Kitchen waste pretreatment intelligent classification treatment method based on AI identification

Abstract

The invention discloses a kitchen waste pretreatment intelligent classification treatment method based on AI identification, and relates to the technical field of building early warning, and the method comprises the following steps: synchronously collecting three types of data of a color image, a near infrared spectrum and surface conductivity of waste on a conveyor belt; performing space-time alignment and data preprocessing; an AI model is used for identifying organic substances in the images, spectral salinity characteristics and conductivity data of corresponding positions are fused and analyzed, and the probability that each organic fragment belongs to the high-salinity garbage is comprehensively judged; the occurrence density of the high-salinity garbage is calculated in real time, and a sorting mode is dynamically decided according to the density; and the mechanical arm executes the sorting action according to the instruction, the high-salt garbage is transferred to a special collecting bin, the sorting effect is verified through a downstream sensor, and meanwhile the system continuously records operation data for optimization. Through multi-mode sensing data fusion and dynamic regulation and control of a dynamic sorting strategy, the organic garbage with the same appearance and the standard-exceeding salinity can be recognized and sorted out.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of building early warning, and particularly relates to an AI recognition-based intelligent classification processing method for kitchen waste pretreatment. BACKGROUND

[0002] Kitchen waste mainly refers to waste food residues and waste edible oil generated in food processing, food service, unit meal service and other activities of the catering industry, and has the characteristics of high organic matter content, high moisture content, easy rotting and stench, and complex and variable composition. If not properly treated, it will not only cause environmental pollution, but also lead to resource waste. Typical components include rice, vegetables, peels, meat, bones and a small amount of inorganic substances such as plastic and paper towels, which bring significant challenges to efficient classification and resource utilization.

[0003] The existing AI recognition-based intelligent classification technology for kitchen waste relies on machine vision to identify and sort the shape, color and texture features of the waste. However, this method has an inherent limitation, which is that it cannot effectively identify the chemical properties of the waste that are not visible inside. In actual application scenarios, especially when dealing with mixed kitchen waste from diversified catering formats, the waste stream often contains a large amount of high-salt substances, which have no significant difference in appearance from ordinary organic waste, making it difficult for pure visual recognition algorithms to accurately distinguish them. If these high-salt waste is mistakenly mixed into the subsequent anaerobic fermentation system, it will continuously inhibit microbial activity and severely damage the stable operation and resource efficiency of the treatment system. Therefore, the following solution is proposed to address the above problems. SUMMARY

[0004] The purpose of the present application is to provide an AI recognition-based intelligent classification processing method for kitchen waste pretreatment, which can identify and sort out organic waste with no difference in appearance but excessive salt content through multi-modal sensor data fusion and dynamic sorting strategy dynamic regulation, solving the problem that the existing machine vision-based recognition method cannot perceive the internal chemical properties of the waste, leading to the mixing of high-salt substances into the downstream biochemical treatment system and severely inhibiting microbial activity.

[0005] To solve the above technical problems, the present application is realized by the following technical scheme:

[0006] The present application is an AI recognition-based intelligent classification processing method for kitchen waste pretreatment, which specifically includes the following steps:

[0007] Step S1, data acquisition: simultaneously acquiring color images, near-infrared spectra and surface conductivity of the waste on the conveyor belt;

[0008] Step S2, data preprocessing: performing spatiotemporal alignment and data preprocessing to provide standardized multi-modal information for subsequent analysis;

[0009] Step S3, feature extraction and fusion recognition: using AI model to recognize organic matter in the image, and fusing and analyzing the spectral salt feature and conductivity data of the corresponding position to comprehensively judge the probability of each organic fragment belonging to high-salt waste;

[0010] Step S4, dynamic sorting decision: real-time calculation of the occurrence density of high-salt waste, and dynamic decision of sorting mode according to the density;

[0011] Step S5, execute sorting: the mechanical arm executes the sorting action according to the instruction, moves the high-salt waste to the special collection bin, and verifies the sorting effect by the downstream sensor, while the system continuously records the operation data for optimization.

[0012] Further, the step S1, data acquisition specifically includes the following steps:

[0013] Step S11: after the kitchen waste is pretreated by crushing and homogenization, it is transmitted by a conveyor belt with a constant width W at a constant speed V, forming a uniform waste thin layer;

[0014] Step S12: a high-speed industrial camera (visual sensor V) is fixedly installed above the conveyor belt, and the frame rate F v The RGB color image of the waste is collected; the line-scan near-infrared spectrometer (NIR sensor) is triggered synchronously, the scanning line direction is parallel to the width direction of the conveyor belt, and the spectral reflectivity data of the waste in the wavelength range λ1 to λ n (For example, 900nm-1700nm) is collected; a non-contact multi-channel conductivity sensor (C sensor) is installed behind the spectrometer scanning line to measure the conductivity distribution of the waste surface.

[0015] Further, the step S2, data preprocessing specifically includes the following steps:

[0016] Step S21: establish a unified space-time coordinate system; define the running direction of the conveyor belt as the time axis t and the direction perpendicular to the conveyor belt as the space axis x; label each frame of visual image, each spectral scanning line and each conductivity measurement data with a unified time stamp and space position label (x, t) to ensure that the three types of data are completely synchronized and aligned in space-time;

[0017] Step S22: color correction and distortion correction are performed on the visual image data; near-infrared spectral data are smoothed (Savitzky-Golay filtering) and standard normal variable transformation (SNV) is performed to eliminate scattering effects; temperature compensation and calibration are performed on the conductivity data to convert them into standard conductivity values σ(x, t).

[0018] Furthermore, step S3, feature extraction and fusion recognition, specifically includes the following steps:

[0019] Step S31: Perform semantic segmentation on the preprocessed visual image data. Use a pre-trained deep learning model (such as a U-Net variant) to identify and segment organic matter fragments (such as vegetable leaves, fruit peels, and meat residue) in the image, while ignoring inorganic matter (such as plastic and tableware) and a large amount of free water. The model outputs a binary mask matrix M. org (x,t), where a value of 1 represents that the pixel position (x,t) is organic matter;

[0020] Step S32: Extract features from near-infrared spectral data; for each spectral curve I(λ,x,t) acquired at position (x,t), calculate its correlation coefficient with the characteristic spectra of standard carbohydrates, proteins, fats, and water; simultaneously calculate its correlation coefficient in the specific salt-sensitive band λ. salt The absorption depth D near (e.g., ≈1450nm, ≈1950nm) salt (x,t):

[0021]

[0022] In the formula, D salt (x,t) represents the salt absorption depth measured at spatial position x and time point t on the conveyor belt, where λ is the wavelength. salt Here, λ is the center wavelength of the salt-sensitive band, Δλ is the wavelength bandwidth, and I(λ,x,t) is the spectral intensity or reflectance value measured by the near-infrared spectral sensor at wavelength λ, spatial location x, and time t. baseline (x,t) represents the wavelength at the baseline λ. baseline Spectral intensity measured at spatial location x and time t;

[0023] Step S33: Mask the obtained organic material M org (x,t) serves as a spatial filter, allowing subsequent analysis only on regions marked as organic matter; the absorption depth D at the corresponding spatial location is then used. salt Feature-level fusion of (x,t) and conductivity data σ(x,t);

[0024] Step S34: Combine the fused feature vector [D] salt [σ] is input to a lightweight machine learning classifier (such as a support vector machine, SVM), which has been trained using sample data with known salt concentrations; the classifier outputs a probability value P. salt (x,t)∈[0,1] represents the probability that the organic debris at position (x,t) belongs to high-salinity waste; a probability threshold T is set. h (e.g., 0.85), when Psalt (x,t)>T h At that time, it was determined that there were high-salinity target fragments at that location.

[0025] Furthermore, step S4, the dynamic sorting decision, specifically includes the following steps:

[0026] Step S41: Obtain the location information (x) of all high-salinity target fragments identified in step S3. i ,t i The spatial extent of the target fragment is mapped from a two-dimensional spatiotemporal coordinate system to a three-dimensional robotic arm coordinate system; combined with the conveyor belt speed V, the exact time for each target fragment to arrive at the working area of ​​the sorting robotic arm is calculated. and center coordinates (X) i ,Y i );

[0027] Step S42: Monitor the number N of high-salinity target fragments identified per unit time in real time, and calculate the dynamic response factor α:

[0028]

[0029] In the formula, N represents the number of high-salinity targets in real time. max This represents the system's maximum processing capacity.

[0030] When α is lower than the threshold α low When α is 0.3, the current waste stream is determined to be in a low-salt load state, and the robotic arm adopts a precise point-grabbing mode, only grabbing the identified independent high-salt fragments; when α is higher than the threshold α high When the value is 0.7, it is determined that the current state is a high salt load, indicating that the content of high salt substances in the garbage flow is large and the distribution is dense. At this time, the system controls the robotic arm to switch to the area scraping mode. Its movement amplitude and speed are positively correlated with the α value. That is, the robotic arm will use a larger movement amplitude and a faster speed to scrape the garbage group suspected of containing high salt in the target area to the removal channel to ensure processing efficiency and avoid omissions.

[0031] Step S43: Based on the decision in step S42, generate the corresponding robotic arm motion trajectory and end effector (such as suction cup or gripper) control instructions; the instructions include target coordinates, motion type (grabbing / scraping) and motion force / amplitude parameters.

[0032] Furthermore, step S5, performing sorting, specifically includes the following steps:

[0033] Step S51: The sorting robot arm receives the control command and completes the sorting at a predetermined time. Reach the target location (X) i ,Y iIt performs precise grabbing or area sweeping actions according to instructions, transferring high-salt waste fragments or waste groups to a dedicated high-salt waste collection bin;

[0034] Step S52: After the sorting action is performed, a verification sensor (a simplified spectral sensor can be used) is installed downstream of the conveyor belt to perform a second rapid scan of the garbage in the sorted area to verify whether the high-salt substances have been effectively removed.

[0035] Step S53: The system continuously records the total amount of waste processed in each shift, the amount of high-salt waste identified, the sorting success rate, and the fluctuation of salt load status (α value), forming a data log for optimizing the model and adjusting operating parameters.

[0036] The present invention has the following beneficial effects:

[0037] 1. This invention, through the introduction of near-infrared spectroscopy analysis and conductivity measurement, can penetrate surface morphology and directly detect the internal chemical composition characteristics of materials. This identification mechanism based on the essential properties of materials greatly reduces the over-reliance on complex surface textures, colors, and shape features, thereby effectively avoiding misjudgments and omissions caused by visual obstruction, adhesion, or similar colors of foreign objects, and achieving high-precision and high-reliability identification and sorting of high-salt organic waste.

[0038] 2. This invention introduces a dynamic response mechanism based on real-time load, enabling the system to intelligently sense changes in the composition of the waste stream and automatically adjust the sorting strategy. When facing low-concentration target materials, a precise grasping mode is adopted to ensure sorting purity. When facing high-concentration, high-density target materials, it automatically switches to a zone scraping mode to prioritize processing throughput. This adaptive capability allows the system to maintain efficient operation under different working conditions, optimizes overall resource allocation and energy consumption, and solves the problem of low efficiency of fixed sorting modes when dealing with food waste with drastic compositional fluctuations.

[0039] 3. This invention improves the purity of the separated material, ensuring that the high-salt waste that was previously rejected is effectively separated. This keeps the salt concentration of the material entering the subsequent mainstream anaerobic fermentation and other biological treatment stages stable below the safe threshold where microbial activity is not inhibited. This provides a consistent and reliable feed guarantee for the downstream core treatment processes, thereby improving the biogas production efficiency and operational stability, reducing the risk of process failures and maintenance costs caused by fluctuations in feed quality, and ensuring the continuity, stability and efficiency of the entire food waste resource utilization chain.

[0040] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0041] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 This is a schematic diagram of the process of the intelligent sorting and treatment method for pretreatment of kitchen waste based on AI recognition according to the present invention. Detailed Implementation

[0043] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] Please see Figure 1 As shown, this invention is an AI-based intelligent sorting and treatment method for pre-treatment of kitchen waste, comprising the following steps:

[0045] Step S1, Data Acquisition: Simultaneously acquire three types of data: color images, near-infrared spectra, and surface conductivity of the waste on the conveyor belt;

[0046] Step S1, data collection specifically includes the following steps:

[0047] Step S11: After being crushed and homogenized, the kitchen waste is transported at a constant speed V through a conveyor belt with a constant width W, forming a thin layer of waste with uniform thickness.

[0048] Step S12: Fix a high-speed industrial camera (vision sensor V) directly above the conveyor belt, at a frame rate F. v Collect RGB color images of the waste; simultaneously trigger a line-scanning near-infrared spectrometer (NIR sensor), with the scan line direction parallel to the width direction of the conveyor belt, collecting images of the waste in the wavelength range λ1 to λ2. n Spectral reflectance data in the range of 900nm to 1700nm; a non-contact multi-channel conductivity sensor (C sensor) is installed behind the spectrometer scanning line to measure the conductivity distribution on the surface of the waste.

[0049] Step S2, Data Preprocessing: Perform spatiotemporal alignment and data preprocessing to provide standardized multimodal information for subsequent analysis;

[0050] Step S2, data preprocessing specifically includes the following steps:

[0051] Step S21: Establish a unified spatiotemporal coordinate system; define the direction of the conveyor belt as the time axis t, and the direction perpendicular to the conveyor belt as the spatial axis x; assign a unified timestamp and spatial location label (x,t) to each frame of visual image, each spectral scan line, and each conductivity measurement data to ensure that the three types of data are completely synchronized and aligned in spatiotemporal space.

[0052] Step S22: Perform color correction and distortion correction on the visual image data; perform smoothing (Savitzky-Golay filtering) and standard normal variable transformation (SNV) on the near-infrared spectral data to eliminate scattering effects; perform temperature compensation and calibration on the conductivity data, converting it into a standard conductivity value σ(x,t).

[0053] Step S3, Feature Extraction and Fusion Recognition: Use an AI model to identify organic matter in the image, and fuse and analyze the spectral salinity features and conductivity data of the corresponding locations to comprehensively determine the probability that each organic fragment belongs to high-salt waste;

[0054] Step S3, feature extraction and fusion recognition, specifically includes the following steps:

[0055] Step S31: Perform semantic segmentation on the preprocessed visual image data. Use a pre-trained deep learning model (such as a U-Net variant) to identify and segment organic matter fragments (such as vegetable leaves, fruit peels, and meat residue) in the image, while ignoring inorganic matter (such as plastic and tableware) and a large amount of free water. The model outputs a binary mask matrix M. org (x,t), where a value of 1 represents that the pixel position (x,t) is organic matter;

[0056] Step S32: Extract features from near-infrared spectral data; for each spectral curve I(λ,x,t) acquired at position (x,t), calculate its correlation coefficient with the characteristic spectra of standard carbohydrates, proteins, fats, and water; simultaneously calculate its correlation coefficient in the specific salt-sensitive band λ. salt The absorption depth D near (e.g., ≈1450nm, ≈1950nm) salt (x,t):

[0057]

[0058] In the formula, D salt (x,t) represents the salt absorption depth measured at spatial position x and time point t on the conveyor belt, where λ is the wavelength. salt Here, λ is the center wavelength of the salt-sensitive band, Δλ is the wavelength bandwidth, and I(λ,x,t) is the spectral intensity or reflectance value measured by the near-infrared spectral sensor at wavelength λ, spatial location x, and time t. baseline(x,t) represents the wavelength at the baseline λ. baseline Spectral intensity measured at spatial location x and time t;

[0059] Step S33: Mask the obtained organic material M org (x,t) serves as a spatial filter, allowing subsequent analysis only on regions marked as organic matter; the absorption depth D at the corresponding spatial location is then used. salt Feature-level fusion of (x,t) and conductivity data σ(x,t);

[0060] Step S34: Combine the fused feature vector [D] salt [σ] is input to a lightweight machine learning classifier (such as a support vector machine, SVM), which has been trained using sample data with known salt concentrations; the classifier outputs a probability value P. salt (x,t)∈[0,1] represents the probability that the organic debris at position (x,t) belongs to high-salinity waste; a probability threshold T is set. h (e.g., 0.85), when P salt (x,t)>T h At that time, it was determined that there were high-salinity target fragments at that location.

[0061] Step S4, Dynamic Sorting Decision: Calculate the density of high-salt waste in real time and dynamically decide the sorting mode based on this density;

[0062] Step S4, the dynamic sorting decision specifically includes the following steps:

[0063] Step S41: Obtain the location information (x) of all high-salinity target fragments identified in step S3. i ,t i The spatial extent of the target fragment is mapped from a two-dimensional spatiotemporal coordinate system to a three-dimensional robotic arm coordinate system; combined with the conveyor belt speed V, the exact time for each target fragment to arrive at the working area of ​​the sorting robotic arm is calculated. and center coordinates (X) i ,Y i );

[0064] Step S42: Monitor the number N of high-salinity target fragments identified per unit time in real time, and calculate the dynamic response factor α:

[0065]

[0066] In the formula, N represents the number of high-salinity targets in real time. max This represents the system's maximum processing capacity.

[0067] When α is lower than the threshold α lowWhen α is 0.3, the current waste stream is determined to be in a low-salt load state, and the robotic arm adopts a precise point-grabbing mode, only grabbing the identified independent high-salt fragments; when α is higher than the threshold α high When the value is 0.7, it is determined that the current state is a high salt load, indicating that the content of high salt substances in the garbage flow is large and the distribution is dense. At this time, the system controls the robotic arm to switch to the area scraping mode. Its movement amplitude and speed are positively correlated with the α value. That is, the robotic arm will use a larger movement amplitude and a faster speed to scrape the garbage group suspected of containing high salt in the target area to the removal channel to ensure processing efficiency and avoid omissions.

[0068] Step S43: Based on the decision in step S42, generate the corresponding robotic arm motion trajectory and end effector (such as suction cup or gripper) control instructions; the instructions include target coordinates, motion type (grabbing / scraping) and motion force / amplitude parameters.

[0069] Step S5: Perform sorting: The robotic arm performs sorting actions according to instructions, transferring high-salt waste to a dedicated collection bin, and the sorting effect is verified by downstream sensors. At the same time, the system continuously records operating data for optimization.

[0070] Step S5, the sorting process specifically includes the following steps:

[0071] Step S51: The sorting robot arm receives the control command and completes the sorting at a predetermined time. Reach the target location (X) i ,Y i It performs precise grabbing or area sweeping actions according to instructions, transferring high-salt waste fragments or waste groups to a dedicated high-salt waste collection bin;

[0072] Step S52: After the sorting action is performed, a verification sensor (a simplified spectral sensor can be used) is installed downstream of the conveyor belt to perform a second rapid scan of the garbage in the sorted area to verify whether the high-salt substances have been effectively removed.

[0073] Step S53: The system continuously records the total amount of waste processed in each shift, the amount of high-salt waste identified, the sorting success rate, and the fluctuation of salt load status (α value), forming a data log for optimizing the model and adjusting operating parameters.

[0074] One specific application of this embodiment is:

[0075] Implementation Background:

[0076] This facility is located on the second basement level of a large urban commercial complex and serves as a centralized pre-treatment center for food waste. The center handles food waste from over 50 restaurants within the complex, including 8 Sichuan and Hunan restaurants, 5 hot pot restaurants, 3 Western restaurants, 2 Japanese restaurants, and several bakeries and fast food outlets. It processes approximately 20 tons of waste daily. The waste is characterized by its complex composition, with salt (NaCl) concentration fluctuating wildly between 0.5% and 8% due to waste from hot pot broth, pickles, sauces, and baking ingredients, threatening the stable operation of the subsequent anaerobic digester system.

[0077] System hardware configuration:

[0078] Conveying system: One stainless steel conveyor belt, width W = 800mm, running speed V = 0.2m / s;

[0079] Sensing system:

[0080] Vision sensor (V): A 2-megapixel high-speed industrial camera with a frame rate of F... v =60fps, equipped with a ring LED light source to reduce glare;

[0081] Near-infrared spectral sensor (NIR): A line-scan hyperspectral camera with a spectral range of 900-1700nm, a spectral resolution of 10nm, and a scanning frequency synchronized with the visual frame rate;

[0082] Conductivity sensor (C): A non-contact multi-channel conductivity measurement array based on the eddy current principle, with a measurement range of 0-100 mS / cm and an accuracy of ±0.5%, consisting of 16 measurement channels arranged along the width of the conveyor belt;

[0083] Processing unit: An industrial computer equipped with an NVIDIA Jetson AGX Orin module, used to run deep learning models and fusion algorithms;

[0084] Actuator: A six-axis high-speed industrial robot with a load capacity of 5kg and a repeatability of ±0.1mm. The end effector is a flexible vacuum suction cup device that combines gripping and scraping functions.

[0085] Implementation steps:

[0086] S1. Data Synchronization Acquisition and Preprocessing:

[0087] After being crushed and initially dehydrated, the kitchen waste is spread on a conveyor belt, forming a material layer approximately 30mm thick. The conveyor belt passes through the sensing area at a uniform speed of 0.2m / s. An industrial camera, a hyperspectral camera, and a conductivity sensor are rigidly fixed on the same crossbeam and synchronized at the microsecond level via a hardware trigger signal. The system assigns a uniform timestamp and spatial location label (x,t) to each frame of image, each spectral scan line, and each set of conductivity data. The visual image undergoes white balance correction, the hyperspectral data undergoes SNV preprocessing, and the conductivity data undergoes 25℃ standard temperature compensation to obtain the standard conductivity value σ(x,t).

[0088] S2. High-salinity target identification and localization:

[0089] Organic matter segmentation: The preprocessed visual image is fed into a trained DeepLabV3+ semantic segmentation model; this model can accurately separate organic fragments such as lettuce leaves, chili peels, and meat scraps from inorganic materials such as plastic bags, napkins, and disposable tableware; the model outputs a binary mask matrix M with the same spatial resolution as the image. org (x,t), where pixels with a value of 1 are marked as organic matter;

[0090] Spectral feature extraction: For hyperspectral data, the system calculates the spectral feature at λ for each spatial location (x,t) of the spectral curve. salt The absorption depth D in the band [1430nm, 1470nm] near 1450nm (the characteristic absorption peak of the mixture of water and salt) is... salt (x,t);

[0091]

[0092] Feature fusion and classification: Integrating organic mask M org (x,t) serves as a spatial filter, performing fusion judgment only on pixels marked as organic matter; the feature vectors corresponding to these points [D] are then used. salt [σ] is input to a pre-trained Support Vector Machine (SVM) classifier; the classifier outputs the probability P that this location is high-salt waste. salt (x,t); Set the threshold T h =0.82, when P salt When (x,t)>0.82, the system determines that there is a high-salt target at that location (such as a piece of vegetable leaf covered in hot pot broth) and records its center coordinates (x,t). i ,t i ) and general outline;

[0093] S3. Dynamic Decision-Making and Adaptive Sorting Control:

[0094] Coordinate mapping and prediction: The system will map and predict the spatial coordinates (x, y) of all identified high-salinity targets. i ,t i Based on the conveyor belt speed V = 0.2 m / s, the time it takes for the conveyor belt to reach the working area of ​​the robotic arm is accurately calculated. and three-dimensional spatial coordinates (X i ,Y i );

[0095] Dynamic strategy selection: The system counts the number N of newly identified high-salt targets in a 1-second window; in this embodiment, the system is designed with a maximum processing capacity of N. max =25 per second;

[0096] Scenario A (Low Salt Load): At a certain moment, the system counts N=6 targets within the past second; the dynamic response factor α is calculated to be 6 / 25 = 0.24; this value is lower than the preset threshold α. low =0.3, the system determines that it is currently in a low-salt load state; the control strategy is precise point grasping, and the robotic arm will move sequentially to each target coordinate (X). i ,Y i The vacuum suction cup is driven to accurately grasp the identified individual fragments with relatively low power (-60kPa);

[0097] Scenario B (High Salt Load): Ten minutes later, due to the centralized processing of garbage from the hot pot restaurant, the system detected N = 20 targets / second; the calculated α = 20 / 25 = 0.8, which is higher than the preset threshold α. high =0.7, the system determines that it has entered a high salt load state; the control strategy immediately switches to area scraping; the movement trajectory of the robotic arm is no longer discrete points, but connects multiple dense target points into a continuous area; the power of the vacuum suction cup is increased to -85kPa, the coverage area is expanded, and the robotic arm performs a large-area scraping action at a faster speed to remove the high salt content substances (such as mixed chili and oil) adhering to an area as a whole.

[0098] Command generation and transmission: Based on the above decisions, the system generates a set of control commands including target coordinates, motion mode (point gripping / scraping), suction cup power and motion amplitude, and sends them to the robot controller via Ethernet;

[0099] S4. Sorting and System Feedback:

[0100] The robot controller receives instructions and precisely controls the robotic arm and end effector to complete the specified sorting actions. The high-salt waste that is grabbed or swept out is sent to a special collection bin labeled "High-Salt Waste" and will then undergo specialized low-load fermentation or safe landfill treatment. About 1 meter behind the sorting arm, a simplified single-point spectral sensor is installed to quickly scan the surface of the waste that has just been sorted. If significant high-salt characteristics are still detected, the system will record a sorting omission event. This data will be used to periodically optimize the AI ​​model and classification threshold. The system's human-machine interface displays the current salt load status (α value), sorting efficiency, and historical data trends of the waste flow in real time, providing operators with intuitive monitoring of the system's operating status.

[0101] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0102] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. An AI-based intelligent sorting and treatment method for pre-treatment of kitchen waste, characterized in that: The processing method specifically includes the following steps: Step S1, Data Acquisition: Simultaneously acquire three types of data: color images, near-infrared spectra, and surface conductivity of the waste on the conveyor belt; Step S2, Data Preprocessing: Perform spatiotemporal alignment and data preprocessing to provide standardized multimodal information for subsequent analysis; Step S3, Feature Extraction and Fusion Recognition: Use an AI model to identify organic matter in the image, and fuse and analyze the spectral salinity features and conductivity data of the corresponding locations to comprehensively determine the probability that each organic fragment belongs to high-salt waste; Step S4, Dynamic Sorting Decision: Calculate the density of high-salt waste in real time and dynamically decide the sorting mode based on this density; Step S5: Perform sorting: The robotic arm performs sorting actions according to instructions, transferring high-salt waste to a dedicated collection bin. Downstream sensors verify the sorting effect, and the system continuously records operating data for optimization.

2. The AI-based intelligent sorting and treatment method for pretreatment of kitchen waste according to claim 1, characterized in that, Step S1, data acquisition, specifically includes the following steps: Step S11: After being crushed and homogenized, the kitchen waste is transported at a constant speed V through a conveyor belt with a constant width W, forming a thin layer of waste with uniform thickness. Step S12: Fix a high-speed industrial camera directly above the conveyor belt, at a frame rate of F... v Collect RGB color images of the waste; simultaneously trigger a line-scanning near-infrared spectrometer, with the scan line direction parallel to the width direction of the conveyor belt, to collect images of the waste in the wavelength range λ1 to λ2. n Spectral reflectance data within the range; a non-contact multi-channel conductivity sensor is installed behind the spectrometer scanning line to measure the conductivity distribution on the surface of the waste.

3. The intelligent sorting and treatment method for pretreatment of kitchen waste based on AI recognition according to claim 1, characterized in that, Step S2, data preprocessing, specifically includes the following steps: Step S21: Establish a unified spatiotemporal coordinate system; define the direction of the conveyor belt as the time axis t, and the direction perpendicular to the conveyor belt as the spatial axis x; assign a unified timestamp and spatial location label (x,t) to each frame of visual image, each spectral scan line, and each conductivity measurement data. Step S22: Perform color correction and distortion correction on the visual image data, smooth and standard normal variable transformation on the near-infrared spectral data to eliminate scattering effects, perform temperature compensation and calibration on the conductivity data, and convert it into a standard conductivity value σ(x,t).

4. The intelligent sorting and treatment method for pretreatment of kitchen waste based on AI recognition according to claim 1, characterized in that, Step S3, feature extraction and fusion recognition, specifically includes the following steps: Step S31: Perform semantic segmentation on the preprocessed visual image data. Use a pre-trained deep learning model to identify and segment organic matter fragments in the image, while ignoring inorganic matter and a large amount of free water. The model outputs a binary mask matrix M. org (x,t), where a value of 1 represents that the pixel position (x,t) is organic matter; Step S32: Extract features from the near-infrared spectral data. For each spectral curve I(λ,x,t) acquired at position (x,t), calculate its correlation coefficient with the characteristic spectra of standard carbohydrates, proteins, fats, and water. Simultaneously, calculate its correlation coefficient in the specific salt-sensitive band λ. salt Nearby absorption depth D salt (x,t): In the formula, D salt (x,t) represents the salt absorption depth measured at spatial position x and time point t on the conveyor belt, where λ is the wavelength. salt Let λ be the center wavelength of the salt-sensitive band, Δλ be the wavelength bandwidth, and I(λ,x,t) be the spectral intensity or reflectance value measured by the near-infrared spectral sensor at wavelength λ, spatial location x, and time t. baseline (x,t) represents the wavelength at the baseline λ. baseline Spectral intensity measured at spatial location x and time t; Step S33: Mask the obtained organic material M org (x,t) serves as a spatial filter, performing subsequent analysis only on regions marked as organic matter, and determining the absorption depth D at the corresponding spatial location. salt Feature-level fusion of (x,t) and conductivity data σ(x,t); Step S34: Combine the fused feature vector [D] salt [σ] is input to a lightweight machine learning classifier that has been trained using sample data with known salt concentrations. The classifier outputs a probability value P. salt (x,t)∈[0,1] represents the probability that the organic debris at position (x,t) belongs to high-salinity waste; a probability threshold T is set. h When P salt (x,t)>T h At that time, it was determined that there were high-salinity target fragments at that location.

5. The AI-based intelligent sorting and treatment method for pretreatment of kitchen waste according to claim 1, characterized in that, Step S4, the dynamic sorting decision, specifically includes the following steps: Step S41: Obtain the location information (x) of all high-salinity target fragments identified in step S3. i ,t i The exact time it takes for each target fragment to arrive at the sorting robot's working area is calculated by mapping its spatial range from a two-dimensional spatiotemporal coordinate system to a three-dimensional robotic arm coordinate system, combined with the conveyor belt speed V. and center coordinates (X) i ,Y i ); Step S42: Monitor the number N of high-salinity target fragments identified per unit time in real time, and calculate the dynamic response factor α: In the formula, N represents the number of high-salinity targets in real time. max This represents the system's maximum processing capacity. When α is lower than the threshold α low When the current waste stream is determined to be in a low-salt load state, the robotic arm adopts a precise point-grabbing mode, only picking up the identified independent high-salt fragments; when α is higher than the threshold α high When the current state is determined to be high salt load, it indicates that the high salt content and dense distribution of high salt substances in the garbage flow are high. At this time, the system controls the robotic arm to switch to the area scraping mode. Its movement amplitude and speed are positively correlated with the α value. That is, the robotic arm will use a larger movement amplitude and a faster speed to scrape the garbage group suspected of containing high salt in the target area to the removal channel to ensure processing efficiency and avoid omissions. Step S43: Based on the decision made in step S42, generate the corresponding robotic arm motion trajectory and end effector control instructions; the instructions include target coordinates, motion type, and motion force / amplitude parameters.

6. The AI-based intelligent sorting and treatment method for pretreatment of kitchen waste according to claim 1, characterized in that, Step S5, the sorting process, specifically includes the following steps: Step S51: The sorting robot arm receives the control command and completes the sorting at a predetermined time. Reach the target location (X) i ,Y i It performs grabbing or area scraping actions according to instructions, transferring high-salt waste fragments or waste groups to a dedicated high-salt waste collection bin; Step S52: After the sorting action is performed, a verification sensor is installed downstream of the conveyor belt to perform a second rapid scan of the garbage in the sorted area to verify whether the high-salt substances have been effectively removed. Step S53: The system continuously records the total amount of waste processed in each shift, the amount of high-salt waste identified, the sorting success rate, and the fluctuation of salt load status, forming a data log for optimizing the model and adjusting operating parameters.

Images