the original signal, positive and negative sample pairs are constructed, learning a general and robust feature representation from massive amounts of unlabeled data. This pre-training method enables the model to capture the essential patterns of rice grain vibration, significantly improving its generalization ability in scenarios with noise interference and signal variations. Thanks to high-quality feature input, the subsequent classifier and counting head can maintain high-precision classification and counting performance even in situations of mixed types and signal distortion caused by minor equipment failures. This is an advantage that traditional feature engineering and supervised learning that relies on a small amount of labeled data cannot achieve.

[0126] 2. Improved Accuracy of Superimposed Signal Counting: Existing rice grain counting methods mostly rely on direct counting of peak values, assuming a single peak corresponds to a single grain. This approach fails completely when multiple grains collide simultaneously, generating superimposed signals, leading to significant counting errors in high-yield scenarios. This invention innovatively designs a quantity classification head training strategy based on synthetic samples: superimposed samples are generated by randomly mixing single-grain signals, and the counting model is trained using deep features extracted by SimCLR, directly achieving end-to-end prediction of the superimposed quantity. This method overcomes the dependence of traditional peak counting on single-grain signals and can effectively identify complex signal patterns generated by multiple grain superimposed signals. Experimental data shows that in dense collision scenarios, the counting error of this invention is significantly smaller than that of the traditional peak method, making it particularly suitable for accurate statistics during high-yield harvesting.

[0127] 3. Adaptability and Precision of Secondary Collision Filtering: Existing vibration signal denoising methods often use fixed thresholds (such as a single time window or amplitude ratio) to filter secondary collisions, which cannot adapt to complex scenarios with different collision intensities and time intervals, and are prone to problems such as mistakenly deleting primary collisions or retaining too much interference signal. This invention proposes a hierarchical dynamic threshold filtering strategy: setting differentiated amplitude ratio thresholds for different time intervals of ≤11ms, 11-45ms, and 45-80ms, and selecting primary collisions by combining amplitude and energy dual evaluation criteria. This method can adapt to the signal characteristics of different collision intensities, and the accuracy of secondary collision recognition is significantly improved compared with the fixed threshold method in dense collision scenarios, providing a cleaner signal source for subsequent classification and counting.

[0128] 4. Deep Integration of Real-Time Analysis and Dynamic Optimization of Production Parameters: Existing rice grain vibration signal analysis technologies mostly rely on offline processing, requiring manual data collection and batch analysis. This lack of real-time feedback on rice grain loss rates during harvesting leads to lags in harvester parameter adjustments, making it difficult to adapt to complex and variable harvesting conditions (such as differences in rice grain maturity and straw moisture). This invention achieves a real-time breakthrough from signal acquisition to result output through lightweight model design and engineering optimization. This real-time capability overcomes the limitations of traditional offline analysis, enabling dynamic optimization of equipment parameters during harvesting, significantly reducing rice grain loss rates, and directly improving harvesting efficiency and economic benefits—a core advantage that existing non-real-time technologies struggle to achieve.

[0129] This invention may have other embodiments. Without departing from the spirit and essence of this invention, those skilled in the art can make various corresponding changes and modifications according to this invention, but these corresponding changes and modifications should all fall within the protection scope of the appended claims.

Claims

1. A method for monitoring rice harvesting loss rate based on vibration signal analysis, characterized in that: The specific process of the method is as follows: Step S1: Deploy a vibration signal acquisition system to collect vibration signals from rice grain collisions; Step S2: Vibration signal preprocessing and standardization; Step S3: Comparative learning of the self-supervised pre-trained encoder; Step S4: Training the rice grain classification and counting model; Step S5: Collision event detection and secondary collision filtering; Step S6: Rice grain sorting and counting; Step S7: Calculation of rice harvesting loss rate.

2. The method for monitoring rice harvesting loss rate based on vibration signal analysis according to claim 1, characterized in that: In step S1, the vibration signal acquisition system is deployed to collect vibration signals from rice grain collisions. The specific process is as follows: S11. Sensor selection and deployment plan; the specific process is as follows: (1) Select a high-precision wireless vibration sensor (sampling rate ≥25600Hz) with anti-interference capability and low power consumption to adapt to complex farmland environments; (2) The sensor is installed at the outlet of the harvester's separating device to ensure that it can effectively capture the vibration signal generated by the collision of rice grains; (3) The sampling data is uploaded to the cloud server in real time via the network; S12. Data acquisition parameter settings and data storage; the specific process is as follows: (1) Set the data collection duration to the entire harvester operation time and the triggering method to continuous data collection (to avoid missing instantaneous collision signals). (2) The data is stored in the original waveform format (such as CSV), and auxiliary information such as collection time, harvester location (in conjunction with GPS module) and operation speed are recorded to provide scene context for subsequent analysis.

3. The method for monitoring rice harvesting loss rate based on vibration signal analysis according to claim 2, characterized in that: The vibration signal preprocessing and standardization in step S2 is as follows: S21. Signal noise reduction and filtering; the specific process is as follows: (1) Baseline correction: Baseline correction is performed on the original vibration signal to eliminate DC offset; (2) Noise Filtering: A fourth-order Butterworth bandpass filter was used to filter environmental noise while preserving the characteristic frequency band signal of rice grain collision. The filtering effect was verified by analyzing the signal spectrum distribution through short-time Fourier transform. S22. Signal segmentation and standardization; the specific process is as follows: (1) Based on the peak detection algorithm, locate collision events and extract window segments for each collision signal; (2) Standardize the fragment data to eliminate amplitude scale differences and improve data consistency.

4. The method for monitoring rice harvesting loss rate based on vibration signal analysis according to claim 3, characterized in that: The specific process of contrastive learning of the self-supervised pre-trained encoder in step S3 is as follows: S31. Vibration signal enhancement and construction of positive and negative sample pairs; the specific process is as follows: A random augmentation strategy is applied to a single collision signal segment to generate positive sample pairs. Specifically, the augmentation strategy employed includes: (1) Amplitude scaling: Randomly scale by 0.8-1.2 times to simulate the vibration signal generated by the collision of rice grains of different sizes; (2) Temporal shift: Randomly shift 20 sampling points forward and backward to randomly cut segments from the long signal. Negative sample pairs are randomly selected from signals of collision events of different materials to ensure that they are not related to the anchored samples; (3) Add Gaussian noise, salt and pepper noise or blurring to simulate machine vibration interference during the actual vibration signal collection process; This invention employs a series of random signal enhancement strategies for individual collision signal segments to simulate the vibration differences generated by collisions of rice grains of different sizes, the selection of different segments within a long signal, and machine vibration interference during actual vibration signal collection. Each anchored collision signal segment and its enhanced version constitute a positive sample pair. Negative samples are randomly selected from the dataset from other signal segments with different origins (i.e., collision events involving different materials), ensuring that there is no correlation between negative samples and anchored samples. By constructing a large number of positive and negative sample pairs, sufficient training data is provided for the contrastive learning self-supervised pre-trained encoder. S32. Select encoder architecture; the specific process is as follows: This invention selects a 1D convolutional neural network as the encoder for self-supervised pre-training. This 1D convolutional neural network mainly consists of three convolutional layers to achieve channel conversion; it is compressed into a 128-dimensional feature vector through adaptive average pooling (AdaptiveAvgPool1D) and then output after L2 normalization. This encoder has efficient feature extraction capabilities for one-dimensional signals, effectively capturing local and global features in vibration signals, and is suitable for feature learning tasks related to rice grain collision signals. S33. Design of the contrastive loss function; the specific process is as follows: This invention employs the NT-Xent loss function as the loss function for contrastive learning to maximize the consistency of feature representations between positive sample pairs while minimizing the consistency of feature representations between negative sample pairs. For a batch of anchored samples... The positive samples obtained through data augmentation are The set of randomly sampled negative samples is The NT-Xent loss function is defined as follows: in and The feature vectors of positive sample pairs It is a temperature coefficient used to adjust the discrimination between different sample pairs and control the degree of separation between positive and negative sample pairs in the feature space. The batch size is specified. The Adam optimizer (learning rate 1e-3) is used, and the pre-trained encoder weights are saved. S34. Self-supervised pre-training; the specific process is as follows: The constructed positive and negative sample pairs are input into a 1D convolutional neural network encoder to obtain feature vectors for the anchored sample and its enhanced version, respectively. The contrastive loss of the positive and negative sample pairs is calculated using the NT-Xent loss function, where the positive sample pair is a feature vector pair obtained from the same collision signal after different enhancements, and the negative sample pair is a feature vector pair from collision signals of different materials. The encoder parameters are updated using the Adam optimizer, and the contrastive loss is minimized through iterative training. Pre-training is performed on a dataset containing a large number of collision vibration signals from different materials, including rice grains (full and shriveled), straw, and light debris, enabling the encoder to learn robust and discriminative feature representations for various collision signals, thus improving its feature extraction capability for vibration signals from different materials. After pre-training, the encoder weights are saved and used to initialize the feature extraction parts of the downstream rice grain classifier and quantity prediction model, realizing knowledge transfer from vibration signal features to features for the full rice grain detection task.

5. The method for monitoring rice harvesting loss rate based on vibration signal analysis according to claim 4, characterized in that: The specific process for training the rice grain classification and counting model in step S4 is as follows: S41. Downstream classifier training (saturated / non-saturated particle recognition); the specific process is as follows: The rice grain classification and counting model of this invention consists of three parts: a feature extraction module, a classification head, and a counting head. The feature extraction module uses a self-supervised pre-trained 1D convolutional neural network encoder in S3 to extract robust feature representations from vibration signals. The classification head uses a random forest classifier to receive the feature vector output by the encoder and achieve binary classification between "full grains" and "other types (shriveled grains, straw, light debris, etc.)". The counting head uses a random forest regressor to predict the number of overlapping grains based on the same feature vector, solving the signal superposition problem caused by simultaneous collisions of multiple grains. This architecture, by sharing the feature extraction capabilities of the pre-trained encoder, balances classification accuracy and counting efficiency, making it suitable for real-time rice grain detection scenarios. S42. Training of the superimposed particle counter head; the specific process is as follows: The encoder weights obtained from self-supervised pre-training with large-scale vibration signals in S3 are used as the initial weights for the feature extraction modules of the downstream classification and counting models. Through knowledge transfer, the general features of vibration signals (such as collision amplitude changes and spectral characteristics) contained in unlabeled data are fully utilized, enabling the model to maintain strong feature discrimination ability even with limited labeled data, accelerating training convergence and improving generalization performance. During the transfer process, the first two convolutional parameters of the encoder are frozen (preserving the basic vibration feature extraction capability), and only the classification head and counting head are trained; in the later stage of training, all layers are unfrozen, and end-to-end fine-tuning is performed to make the feature extraction more suitable for the downstream task. S43. Loss function design; the specific process is as follows: The cross-entropy loss function is used to optimize the classification head, and the mean squared error (MSE) loss function is used to optimize the counting head. The specific loss function forms are as follows: in, The true label for sample i (1 represents full grains, 0 represents other types). N represents the probability that the model predicts a saturated grain, and N is the number of samples. in, The true number of particles in sample i. To predict the number of particles for the model, The number of samples; S44. Optimizer selection and parameter setting; the specific process is as follows: Adam was chosen as the deep learning optimizer. A segmented learning rate decay strategy was employed, using a higher learning rate in the early stages of training to accelerate convergence and then reducing the learning rate in the later stages to improve accuracy. Early stopping was used during training to prevent overfitting. S45, Training Process: (1) Divide the constructed classification and counting datasets into training set, validation set and test set respectively, with the training set, validation set and test set divided in a ratio of 7:2:1; (2) Load the pre-trained encoder weights, initialize the feature extraction module, and randomly initialize the classification head and counting head parameters; (3) Input the training set into the model, extract features through the encoder, input the classification head and the counting head respectively to obtain the prediction results, calculate the classification loss and the counting loss and sum them; (4) Calculate the gradient using the backpropagation algorithm and update the model parameters using the Adam optimizer; (5) After each round of training, evaluate the model performance on the validation set and save the model parameters with the highest classification F1 value and the smallest count MAE.

6. The method for monitoring rice harvesting loss rate based on vibration signal analysis according to claim 5, characterized in that: The collision event detection and secondary collision filtering in step S5 are as follows: S51, Peak Detection Algorithm; the specific process is as follows: The following peak detection procedure is performed on the pre-processed vibration signal (bandpass filtering, baseline correction): (1) Amplitude envelope calculation: The instantaneous amplitude envelope of the signal is extracted using Hilbert transform; in This is the preprocessed vibration signal. This represents the Hilbert transform, which highlights the energy concentration characteristics of collision events through envelope calculation; (2) Adaptive threshold setting: The detection threshold is set based on the statistical characteristics of the envelope signal; in The mean of the envelope signal is denoted as . The threshold is the standard deviation of the envelope signal, and it can dynamically adapt to the signal noise level under different operating conditions. (3) Peak filtering: The local maximum detection algorithm is used to identify peaks in the envelope signal that exceed the threshold. The potential peak points are identified, and a peak spacing constraint is applied. The time interval between any two peak points must be ≥0.005 seconds to avoid repeated detections caused by high-frequency signal oscillations. The final output is a set of time indexes of potential collision peaks. ,in This represents the timestamp of the i-th peak. S52, Secondary Collision Filtering; The specific process is as follows: To address the redundant peak values ​​of secondary collisions caused by factors such as rice grain bouncing, equipment resonance, and straw friction, a hierarchical filtering strategy is adopted to remove secondary collisions (redundant peak values ​​caused by rice grain bouncing, equipment resonance, etc.) while retaining the true primary collision events. (1) Time grouping: for peak sets Arranged chronologically, starting from the first peak, consecutive peaks with a time difference ≤ 0.2 seconds are grouped together to form several peak groups. ; (2) Amplitude ratio screening: For each group Calculate the maximum amplitude peak (main peak) within the group. in Peak value The corresponding signal amplitude is calculated using the maximum absolute value within a 30ms window. The amplitude ratio of the remaining peaks (secondary peaks) to the main peak is... ; (3) Retention rules within a group: If the time difference between the secondary peak and the primary peak is ≤0.2 seconds and r <0.6, it is judged as a secondary collision and is removed. Only the primary peak is retained in each group. This forms a set of peak values ​​after initial filtering.

1. Focus on removing strong redundant signals with significant energy decay; (4) Domain Relationship Modeling: This involves transforming the initially filtered peak data set... Mapped to a set of points in the time-amplitude feature space; ( Peak value The timestamp is calculated using the sampling point index and sampling interval. A sliding window grouping strategy is used to build neighborhood relationships, that is, the peak value is dynamically determined by the time difference to ensure that adjacent real collision events are not mistakenly classified into the same neighborhood. (5) Dynamic threshold filtering: For filters that satisfy a time difference of 0.011 ≤ ≤0.08 seconds ( The peak value of (p,q) is set using a piecewise dynamic amplitude ratio threshold: when 0.011 ≤ When the time is ≤0.0045 seconds, the threshold is 0.85, that is, if If q ≤ 0.85, then it is a second-order collision; if 0.045 ≤ When the time is ≤0.08 seconds, the threshold linearly decreases to 0.55, that is, if If q ≤ 0.55, then it is a second-order collision; (6) Final main collision determination: For the remaining peaks after dynamic threshold filtering, the principle of "amplitude priority, energy secondary selection" is adopted. In each neighborhood, the peak with the largest amplitude is selected first. If the amplitudes are close (difference <5%), the energy is determined again through a 30ms window, and the final set of main collision events is output to ensure that the retained peaks truly reflect the characteristics of rice grain collision.

7. The method for monitoring rice harvesting loss rate based on vibration signal analysis according to claim 6, characterized in that: The specific process of classifying and counting rice grains in step S6 is as follows: S61. Single-peak signal classification; the specific process is as follows: For each signal segment of a primary collision event, features are extracted using a pre-trained encoder and input into a downstream random forest classifier to predict whether the signal is "full" or "other". The classification confidence score is then output to quantify the reliability of the prediction. When the confidence score is ≥0.7, the classification result is used directly; otherwise, it is marked as "low confidence" and subsequently validated using energy features. S62. Prediction of the number of superimposed particles; the specific process is as follows: For signals classified as "full grains", a pre-trained quantity classification head is invoked, and the signal energy and peak amplitude are used for verification (energy is positively correlated with the number of grains); for signals classified as "other", a heuristic rule (energy / single grain energy threshold) is used to estimate the quantity. S62, Rice grain classification and counting model test; the specific process is as follows: (1) Use a test dataset that is independent of the training and validation sets to ensure the objectivity and reliability of the test results. The test dataset should cover collision vibration signals of different material combinations, such as plump rice grains and shriveled rice grains, plump rice grains and straw, etc. (2) Select appropriate model performance evaluation metrics to comprehensively evaluate the model's detection accuracy. Evaluation metrics for classification tasks include: precision, recall, and F1 score; evaluation metrics for counting tasks include: mean absolute error (MAE), root mean square error (RMSE), and counting accuracy. (3) Test the trained model using the test dataset and calculate various performance evaluation metrics. Based on the test results, further optimize the model or adjust the training strategy.

8. The method for monitoring rice harvesting loss rate based on vibration signal analysis according to claim 7, characterized in that: The calculation of rice harvesting loss rate in step S7 is as follows: The total number of rice grains predicted as "full" in all main collision events is accumulated, and the loss rate is calculated by combining the total number of lost rice grains and the total number of harvested rice grains.

Images