Search

CN-122024887-A - Durian sugar degree intelligent detection sorting equipment based on spectrum and algorithm

CN122024887ACN 122024887 ACN122024887 ACN 122024887ACN-122024887-A

Abstract

The invention discloses intelligent detection and separation equipment for the sugar content of durian based on a spectrum and an algorithm, belongs to the technical field of nondestructive detection of fruit quality, and mainly solves the technical problem that the sugar content in durian with a shell cannot be accurately detected in real time. The method comprises the steps of continuously conveying a durian sample with a shell by adopting a variable-frequency speed-regulating conveying system, collecting multidimensional spectral information penetrating through the surface and the interior of a durian shell by a hyperspectral camera with a spectral range of 900nm to 1700nm and a uniform diffusion light source system, utilizing a computing platform formed by an AI acceleration chip with FP8 data format support and a peak computing power of not less than 200TFLOPS to run a pre-trained deep neural network model in real time for predicting a sugar degree value, and splitting the sample according to a sugar degree prediction result by combining an automatic sorting mechanism, wherein hyperspectral collection frequency is synchronously matched with conveying speed, and the deep neural network model is trained by adopting a small sample modeling technology. The device can be used for fresh meat sorting in durian processing factories, quality control of high-end fruit supply chains and variety analysis of scientific research institutions.

Inventors

  • CHEN ZHUOYI
  • WU JUNYI

Assignees

  • 蓝曜智鑫科技(广西)有限公司

Dates

Publication Date
20260512
Application Date
20250825

Claims (10)

  1. 1. The utility model provides a durian sugar degree intellectual detection system sorting facilities based on spectrum and algorithm which characterized in that includes: the variable-frequency speed-regulating conveying system is used for continuously conveying the durian or fresh durian meat samples with shells; the hyperspectral imaging module comprises a hyperspectral camera with a spectral range of 900-1700 nm and a uniform diffusion light source system, wherein the uniform diffusion light source system provides illumination at a detection position, so that the hyperspectral camera can acquire multidimensional spectral information on the surface and the inside of a sample penetrating through the durian shell; The computing platform is composed of an AI acceleration chip which is supported by an FP8 data format and has a peak computing power not lower than 200TFLOPS, receives spectrum data transmitted by a hyperspectral imaging module in real time, performs feature extraction and sugar degree value prediction calculation on the spectrum data through a pre-trained deep neural network model, and outputs a sugar degree prediction result; The automatic sorting mechanism is used for shunting the durian samples to corresponding quality grade channels according to the sugar degree prediction result output by the computing platform and a preset grading threshold value; The collection frequency of the hyperspectral imaging module is synchronously matched with the conveying speed of the variable-frequency speed-regulating conveying system, so that complete spectrum data of not less than 4000 samples are collected in each hour; The deep neural network model is trained by adopting a small sample modeling technology, and a prediction model is built by utilizing a nonlinear mapping relation between sample spectral characteristics and a sugar degree value; The uniform diffusion light source system realizes the uniformity of illumination of a detection surface through the annularly distributed light source assemblies and the scattering plates, and the illumination unevenness is controlled within 5%.
  2. 2. The intelligent detection and sorting equipment for the durian sugar degree based on spectrum and algorithm as claimed in claim 1, wherein: The deep neural network model performs the following steps when processing spectral data: preprocessing the collected spectrum data from 900nm to 1700nm to eliminate noise interference caused by illumination unevenness; Extracting characteristic wave band combinations which are strongly related to the sugar degree of durian in the pretreated spectrum data; Inputting the characteristic wave band combination into a fully-connected network layer trained based on a small sample modeling technology, and calculating a sugar degree predicted value through a nonlinear mapping relation; Wherein the characteristic band combination comprises an internal tissue absorption characteristic band penetrating the durian shell and a surface reflection characteristic band.
  3. 3. The intelligent detection and sorting equipment for the durian sugar degree based on spectrum and algorithm as claimed in claim 2, wherein: The preset grading threshold is a dynamic adjustment threshold, and the grading interval is automatically optimized according to the real-time sugar degree prediction result distribution; The decision logic for the automated sorting mechanism to perform the diversion includes the steps of: performing interval matching on the sugar degree predicted value and the dynamic grading threshold value; Triggering a grading threshold self-adaptive calibration mechanism when the sugar degree predicted value exceeds the same grade threshold range for 3 times continuously; and updating the sample distribution instruction according to the calibrated grading threshold value, and ensuring the corresponding relation between the sorting action and the quality grade channel.
  4. 4. The intelligent detection and sorting equipment for the durian sugar degree based on spectrum and algorithm as claimed in claim 1, wherein: the training process of the small sample modeling technique comprises the following steps: Carrying out data enhancement on the original sample set through spectral feature transformation to generate an enhanced data set with training sample size not less than 5 times of the original sample size; Initializing the weight of a deep neural network model by adopting a transfer learning mode, wherein a pre-training model is obtained by training a general data set based on fruit sugar detection; freezing shallow network weights when performing end-to-end training on the enhanced data set, and only optimizing full-connection layer weight parameters; After training, the root mean square error of the set sugar degree predicted value is verified to be controlled within 0.5 Brix.
  5. 5. The intelligent detection and sorting equipment for the durian sugar degree based on spectrum and algorithm as claimed in claim 1, wherein: The synchronous matching of the variable-frequency speed-regulating conveying system and the hyperspectral imaging module is realized through a closed-loop control strategy, and the variable-frequency speed-regulating conveying system specifically comprises: Monitoring the actual running speed of the conveying system in real time, and calculating the theoretical conveying speed according to 4000 samples/hour of the target processing amount; dynamically adjusting the conveying speed based on the exposure time of the hyperspectral camera to ensure that the residence time of the sample at the detection position covers the complete exposure period; when the sample interval is smaller than a preset safety distance threshold, automatically reducing the conveying speed and connecting with the long hyperspectral camera integration time; Wherein the dwell time of the detection bit is not less than 0.5 seconds and the sample displacement in the exposure period is not more than 0.1 mm.
  6. 6. The intelligent detection and sorting equipment for the durian sugar degree based on spectrum and algorithm as claimed in claim 1, wherein: The data storage and management system performs the following integrated operations of the internet of things: Packaging the sugar degree predicted value and the corresponding quality grade of each sample into a structured data unit; Synchronously writing a time stamp and a sorting channel number when a sample passes through an automatic sorting mechanism; Transmitting the data unit to a blockchain certification platform in real time through an MQTT protocol, and generating a data block according to every 10-100 samples; Each data block comprises a sample spectral feature hash value, a sugar degree predicted value, a sorting action time stamp and equipment operation environment parameters.
  7. 7. The intelligent detection and sorting equipment for the durian sugar degree based on spectrum and algorithm as claimed in claim 6, wherein: The user interaction interface realizes the following real-time decision support functions: Dynamically displaying a current sample sugar degree predicted value and a sugar degree distribution probability density map of 10-100 historical samples; generating a sample tracing path three-dimensional point cloud reconstruction model based on the transmitted blockchain certification data; When the sugar degree predicted value exceeds the dynamic grading threshold range set by the weight 3 for 3 times continuously, automatically popping up a threshold calibration suggestion window; And when the grading threshold adjustment is performed in response to an operator instruction, the shunt control parameters of the automatic sorting mechanism are updated in a linkage manner and are effective within 1 second.
  8. 8. The intelligent detection and sorting equipment for the durian sugar degree based on spectrum and algorithm as claimed in claim 3, wherein: The AI analysis and decision system integrates an environmental parameter compensation mechanism, and specifically performs: collecting and detecting the temperature and the relative humidity of the position environment in real time, wherein the temperature monitoring range is 15-40 ℃ and the humidity monitoring range is 60-90%; Establishing a temperature-sugar degree predicted value compensation mapping matrix, and compensating sugar degree predicted values by 0.1-0.3Brix per 1 ℃ when the temperature exceeds 25 ℃; constructing a humidity-characteristic wave band weight adjustment model, and reducing the wave band weight coefficient of 900-1200nm to 0.6-0.8 times when the relative humidity exceeds 80%; and inputting the compensated sugar degree predicted value into a grading decision flow, and updating a shunting instruction of the automatic sorting mechanism in a linkage way.
  9. 9. The intelligent detection and sorting equipment for the durian sugar degree based on spectrum and algorithm as claimed in claim 2, wherein: The integrated control system performs a periodic self-calibration operation comprising the steps of: automatically switching to a calibration mode after processing 100-500 samples each; placing a standard sugar degree calibration block at a detection position, and collecting calibration spectrum data through a hyperspectral imaging module; Comparing the average deviation value of the measured spectrum and the pre-stored reference spectrum in the characteristic wave band; Triggering a light source intensity compensation mechanism when the deviation of the 900-1000nm wave band exceeds 5% or the deviation of the 1400-1500nm wave band exceeds 8%; and after the calibration is finished, automatically updating the input data normalization parameters of the deep neural network model.
  10. 10. The intelligent detection and sorting equipment for the durian sugar degree based on spectrum and algorithm as claimed in claim 8, wherein: the AI analysis and decision system integrates a variety self-adaptation module, and performs the following operations: Establishing a durian variety characteristic spectrum library which comprises characteristic absorption peak positions of Jin Zhen, cat mountain king and Sudan king 3-5 main cultivars in a 900-1700nm wave band; Identifying the surface texture features of the sample in real time through a convolutional neural network, and calculating the confidence coefficient of the variety by combining the extracted feature wave band combination; when the confidence coefficient of the variety is more than or equal to 0.7, the sugar degree-spectrum mapping relation of the corresponding variety is called to correct the sugar degree predicted value compensation quantity; and the linkage grading decision module dynamically adjusts the grading threshold interval + -0.3-0.8 Brix according to the variety characteristics.

Description

Durian sugar degree intelligent detection sorting equipment based on spectrum and algorithm Technical Field The invention belongs to the technical field of nondestructive detection of fruit quality, and particularly relates to intelligent detection and separation equipment for durian sugar degree based on a spectrum and an algorithm. Background The nondestructive detection of the internal sugar degree of durian faces a technical bottleneck for a long time. Traditional destructive sampling methods require puncturing the pulp to obtain juice, which results in the inability of the sample to enter subsequent processing steps. Near infrared spectrum technology is applied to partial fruit detection, but the physiological structure of durian causes significant interference. The thickness difference of the shell reaches 5 to 15 millimeters, sharp protrusions are distributed on the surface, multiple scattering of an incident spectrum signal is caused, and the effective penetration rate is less than 30%. The heterogeneous medium formed by the interleaving of the internal pulp and the air chamber distorts the characteristic absorption peak, and particularly, the absorption band of water molecules is obvious in the range of 1200 to 1400 nanometers. The existing detection equipment mostly adopts a spectrum range of 600 to 1100 nanometers, and is difficult to cover a characteristic wave band related to the sugar degree of durian, such as a C-H bond frequency tripled absorption peak in a range of 1300 to 1600 nanometers. Together, these factors result in a prediction error range of plus or minus 2.0 Brix for conventional equipment, and cannot meet the plus or minus 0.8 Brix accuracy requirements required for commercial classification. The industrialized sorting scene requires more than 3000 samples per hour, and the prior art has double efficiency limitations. The floating point computing capability of the traditional graphics processor at the hardware level is about 60 trillion times per second, and the hyperspectral data cube of about 1GB of a single sample is difficult to process in real time, so that the single sample detection delay exceeds 1 second. The algorithm level depends on large-scale sample modeling, more than 5000 groups of calibration data are usually needed, the durian variety is high in diversity and centralized in season production, and sufficient training samples are difficult to obtain in practice. The prior art improvement attempts encounter inherent contradiction that the detection precision can be improved by expanding the spectrum range to 1700 nanometers, the processing speed is further reduced by the rapid increase of the data quantity, and the prediction deviation caused by variety difference can be amplified by simplifying the calculation model while the processing can be accelerated. The technical conflict causes that the automatic sorting equipment for durian with shell can not realize industrialized application for a long time. Disclosure of Invention The invention aims to solve the problems that the sugar content in durian with a shell cannot be detected in real time and the industrialized separation efficiency is low. The traditional method has the advantages that the spectral transmittance is lower than 30% due to the uneven thickness (5-15 mm) of the durian shell and the spike structure, the characteristic absorption peak is distorted due to the internal heterogeneous structure, the detection error reaches +/-2.0 Brix, meanwhile, the traditional graphic processor computing power (about 60 TFLOPS) cannot process hyperspectral data (1 GB of single sample) in real time, more than 5000 groups of samples are needed for modeling, and the sorting requirement of more than 3000 samples per hour is difficult to meet. The problem of inaccurate extraction of sugar degree related features in hyperspectral data is solved. The conventional method is difficult to separate effective sugar degree signals due to illumination residual noise and surface/internal characteristic coupling, and especially has insufficient identification capability on internal tissue information penetrating through the shell. The problem of sorting misalignment caused by sorting threshold drift in long-term operation is solved. The fixed threshold cannot adapt to environmental fluctuation and sample difference, manual frequent intervention and calibration are needed, and the continuous sorting flow is interrupted. The problem of poor model generalization under the condition of a small sample is solved. Durian varieties are various, calibration data are scarce, the traditional deep learning model needs more than 5000 groups of samples, and sufficient training data are difficult to obtain in actual season. The method solves the problem of cooperative control of high-speed conveying and hyperspectral imaging. Insufficient residence time or excessive displacement of the sample at the detection position causes imaging blurring, affec