Corn kernel moisture on-line detection device
By combining a deep belief network-long short-term memory adaptive weighted fusion algorithm with a dual-channel symmetrical dielectric constant sensor, temperature sensor, and weighing module with an embedded AI processing module, the accuracy and anti-interference issues of corn kernel moisture detection in complex environments are solved, achieving high-precision real-time online detection.
Patent Information
- Application Number
- CN202610085024.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-22
- Publication Date
- 2026-02-24
AI Technical Summary
Existing corn kernel moisture detection technologies suffer from poor adaptability to environmental changes, weak anti-interference capabilities, and insufficient accuracy in online detection scenarios such as combine harvesters, especially with poor measurement repeatability under mechanical vibration and temperature fluctuations.
It employs a dual-channel symmetrical dielectric constant sensor module, a temperature sensing module, and a weighing module, combined with a high-frequency signal source and an embedded AI processing module. Through a deep belief network-long short-term memory adaptive weighted fusion algorithm, it achieves the fusion processing of multi-sensor data and real-time intelligent decision-making.
It achieves high-precision real-time detection of corn kernel moisture under complex working conditions, has adaptive capabilities, can suppress mechanical vibration and temperature interference, reduce measurement errors, and meet the needs of online real-time detection.
Smart Images
Figure CN121558828A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of agricultural testing technology, and in particular to an online corn kernel moisture detection device. Background Technology
[0002] Corn kernel moisture detection is a crucial step in grain harvesting, storage, processing, and trade, directly impacting corn quality assessment, pricing, and storage safety. Currently, existing corn moisture detection technologies suffer from the following limitations: while traditional drying methods are used as a benchmark, the detection time is several hours, failing to meet the demands of online real-time monitoring; electrical-based detection technologies, such as capacitance methods, while offering faster detection speeds, are highly susceptible to significant influences from temperature, grain density, and varietal differences in accuracy. Particularly in online detection scenarios such as those using combine harvesters, mechanical vibration, ambient temperature fluctuations, and changes in grain density can introduce significant errors, resulting in poor measurement repeatability.
[0003] In existing technologies, some improved solutions attempt to compensate by adding temperature sensors or employing dual-sensor designs to suppress common-mode interference. However, these methods mostly rely on linear compensation or simple lookup tables, which are insufficient for handling the complex nonlinear relationship between corn moisture and dielectric constant, as well as situations where multiple environmental factors are coupled influencing the process, resulting in poor adaptability. Furthermore, some technical solutions employ radio frequency impedance measurement or near-infrared spectroscopy. The former is ineffective for detecting grain moisture below the husk layer, while the latter involves expensive equipment and requires complex calibration, making it difficult to popularize in low-cost, industrialized online detection scenarios.
[0004] Therefore, there is an urgent need in this field for a device that can adapt to environmental changes, has strong anti-interference capabilities, and can detect the moisture content of corn kernels in real time with high precision. Summary of the Invention
[0005] To address the shortcomings of existing technologies in terms of environmental adaptability, anti-interference capability, and accuracy, this invention provides an online corn kernel moisture detection device.
[0006] The technical solution adopted in this invention is:
[0007] An online corn kernel moisture detection device includes:
[0008] A dual-channel symmetrical dielectric constant sensor module is used to synchronously acquire the dielectric constant signal of corn kernels;
[0009] Temperature sensing module for real-time monitoring of corn kernel temperature;
[0010] The weighing module is used to estimate the instantaneous mass of the corn kernels being tested;
[0011] A high-frequency signal source module is used to generate frequency-adjustable excitation signals;
[0012] An embedded AI processing module is used to run a deep belief network-long short-term memory adaptive weighted fusion algorithm to fuse data from the dual-path symmetric dielectric constant sensor module, the temperature sensing module, and the weighing module, and output the moisture content value.
[0013] Preferably, the dual-path symmetrical dielectric constant sensor module adopts a concentric circular planar capacitor design, including two sensing units with completely identical geometric dimensions and electrical parameters. The sensing unit consists of a driving electrode and a ground electrode, with the ground electrode arranged around the driving electrode.
[0014] Preferably, the high-frequency signal source module employs a direct digital frequency synthesizer, whose output frequency is... The adjustment is dynamically adjusted based on the dielectric properties of corn kernels, using the following formula: ,in, As the reference frequency, This is the frequency adjustment factor. This represents the deviation between the current dielectric constant and the calibrated value.
[0015] Preferably, the embedded AI processing module runs the Deep Belief Network-Long Short-Term Memory Adaptive Weighted Fusion Algorithm, which includes the following steps:
[0016] S10. Data preprocessing: Denoise the original dielectric constant signal using wavelet transform and perform nonlinear compensation on the temperature signal.
[0017] S11. Deep Belief Network Feature Extraction: High-order nonlinear features are extracted from preprocessed multi-source data using deep belief networks;
[0018] S12. Temporal Modeling of Long Short-Term Memory Networks: Using Long Short-Term Memory Networks to perform temporal modeling of feature sequences and capture dynamic change patterns;
[0019] S13, Adaptive Weighted Fusion: The weights of the dual-channel sensor data are dynamically calculated based on the real-time signal-to-noise ratio and confidence level, and the final moisture value is output by weighted averaging.
[0020] Preferably, in step S10, wavelet transform denoising specifically includes:
[0021] The dielectric constant signal was decomposed into three levels using the Daubechies 4 wavelet basis to obtain the approximation coefficient A3 and detail coefficients D1, D2, and D3.
[0022] An adaptive thresholding method is used to filter the detail coefficients D1 and D2. The threshold calculation formula is as follows: ,in, The noise standard deviation is estimated using the median of the D1 coefficient. The signal length;
[0023] The filtered coefficients are reconstructed to obtain the denoised dielectric constant signal.
[0024] Preferably, in step S10, the formula for nonlinear compensation of the temperature signal is as follows: ,in, The compensated dielectric constant, For measured values, The current temperature. For reference temperature, and The temperature coefficient of the dielectric constant of corn kernels, determined through experimental data.
[0025] Preferably, in step S11, the deep belief network comprises an input layer, three hidden layers, and an output layer, wherein:
[0026] The number of nodes in the input layer corresponds to the preprocessed feature dimensions, including dielectric constant, temperature, and mass.
[0027] The hidden layers are stacked using restricted Boltzmann machines, and the pre-training uses the contrastive divergence algorithm, while the fine-tuning uses the backpropagation algorithm.
[0028] The output layer is a 128-dimensional feature vector, and the feature dimensionality is reduced by a sparse autoencoder to retain the 20 core features that are most correlated with water content.
[0029] Preferably, in step S13, the formula for adaptive weighted fusion is: ,in, and The dynamic weighting coefficients satisfy the following conditions: The weighting coefficients are calculated in real time by a small neural network based on the signal-to-noise ratio of the two current data streams and the prediction confidence of the long short-term memory network output. and These are the preliminary moisture content predictions provided by two independent sensing units in the dual-path symmetrical dielectric constant sensor module, obtained after feature extraction via deep belief network and time-series modeling via long short-term memory network.
[0030] Preferably, it also includes a real-time model update module, which is configured as follows:
[0031] When the variance of multiple consecutive prediction results exceeds the threshold, incremental learning of the model is automatically triggered.
[0032] Incremental learning employs an online sequence extreme learning machine, updating only the output layer weights to avoid global retraining.
[0033] The updated model parameters are synchronized to the cloud database via an encrypted channel for other devices of the same model to download and use.
[0034] The beneficial effects of this invention are:
[0035] By simultaneously sensing physical information through multiple sensors (dielectric, temperature, and weight), rich and multidimensional input data is provided for AI algorithms. The embedded AI processing module pushes complex cloud computing capabilities down to edge devices, enabling real-time intelligent decision-making. Compared with traditional single-sensor and fixed-formula calculations, data fusion and intelligent algorithms achieve adaptability to complex working conditions (such as temperature changes, vibration interference, and product differences), upgrading detection from "single-point measurement" to "systematic perception and decision-making," thereby fundamentally solving the persistent problems of low accuracy and poor anti-interference ability of traditional methods. Attached Figure Description
[0036] Figure 1 This is a schematic diagram of the module structure of the online corn kernel moisture detection device in an embodiment of the present invention;
[0037] Figure 2 This is a flowchart illustrating the Deep Belief Network-Long Short-Term Memory Adaptive Weighted Fusion Algorithm in an embodiment of the present invention.
[0038] Reference numerals: 1. Dual-channel symmetrical dielectric constant sensor module; 2. Temperature sensing module; 3. Weighing module; 4. High-frequency signal source module; 5. Embedded AI processing module; 6. Real-time model update module. Detailed Implementation
[0039] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.
[0040] Example
[0041] like Figure 1 As shown, an online corn kernel moisture detection device includes:
[0042] A dual-channel symmetrical dielectric constant sensor module is used to synchronously acquire the dielectric constant signal of corn kernels;
[0043] Temperature sensing module for real-time monitoring of corn kernel temperature;
[0044] The weighing module is used to estimate the instantaneous mass of the corn kernels being tested;
[0045] A high-frequency signal source module is used to generate frequency-adjustable excitation signals;
[0046] An embedded AI processing module is used to run a deep belief network-long short-term memory adaptive weighted fusion algorithm to fuse data from the dual-path symmetric dielectric constant sensor module, the temperature sensing module, and the weighing module, and output the moisture content value.
[0047] The device runs a deep belief network-long short-term memory adaptive weighted fusion algorithm through an embedded AI processing module, which can accurately characterize the complex nonlinear relationship between moisture and multiple parameters, suppress random errors, and reduce mean absolute error and repeated measurement standard deviation.
[0048] The combination of dual-sensor design and adaptive weighted fusion algorithm enables the device to be fault-tolerant to single-point failures or transient disturbances, making it particularly suitable for the environment of combine harvesters with severe vibrations.
[0049] Real-time online detection: The entire processing flow is optimized and runs on the embedded AI module, with a single detection cycle of less than 2 seconds, which fully meets the requirements of online real-time detection.
[0050] In one embodiment, in order to provide a cleaner and more reliable raw data source to the embedded AI module and achieve high-precision detection, the dual-path symmetrical dielectric constant sensor module adopts a concentric circular planar capacitor design, including two sensing units with completely identical geometric dimensions and electrical parameters. The sensing unit consists of a driving electrode and a ground electrode, with the ground electrode arranged around the driving electrode.
[0051] The concentric planar capacitor design generates a uniform electric field distribution, and the dual-path symmetrical design provides a physical basis for subsequent differential measurements and data fusion. Its working principle utilizes the significant difference in dielectric properties between corn kernels (dielectric constant approximately 3–80) and air (dielectric constant approximately 1). When a kernel passes through the sensor surface, it causes a significant change in capacitance. Two symmetrical units measure simultaneously; one serves as the measurement unit, while the other can be used as a reference or for differential calculations. This design firstly reduces edge effects and improves measurement stability; secondly, the dual-path design allows for real-time self-calibration and differential signal output, effectively suppressing common-mode interference (such as ambient temperature drift and power supply fluctuations), thereby reducing the sensor's inherent system error.
[0052] The dielectric properties of corn exhibit a non-linear relationship with frequency and moisture content (dispersion effect). Sensitivity decreases at both ends of the moisture range when excited at a fixed frequency. To ensure the sensor always operates at the frequency point where the dielectric response is most sensitive, thereby improving detection sensitivity and adaptability to different corn varieties, in one embodiment, the high-frequency signal source module employs a direct digital frequency synthesizer, whose output frequency... The adjustment is dynamically adjusted based on the dielectric properties of corn kernels, using the following formula: ,in, As the reference frequency, This is the frequency adjustment factor. This represents the deviation between the current dielectric constant and the calibrated value.
[0053] Direct digital frequency synthesizers based on dielectric constant deviation Dynamically adjust excitation frequency When the current dielectric value deviates from the calibration reference (such as the dielectric constant of dried corn at 25°C), the system adjusts the dielectric constant according to the proportional coefficient. This ensures the sensor always operates at its maximum dielectric sensitivity frequency. For example, the dielectric relaxation frequency of high-moisture corn (moisture content >20%) is between 80 and 120 MHz, while that of low-moisture corn (moisture content <14%) shifts to above 200 MHz. Dynamic tracking avoids sensitivity degradation at a fixed frequency. Furthermore, changes in ambient temperature cause frequency shifts in the dielectric curve, requiring complex temperature compensation for fixed-frequency solutions. This solution indirectly compensates for temperature effects through frequency adaptation (because...). (Already includes dielectric changes caused by temperature), simplifying the algorithm structure.
[0054] To create a detection system capable of self-optimization and self-correction, thereby improving detection accuracy and robustness, in one embodiment, such as Figure 2 As shown, the embedded AI processing module runs the Deep Belief Network-Long Short-Term Memory Adaptive Weighted Fusion Algorithm, which includes the following steps:
[0055] S10. Data preprocessing: Denoise the original dielectric constant signal using wavelet transform and perform nonlinear compensation on the temperature signal.
[0056] S11. Deep Belief Network Feature Extraction: High-order nonlinear features are extracted from preprocessed multi-source data using deep belief networks;
[0057] S12. Temporal Modeling of Long Short-Term Memory Networks: Using Long Short-Term Memory Networks to perform temporal modeling of feature sequences and capture dynamic change patterns;
[0058] S13, Adaptive Weighted Fusion: The weights of the dual-channel sensor data are dynamically calculated based on the real-time signal-to-noise ratio and confidence level, and the final moisture value is output by weighted averaging.
[0059] The above steps provide a hierarchical data processing pipeline. The preprocessing stage is data cleaning, and wavelet transform can accurately separate high-frequency components such as vibration noise. Nonlinear temperature compensation corrects a major source of error at the physical level. The deep belief network feature extraction stage automatically learns the complex and difficult-to-describe mapping relationship between dielectric constant, temperature, mass and moisture, replacing the traditional rudimentary linear formula or lookup table method. The long short-term memory network temporal modeling stage remembers the correlation between data points before and after, which can smooth the instantaneous fluctuations caused by uneven grain distribution and make the output more stable. The adaptive weighted fusion stage increases the system's ability to automatically reduce weights when one sensor data is abnormal, realizing active fault tolerance at both the hardware and software levels.
[0060] To achieve accurate quantization for denoising the dielectric constant signal, in one embodiment, wavelet transform denoising specifically includes the following in step S10:
[0061] The dielectric constant signal was decomposed into three levels using the Daubechies 4 wavelet basis to obtain the approximation coefficient A3 and detail coefficients D1, D2, and D3.
[0062] An adaptive thresholding method is used to filter the detail coefficients D1 and D2. The threshold calculation formula is as follows: ,in, The noise standard deviation is estimated by the median of the D1 coefficient. ), The signal length;
[0063] The filtered coefficients are reconstructed to obtain the denoised dielectric constant signal.
[0064] Among them, detail coefficients D1 and D2 mainly contain high-frequency noise caused by mechanical vibration, which is filtered by an adaptive thresholding method. The threshold is calculated using the noise standard deviation. and signal length The relationship ensures that the threshold can be automatically adjusted according to the signal characteristics. During reconstruction, only the filtered D1 and D2 and the unprocessed A3 and D3 are retained. While removing noise, the original abrupt change characteristics of the dielectric signal (such as the frequency domain change corresponding to the moisture critical point) are preserved.
[0065] Traditional filtering methods (such as mean filtering) smooth out the effective signal, leading to the loss of moisture characteristics. Wavelet transform's multi-resolution analysis capabilities can concentrate noise in high-frequency detail coefficients, achieving selective filtering through thresholding. For example, the dominant frequency of vibration noise in a combine harvester is 200–500 Hz, while the effective component of the dielectric signal is in the 0–50 Hz range. The three-level decomposition of the Daubechies 4 wavelet precisely preserves the signal below 50 Hz in the A3 coefficients, achieving noise stripping.
[0066] Fixed thresholds can easily cause filtering or under-filtering when flow rates change. The above steps estimate the noise level by using the median of the D1 coefficient. When the seed flow density changes abruptly, causing the signal amplitude to fluctuate, the threshold is automatically adjusted to avoid the blindness of manually setting the threshold. This improves the signal-to-noise ratio and reduces signal reconstruction error in vibration environments.
[0067] To reduce the difficulty of deep belief networks learning temperature-independent features and improve the model's convergence speed and generalization ability, especially given the significant temperature differences between northern and southern my country, and to ensure consistent detection accuracy, in one embodiment, the formula for nonlinear compensation of the temperature signal in step S10 is as follows: ,in, The compensated dielectric constant, For measured values, The current temperature. For reference temperature, and The temperature coefficient of the dielectric constant of corn kernels, determined through experimental data.
[0068] A second-order polynomial model is used to fit the nonlinear relationship between the dielectric constant and temperature, because the effect of temperature on moisture polarization and dielectric constant is inherently nonlinear, and first-order linear compensation would introduce residual errors; the formula introduces a quadratic term... This allows for a more accurate description of the different rates of change in the high and low temperature ranges. The refined compensation significantly eliminates a major interfering factor during the data preprocessing stage, essentially providing a "temperature-constant" or "temperature-normalized" feature input for subsequent AI models.
[0069] To avoid overfitting in the deep belief network, in one embodiment, in step S11, the deep belief network includes an input layer, three hidden layers, and an output layer, wherein:
[0070] The number of nodes in the input layer corresponds to the preprocessed feature dimensions, including dielectric constant, temperature, and mass.
[0071] The hidden layers are stacked using restricted Boltzmann machines, and the pre-training uses the contrastive divergence algorithm, while the fine-tuning uses the backpropagation algorithm.
[0072] The output layer is a 128-dimensional feature vector, and the feature dimensionality is reduced by a sparse autoencoder to retain the 20 core features that are most correlated with water content.
[0073] The deep belief network employs a stack of three layers of restricted Boltzmann machines. Each layer undergoes unsupervised pre-training using a contrastive divergence algorithm to learn high-order statistical features from multiple sources, including dielectric constant, temperature, and mass. After pre-training, a sparse autoencoder is used to reduce the dimensionality of the output 128-dimensional feature vector, selecting the 20 core features most strongly correlated with moisture content. Sparsity constraints are implemented by adding a KL divergence penalty term, forcing the network to learn sparse representations and preventing overfitting.
[0074] Traditional linear dimensionality reduction methods such as PCA cannot capture the nonlinear relationships between multi-sensor data (such as the coupling effect between dielectric constant and temperature). Deep belief networks map the original signal to a highly decoupled feature space through multi-layer nonlinear transformation. For example, the second-layer restricted Boltzmann machine may learn the dielectric characteristics of "temperature invariance", and the third layer further integrates mass information to generate "density normalized" features.
[0075] The 128-dimensional features contain a large amount of redundancy (such as the slowly varying signals from temperature sensors). The sparse autoencoder sorts the features by relevance weights, retaining the 20 features with the highest mutual information with the moisture label (such as the responses of the real and imaginary dielectric parts at specific frequencies), and eliminating irrelevant dimensions. This allows the model to be transferred to different varieties of maize with only the output layer fine-tuned, without retraining the feature extraction network, thus reducing the workload of calibration.
[0076] The contribution of 20 core features can be analyzed through SHAP values, which helps to understand key detection parameters (such as the strong correlation between "dielectric loss angle at 10MHz" and moisture content), and is superior to black-box models.
[0077] To improve the system's reliability and stability during continuous operation, in one embodiment, the adaptive weighted fusion formula in step S13 is as follows: ,in, and The dynamic weighting coefficients satisfy the following conditions: The weighting coefficients are calculated in real time by a small neural network based on the signal-to-noise ratio of the two current data streams and the prediction confidence of the long short-term memory network output. and These are the preliminary moisture content predictions provided by two independent sensing units in the dual-path symmetrical dielectric constant sensor module, obtained after feature extraction via deep belief network and time-series modeling via long short-term memory network.
[0078] A feedback control loop based on real-time performance evaluation is introduced into the adaptive weighted fusion. A small neural network acts as a "weight calculator," continuously evaluating the quality (signal-to-noise ratio) of the two sensor data streams and the model's confidence level in predicting the current data. If a sensor is briefly covered by a foreign object, causing a sharp drop in the signal-to-noise ratio, or if its prediction deviates significantly from the other sensor, resulting in low confidence, the system automatically reduces the weight of that sensor. These steps endow the detection system with strong fault tolerance and adaptability. When faced with unavoidable transient anomalies in the real world (such as seed blockage or water splashing), the device does not simply produce incorrect results but can "intelligently" rely on higher-quality data channels. This significantly improves the system's reliability and the stability of continuous operation, which is crucial for industrial-grade online detection equipment.
[0079] To overcome the performance degradation of the model (i.e., data distribution drift problem) caused by gradual changes in environmental temperature and humidity and switching of corn varieties, and to maintain long-term detection accuracy, in one embodiment, the device further includes a real-time model update module, which is configured to:
[0080] When the variance of multiple consecutive prediction results exceeds the threshold, incremental learning of the model is automatically triggered.
[0081] Incremental learning employs an online sequence extreme learning machine, updating only the output layer weights to avoid global retraining.
[0082] The updated model parameters are synchronized to the cloud database via an encrypted channel for other devices of the same model to download and use.
[0083] By introducing an online sequential extreme learning machine (SEM) incremental learning mechanism, the system automatically triggers a model update when it detects that the prediction variance exceeds a threshold (e.g., 0.1%) for 10 consecutive times. Mathematically, this mechanism updates only the output layer weights using recursive least squares, enabling the model to continuously evolve. Compared to traditional periodic retraining, the SEM incremental learning mechanism only requires storing the current weight matrix, resulting in low memory consumption and suitability for edge devices. For example, during the rainy season when air humidity rises continuously, the model automatically adjusts the weights through incremental learning to avoid accuracy drift. Simultaneously, the locally updated gradients can be encrypted and uploaded to the cloud platform, aggregating global knowledge through federated learning to form a "swarm intelligence" evolutionary ecosystem, allowing devices deployed in different climate zones to share optimization results.
[0084] The embodiments described above are merely illustrative of specific implementations of the present invention, and while the descriptions are detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.
Claims
1. An online moisture detection device for corn kernels, characterized in that, include: A dual-channel symmetrical dielectric constant sensor module is used to synchronously acquire the dielectric constant signal of corn kernels; Temperature sensing module for real-time monitoring of corn kernel temperature; The weighing module is used to estimate the instantaneous mass of the corn kernels being tested; A high-frequency signal source module is used to generate frequency-adjustable excitation signals; An embedded AI processing module is used to run a deep belief network-long short-term memory adaptive weighted fusion algorithm to fuse data from the dual-path symmetric dielectric constant sensor module, the temperature sensing module, and the weighing module, and output the moisture content value.
2. The online corn kernel moisture detection device according to claim 1, characterized in that, The dual-path symmetrical dielectric constant sensor module adopts a concentric circular planar capacitor design and includes two sensing units with identical geometric dimensions and electrical parameters. The sensing unit consists of a driving electrode and a ground electrode, with the ground electrode arranged around the driving electrode.
3. The online corn kernel moisture detection device according to claim 2, characterized in that, The high-frequency signal source module employs a direct digital frequency synthesizer, whose output frequency... The adjustment is dynamically adjusted based on the dielectric properties of corn kernels, using the following formula: ,in, As the reference frequency, This is the frequency adjustment factor. This represents the deviation between the current dielectric constant and the calibrated value.
4. The online corn kernel moisture detection device according to claim 3, characterized in that, The embedded AI processing module runs the Deep Belief Network-Long Short-Term Memory Adaptive Weighted Fusion Algorithm, which includes the following steps: S10. Data preprocessing: Denoise the original dielectric constant signal using wavelet transform and perform nonlinear compensation on the temperature signal. S11. Deep Belief Network Feature Extraction: High-order nonlinear features are extracted from preprocessed multi-source data using deep belief networks; S12. Temporal Modeling of Long Short-Term Memory Networks: Using Long Short-Term Memory Networks to perform temporal modeling of feature sequences and capture dynamic change patterns; S13, Adaptive Weighted Fusion: The weights of the dual-channel sensor data are dynamically calculated based on the real-time signal-to-noise ratio and confidence level, and the final moisture value is output by weighted averaging.
5. The online corn kernel moisture detection device according to claim 4, characterized in that, In step S10, wavelet transform denoising specifically includes: The dielectric constant signal was decomposed into three levels using the Daubechies 4 wavelet basis to obtain the approximation coefficient A3 and detail coefficients D1, D2, and D3. An adaptive thresholding method is used to filter the detail coefficients D1 and D2. The threshold calculation formula is as follows: ,in, The noise standard deviation is estimated using the median of the D1 coefficient. The signal length; The filtered coefficients are reconstructed to obtain the denoised dielectric constant signal.
6. The online corn kernel moisture detection device according to claim 5, characterized in that, In step S10, the formula for nonlinear compensation of the temperature signal is as follows: ,in, The compensated dielectric constant, For measured values, The current temperature. For reference temperature, and The temperature coefficient of the dielectric constant of corn kernels, determined through experimental data.
7. The online corn kernel moisture detection device according to claim 6, characterized in that, In step S11, the deep belief network comprises an input layer, three hidden layers, and an output layer, wherein: The number of nodes in the input layer corresponds to the preprocessed feature dimensions, including dielectric constant, temperature, and mass. The hidden layers are stacked using restricted Boltzmann machines, and the pre-training uses the contrastive divergence algorithm, while the fine-tuning uses the backpropagation algorithm. The output layer is a 128-dimensional feature vector, and the feature dimensionality is reduced by a sparse autoencoder to retain the 20 core features that are most correlated with water content.
8. The online corn kernel moisture detection device according to claim 7, characterized in that, In step S13, the formula for adaptive weighted fusion is: ,in, and The dynamic weighting coefficients satisfy the following conditions: The weighting coefficients are calculated in real time by a small neural network based on the signal-to-noise ratio of the two current data streams and the prediction confidence of the long short-term memory network output. and These are the preliminary moisture content predictions provided by two independent sensing units in the dual-path symmetrical dielectric constant sensor module, obtained after feature extraction via deep belief network and time-series modeling via long short-term memory network.
9. The online corn kernel moisture detection device according to claim 8, characterized in that, It also includes a real-time model update module, which is configured as follows: When the variance of multiple consecutive prediction results exceeds the threshold, incremental learning of the model is automatically triggered. Incremental learning employs an online sequence extreme learning machine, updating only the output layer weights to avoid global retraining. The updated model parameters are synchronized to the cloud database via an encrypted channel for other devices of the same model to download and use.