Facility rotten vegetable processing method based on AI intelligent regulation and control

Abstract

The invention discloses an AI intelligent regulation and control facility rotten vegetable processing method, and aims to solve the problems that existing rotten vegetable processing equipment is easy to block due to fixed parameters and rotten degree judgment is disjointed with plant nutrition standards, the method comprises the following steps: acquiring rotten vegetable material multi-view images through a visual acquisition device, calculating a fiber winding index in combination with a physical correction action, and identifying a category label; the crushing particle size and the squeezing rotating speed are regulated and controlled according to category labels and winding index differences, and loss of water-soluble nitrogen and potassium nutrients is reduced while fiber winding is prevented; conveying the dehydrated rotten vegetable leaves to an aerobic fermentation device, collecting a compost surface image, inputting the compost surface image into a decomposition degree identification model associated with a plant nutrition index, outputting a decomposition degree probability value, triggering verification sampling when the decomposition degree probability value is greater than 0.95 for three continuous times, and judging a fermentation end point and starting discharging if a seed germination index is greater than or equal to 95%. Pretreatment anti-blocking, fermentation end point judgment intelligence and plant nutrition adaptability are achieved, and the nutrient retention rate and quality stability of the waste vegetable organic fertilizer are improved.

Description

Technical Field

[0001] This invention belongs to the field of organic solid waste resource utilization technology, specifically relating to an AI-intelligent controlled method for processing waste vegetables from facilities. Background Technology

[0002] In the field of resource utilization of agricultural waste, existing technologies typically include two stages: dehydration and aerobic fermentation. The dehydration stage generally employs a combination of crushing and mechanical pressing to reduce the moisture content of the waste, while the fermentation stage relies on natural decomposition or manual turning to promote biotransformation. However, in actual production, existing dehydration equipment often uses fixed process parameters, making it difficult to adapt to the significant differences in fiber content and mechanical entanglement characteristics among different types of waste. When processing vine-like or high-fiber fruit and vegetable waste, long fibrous materials easily become entangled on the crushing rollers or the main shaft of the screw press, causing equipment overload and shutdown, necessitating manual cleaning. This reliability issue caused by fluctuations in material characteristics has long remained unresolved. Furthermore, existing dehydration process designs primarily focus on dehydration efficiency, lacking specific consideration for the retention of water-soluble nitrogen, potassium, and other plant nutrients in the waste. Unnecessary leaching and loss of nutrients often occurs during the pressing and dehydration process, affecting the fertilizer efficacy of subsequent fermentation products.

[0003] Determining the maturity of fermentation also faces technical bottlenecks. Currently, production sites generally rely on manual experience or timed sampling and testing to determine the timing of discharge. The former is highly subjective and its accuracy is difficult to guarantee, while the latter requires frequent manual sampling and laboratory testing, which is cumbersome, time-consuming, and cannot achieve real-time monitoring of the fermentation process. Discharging too early may lead to incomplete fermentation and incomplete degradation of phytotoxic substances in the organic material; discharging too late will prolong the processing cycle, reduce equipment utilization, and excessive fermentation may lead to nutrient loss. Existing methods for determining maturity have not yet established an effective correlation with core plant nutrition indicators. Key quality parameters such as the seed germination index of fermentation products can often only be obtained through laboratory testing after discharge, making it impossible to form an effective closed-loop control during production and affecting the quality stability of the vegetable waste organic fertilizer product. Summary of the Invention

[0004] One object of the present invention is to solve at least the above-mentioned problems and to provide at least the advantages that will be described later.

[0005] Another objective of this invention is to provide an AI-intelligent controlled method for processing waste vegetables in facilities. This method can reduce the loss of water-soluble nitrogen and potassium nutrients in waste vegetables while preventing fiber entanglement and blockage through material characteristic perception and adaptive control of process parameters. Furthermore, it can automatically determine the fermentation endpoint based on a visual model associated with plant nutrient indicators, thereby improving the quality stability of waste vegetable organic fertilizer products and the automation level of the processing process.

[0006] This invention discloses an AI-controlled method for processing waste vegetables from agricultural facilities, comprising the following steps: 1) Intelligent perception of material characteristics: Multi-view images of the waste vegetable material are acquired through a visual acquisition device; the AI ​​control unit calculates the fiber entanglement index, which comprehensively reflects the fiber characteristics and entanglement risk of the material, based on the multi-view images and preset physical correction actions, and identifies the category label of the waste vegetable material. 2) Differentiated adaptive control of dehydration process: The AI ​​control unit differentiates and controls the process parameters of the crushing and pressing devices according to the category labels and fiber entanglement index, so as to prevent fiber entanglement and reduce the loss of water-soluble nitrogen and potassium nutrients in the waste vegetables during the dehydration stage. 3) Determination of fermentation endpoint based on plant nutrient indicators: The dehydrated vegetable waste from step 2) is transported to an aerobic fermentation device. Images of the compost surface are collected by a visual acquisition device and input into a maturity recognition model that is constructed based on core plant nutrient indicators of compost and can be dynamically adjusted with the fermentation process. The model outputs a maturity probability value. When the probability value meets the preset conditions, the AI ​​control unit determines that the fermentation endpoint has been reached and initiates the discharge command.

[0007] Preferably, the category label includes leafy vegetables, fruit vegetables, or vines; the fiber entanglement index is a dimensionless value between 0 and 1, which is calculated by the AI ​​control unit based on multi-view images by fusing the area ratio of fibrous materials in the image and the degree of out-of-surface warping.

[0008] Preferably, the process parameters of the differentiated control of the pulverizing device and the pressing device in step 2) specifically include: Set the crushing particle size range and the reference value of the main shaft speed of the pressing device according to the category label; The spindle speed reference value is adjusted proportionally according to the fiber entanglement index. The higher the fiber entanglement index, the greater the reduction in spindle speed. The adjusted spindle speed is not lower than the preset safe operating threshold to reduce nutrient loss and prevent fiber entanglement.

[0009] Preferably, if the category label is leafy vegetables, the crushing particle size is controlled at 10-25mm, and the spindle speed reference value is set at 30-50r / min; if the category label is fruit vegetables, the crushing particle size is controlled at 10-25mm, and the spindle speed reference value is set at 20-40r / min; if the category label is vines, the crushing particle size is controlled at 2-5mm, and the spindle speed reference value is set at 20-40r / min. The proportional reduction rule is as follows: for every 0.1 increase in the fiber winding index, the final target spindle speed is reduced by 10% based on the reference value of the corresponding category, and the final spindle speed is not lower than 10r / min.

[0010] Preferably, the visual acquisition device includes a first industrial camera deployed above the conveying device and a second industrial camera deployed on one side of the conveying device, as well as a third industrial camera and a laser rangefinder fixed above the aerobic fermentation device; the measurement spot of the laser rangefinder is matched with the projection position of the center point of the field of view of the third industrial camera, and is used to monitor the change in object distance caused by the settlement of the pile in real time.

[0011] Preferably, the maturity recognition model in step 3) adopts a multi-branch parallel convolutional neural network architecture that supports dynamic weight fusion, including a first convolutional branch, a second convolutional branch and a third convolutional branch, which are used to extract coarse-grained, medium-grained and fine-grained texture features of the compost surface, respectively; after the collected compost surface image is input into the model, the features are extracted and fused by each branch, and finally the maturity probability value is output by the Softmax function.

[0012] Preferably, the preset conditions in step 3) include: when the compost maturity probability value corresponding to the three consecutive collected compost surface images is greater than 0.95, the AI ​​control unit performs verification sampling; if the GI is determined to be ≥95%, the fermentation endpoint is determined and the discharge command is initiated.

[0013] Preferably, the maturity identification model dynamically adjusts the fusion weights of the output features of each convolutional branch based on the fermentation process coefficient; the fermentation process coefficient is calculated based on the ratio of the current fermentation days to the historical average fermentation cycle of this type of leftover vegetable, and the value range is 0 to 1; the historical average fermentation cycle is obtained by taking the arithmetic mean of the fermentation days of no less than 30 historical batches that were determined to be mature by the method described in claim 7; the weight adjustment rule is: the output feature weight of the coarse-grained branch is higher in the early stage of fermentation, and the output feature weight of the fine-grained branch is higher in the later stage of fermentation.

[0014] Preferably, step 3) also includes steps for pile settlement compensation and lens anti-fog correction: The AI ​​control unit dynamically adjusts the aperture coefficient of the third industrial camera based on the object distance value obtained in real time by the laser rangefinder, so that the compost surface is always within the depth of field. The AI ​​control unit calculates the dew point temperature based on the lens surface temperature and humidity data, and controls the lens surface temperature to be 3-8°C higher than the dew point temperature through a ring-shaped electric heating coil to prevent the lens from fogging.

[0015] Preferably, the preset physical correction action in step 1) includes: when the area ratio of fibrous material detected by the image exceeds the preset risk threshold, the AI ​​control unit pauses the conveying device and drives the horizontal pressure plate to press down the material, and after compaction and smoothing, it collects the corrected multi-view image for calculating the fiber entanglement index.

[0016] The present invention has at least the following beneficial effects: This invention introduces a material characteristic identification mechanism combining physical correction and multi-view visual perception, enabling more accurate acquisition of the category label and fiber entanglement index of waste vegetables, providing a reliable data foundation for differentiated control of subsequent processes. Based on real-time identification of material category and entanglement index, this invention achieves adaptive control of crushing particle size and pressing speed, allowing equipment parameters to be matched in real time with the fiber characteristics of different types of waste vegetables. This effectively avoids equipment blockage caused by long-fiber materials entangled in the main shaft, while reducing unnecessary leaching and loss of water-soluble nitrogen and potassium nutrients, increasing the nutrient retention rate of dehydrated waste vegetables to over 85%, which is beneficial for improving the continuity of the dehydration process and nutrient retention rate. In the fermentation stage, this invention constructs a maturity identification model associated with core plant nutrition indicators. By collecting images of the compost surface, it can output real-time maturity probability values ​​corresponding to key quality parameters such as seed germination index. When the probability value continuously meets the preset threshold, the AI ​​control unit triggers targeted verification sampling. Only a small number of suspected fermented batches need to be tested using the standard method. This transforms the traditional tedious timed sampling and testing into verification after intelligent early warning. While ensuring that the germination index (GI) of the fermented product seeds is ≥95% and the carbon-nitrogen ratio (C / N) is ≤18 (compliant with the NY / T 525-2021 organic fertilizer standard), it significantly reduces the workload of manual sampling and laboratory testing, realizes intelligent auxiliary determination of the fermentation endpoint, and avoids insufficient fermentation or nutrient loss caused by improper discharge timing.

[0017] Through the synergistic effect of the above-mentioned material sensing, dehydration control and fermentation determination, the present invention improves the automation level of the entire process of resource utilization of vegetable waste, significantly reduces the frequency of manual intervention while ensuring that the product quality meets the standards, and helps to ensure the quality stability of vegetable waste organic fertilizer products.

[0018] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the overall process flow for processing vegetable waste according to the present invention. Figure 2 This is a schematic diagram of the material property sensing device of the present invention.

[0020] Among them, there is a conveying device 1; a horizontal pressure plate 2; a linear module 3; a gantry bracket 4; a first industrial camera 5; and a second industrial camera 6. Detailed Implementation

[0021] The present invention will now be described in further detail so that those skilled in the art can implement it based on the description.

[0022] It should be understood that terms such as “having,” “comprising,” and “including” as used herein do not exclude the presence or addition of one or more other elements or combinations thereof.

[0023] It should be noted that, unless otherwise specified, the experimental methods described in the following implementation plan are all conventional methods, and the reagents and materials described are all commercially available unless otherwise specified.

[0024] This invention discloses an AI-controlled method for processing waste vegetables from agricultural facilities, comprising the following steps: 1) Intelligent perception of material characteristics: Multi-view images of the waste vegetable material are acquired through a visual acquisition device; the AI ​​control unit calculates the fiber entanglement index, which comprehensively reflects the fiber characteristics and entanglement risk of the material, based on the multi-view images and preset physical correction actions, and identifies the category label of the waste vegetable material. 2) Differentiated adaptive control of dehydration process: The AI ​​control unit differentiates and controls the process parameters of the crushing and pressing devices according to the category labels and fiber entanglement index, so as to prevent fiber entanglement and reduce the loss of water-soluble nitrogen and potassium nutrients in the waste vegetables during the dehydration stage. 3) Determination of fermentation endpoint based on plant nutrient indicators: The dehydrated vegetable waste from step 2) is transported to an aerobic fermentation device. Images of the compost surface are collected by a visual acquisition device and input into a maturity recognition model that is constructed based on core plant nutrient indicators of compost and can be dynamically adjusted with the fermentation process. The model outputs a maturity probability value. When the probability value meets the preset conditions, the AI ​​control unit determines that the fermentation endpoint has been reached and initiates the discharge command.

[0025] In existing waste vegetable processing technology, dehydration equipment usually operates with fixed parameters. When processing vine-type waste vegetables with high fiber content, long fiber materials are prone to getting tangled in the crushing rollers or pressing shaft, causing the equipment to stop frequently for cleaning. At the same time, the degree of fermentation and decomposition depends on human experience, making it difficult to effectively control the timing of discharge. Furthermore, the traditional dehydration process does not take into account the loss of water-soluble nitrogen and potassium nutrients.

[0026] In the embodiments of the present invention, see Figure 1-2 As shown, a conveyor device 1 is installed on the processing line. A first industrial camera 5 and a second industrial camera 6 are respectively installed above and on one side of the conveyor device 1 to collect multi-view images of the waste material. The AI ​​control unit receives the images from the first camera in real time and identifies the area ratio of fibrous materials with an aspect ratio greater than 10 in the image through a trained convolutional neural network. When this ratio exceeds a preset threshold (e.g., 5% to 15%), the AI ​​control unit determines that the current material has a high risk of fiber entanglement, and then stops the drive motor of the conveyor device and sends a downward pressing command to the linear module 3. The module drives the horizontal pressure plate 2 to move downward at an appropriate speed to compact and flatten the loosely piled fibrous material, and then resets it after a few seconds (the horizontal pressure plate is located above the first industrial camera in both the initial state and the reset state to avoid obstructing the image collected by the first industrial camera). Figure 2 The image shows the horizontal pressure plate 2 moving downwards and pressing down. The AI ​​control unit then acquires top-view and side-view images of the flattened material. Based on these multi-view images, the AI ​​control unit extracts fibrous material regions using a semantic segmentation model, calculates the area and aspect ratio of each region, and calculates the warpage height of the fibrous material relative to the conveyor belt surface from the side-view image. These parameters are then fused into a fiber entanglement index between 0 and 1, which comprehensively reflects the degree of entanglement risk caused by the material. Simultaneously, the AI ​​control unit inputs the top-view image into a classification model and outputs the category label of the material, such as leafy vegetables, fruit vegetables, or vines.

[0027] To verify the nutrient retention effect of the dehydration process, this method can optionally include a step for detecting the nutrient content of the vegetable waste: the Kjeldahl method is used to determine the total nitrogen content of the waste, and the potassium dichromate titration method is used to determine the organic carbon content, and the carbon-to-nitrogen ratio is calculated. The detection frequency is one sample before and after dehydration for each batch, with no fewer than three sampling points. The average value is used as the basis for calculating the nutrient retention rate for that batch. During aerobic fermentation, the carbon-to-nitrogen ratio is adjusted to 25:1 to 30:1 before the dehydrated waste enters the fermentation device. The fermentation device uses forced ventilation for oxygen supply, with the ventilation volume controlled at 0.1 m³ / s. 3 / (m 3 ·min) to 0.3 m 3 / (m 3 ·min), the fermentation temperature is maintained in the high-temperature composting range of 55℃ to 65℃.

[0028] After obtaining the category label and fiber entanglement index, the AI ​​control unit retrieves the preset process parameter baseline values ​​based on the category label. For leafy vegetables, the particle size is set to 10-25 mm, and the base speed of the pressing device's main shaft is 30-50 r / min; for fruit vegetables, the particle size is also 10-25 mm, and the base speed is 20-40 r / min; for vine vegetables, the particle size is refined to 2-5 mm, and the base speed is 20-40 r / min. Based on this, the AI ​​control unit linearly adjusts the speed according to the fiber entanglement index. For every 0.1 increase in the entanglement index, the final target speed decreases by 10% from the baseline value, and the adjusted speed is no lower than 10 r / min. This reduces the risk of long fibers entangled in the main shaft and avoids excessive compression leading to cell rupture in the vegetable residue and significant loss of water-soluble nitrogen and potassium nutrients. The adjusted particle size and speed commands are sent to the gap adjustment mechanism of the crushing device and the frequency converter of the pressing device, respectively, to complete the dehydration process.

[0029] The dehydrated compost is transported to an aerobic fermentation unit. A third industrial camera and a laser rangefinder are fixedly installed above the fermentation unit. The laser rangefinder's measurement spot corresponds to the center point of the camera's field of view, used to monitor changes in object distance caused by surface settlement in the compost pile in real time. The AI ​​control unit triggers the camera to acquire images of the compost surface at a frequency of once per hour, and obtains the current object distance value through laser rangefinder during the acquisition interval. When the object distance value exceeds the depth of field range corresponding to the current aperture value, the AI ​​control unit automatically adjusts the camera aperture to an appropriate setting according to the pre-stored aperture-depth of field parameter table to ensure that the compost surface is always clearly imaged. A ring-shaped electric heating coil is nested at the front of the camera lens. The AI ​​control unit calculates the dew point temperature based on the temperature and humidity sensor data near the lens. When the lens surface temperature is 2°C lower than the dew point temperature plus 2°C, the heating coil is activated to maintain the lens temperature at 3-8°C above the dew point, preventing water vapor condensation and fogging.

[0030] The collected surface images of the compost were input into a pre-trained maturity recognition model. This model employed a multi-branch parallel convolutional neural network architecture, comprising three branches that extracted coarse-grained, medium-grained, and fine-grained texture features, respectively. During the training phase, the model used core plant nutrient indicators such as seed germination index and carbon-nitrogen ratio of the compost samples as supervisory labels, closely correlating the output probability value with the actual degree of maturity. Simultaneously, the AI ​​control unit calculated a fermentation progress coefficient based on the ratio of the current fermentation days to the historical average fermentation cycle for this type of vegetable waste. This coefficient, varying between 0 and 1, was used to dynamically adjust the fusion weights of the output features from the three branches: the coarse-grained branch had a higher weight in the early stages of fermentation, while the fine-grained branch had a higher weight in the later stages. The model ultimately output a maturity probability value between 0 and 1 through a Softmax layer. Based on the high linear correlation between the model's output maturity probability value and the measured GI value, when the maturity probability values ​​corresponding to three consecutive collected images were all greater than 0.95, this could be used as a criterion for the AI ​​control unit to determine that the fermentation endpoint had been reached, automatically activating the discharge mechanism to remove the fully decomposed material. Alternatively: When the probability value of compost maturity corresponding to three consecutive collected compost surface images is greater than 0.95, the AI ​​control unit performs verification sampling; the verification sampling is: the compost seed germination index of the sampled sample determined according to the NY / T 525-2021 standard; if the GI is ≥ 95%, it is determined that the fermentation endpoint has been reached, and the discharge is started again.

[0031] Compared to existing dehydration methods with fixed parameters and methods relying on manual experience to judge compost maturity, this implementation combines multi-view visual perception with physical correction to identify material type and entanglement risk in real time. Based on this, it differentiates and adjusts the crushing particle size and pressing speed, effectively avoiding equipment downtime caused by fiber entanglement and reducing the loss of water-soluble nitrogen and potassium nutrients. During the fermentation stage, a visual model associated with plant nutrient indicators monitors the composting process in real time. When a preset maturity probability value is reached, it can be triggered as the fermentation endpoint or as a verification step, significantly reducing reliance on manual experience and frequent sampling. This facilitates the determination of the discharge timing and ensures the quality stability of the vegetable waste organic fertilizer product. The entire process is coordinated by an AI control unit, achieving intelligent auxiliary control from pretreatment to fermentation discharge.

[0032] In another embodiment, the fiber entanglement index is a dimensionless value between 0 and 1, which is calculated by the AI ​​control unit based on multi-view images by fusing the area ratio of fibrous material in the image and the degree of off-surface warping.

[0033] In existing vegetable waste dehydration processes, the assessment of fiber entanglement risk is usually based solely on the area ratio of fibrous material in a two-dimensional image. This single indicator ignores the actual impact of the height of the fibrous material's rise on the entanglement risk, resulting in inaccurate risk quantification and potentially leading to insufficient or excessive control.

[0034] In this embodiment, as part of the physical correction process, after the pressure plate is reset and before the conveying device resumes operation, the AI ​​control unit triggers the first industrial camera to capture a top-view image, and simultaneously triggers the second industrial camera to capture a side-view image. These two images are acquired after the material has been compacted and smoothed by the horizontal pressure plate, but before it has moved, and can accurately reflect the fiber distribution and warping state of the material.

[0035] The AI ​​control unit first inputs the top-down image into a convolutional neural network recognition model trained through transfer learning. This model includes an encoder-decoder semantic segmentation branch. The semantic segmentation branch outputs a pixel-level classification feature map with the same resolution as the input image, where each pixel is labeled as either background or fibrous material. The AI ​​control unit then performs connected component analysis on all pixels labeled as fibrous material, extracting each connected region and calculating the minimum bounding rectangle and minimum bounding ellipse for each region. The ratio of the major axis to the minor axis of the minimum bounding ellipse is used as the aspect ratio of the fibrous material, and connected regions with an aspect ratio greater than 10 are selected. These regions correspond to long fibers that may pose a risk of entanglement. The AI ​​control unit calculates the sum of the pixel areas of all selected connected regions, which reflects the total amount of long fibers within the field of view.

[0036] Simultaneously, the AI ​​control unit inputs the side view image captured by the second industrial camera into the same convolutional neural network recognition model to perform the second semantic segmentation task, identifying pixel regions of all fibrous materials in the side view image. For each fibrous material region, its upper edge contour is extracted, and the pixel coordinates of each pixel on the contour line in the normal direction of the conveying device are calculated. The pre-calibrated reference coordinates of the upper surface of the conveying device corresponding to that pixel column are subtracted to obtain the off-surface offset of each point. The off-surface offset reflects the height of the fibrous material warping. The AI ​​control unit calculates the arithmetic mean of the off-surface offsets of all pixels in all fibrous material regions, and uses the distance from the end point of the pressure plate as described in the first embodiment (30 mm to 50 mm) as the normalization upper limit. This mean is divided by the upper limit and truncated to the interval of 0 to 1 to obtain the global off-surface warping coefficient at the current moment. This coefficient reflects the overall degree of warping of the fibrous material within the entire field of view; the higher the warping, the greater the risk of entanglement.

[0037] After completing the above calculations, the AI ​​control unit multiplies the sum of the pixel areas of all connected regions with an aspect ratio greater than 10 in the top-view image by the global out-of-plane warping coefficient, and then divides it by the total pixel area of ​​the image captured by the first industrial camera to obtain a dimensionless value between 0 and 1. This value is the fiber entanglement index at the current moment. This index comprehensively reflects the content of long fibers in the material and the degree of fiber warping; the higher the value, the greater the overall entanglement risk.

[0038] Compared to existing technologies that assess entanglement risk solely based on two-dimensional area ratio, this implementation introduces a side-view image to obtain fiber warping information. This allows the entanglement index to more comprehensively reflect the entanglement risk under actual working conditions, providing a more reliable basis for precise control of subsequent process parameters. The AI ​​control unit outputs this fiber entanglement index to the differentiated adaptive control step of the dewatering process for dynamic correction of the spindle speed of the variable-diameter screw press.

[0039] In another embodiment, the process parameters of the differentially controlled crushing device and pressing device in step 2) specifically include: setting the crushing particle size range and the reference value of the main shaft speed of the pressing device according to the category label; adjusting the reference value of the main shaft speed proportionally according to the fiber entanglement index. The higher the fiber entanglement index, the greater the reduction in the main shaft speed, and the reduced main shaft speed is not lower than the preset safe operating threshold, so as to reduce nutrient loss and prevent fiber entanglement.

[0040] After acquiring the category label and fiber entanglement index of the current batch of leftover vegetables, the AI ​​control unit retrieves the corresponding particle size range and the benchmark value of the pressing device's main shaft speed from an internally preset mapping relationship based on the category label. Leafy vegetables correspond to a particle size of 10-25 mm and a benchmark speed of 30-50 r / min; fruit vegetables correspond to a particle size of 10-25 mm and a benchmark speed of 20-40 r / min; and vine vegetables correspond to a particle size of 2-5 mm and a benchmark speed of 20-40 r / min. Based on the obtained benchmark values, the AI ​​control unit uses the fiber entanglement index as a risk quantification basis and linearly reduces the main shaft speed. For every 0.1 increase in the entanglement index, the final target speed decreases by 10% from the benchmark value for the corresponding category. If the target speed calculated according to this rule is lower than the preset safe operating threshold of 10 r / min, the AI ​​control unit automatically locks the final command speed to this threshold, ensuring that the equipment can maintain basic operation even under extremely high entanglement risk. The adjusted particle size and rotation speed commands are sent to the gap adjustment mechanism of the crushing device and the frequency converter of the pressing device, respectively, so that the process parameters matching the current batch of leftover vegetables are used when they enter the dewatering section. Category labels ensure that the basic parameters are compatible with the material category, and the rotation speed is dynamically reduced according to actual risks by using the winding index. This effectively avoids equipment failures caused by long fibers winding around the main shaft and prevents unnecessary loss of water-soluble nitrogen and potassium nutrients due to excessive compression.

[0041] If the category label is leafy vegetables, the particle size should be controlled at 10-25mm, and the spindle speed should be set at 30-50 r / min. If the category label is fruit vegetables, the particle size should be controlled at 10-25mm, and the spindle speed should be set at 20-40 r / min. If the category label is vines, the particle size should be controlled at 2-5mm, and the spindle speed should be set at 20-40 r / min. The proportional reduction rule is as follows: for every 0.1 increase in the fiber winding index, the final target spindle speed should be reduced by 10% based on the corresponding category's benchmark value, and the final spindle speed should not be lower than 10 r / min.

[0042] In existing dehydration processes for vegetable waste, equipment typically operates with fixed particle sizes and pressing speeds, making it difficult for operators to adjust parameters in a timely manner based on material characteristics. When processing leafy vegetable waste, if the speed is set too high, although the dehydration efficiency is acceptable, it can easily cause cell rupture and loss of nitrogen and potassium nutrients. When processing vine-like vegetable waste, if the speed is too low, dehydration will be insufficient, while if it is too high, it can easily cause fiber entanglement.

[0043] In this embodiment, after the AI ​​control unit obtains the category label and fiber entanglement index of the current batch of vegetable waste, it retrieves the corresponding particle size range and the reference value of the main shaft speed of the pressing device from an internally preset category parameter mapping table based on the category label. This mapping table is pre-calibrated and fixed in the control unit based on a large amount of experimental data. If the category label is leafy vegetables, the particle size is controlled to be 10 mm to 25 mm, and the reference value of the main shaft speed is set to 30 r / min to 50 r / min; if it is fruit vegetables, the particle size is also 10 mm to 25 mm, and the reference value of the speed is set to 20 r / min to 40 r / min; if it is vine vegetables, the particle size is refined to 2 mm to 5 mm, and the reference value of the speed is set to 20 r / min to 40 r / min. These reference values ​​enable various types of vegetable waste to achieve effective dehydration under normal operating conditions with a low risk of entanglement.

[0044] After obtaining the baseline value, the AI ​​control unit linearly adjusts the spindle speed based on the fiber entanglement index. Specifically, for every 0.1 increase in the entanglement index, the final target spindle speed decreases by 10% from the baseline value for that category. For example, for vine-like vegetables, if the baseline speed is the median of 30 r / min, when the fiber entanglement index is 0.2, the target speed is calculated as 30 multiplied by 0.8 equals 24 r / min; when the entanglement index is 0.5, the target speed is 30 multiplied by 0.5 equals 15 r / min; when the entanglement index reaches 0.8, the target speed is 30 multiplied by 0.2 equals 6 r / min. Since this is below the preset safe operating threshold of 10 r / min, the AI ​​control unit automatically locks the final command speed at 10 r / min to ensure the pressing device can maintain basic conveying and dehydration functions and prevent material accumulation. Similarly, this rule applies to leafy vegetables and fruit vegetables; if the calculated result is below 10 r / min, then 10 r / min is used as the final speed.

[0045] After completing the rotational speed calculation, the AI ​​control unit converts the determined particle size range and target spindle speed into control commands, which are then sent via fieldbus to the gap adjustment mechanism of the crushing unit and the frequency converter driver of the pressing unit. The crushing unit adjusts the blade gap according to the commands to ensure the output particle size falls within the specified range; the pressing unit adjusts the spindle motor frequency according to the commands to precisely control the rotational speed at the target value. The entire parameter adjustment process is completed within the time window from when the material is transported from the image acquisition position to the inlet of the crushing unit, ensuring that the matching process parameters are used when the current batch of leftover material enters the dewatering section.

[0046] Compared to existing fixed-parameter dehydration methods, this implementation method matches basic parameters with category labels, enabling different categories of waste vegetables to obtain suitable initial crushing particle size and rotation speed. It dynamically adjusts the rotation speed by controlling the entanglement index, automatically reducing the speed to avoid entanglement in high-risk situations and maintaining a higher speed to ensure dehydration efficiency in low-risk situations. Simultaneously, it sets a safe operating threshold to prevent equipment blockage caused by excessively low rotation speeds in extreme cases. This differentiated control strategy based on category and risk level effectively resolves the contradiction between fiber entanglement and nutrient loss, improving the adaptability and stability of the dehydration process.

[0047] In another embodiment, the visual acquisition device includes a first industrial camera deployed above the conveying device and a second industrial camera deployed on one side of the conveying device, as well as a third industrial camera and a laser rangefinder fixed above the aerobic fermentation device; the measurement spot of the laser rangefinder is matched with the projection position of the center point of the field of view of the third industrial camera, and is used to monitor the change in object distance caused by the settlement of the pile in real time.

[0048] In this embodiment, the material conveying section of the processing line is equipped with a first industrial camera 5 and a second industrial camera 6. The first industrial camera 5 is mounted directly above the conveyor device 1, with its main optical axis perpendicular to the surface of the conveyor belt, and is used to capture top-view images of the waste material. The second industrial camera 6 is mounted on one side of the conveyor device, with its main optical axis parallel to the surface of the conveyor belt and perpendicular to the direction of operation of the conveyor device, and is used to capture side-view images to obtain the warping state of the fibrous material. Both cameras are fixed on the gantry bracket 4 to ensure stable shooting positions. Above the aerobic fermentation device, a third industrial camera is fixedly installed to capture images of the compost surface at a preset frequency throughout the entire fermentation cycle. The third industrial camera uses a fixed focal length lens, and a laser rangefinder sensor is fixedly mounted on its side. The main optical axis of the measuring beam of the laser rangefinder is parallel to the main optical axis of the third industrial camera, and the projection position of its measuring spot in the compost surface image corresponds to the center point of the camera's field of view (those skilled in the art can understand this as: the laser ranging point falls within a sufficiently close range near the center point of the camera's field of view, such as ±5cm, so that the distance value of this point can represent the object distance of the entire imaging area), ensuring that the ranging result can accurately reflect the object distance of the imaging area.

[0049] During the interval between two adjacent image acquisitions, the AI ​​control unit sends ranging commands to the laser rangefinder at an appropriate frequency. The sensor emits laser pulses towards the compost surface and receives reflected pulses. It calculates the vertical distance between the camera's current position and the compost surface using the pulse flight time and records this distance as the current object distance value. The AI ​​control unit internally stores at least three sets of depth-of-field parameter tables for the camera at the current aperture value. Each table includes the aperture value and the corresponding near and far distance values. The AI ​​control unit compares the current object distance value with the corresponding near and far distance values. If the current object distance value is less than the near distance value or greater than the far distance value, it indicates that the compost surface has subsided or bulged beyond the current depth-of-field range, and the image will tend to be blurry. At this time, the AI ​​control unit retrieves the target aperture value from the depth-of-field parameter table based on the current object distance value. This target aperture value is the aperture value that places the current object distance value between its corresponding near and far distance values. If multiple aperture values ​​meet the requirements, the one with the widest depth of field is selected. If no pre-stored aperture value can ensure the current object distance falls within the depth of field, the AI ​​control unit selects the one with the largest aperture value (i.e., the smallest aperture diameter) to obtain the maximum depth of field. After determining the target aperture value, the AI ​​control unit sends an adjustment command to the automatic aperture drive circuit of the third industrial camera, driving the aperture blades to adjust the camera aperture to the target setting, ensuring the compost surface falls back into the clear imaging range.

[0050] A ring-shaped electric heating coil is fixedly nested at the front end of the third industrial camera lens. The inner diameter of the coil fits tightly with the outer diameter of the lens front end, and the outer side is covered with a polyimide heat insulation layer to reduce heat loss. Temperature and humidity sensors are deployed on the side wall of the lens front end to acquire the lens surface temperature and relative humidity in real time. The AI ​​control unit calculates the dew point temperature under the current environmental conditions based on an air enthalpy-humidity chart. When the lens surface temperature is 2°C below the dew point temperature, it indicates that the lens surface temperature is approaching or below the dew point, and water vapor condensation is about to occur or has already occurred. The AI ​​control unit then sends a conduction command to the solid-state relay of the ring-shaped electric heating coil. The solid-state relay controls the on / off state of the heating coil using pulse width modulation at an appropriate frequency, causing the lens surface temperature to gradually rise and be maintained within a safe range of 3°C to 8°C above the current dew point temperature, thus thermally preventing water vapor from condensing into fog on the lens surface.

[0051] After aperture adjustment and lens heating are completed, the AI ​​control unit triggers the third industrial camera to acquire images of the compost surface and inputs these images into the maturity recognition model for subsequent determination. Compared with existing fixed-parameter imaging methods, this implementation method uses laser ranging to monitor the object distance in real time and dynamically adjust the aperture, ensuring that the camera can always acquire clear images of the compost surface throughout the entire fermentation cycle. Simultaneously, active lens temperature control technology continuously prevents droplet formation, ensuring distortion-free images. These two active correction mechanisms guarantee stable acquisition of visual perception data during long-cycle fermentation, providing a reliable image data foundation for visual determination of maturity.

[0052] In another embodiment, the maturity recognition model in step 3) adopts a multi-branch parallel convolutional neural network architecture that supports dynamic weight fusion, including a first convolutional branch, a second convolutional branch and a third convolutional branch, which are used to extract coarse-grained, medium-grained and fine-grained texture features of the compost surface, respectively; after the collected compost surface image is input into the model, the features are extracted and fused by each branch, and finally the maturity probability value is output by the Softmax function.

[0053] Current visual assessment techniques for compost maturity typically employ single-architecture convolutional neural network models for feature extraction and classification of compost surface images. These models use fixed-size convolutional kernels and can only extract image features at a single scale. However, the surface texture of compost changes during fermentation: in the early stages, the material is coarse, exhibiting distinct straw and leaf outlines; in the middle stages, the material gradually refines, forming a uniform granular texture; and in the final stages, white actinomycete colonies and fine humic particles appear. Single-scale feature extraction struggles to simultaneously capture these cross-scale texture changes, leading to a decrease in model accuracy in the later stages of fermentation.

[0054] In this embodiment, after completing the aforementioned aperture adjustment and lens heating, the AI ​​control unit triggers the third industrial camera to acquire an image of the compost surface. This image is input into a pre-trained maturity recognition model, which employs a multi-branch parallel convolutional neural network architecture containing three parallel convolutional branches. The first convolutional branch uses a larger kernel to extract coarse-grained texture features from the compost surface, such as the outlines of thick stems and leaves that are not yet fully decomposed in the early stages of fermentation. The second convolutional branch uses a medium-sized kernel to extract medium-grained texture features, such as the mesoscopic texture formed after the material particles are homogenized in the middle stages of fermentation. The third convolutional branch uses a smaller kernel to extract fine-grained texture features, such as the microscopic morphology of white actinomycete colonies and fine humic particles in the late stages of fermentation.

[0055] The input image is processed simultaneously through three convolutional branches. Each branch consists of a concatenated convolutional layer, a pooling layer, and a fully connected layer, independently extracting texture features at their respective scales and outputting a feature vector representing the coarse-grained, medium-grained, and fine-grained texture information of the compost surface. These three feature vectors are then merged in the model's fusion layer to form a fused feature vector, which integrates image information from different scales. The fused feature vector is then input to the output fully connected layer and finally processed by the Softmax activation function, outputting a maturity probability value between 0 and 1, representing the degree of maturity corresponding to the current compost surface image.

[0056] Compared to existing convolutional neural network models that use a single fixed receptive field, this implementation uses a multi-branch parallel architecture to simultaneously extract texture features at three scales: coarse, medium, and fine. This enables the model to perceive multi-scale texture changes throughout the entire composting fermentation process. Regardless of whether fermentation is in the early, middle, or late stages, the model can extract feature information from the image that is appropriate to that stage, avoiding the problem of decreased recognition ability in the later stages caused by a single feature scale. This provides a more accurate and stable input for subsequent determination of the fermentation endpoint based on the probability value of maturity.

[0057] In another embodiment, the preset conditions in step 3) include: when the probability value of compost maturity corresponding to three consecutive collected compost surface images is greater than 0.95, the AI ​​control unit performs verification sampling; the verification sampling is: the compost seed germination index of the sampled sample measured according to the NY / T 525-2021 standard; if the GI is ≥ 95%, the fermentation endpoint is determined and the discharge command is initiated.

[0058] In this embodiment, after aperture adjustment and lens heating are completed, the AI ​​control unit triggers the third industrial camera to acquire images of the compost surface at a preset frequency and inputs the images into the aforementioned multi-branch parallel convolutional neural network model. The model outputs a maturity probability value between 0 and 1 through the Softmax layer. During the training phase, the model employs a special supervised labeling method: for each training image, not only are technical personnel visually labeled, but multiple points are sampled from the compost pile position corresponding to the camera's field of view, and the seed germination index of the samples is tested according to the NY / T 525-2021 organic fertilizer industry standard. Only when the seed germination index of the sample is greater than or equal to 95% is the image labeled as fully decomposed and used as a positive sample for model training. In this way, the maturity probability value learned by the model forms an intrinsic correlation with the seed germination index, a core indicator of plant nutrition.

[0059] During fermentation, the AI ​​control unit continuously monitors the maturity probability value output by the model. When the maturity probability value corresponding to three consecutive collected compost surface images is greater than 0.95, the AI ​​control unit triggers a verification sampling. The sampling location is the compost pile area corresponding to the center of the camera's field of view, and the sampling depth is 10 cm to 20 cm below the surface. The seed germination index of the sample is determined according to the NY / T 525-2021 standard. If the measurement result shows that the seed germination index is greater than or equal to 95%, the AI ​​control unit finally confirms that the fermentation endpoint has been reached and automatically initiates the discharge command to discharge the fully decomposed material. If the measurement result is lower than this threshold, the AI ​​control unit determines that there is a deviation in the model output, continues fermentation, and records the anomaly for subsequent model optimization.

[0060] Compared to existing methods that rely solely on model probability values, this implementation substantially links the model output to industry standards. During the training phase, the model is aligned with the seed germination index, ensuring its output probability values ​​have plant nutritional significance. In the application phase, a sampling verification process is retained to ensure that every discharge decision is confirmed by industry standards. This dual-guarantee mechanism leverages the advantages of real-time monitoring by the visual model while eliminating the uncertainty of black-box model decisions. This allows the fermentation endpoint determination to be directly used for discharge control after verification, avoiding misjudgments caused by model bias. Furthermore, it eliminates the need for additional waiting time for laboratory testing, achieving automation and standardization in the determination of fermentation maturity.

[0061] In another embodiment, the maturity identification model dynamically adjusts the fusion weights of the output features of each convolutional branch based on the fermentation process coefficient; the fermentation process coefficient is calculated based on the ratio of the current fermentation days to the historical average fermentation cycle of this type of leftover vegetable, and the value ranges from 0 to 1; the historical average fermentation cycle is obtained by taking the arithmetic mean of the fermentation days of no less than 30 historical batches that were determined to be mature by the aforementioned method; the weight adjustment rule is: the output feature weight of the coarse-grained branch is higher in the early stage of fermentation, and the output feature weight of the fine-grained branch is higher in the later stage of fermentation.

[0062] In this embodiment, before each acquisition of compost surface images and input into the maturity recognition model, the AI ​​control unit first obtains the current number of natural days the aerobic fermentation device has been running. Simultaneously, the AI ​​control unit internally stores the historical average fermentation cycle days for this type of waste under standard operating conditions. This cycle day number is obtained by taking the arithmetic mean of the fermentation days at which the maturity probability value of at least 30 historical batches first consecutively exceeds 0.95 three times. The AI ​​control unit divides the current fermentation days by the historical average fermentation cycle days to obtain a value between 0 and 1 as the fermentation progress coefficient. This coefficient is close to 0 on the fermentation start day and gradually increases as fermentation progresses, approaching 1 when the historical average fermentation cycle is reached.

[0063] After obtaining the fermentation process coefficient, the AI ​​control unit inputs this coefficient into the weight generation module to dynamically adjust the fusion weights of the output features of the three convolutional branches in the multi-branch parallel convolutional network. The weight generation rules are as follows: The first convolutional branch is responsible for extracting coarse-grained texture features, and its weight is set to 1 minus the fermentation process coefficient. That is, when the coefficient is close to 0 in the early stage of fermentation, the weight of this branch is close to 1, and when the coefficient is close to 1 in the later stage of fermentation, the weight of this branch is close to 0. The second convolutional branch is responsible for extracting medium-grained texture features, and its weight is always kept at a constant value of 0.5. The third convolutional branch is responsible for extracting fine-grained texture features, and its weight is directly set to the fermentation process coefficient. That is, the weight is lower in the early stage of fermentation and gradually increases to close to 1 in the later stage of fermentation.

[0064] After weight calculation, the decay maturity recognition model multiplies the feature vector output from the first convolutional branch by its corresponding weight, the feature vector output from the second convolutional branch by 0.5, and the feature vector output from the third convolutional branch by the fermentation process coefficient. Then, it performs element-wise summation on the three weighted feature vectors to obtain a fused feature vector. This fused feature vector contains the main scale features relevant to the current fermentation stage while retaining auxiliary information from other scales. The fused feature vector is then input to the output fully connected layer, where it is calculated using the Softmax function and outputs the current decay maturity probability value.

[0065] Compared to existing models that use fixed feature weights, this implementation introduces a fermentation process coefficient to achieve dynamic adjustment of feature weights. In the early stages of fermentation, coarse-grained features have higher weights, allowing the model to focus on changes in the macroscopic morphology of the material; in the middle stages of fermentation, medium-grained features maintain a stable contribution; and in the later stages of fermentation, fine-grained features have higher weights, enabling the model to capture microscopic ripening characteristics. This weight adjustment mechanism, which matches the fermentation process, ensures that the model consistently focuses on the texture scale most relevant to the current stage throughout the entire fermentation cycle. This avoids the decline in recognition ability in the middle and later stages caused by fixed feature weights, providing a more stable and reliable feature basis for the output of ripening probability values.

[0066] In another embodiment, step 3) further includes refined parameters for the pile settling compensation and lens anti-fogging correction steps: the measurement frequency of the laser rangefinder can be set to 1 Hz to 2 Hz, the pulse width modulation frequency of the annular electric heating coil can be set to 20 Hz to 50 Hz, and the lens surface temperature is maintained at 3°C ​​to 8°C above the dew point temperature. These parameters can be adaptively adjusted according to the actual fermentation environment to further optimize imaging stability.

[0067] In another embodiment, the preset physical correction action in step 1) includes: when the area ratio of fibrous material detected by the image exceeds a preset risk threshold, the AI ​​control unit pauses the conveying device and drives the horizontal pressure plate to press down the material, and after compaction and smoothing, it collects the corrected multi-view image to calculate the fiber entanglement index.

[0068] In existing visual inspection systems for waste vegetable processing lines, industrial cameras typically capture raw images of the materials on the conveyor belt, and then image processing algorithms are used to identify material features. However, waste vegetables, especially vine-like materials, are often loosely piled up, with fibrous materials intertwined and layered. In images taken under natural conditions, a large amount of fiber is obscured by the upper layers of material, failing to accurately reflect the actual fiber content and distribution. This visual occlusion caused by the material's pile-up shape results in the fibrous material area ratio calculated based on the raw image being far lower than the actual value, leading to severely distorted entanglement risk assessments and a lack of reliable basis for subsequent process parameter adjustments.

[0069] In this embodiment, a horizontal pressure plate that can move up and down is installed above the conveyor of the processing line. The long side of the pressure plate is parallel to the running direction of the conveyor, and its lower surface is covered with a polytetrafluoroethylene (PTFE) anti-stick coating to prevent material adhesion. The initial distance between the pressure plate and the upper surface of the conveyor is maintained at 200 mm to 300 mm to ensure that the material can pass smoothly during normal conveying. The pressure plate is driven by a linear module and can move up and down in response to control commands. A first industrial camera is arranged close to the rear of the pressure plate, with its main optical axis perpendicular to the upper surface of the conveyor, for capturing top-view images; a second industrial camera is installed on one side of the conveyor, with its main optical axis parallel to the upper surface of the conveyor and perpendicular to the running direction, for capturing side-view images.

[0070] The AI ​​control unit receives real-time image data from the first industrial camera, and calculates the area ratio of fibrous materials with an aspect ratio greater than 10 in the image using a convolutional neural network recognition model trained through transfer learning. When this area ratio exceeds a preset risk threshold, the AI ​​control unit determines that the current material has a high fiber content and is loosely packed, and the directly acquired image cannot accurately reflect the risk of fiber entanglement. It then issues a stop operation command to the drive motor of the conveyor and a downward pressure command to the linear module.

[0071] Upon receiving the command, the linear module drives the horizontal pressure plate downwards at a speed of 50 mm / s to 80 mm / s until the distance between the lower edge of the pressure plate and the upper surface of the conveyor reaches a final position of 30 mm to 50 mm. At this position, the pressure plate applies pressure to the loosely piled fibrous material, compacting and smoothing it, causing the originally layered and raised fibrous material to unfold and adhere to the surface of the conveyor belt. The pressure plate is held at this final position for 1.5 s to 2.5 s to ensure that the material is fully compacted, and then the linear module drives the pressure plate to quickly return to the initial distance position at a speed of 80 mm / s to 120 mm / s.

[0072] After the pressure plate resets, the conveyor remains paused. The AI ​​control unit immediately triggers the first and second industrial cameras to capture the first top-down view and one side-view image after flattening, respectively. In these images, the fibrous material has been freed from the mutual obstruction caused by its loose accumulation, accurately reflecting the fiber content, distribution, and any remaining warping after compaction. After image acquisition, the AI ​​control unit waits 1.5 to 2.5 seconds before sending a command to the conveyor's drive motor to resume operation, allowing the material to continue forward into subsequent processes. The entire physical correction process is completed within the few seconds of the conveyor's brief pause, without significantly impacting overall processing efficiency.

[0073] The flattened top and side views were used to calculate the fiber entanglement index. The top view was used for semantic segmentation and connected component analysis to obtain the regions of fibrous material with an aspect ratio greater than 10 and their total area; the side view was used for semantic segmentation and edge extraction to obtain the out-of-plane warpage coefficient of the fibrous material. The fiber entanglement index obtained by fusing these two images more accurately reflects the actual degree of entanglement risk caused by the material.

[0074] Compared to existing methods that directly acquire raw images for feature recognition, this implementation uses a horizontal pressure plate to physically correct the loose material, eliminating visual occlusion and morphological distortion caused by material accumulation. This ensures that the acquired images accurately reflect the actual content and warping state of the fibrous material. The fiber entanglement index calculated based on the corrected image is more accurate and reliable, providing a real data foundation for differentiated control of subsequent dewatering process parameters and avoiding the underestimation of risks caused by material accumulation.

[0075] Example 1: Construction and Training of a Compost Maturity Recognition Model This embodiment details the specific construction method and training process of the compost maturity identification model of the present invention.

[0076] 1. Dataset Construction and Preprocessing Data source: At a vegetable waste resource utilization center, the same industrial camera and imaging system (including sedimentation compensation and anti-fogging functions) as described in this invention were deployed to monitor the entire aerobic fermentation cycle of dehydrated leafy vegetables, fruit vegetables, and vine-type vegetable waste. Images of the compost surface were acquired every hour, with a resolution of 1920×1080 pixels and RGB three-channel imaging. The acquisition period was from day 0 of fermentation start to day 3 of fermentation end (after artificial chemical testing confirms maturity), resulting in 15,000 valid images.

[0077] Labeling Rules: A dual labeling strategy of "human expert labeling + chemical testing verification" was adopted. Simultaneously with each image acquisition, multiple samples were taken from the corresponding pile location within the camera's field of view. Samples were tested according to the NY / T 525-2021 organic fertilizer industry standard for key chemical indicators such as seed germination index, carbon-nitrogen ratio, and ammonium nitrogen / nitrate nitrogen ratio. The labeling rules were as follows: when the seed germination index ≥ 95% and the carbon-nitrogen ratio ≤ 18, the image was labeled "fully decomposed"; when the seed germination index ≥ 50% but < 95%, it was labeled "semi-decomposed"; and when the seed germination index < 50%, it was labeled "undecomposed". Each image was independently labeled by three qualified agricultural technicians, and the final label was determined by majority voting. Disputed samples were reviewed and removed to ensure data quality.

[0078] Data preprocessing: Divide the image pixel values ​​by 255 to normalize them to the range of 0 to 1. During training, the image is randomly flipped horizontally, randomly flipped vertically, and randomly rotated between -10° and +10° with a 50% probability to enhance the model's generalization ability.

[0079] 2. Model Structure and Parameter Settings The compost maturity recognition model adopts a multi-branch parallel convolutional neural network architecture, which contains three parallel convolutional branches to extract coarse-grained, medium-grained, and fine-grained texture features, respectively.

[0080] The first convolutional branch consists of a series of convolutional layers, pooling layers, and fully connected layers. The convolutional layers are configured with 5×5 kernels, a stride of 2, and 2 zero-padding elements, for a total of 32 kernels. These kernels are used to extract macroscopic textures such as the outlines of coarse materials in the early stages of fermentation. This branch outputs a 256-dimensional feature vector.

[0081] The second convolutional branch consists of a series of convolutional layers, pooling layers, and fully connected layers. The convolutional layers are configured with 3×3 kernels, a stride of 1, and zero-padding of 1, for a total of 64 kernels. These kernels are used to extract the mesoscopic texture of the material particles after homogenization during the mid-fermentation stage. This branch outputs a 256-dimensional feature vector.

[0082] The third convolutional branch consists of a series of convolutional layers, pooling layers, and fully connected layers. Each convolutional layer has a 1×1 kernel with a stride of 1 and zero padding, totaling 128 kernels. These kernels are used to extract the microscopic texture of white actinomycete colonies and fine humic particles at the end of fermentation. This branch outputs a 256-dimensional feature vector.

[0083] The three branches process the same input image in parallel, extract features independently, and then compress the features to a uniform 256 dimensions through a fully connected layer to obtain the first feature vector, the second feature vector, and the third feature vector.

[0084] 3. Training Process and Algorithm Training Strategy: The model employs a two-stage training method. First, it is pre-trained on the ImageNet dataset to enable the convolutional kernels to extract basic visual features. Then, the model is fine-tuned using transfer learning on the constructed vegetable waste compost image dataset, freezing the parameters of the first two convolutional layers and fine-tuning the high-level semantic features.

[0085] Loss function: The weighted cross-entropy loss function is used. Since the number of samples in the "fully decomposed" category at the end of fermentation is relatively small, it is assigned a weight coefficient of 2.0 to alleviate the class imbalance problem.

[0086] Optimization algorithm: The Adam optimizer is used, with an initial learning rate of 0.0001 and momentum parameters β1=0.9 and β2=0.999. The learning rate decays to 0.1 times its current value every 20 iterations.

[0087] Hyperparameter settings: Batch size set to 32, total training epochs 200. Early stopping is used; training stops when the validation set loss no longer decreases for 15 consecutive epochs. The dataset is randomly divided into 70% training, 15% validation, and 15% test datasets.

[0088] 4. Input / Output and Relationship Description Input: Compost surface image after aperture correction and anti-fogging heating, size 1920×1080 pixels, RGB three channels, pixel values ​​normalized to 0 to 1.

[0089] Output: The model outputs a three-dimensional vector after passing through the Softmax activation function, representing the probability values ​​of "unripe", "semi-ripe", and "fully ripe". The ripeness probability value mentioned in this invention specifically refers to the probability value of the "fully ripe" category.

[0090] The key innovation of this model lies in using core plant nutrition indicators such as seed germination index as supervision labels during the training phase, thus establishing an intrinsic correlation between the output probability value and the actual degree of decomposition. On the test set, the model achieved an AUC of 0.97 for the "fully decomposed" category, with an accuracy of 94.2% and a recall of 93.5%, validating the model's effectiveness.

[0091] 5. Implementation of Dynamic Weight Fusion During the model application phase, the fusion weights of the three branch output features are dynamically adjusted based on the fermentation process coefficient. The fermentation process coefficient is calculated based on the ratio of the current fermentation days to the historical average fermentation cycle of this type of vegetable waste, with a value ranging from 0 to 1. The weight generation rules are as follows: the weight of the coarse-grained branch is equal to 1 minus the fermentation process coefficient, the weight of the medium-grained branch is always 0.5, and the weight of the fine-grained branch is equal to the fermentation process coefficient. The three weighted feature vectors are summed element-wise to obtain the fusion feature vector, which is then input into the output layer to calculate the maturity probability value. This mechanism enables the model to adaptively adjust the focus scale as the fermentation process progresses, ensuring the accuracy of maturity determination.

[0092] Application Example 1 This method was applied to treat a batch of vine-like vegetable waste at a vegetable growing base. The batch consisted of a mixture of bean vines and cucumber vines, with an initial moisture content of 92% and an initial carbon-to-nitrogen ratio of 22:1.

[0093] During the material characteristic perception stage, the AI ​​control unit identifies the category label as vine-like through physical correction and multi-view image analysis, calculating the fiber entanglement index as 0.45. Based on the category label, it retrieves the grinding particle size of 2-5 mm and the spindle speed baseline value of 30 r / min. Based on the entanglement index of 0.45, it linearly adjusts the speed downwards, and the final target speed is calculated as 30 multiplied by 1 minus 0.45, which equals 16.5 r / min, not lower than the safety threshold of 10 r / min. The actual setting is 17 r / min.

[0094] Sampling and testing after dehydration: Total nitrogen content was 1.82% (dry basis) determined by Kjeldahl method, and organic carbon content was 32.5% determined by potassium dichromate titration method, with a calculated carbon-to-nitrogen ratio of 17.9:1. Compared with before dehydration, nitrogen retention rate was 89%, potassium retention rate was 93%, and moisture content was reduced to 63%, meeting the optimal fermentation moisture content requirements for plant nutrition.

[0095] During aerobic fermentation, the AI ​​control unit automatically replenishes corn stalks based on the initial carbon-to-nitrogen ratio, adjusting it to 28:1. The ventilation volume of the fermentation unit is controlled at 0.2 m³ / s. 3 / (m 3 The temperature was maintained at 58-62℃. On the 18th day of fermentation, the maturity recognition model output a probability value greater than 0.95 for the first three consecutive times, triggering confirmatory sampling. According to NY / T 525-2021 standards, the seed germination index (GI) was 98%, the carbon-to-nitrogen ratio was 16.5:1, and the ammonium nitrogen / nitrate nitrogen ratio was 0.8, all meeting the standards for organic fertilizer. The AI ​​control unit determined that the fermentation endpoint had been reached and automatically started discharging.

[0096] Compared with traditional fixed-parameter processing methods, this method did not cause fiber entanglement shutdowns during the batch processing, shortened the dehydration time by 20%, increased the nutrient retention rate by 25%, shortened the fermentation cycle from 25 days to 18 days, and ensured stable and qualified product quality.

[0097] Verification Example 1: Experiment verifying the correlation between the probability value of decomposition and the seed germination index Based on the method of the present invention, this embodiment conducts a verification experiment at a vegetable waste resource utilization center to establish a quantitative correlation between the output probability value of the compost maturity identification model and the actual seed germination index.

[0098] 1. Experimental Design Experimental subject: A mixture of bean vines and cucumber vines was selected as the experimental material, with a batch size of 5 tons. The initial moisture content of the material was 92%, and the initial carbon-to-nitrogen ratio was 22:1.

[0099] Experimental equipment: A visual acquisition device and AI control unit identical to those described in this invention were deployed. A third industrial camera and a laser rangefinder were fixedly installed above the aerobic fermentation device, with camera aperture adjustment and lens anti-fog functions enabled throughout the process. The fermentation device was equipped with a forced ventilation system, with the ventilation volume controlled at 0.2 m³ / s. 3 / (m 3 (min), the temperature is maintained at 55-65℃.

[0100] Experimental period: from day 0 of fermentation to day 25 (exceeding the historical average fermentation period of 22 days, in order to observe the state of fermentation).

[0101] 2. Data Acquisition Methods Image Acquisition: The AI ​​control unit triggers the third industrial camera to acquire images of the compost surface twice daily (9:00 AM and 3:00 PM). Before each acquisition, the laser rangefinder automatically measures the object distance, and the AI ​​control unit dynamically adjusts the aperture to the appropriate setting to ensure image clarity. The lens heating function is on throughout the process to prevent fogging. A total of 50 frames are acquired, each with a resolution of 1920×1080 pixels and RGB three-channel resolution.

[0102] Simultaneous sampling: Immediately after each image acquisition, multiple points are sampled from the pile area corresponding to the center of the camera's field of view, with a sampling depth of 10 cm to 20 cm below the surface. The samples are mixed and divided into two portions: one portion is used immediately for seed germination index determination, and the other portion is stored for future reference.

[0103] 3. Methods for determining seed germination index Seed germination index was determined according to the NY / T 525-2021 organic fertilizer industry standard: Test seeds: Chinese cabbage seeds (Brassica chinensis), germination rate ≥95% Preparation of extract: Freshly taken compost sample was added to deionized water at a mass ratio of 1:10, shaken for 1 hour, and then filtered. Culture conditions: Place filter paper in a petri dish, add 5 mL of extraction solution, evenly distribute 20 seeds, and incubate at 25℃ for 48 hours. Control treatment: Deionized water was used instead of the extract as a blank control. Calculation formula: Germination Index (GI) = (Germination rate of treatment group × Root length of treatment group) / (Germination rate of control group × Root length of control group) × 100% Each sample was replicated in 3 places, and the average value was taken as the GI value at that time.

[0104] 4. Method for obtaining the probability value of decay. Each collected compost surface image is input into a multi-branch parallel convolutional neural network model. The model is constructed according to the training method described in Example 1 above, and has been intrinsically correlated with the seed germination index during the training phase. The model output, after being calculated using the Softmax function, yields a maturity probability value between 0 and 1, which is recorded as the output probability at that moment.

[0105] 5. Experimental Data Recording The experiment lasted for 25 days, and the data at key time points are shown in the table below: Note: On day 18, the probability value of maturity was greater than 0.95 for the first time three times in a row, corresponding to a GI value of 98%; on day 25, the probability value was still close to 1, but the GI value decreased slightly, indicating that over-fermentation may lead to nutrient loss. This shows the scientific nature of the preset conditions of this invention: when the value is greater than 0.95 for three consecutive times, verification sampling is triggered, and the measured GI values ​​at the corresponding times are all ≥95%, indicating that the triggering condition is reliable and can ensure the accuracy of the discharge decision.

[0106] 6. Data Analysis and Correlation Verification Linear regression analysis was performed on the above data points to obtain the regression equation between the probability value of decomposition x and the seed germination index GI y: y = 95x + 5, with a coefficient of determination R. 2 = 0.98, indicating a strong positive correlation between the two.

[0107] When the probability value of decomposition reaches 0.95, the corresponding GI value calculated by the regression equation is 95 × 0.95 + 5 = 95.25%, which is close to the measured value of 98%. When the probability value is greater than 0.95 three times consecutively, the corresponding measured GI values ​​are all ≥95%, verifying the scientific validity and reliability of the preset conditions of this invention.

[0108] This verification experiment shows that the output probability value of the compost maturity identification model constructed using the method of this invention has a significant linear positive correlation with the seed germination index measured according to the NY / T 525-2021 standard. It can be seen that the model can accurately reflect the plant nutrient status of compost through visual features, providing a reliable basis for the automatic determination of the fermentation endpoint.

[0109] Although the embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details.

Claims

1. A method for processing waste vegetables from facilities using AI-controlled intelligent regulation, characterized in that, Includes the following steps: 1) Intelligent perception of material characteristics: Multi-view images of the waste vegetable material are acquired through a visual acquisition device; the AI ​​control unit calculates the fiber entanglement index, which comprehensively reflects the fiber characteristics and entanglement risk of the material, based on the multi-view images and preset physical correction actions, and identifies the category label of the waste vegetable material. 2) Differentiated adaptive control of dehydration process: The AI ​​control unit differentiates and controls the process parameters of the crushing and pressing devices according to the category labels and fiber entanglement index; 3) Determination of fermentation endpoint based on plant nutrient indicators: The dehydrated vegetable waste from step 2) is transported to an aerobic fermentation device. Images of the compost surface are collected by a visual acquisition device and input into a maturity recognition model that is constructed based on core plant nutrient indicators of compost and can be dynamically adjusted with the fermentation process. The model outputs a maturity probability value. When the probability value meets the preset conditions, the AI ​​control unit determines that the fermentation endpoint has been reached and initiates the discharge command.

2. The method for processing facility waste vegetables using AI intelligent control according to claim 1, characterized in that, The category labels include leafy vegetables, fruit vegetables, or vines; the fiber entanglement index is a dimensionless value between 0 and 1, which is calculated by the AI ​​control unit based on multi-view images by fusing the area ratio of fibrous materials in the images and the degree of out-of-surface warping.

3. The method for processing facility waste vegetables using AI intelligent control according to claim 2, characterized in that, The process parameters for the differentiated control of the pulverizing and pressing devices mentioned in step 2) specifically include: Set the crushing particle size range and the reference value of the main shaft speed of the pressing device according to the category label; The spindle speed reference value is adjusted proportionally according to the fiber winding index. The higher the fiber winding index, the greater the reduction in spindle speed, and the reduced spindle speed is not lower than the preset safe operating threshold.

4. The method for processing facility waste vegetables using AI intelligent control according to claim 3, characterized in that, If the category label is leafy vegetables, the particle size should be controlled at 10-25mm, and the spindle speed should be set at 30-50 r / min. If the category label is fruit vegetables, the particle size should be controlled at 10-25mm, and the spindle speed should be set at 20-40 r / min. If the category label is vines, the particle size should be controlled at 2-5mm, and the spindle speed should be set at 20-40 r / min. The proportional reduction rule is as follows: for every 0.1 increase in the fiber winding index, the final target spindle speed should be reduced by 10% based on the corresponding category's benchmark value, and the final spindle speed should not be lower than 10 r / min.

5. The method for processing facility waste vegetables using AI intelligent control according to claim 1, characterized in that, The visual acquisition device includes a first industrial camera deployed above the conveying device, a second industrial camera deployed on one side of the conveying device, a third industrial camera fixed above the aerobic fermentation device, and a laser rangefinder sensor; the measurement spot of the laser rangefinder sensor is matched with the projection position of the center point of the field of view of the third industrial camera, and is used to monitor the change in object distance caused by the settlement of the pile in real time.

6. The method for processing facility waste vegetables using AI intelligent control according to claim 1, characterized in that, The maturity identification model described in step 3) adopts a multi-branch parallel convolutional neural network architecture that supports dynamic weight fusion, including a first convolutional branch, a second convolutional branch and a third convolutional branch, which are used to extract coarse-grained, medium-grained and fine-grained texture features of the compost surface, respectively. After the collected compost surface image is input into the model, features are extracted and fused through each branch, and finally the maturity probability value is output through the Softmax function.

7. The method for processing facility waste vegetables using AI intelligent control according to claim 6, characterized in that, The preset conditions mentioned in step 3) include: when the probability value of compost maturity corresponding to three consecutive collected compost surface images is greater than 0.95, the AI ​​control unit performs verification sampling; if the compost seed germination index GI of the sampled sample is ≥95%, the fermentation endpoint is determined to have been reached and the discharge command is initiated.

8. The method for processing facility waste vegetables using AI intelligent control according to claim 7, characterized in that, The maturity recognition model dynamically adjusts the fusion weights of the output features of each convolutional branch based on the fermentation process coefficient; The fermentation process coefficient is calculated based on the ratio of the current fermentation days to the historical average fermentation cycle of this type of vegetable waste, and the value range is 0 to 1; the historical average fermentation cycle is obtained by taking the arithmetic mean of the fermentation days of no less than 30 historical batches when they are determined to be mature by the method described in claim 7; the weight adjustment rule is: the output feature weight of the coarse-grained branch is higher in the early stage of fermentation, and the output feature weight of the fine-grained branch is higher in the later stage of fermentation.

9. The method for processing facility waste vegetables using AI intelligent control according to claim 1, characterized in that, Step 3) also includes steps for pile settlement compensation and lens anti-fog correction: The AI ​​control unit dynamically adjusts the aperture coefficient of the third industrial camera based on the object distance value obtained in real time by the laser rangefinder, so that the compost surface is always within the depth of field. The AI ​​control unit calculates the dew point temperature based on the lens surface temperature and humidity data, and controls the lens surface temperature to be 3-8°C higher than the dew point temperature through a ring-shaped electric heating coil to prevent the lens from fogging.

10. The method for processing facility waste vegetables using AI intelligent control according to claim 1, characterized in that, The preset physical correction action mentioned in step 1) includes: when the area ratio of fibrous material detected by the image exceeds the preset risk threshold, the AI ​​control unit pauses the conveying device and drives the horizontal pressure plate to press down the material. After compaction and smoothing, the corrected multi-view image is collected to calculate the fiber entanglement index.

Images