Kitchen garbage dynamic processing method and system fusing deep learning

Abstract

This application relates to a method and system for dynamic treatment of kitchen waste integrating deep learning. The method includes: collecting multimodal sensing data during the kitchen waste disposal process; performing spatiotemporal alignment and noise suppression on the multimodal sensing data to generate a structured feature vector set; inputting the structured feature vector set into a pre-set deep learning model to perform waste component distribution inference and degradation potential assessment, outputting a component proportion matrix and a bioactivity index; dynamically generating a phased treatment instruction sequence based on the component proportion matrix and bioactivity index, combined with the current operating conditions of the treatment equipment; and sending the phased treatment instruction sequence to the kitchen waste treatment terminal to complete the dynamic treatment decision identification for the disposed waste. This solution can achieve accurate assessment and dynamic treatment of kitchen waste, improving treatment efficiency and resource utilization efficiency.

Description

Technical Field

[0001] This application relates to the fields of environmental protection and information technology, and in particular to a method and system for dynamic treatment of kitchen waste that integrates deep learning. Background Technology

[0002] In the field of food waste treatment, with the continuous increase in waste generation and increasingly stringent environmental protection requirements, the efficient treatment of food waste has become an important research direction. Currently, traditional food waste treatment methods are mainly based on experience and fixed processing procedures, lacking precise assessment of waste composition and degradation potential. For example, some treatment methods simply involve uniformly crushing and fermenting food waste without considering the differences in the processing of waste with different compositions.

[0003] To improve this situation, some researchers have proposed methods for classifying waste based on certain basic characteristics. For example, waste can be categorized into different types based on simple weight and appearance, and then different processing parameters can be applied. This method improves processing efficiency to some extent, but it still has limitations.

[0004] However, this classification and treatment method based on basic characteristics has a key problem: it cannot accurately obtain detailed information on the composition and degradation potential of kitchen waste. Because the composition of kitchen waste is complex and variable, relying solely on simple weight and appearance is insufficient to fully understand the specific proportions of organic matter, fibers, and oils, as well as the degree of microbial availability. This makes it impossible to dynamically adjust treatment strategies according to the actual condition of the waste during processing, potentially leading to insufficient or excessive treatment, thus affecting treatment effectiveness and resource utilization efficiency. Summary of the Invention

[0005] The main purpose of this application is to provide a method and system for dynamic treatment of kitchen waste that integrates deep learning, which can accurately assess the composition and degradation potential of kitchen waste, and dynamically adjust the treatment strategy based on the assessment results, thereby improving the treatment efficiency and resource utilization efficiency of kitchen waste.

[0006] To achieve the above objectives, embodiments of the present invention provide a method for dynamic treatment of kitchen waste integrating deep learning, the method comprising the following steps: Collect multimodal sensing data during the disposal of kitchen waste, wherein the multimodal sensing data includes visible light image sequences, near-infrared spectral signals and weight change time series, environmental temperature and humidity parameters, and the current filling status of the waste container; The multimodal sensing data is spatiotemporally aligned and noise suppressed to generate a structured feature vector set, wherein the structured feature vector set includes a fusion representation after timestamp registration, spatial coordinate normalization and outlier removal. The structured feature vector set is input into a pre-set deep learning model to perform waste component distribution inference and degradation potential assessment, and output a component ratio matrix and a bioactivity index, wherein the component ratio matrix represents the relative proportion of organic matter, fibrous and oily substances, and the bioactivity index reflects the degree of microbial availability. Based on the component ratio matrix and bioactivity index, and combined with the current operating conditions of the processing equipment, a phased processing instruction sequence is dynamically generated, wherein the phased processing instruction sequence includes crushing particle size setting, fermentation temperature range and stirring frequency adjustment strategy. The phased processing instruction sequence is sent to the food waste treatment terminal to complete the dynamic processing decision identification for the current waste disposal, wherein the dynamic processing decision identification is an executable process control signal.

[0007] In summary, the technical solution of this application, by collecting multimodal sensing data during the disposal of kitchen waste, including visible light image sequences, near-infrared spectral signals and weight change time series, current filling status of waste containers, and environmental temperature and humidity parameters, can comprehensively acquire relevant information about kitchen waste. Spatiotemporal alignment and noise suppression of the multimodal sensing data generate a structured feature vector set, improving data quality and usability. Inputting the structured feature vector set into a pre-set deep learning model performs waste component distribution inference and degradation potential assessment, outputting a component proportion matrix and a bioactivity index, accurately understanding the composition and degradation potential of kitchen waste. Based on the component proportion matrix and bioactivity index, combined with the current operating conditions of the processing equipment, a phased processing instruction sequence is dynamically generated, allowing for the formulation of reasonable processing strategies according to the actual situation of the waste. Sending the phased processing instruction sequence to the kitchen waste processing terminal completes the dynamic processing decision identification for this waste disposal, achieving efficient processing of kitchen waste, avoiding insufficient or excessive processing, and improving processing effect and resource utilization efficiency. Attached Figure Description

[0008] Figure 1 This is a schematic diagram of a scenario for a dynamic kitchen waste treatment method that integrates deep learning, as described in an embodiment of this application. Figure 2 A flowchart illustrating the dynamic processing method for kitchen waste incorporating deep learning provided in this application embodiment; Figure 3 This is a schematic diagram of the structured feature vector processing provided in the embodiments of this application; Figure 4 This is a schematic flowchart of band signal processing provided in an embodiment of this application; Figure 5 A schematic diagram illustrating the process of forming the component ratio matrix and bioactivity index provided in the embodiments of this application; Figure 6 This is a schematic diagram illustrating the process of forming the component ratio matrix provided in the embodiments of this application; Figure 7 This is a flowchart illustrating the generation of the stage processing instruction sequence provided in an embodiment of this application. Figure 8 This is a schematic diagram of the process for generating a temperature control curve provided in an embodiment of this application; Figure 9 A schematic diagram of the structure of the dynamic kitchen waste treatment system integrating deep learning provided in the embodiments of this application; Figure 10 A schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0009] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings.

[0010] This application provides a method and system for dynamic processing of kitchen waste that integrates deep learning, which will be described in detail below.

[0011] In this embodiment, the deep learning-integrated dynamic food waste treatment method is a comprehensive approach to food waste management. It involves the collection, processing, and analysis of various aspects of food waste information, aiming to dynamically adjust treatment strategies based on the actual composition and degradation potential of the waste to achieve efficient processing.

[0012] As shown in Figure 1, a scenario for a dynamic treatment method for kitchen waste integrating deep learning is provided. This scenario may include a kitchen waste disposal area, multimodal sensing devices, processing devices, and a control platform; wherein the multimodal sensing devices, processing devices, and control platform are connected through a network.

[0013] Taking the food waste disposal scenario of a large restaurant as an example, a large amount of food waste is generated every day in this scenario. The food waste disposal area is a place in the restaurant specifically for disposing of food waste, such as the area where the trash cans are located.

[0014] Multimodal sensing devices are distributed at key locations throughout the food waste disposal area. For example, a high-definition camera is installed at the inlet of the waste container to trigger continuous frame image capture, obtaining a visible light image sequence covering the complete falling trajectory. For instance, when restaurant staff pour food waste into the bin, the camera continuously records the appearance and falling process of the waste. Simultaneously, a near-infrared sensor array is deployed around the bin to synchronously scan the surface reflectance spectrum of the waste, collecting raw spectral response values ​​according to preset wavelengths and constructing a spectral-spatial mapping table. Near-infrared spectral analysis reveals the composition of different substances in the waste. Furthermore, a weighing unit is installed at the bottom of the container to monitor changes in mass increment in real time, recording the weight time-series curve from initial rest to a steady state and marking the timestamps corresponding to abrupt changes. Weight changes indicate the amount and speed of waste disposal. Additionally, a built-in liquid level or ultrasonic ranging module reads the filling height data and converts it into a volume occupancy rate value as the current filling status. An environmental monitoring unit acquires current air temperature and relative humidity readings as environmental temperature and humidity parameters. These multimodal sensing devices transmit the collected data to the control platform.

[0015] The control platform processes and analyzes the data uploaded by the multimodal sensing devices. First, it performs spatiotemporal alignment and noise suppression on the multimodal sensing data to generate a structured feature vector set.

[0016] Next, the control platform inputs the structured feature vector set into the pre-set deep learning model to perform waste composition distribution inference and degradation potential assessment, and outputs the composition ratio matrix and bioactivity index.

[0017] Finally, the control platform dynamically generates a phased processing instruction sequence based on the component ratio matrix and bioactivity index, combined with the current operating conditions of the processing equipment. It queries the current motor load, chamber temperature, and remaining processing capacity of the processing equipment to create a snapshot of the equipment's operating status. When the proportion of fiber in the component ratio matrix is ​​higher than a preset threshold and the bioactivity index is lower than the median level, a coarse crushing mode is set and the premixing time is extended. The platform dynamically compresses the fermentation temperature rise slope according to the proportion of oily substances to avoid localized anaerobic acidification, generating a stepped temperature control curve. The stirring frequency is matched according to the bioactivity index, with high activity corresponding to high-frequency intermittent stirring and low activity corresponding to low-frequency continuous stirring. The crushing, temperature control, and stirring strategies are arranged in chronological order into an irreversible execution queue, with added safety interlock conditions, forming a phased processing instruction sequence. This phased processing instruction sequence is sent to the processing equipment, which processes the kitchen waste according to the instruction sequence, completing the dynamic processing decision identification for this waste disposal.

[0018] refer to Figure 2 , Figure 2 This is a flowchart illustrating a method for dynamically processing kitchen waste using deep learning, as provided in an embodiment of this application. The execution entity of this method can be a computer device (such as a control platform), such as a server, etc. Specifically, the method for dynamically processing kitchen waste using deep learning, as provided in this embodiment, includes:

[0019] S10: Collect multimodal sensing data during the disposal of kitchen waste, wherein the multimodal sensing data includes visible light image sequences, near-infrared spectral signals and weight change time series, environmental temperature and humidity parameters, and the current filling status of the waste container.

[0020] In this embodiment, the multimodal sensing data is a dataset comprehensively describing various characteristics of kitchen waste during the disposal process. The visible light image sequence is a series of images continuously captured by a camera, which can intuitively reflect the appearance characteristics of kitchen waste and the disposal process. The near-infrared spectral signal is a signal obtained by scanning the reflectance spectrum of the waste surface using a near-infrared sensor, and can be used to analyze the composition information of different substances in the waste. The weight change time series is a curve obtained by real-time monitoring of waste mass changes by a weighing unit, which can reflect the amount and speed of waste disposal. The current filling status of the waste container refers to the volume occupancy rate of the waste inside the container, which can be obtained through a liquid level or ultrasonic ranging module. The environmental temperature and humidity parameters refer to the air temperature and relative humidity in the waste disposal environment, which have a significant impact on the waste degradation process.

[0021] In one embodiment, a high-definition camera is installed at the inlet of the waste disposal container, and appropriate shooting parameters, such as frame rate and resolution, are set to trigger continuous frame image capture, obtaining a visible light image sequence covering the complete falling trajectory. For the near-infrared sensor array, the scanning frequency and preset band division are set, and scanning of the waste surface reflectance spectrum is started synchronously. The original spectral response values ​​are collected according to the preset band division, and a spectral-spatial mapping relationship table is constructed. A high-precision weighing unit is installed at the bottom of the container to monitor the change in mass increment in real time, record the weight time-series curve from the initial static state to the steady state, and mark the timestamps corresponding to the abrupt change points. The filling height data is read using the container's built-in liquid level or ultrasonic ranging module and converted into a volume occupancy rate value as the current filling status. The environmental monitoring unit is called to obtain the current air temperature and relative humidity readings as environmental temperature and humidity parameters. Using the reference timestamp as a unified time reference, the visible light image sequence, near-infrared spectral response value, weight time-series curve, volume occupancy rate value, and temperature and humidity readings are bound together to form a synchronous sensing data package with environmental context. By collecting multimodal sensing data and synchronously recording relevant parameters, comprehensive information about kitchen waste can be obtained, providing a foundation for subsequent processing and analysis.

[0022] In one embodiment, step S10 can be implemented as follows: A1: Trigger continuous frame image capture at the inlet of the waste disposal container to obtain a sequence of visible light images covering the complete falling trajectory.

[0023] In this embodiment, continuous frame image capture refers to continuously capturing images at a certain frame rate using a camera at the inlet of the waste disposal container to record the waste disposal process. The visible light image sequence is a sequence composed of these continuously captured images, capable of reflecting the appearance characteristics and trajectory of the waste.

[0024] In one embodiment, a high-speed, high-definition camera is installed at the inlet of the waste disposal container, with a frame rate of 30 frames per second and a resolution of 1920×1080. When waste is detected, the camera immediately triggers continuous frame image capture. For example, at the inlet of a restaurant's food waste bin, when a worker pours a plate of leftover food into the bin, the camera begins continuous shooting, recording the entire process of the food sliding off the plate, falling through the air, and landing in the bin, acquiring a visible light image sequence covering the complete trajectory. By acquiring this visible light image sequence, the appearance and disposal status of the food waste can be intuitively understood, providing visual information for subsequent analysis.

[0025] A2: Simultaneously start the near-infrared sensor array to scan the reflectance spectrum of the garbage surface, collect the original spectral response values ​​according to the preset band division, and construct the spectral-spatial mapping relationship table.

[0026] In this embodiment, the near-infrared sensor array is an array composed of multiple near-infrared sensors used to scan the reflectance spectrum of the waste surface. The raw spectral response value is the numerical value of the reflectance spectral signal received by the sensor. The spectral-spatial mapping table is a table that associates spectral information with the spatial location of the waste.

[0027] In one embodiment, a near-infrared sensor array is evenly arranged around the trash can, with a scanning frequency of 10 scans per second and a preset wavelength division of 10 bands. When acquiring visible light image sequences begins, the near-infrared sensor array is simultaneously activated to scan the reflectance spectrum of the trash surface. For example, during trash disposal, the near-infrared sensor array continuously scans the trash surface, acquiring the raw spectral response values ​​for each band. The acquired spectral response values ​​are correlated with the spatial location information of the trash to construct a spectral-spatial mapping table. By constructing this table, spectral information can be combined with the actual location of the trash, allowing for a more accurate analysis of its composition.

[0028] A3: Monitor the change in mass increment in real time through the weighing unit at the bottom of the container, record the weight time series curve from the initial static state to the steady state, and mark the timestamps corresponding to the abrupt change points.

[0029] In this embodiment, the weighing unit at the bottom of the container is a weighing device installed at the bottom of the trash can, used to monitor changes in the weight of the trash in real time. The weight time-series curve records the change in trash weight over time. A sudden change point is a point in the weight time-series curve where a significant change occurs, and the corresponding timestamp is the time when that point occurred.

[0030] In one embodiment, a high-precision weighing unit with an accuracy of 0.1 grams is installed at the bottom of the trash can. When trash is first deposited, the weighing unit begins to monitor changes in mass increment in real time. For example, as restaurant staff sequentially pour multiple plates of leftover food into the trash can, the weighing unit continuously records the increase in mass, forming a weight-time series curve. When a large chunk of trash is deposited, the weight-time series curve shows a significant abrupt change, and the timestamp corresponding to this abrupt change is marked. By recording the weight-time series curve and marking the timestamps of abrupt changes, the amount and speed of trash deposited can be determined, providing weight information for subsequent analysis.

[0031] A4: Read the filling height data output by the liquid level or ultrasonic ranging module built into the container, and convert it into a volume occupancy rate value as the current filling status.

[0032] In this embodiment, the liquid level or ultrasonic ranging module built into the container is a device used to measure the garbage filling height inside the trash can. The filling height data is a numerical value output by the module representing the garbage filling height. The volume occupancy rate value is a value converted from the filling height data to represent the percentage of garbage volume within the trash can.

[0033] In one embodiment, a liquid level sensor or ultrasonic ranging module is installed inside the trash can. During trash disposal, the module outputs real-time filling height data. For example, if the liquid level sensor measures 20 cm of liquid level inside the trash can, and the total height of the trash can is known to be 50 cm, the filling height data is converted into a volume occupancy rate of 40%. This volume occupancy rate is used as the current filling status. By obtaining the current filling status, the space utilization inside the trash can can be understood, providing a reference for subsequent processing decisions.

[0034] A5: Call the environmental monitoring unit to obtain the current air temperature and relative humidity readings, which are used as the environmental temperature and humidity parameters.

[0035] In this embodiment, the environmental monitoring unit is a device for measuring ambient air temperature and relative humidity. The air temperature and relative humidity readings are numerical values ​​output by the unit representing the ambient temperature and humidity.

[0036] In one embodiment, an environmental monitoring unit, such as a temperature and humidity sensor, is installed in the waste disposal area. While collecting multimodal sensing data, the environmental monitoring unit obtains the current air temperature and relative humidity readings. For example, in the food waste disposal area of ​​a restaurant, the environmental monitoring unit displays a current air temperature of 25°C and a relative humidity of 60%, and these two values ​​are used as environmental temperature and humidity parameters. These parameters have a significant impact on the degradation process of food waste, and obtaining them can provide a basis for formulating subsequent treatment strategies.

[0037] A6: Using the reference timestamp as a unified time reference, the visible light image sequence, near-infrared spectral response value, weight time series curve, volume occupancy rate value, and temperature and humidity readings are bound together to form a synchronous sensing data package with environmental context.

[0038] In this embodiment, the reference timestamp is a reference time point for determining the synchronization of different data. The synchronization sensing data packet is a data packet formed by binding multimodal sensing data according to a unified time reference, and it contains environmental context information.

[0039] In one embodiment, the start time of the weight mutation is used as the reference timestamp. Each frame of the visible light image sequence, the near-infrared spectral response value, each data point in the weight time-series curve, the volume occupancy rate value, and the temperature and humidity readings are associated with the reference timestamp based on their corresponding time. For example, the visible light image, near-infrared spectral response value, weight data, volume occupancy rate value, and temperature and humidity readings acquired 5 seconds after the reference timestamp are bound together. This forms a synchronous sensing data package with environmental context. By forming a synchronous sensing data package, multimodal sensing data can be integrated, facilitating subsequent processing and analysis.

[0040] S20: Perform spatiotemporal alignment and noise suppression on the multimodal sensing data to generate a structured feature vector set, wherein the structured feature vector set includes a fusion representation after timestamp registration, spatial coordinate normalization, and outlier removal.

[0041] In this embodiment, spatiotemporal alignment refers to uniformly aligning sensing data from different modalities in time and space to ensure data consistency and comparability. Noise suppression refers to removing interference and outliers from the data to improve data quality. Structured feature vector sets convert processed data into fixed-dimensional feature vectors for easier processing by subsequent deep learning models. Timestamp registration unifies the timestamps of different data to ensure data synchronization in time. Spatial coordinate normalization standardizes the spatial coordinates of the data to eliminate differences caused by different devices or measurement methods. Outlier removal refers to removing abnormal data points that deviate from the normal range.

[0042] S30: Input the structured feature vector set into a preset deep learning model, perform waste component distribution inference and degradation potential assessment, and output a component proportion matrix and a bioactivity index, wherein the component proportion matrix represents the relative proportion of organic matter, fibrous and oily substances, and the bioactivity index reflects the degree of microbial availability.

[0043] In this embodiment, the deep learning model is a neural network-based machine learning model with powerful feature learning and classification capabilities. Waste composition distribution inference refers to analyzing a structured feature vector set using a deep learning model to infer the relative proportions of organic matter, fiber, and oils in kitchen waste. Degradation potential assessment refers to evaluating the degree of microbial availability in the waste, i.e., the bioactivity index. The composition proportion matrix is ​​a matrix used to represent the relative proportions of different substances. The bioactivity index is a numerical value reflecting the degree of microbial availability in the waste.

[0044] S40: Based on the component ratio matrix and bioactivity index, and combined with the current operating conditions of the processing equipment, dynamically generate a phased processing instruction sequence, wherein the phased processing instruction sequence includes crushing particle size setting, fermentation temperature range and stirring frequency adjustment strategy.

[0045] In this embodiment, the component ratio matrix and bioactivity index are important bases for formulating the treatment strategy. The current operating condition of the treatment equipment refers to the current motor load, chamber temperature, and remaining processing capacity of the equipment. The phased treatment instruction sequence is a series of dynamically generated treatment instructions based on the actual situation of the waste and the operating status of the treatment equipment, including particle size setting, fermentation temperature range, and stirring frequency adjustment strategy. Particle size setting refers to determining the particle size when crushing kitchen waste. Fermentation temperature range refers to the temperature range that needs to be controlled during fermentation. Stirring frequency adjustment strategy refers to adjusting the stirring frequency according to the degradation of the waste.

[0046] S50: Send the phased processing instruction sequence to the food waste treatment terminal to complete the dynamic processing decision identifier for the current waste disposal, wherein the dynamic processing decision identifier is an executable process control signal.

[0047] In this embodiment, the food waste treatment terminal is the equipment that actually performs the food waste treatment operation, such as a crusher or fermentation tank. The phased processing instruction sequence is a series of instructions used to guide the operation of the treatment equipment. The dynamic processing decision identifier converts the instruction sequence into executable process control signals to control the operation of the treatment equipment.

[0048] In one embodiment, the control platform sends a sequence of phased processing instructions to the food waste treatment terminal via a network. Upon receiving the instruction sequence, the treatment terminal parses it into executable process control signals, such as controlling the crusher's particle size, the fermentation tank's temperature, and the stirring frequency. The treatment equipment then performs corresponding operations based on the process control signals to complete the processing of the submitted waste. By sending the phased processing instruction sequence to the treatment terminal and converting it into process control signals, automated processing of food waste can be achieved, improving the accuracy and efficiency of the processing.

[0049] In one embodiment, reference Figure 3 Step S20 may include steps S21-S25, which will be described in detail below: S21: Using the weight mutation start time in the multimodal sensing data as the reference time origin, the image frames in the visible light image sequence and the spectral sampling points in the near-infrared spectral signal are resampled and aligned with millisecond-level precision to establish a spatiotemporal synchronization data block.

[0050] In this embodiment, the weight mutation start time refers to the moment when a significant change begins to appear in the weight time-series curve. Millisecond-level precision resampling alignment refers to precisely aligning different data points in time with millisecond-level precision. A spatiotemporal synchronized data block is a data block formed by integrating the resampling-aligned data, possessing temporal and spatial consistency.

[0051] In one embodiment, after determining the start time of the weight mutation, this time is used as the reference time origin. For visible light image sequences, based on their frame rate and capture time, the timestamp of each frame is compared with the reference time origin, and millisecond-level resampling alignment is performed. For near-infrared spectral signals, the timestamp of each spectral sampling point is similarly adjusted to align with the reference time origin. For example, assuming the start time of the weight mutation is 0 milliseconds and the frame rate of the visible light image sequence is 30 frames per second, then the timestamp of the first frame is 0 milliseconds, the timestamp of the second frame is 33.3 milliseconds (1000 milliseconds / 30 frames), and so on. The resampled and aligned visible light image sequences and near-infrared spectral signals are integrated together to establish a spatiotemporally synchronized data block. By establishing a spatiotemporally synchronized data block, the temporal consistency of data from different modalities can be ensured, improving the accuracy of subsequent analysis.

[0052] S22: Perform background segmentation on the visible light image sequence in the spatiotemporal synchronization data block, and use the fixed reference object on the inner wall of the container to correct the viewpoint distortion, and output a background-removed and geometrically normalized garbage area mask.

[0053] In this embodiment, background segmentation refers to separating the background portion from the garbage portion in a visible light image sequence. Fixed reference objects on the inner wall of the container refer to fixed markers placed on the inner wall of the garbage bin to correct for angular distortion. The garbage region mask is an image mask containing only the garbage region, which has undergone background removal and geometric normalization.

[0054] In one embodiment, an image segmentation algorithm is used to perform background segmentation on the visible light image sequence in the spatiotemporal synchronization data block. For example, a deep learning-based semantic segmentation algorithm is used to classify the garbage region and background region in the image. Simultaneously, some obvious fixed reference objects, such as cross marks, are placed on the inner wall of the garbage can. By identifying the positions of these fixed reference objects in the image, the viewpoint distortion parameters of the image are calculated. These parameters are used to correct the image and eliminate viewpoint distortion. For example, the actual position and positional deviation of the fixed reference objects in the image are identified, and the image is corrected using methods such as affine transformation. Finally, a background-removed and geometrically normalized garbage region mask is output. By outputting the garbage region mask, the feature information of the garbage can be extracted more accurately, and background interference can be reduced.

[0055] S23: Group the near-infrared spectral signals in the spatiotemporal synchronization data block according to the material absorption characteristic bands, calculate the mean and variance of each group, and remove abnormal channels that deviate from the global statistical distribution by more than three standard deviations.

[0056] In this embodiment, the characteristic absorption bands of a substance refer to the bands in the near-infrared spectrum where different substances have specific absorption peaks. Grouping refers to dividing the near-infrared spectral signals according to the characteristic absorption bands of the substances. Mean and variance are statistical quantities used to describe the distribution characteristics of the data. Outlier channels refer to channels that deviate from the global statistical distribution by more than three standard deviations.

[0057] In one embodiment, reference Figure 4 Step S23 may include steps S231-S235, which will be described in detail below: S231: Divide the full-band near-infrared signal in the spatiotemporal synchronization data block into three logical groups: water molecule absorption region, carbon-hydrogen bond vibration region, and stray light interference region, and calculate the arithmetic mean of all channels in each group.

[0058] In this embodiment, the full-band near-infrared signal refers to the signal collected by the near-infrared sensor across the entire band. The water molecule absorption region, the carbon-hydrogen bond vibration region, and the stray light interference region are areas defined based on the absorption characteristics of different substances in the near-infrared spectrum. The arithmetic mean is the sum of a set of data divided by the number of data points.

[0059] In one embodiment, based on the characteristics of near-infrared spectroscopy, the corresponding wavelength ranges for the water molecule absorption region, the carbon-hydrogen bond vibration region, and the stray light interference region are determined. For example, the water molecule absorption region corresponds to the 1400-1900 nm wavelength range, the carbon-hydrogen bond vibration region corresponds to the 2100-2300 nm wavelength range, and the stray light interference region corresponds to other wavelength ranges. The full-band near-infrared signal in the spatiotemporal synchronization data block is divided into three logical groups according to these wavelength ranges. For each logical group, the arithmetic mean of all its channels is calculated. For example, in the water molecule absorption region, the signal values ​​of all channels in this region are added together and then divided by the number of channels to obtain the arithmetic mean of the group. By dividing into logical groups and calculating the arithmetic mean, preliminary statistical analysis of the near-infrared spectral signal can be performed, providing a basis for subsequent processing.

[0060] S232: Construct a local Gaussian distribution model based on the average values ​​of each group, and determine the absolute deviation of each channel from the mean of its group.

[0061] In this embodiment, the local Gaussian distribution model is a statistical model based on the normal distribution, used to describe the distribution of signals in each group of channels. The absolute deviation refers to the absolute value of the difference between the signal value of each channel and the mean of its group.

[0062] In one embodiment, a local Gaussian distribution model is constructed centered on the arithmetic mean of each group and based on the standard deviation of the channel signals within that group. For example, for the water molecule absorption region, a Gaussian distribution model is constructed using the arithmetic mean of the group as the mean and the standard deviation of the channel signals within that group as the standard deviation. Then, the absolute deviation of each channel from the mean of its group is calculated. For example, for a channel within the water molecule absorption region with a signal value of 100 and the arithmetic mean of the group being 90, the absolute deviation of that channel is 10. By constructing the local Gaussian distribution model and determining the absolute deviation, the degree of deviation of each channel signal from the mean of its group can be understood.

[0063] S233: Calculate the standard deviation of the deviation of all channels globally, and set the elimination threshold to three times the value of this standard deviation.

[0064] In this embodiment, the global channel deviation refers to the set of absolute deviations of all channels from the mean of their respective groups. Standard deviation is a statistic used to measure the dispersion of data. The elimination threshold is a criterion used to determine whether a channel is an abnormal channel.

[0065] In one embodiment, the absolute deviations of all channels from the mean of their respective groups are collected to form a deviation array. The deviation array is sorted, and the top 5% and bottom 5% of extreme points are removed to prevent extreme anomalies from affecting statistical stability. For example, if the deviation array has 100 data points, the top 5 and bottom 5 data points are removed. The sample variance is calculated based on the remaining deviations, and the square root is taken to obtain the corrected standard deviation. For example, the arithmetic mean of the deviation array after removing extreme values ​​is calculated as the center estimate. The squared difference between each deviation and the center estimate is calculated, and all squared differences are summed to obtain the total sum of squares. The total sum of squares is divided by the effective sample size minus one to obtain the unbiased sample variance. A non-negativity check is performed on the variance values; if the result is negative, it is set to zero; otherwise, the original value is retained. The square root of the checked variance values ​​is calculated, and the corrected standard deviation is output. The corrected standard deviation is multiplied by three to obtain the final anomaly detection threshold. This threshold is stored for subsequent channel screening, and the removal ratio is recorded for system health assessment.

[0066] In one embodiment, the sample variance is calculated based on the residual bias, and the square root of the result is used to obtain the corrected standard deviation, which can be achieved as follows: H1: Calculate the arithmetic mean of the deviation array after removing extreme values, and use it as the central estimate.

[0067] The center estimate is an estimate of the location of this group of data centers.

[0068] In one embodiment, assume the deviation array after removing extreme values ​​is [10, 12, 15, 18, 20]. Summing these data gives a total of 75. Since there are 5 data points in the array, dividing the sum by the number of data points yields an arithmetic mean of 75 / 5 = 15. This arithmetic mean of 15 is used as the central mean estimate. Obtaining the central mean estimate provides a basis for subsequent calculations of statistics such as variance.

[0069] H2: Calculate the squared difference between each deviation and the central estimate, and sum all the squared differences to obtain the total sum of squared deviations.

[0070] In this embodiment, the squared difference refers to the square of the difference between each deviation and the central estimate. The total sum of squares is the value obtained by summing the squared differences of all deviations, and is used to measure the dispersion of the data.

[0071] In one embodiment, given a central mean estimate of 15, for each deviation in the deviation array [10, 12, 15, 18, 20], the squared difference between it and the central mean estimate is calculated. For deviation 10, the squared difference is (10 - 15)^2 = 25; for deviation 12, it is (12 - 15)^2 = 9; for deviation 15, it is (15 - 15)^2 = 0; for deviation 18, it is (18 - 15)^2 = 9; and for deviation 20, it is (20 - 15)^2 = 25. Summing these squared differences yields a total sum of squared deviations of 25 + 9 + 0 + 9 + 25 = 68. Calculating this total sum of squared deviations provides a more accurate understanding of the data's dispersion.

[0072] H3: Divide the total sum of squared deviations by the number of valid samples minus one to obtain the unbiased sample variance.

[0073] In this embodiment, the effective sample size refers to the number of data points in the deviation array after removing extreme values. Unbiased sample variance is an unbiased estimate of the population variance, obtained by dividing the total sum of squared deviations by the effective sample size minus one.

[0074] In one embodiment, the total sum of squares is known to be 68, and the effective sample size is 5. Dividing the total sum of squares by the effective sample size minus one, i.e., 68 / (5 - 1) = 17, yields an unbiased sample variance of 17. By calculating the unbiased sample variance, the dispersion of the population data can be estimated more accurately.

[0075] H4: Performs a non-negativity check on the difference. If the result is negative, it is set to zero; otherwise, the original value is retained.

[0076] In this embodiment, the nonnegativity check refers to checking whether the variance value is negative. Since variance is a statistic that measures the dispersion of data, its value cannot be negative. If the result is negative, it may be due to an error in the calculation process, and it needs to be set to zero.

[0077] In one embodiment, the obtained variance value is 17. Since 17 is a positive number, it satisfies the non-negativity requirement, so the original value of 17 is retained. If the calculated variance value is -5, it is negative and does not meet the non-negativity requirement, so it is set to zero. By performing a non-negativity check, the reasonableness and validity of the variance value can be guaranteed.

[0078] H5: Performs square root operation on the verified variance value and outputs the corrected standard deviation for threshold generation.

[0079] In this embodiment, the square root operation refers to taking the square root of the verified variance value. The corrected standard deviation is the value obtained by performing the square root operation on the verified variance value, and is used to generate the final anomaly detection threshold.

[0080] In one embodiment, the variance after verification is 17. Taking the square root of this variance yields a corrected standard deviation of √17 ≈ 4.12. This corrected standard deviation is used for subsequent threshold generation, for example, multiplying it by three to obtain the anomaly detection threshold. By outputting the corrected standard deviation, an accurate basis for judging abnormal channels can be provided, improving the accuracy and reliability of data processing.

[0081] S234: Mark channels whose deviation exceeds the rejection threshold as abnormal channels and generate an index list of abnormal channels.

[0082] In this embodiment, an abnormal channel refers to a channel whose deviation exceeds the rejection threshold. The abnormal channel index list is a list that records the locations of abnormal channels.

[0083] In one embodiment, the deviation of each channel is compared with a rejection threshold. If the deviation of a channel exceeds the rejection threshold, it is marked as an abnormal channel. For example, if the rejection threshold is 20 and the deviation of a channel is 25, then that channel is marked as an abnormal channel. The indices of all channels marked as abnormal are recorded to generate an abnormal channel index list. By generating an abnormal channel index list, abnormal channels can be easily identified and processed.

[0084] S235: Remove the band data corresponding to the abnormal channel index from the original spectral signal, and retain the valid channels.

[0085] In this embodiment, the original spectral signal refers to the unprocessed spectral signal acquired by the near-infrared sensor. The band data corresponding to the abnormal channel index refers to the spectral signal data corresponding to the abnormal channel. The valid channel refers to the channel whose deviation does not exceed the rejection threshold.

[0086] In one embodiment, the band data corresponding to abnormal channels is removed from the original spectral signal based on an abnormal channel index list. For example, if the abnormal channel index list records channels 10, 20, and 30 as abnormal channels, then the band data corresponding to these three channels are deleted from the original spectral signal. The remaining valid channels are retained. By removing abnormal channel data, the quality of the near-infrared spectral signal can be improved, noise interference can be reduced, and more accurate data can be provided for subsequent analysis.

[0087] S24: Perform sliding window smoothing filtering on the weight change time-series curve in the multimodal sensing data, identify and remove high-frequency jitter components caused by vibration or airflow disturbance, and extract the smoothed mass accumulation trend.

[0088] In this embodiment, sliding window smoothing filtering is a signal processing method that smooths the signal by sliding a fixed-size window across the time series and averaging or weighted averaging the data within the window. High-frequency jitter refers to rapidly changing components in the weight change time series curve caused by factors such as vibration or airflow disturbance. The cumulative mass trend refers to the cumulative change trend of weight over time after smoothing.

[0089] In one embodiment, a suitable sliding window size is selected, such as 5 data points. The weight change time-series curve is then smoothed using a sliding window filter. For example, for each data point in the weight change time-series curve, the average of the two data points before and after it is calculated as the smoothed value for that point. In this way, high-frequency jitter components caused by vibration or airflow disturbances are identified and removed. The smoothed cumulative mass trend is extracted, i.e., the overall change in weight over time is observed. By extracting the smoothed cumulative mass trend, the changing trends of waste disposal volume and disposal speed can be more clearly understood.

[0090] S25: Extract image texture features from the background-removed and geometrically normalized garbage area mask, extract spectral grouping statistics from the processed near-infrared spectral signal, extract the instantaneous slope of the mass accumulation trend from the smoothed weight time-series curve, and obtain the environmental temperature and humidity parameters and volume occupancy rate values ​​contained in the multimodal sensing data.

[0091] In this embodiment, image texture features refer to the features reflecting the surface texture of waste in the waste area mask, such as gray-level co-occurrence matrix features and local binary mode features. Spectral group statistics refer to the statistics calculated by grouping the processed near-infrared spectral signals, such as mean and variance. The instantaneous slope of the mass accumulation trend refers to the slope of the smoothed weight change time series curve at a certain moment, reflecting the rate of weight change. Environmental temperature and humidity parameters refer to the air temperature and relative humidity recorded in the multimodal sensing data.

[0092] In one embodiment, image texture features are extracted from a background-removed and geometrically normalized mask of waste areas. For example, features such as contrast and correlation of the gray-level co-occurrence matrix are calculated. Spectral grouping statistics are extracted from the processed near-infrared spectral signal, such as calculating the mean and variance of each group. The instantaneous slope of the mass accumulation trend is extracted from the smoothed weight change time-series curve. For example, the instantaneous slope is obtained by calculating the difference between two adjacent data points divided by the time interval. Simultaneously, environmental temperature and humidity parameters and volume occupancy rates contained in the multimodal sensing data are acquired. By extracting these features and parameters, a comprehensive description of the characteristics of kitchen waste can be achieved.

[0093] S26: The extracted image texture features, spectral group statistics, instantaneous slope of the mass accumulation trend, volume occupancy rate, and environmental temperature and humidity parameters are concatenated into a fixed-dimensional feature vector to form the structured feature vector set.

[0094] In this embodiment, a fixed-dimensional feature vector refers to a vector of fixed length formed by arranging different types of features in a certain order. A structured feature vector set is a collection of multiple such feature vectors, which facilitates subsequent processing by deep learning models.

[0095] In one embodiment, the extracted image texture features, spectral grouping statistics, instantaneous slope of the mass accumulation trend, volume occupancy rate, and environmental temperature and humidity parameters are concatenated in a predefined order. For example, the dimensions of the image texture features are arranged sequentially first, followed by the spectral grouping statistics, then the instantaneous slope of the mass accumulation trend, and finally the environmental temperature and humidity parameters. These features are concatenated into a fixed-dimensional feature vector. Multiple such feature vectors are combined to form a structured feature vector set. By forming a structured feature vector set, multimodal feature information can be integrated, facilitating analysis and processing by deep learning models.

[0096] In one embodiment, reference Figure 5 Step S30 may include steps S31-S35, which will be described in detail below: S31: Input the structured feature vector into the multi-branch attention mechanism of the deep learning model, perform feature weighting and fusion based on the correlation between different data modalities and waste components and degradation characteristics, and generate an optimized fusion feature representation.

[0097] In this embodiment, the multi-branch attention mechanism is a mechanism in deep learning models used to focus on and weight different data modalities. Different data modalities refer to different types of data contained in the structured feature vector, such as image texture features and spectral grouping statistics. Correlation refers to the degree of correlation between different data modalities and waste components and degradation characteristics. Optimized fusion feature representation refers to a more representative feature representation obtained after weighting and fusion.

[0098] In one embodiment, structured feature vectors are input into a multi-branch attention mechanism of the deep learning model. This mechanism assigns different weights to each data modality based on its correlation with waste components and degradation characteristics. For example, higher weights are assigned to spectral grouping statistics that are highly correlated with waste components, while lower weights are assigned to data modalities with lower correlation. The features from different data modalities are then weighted and fused to generate an optimized fused feature representation. This multi-branch attention mechanism highlights important data modalities and improves the expressive power of the features.

[0099] S32: Using the fusion feature representation, the soft classification probability of each substance category is output through the component classification branch of the deep learning model, and the component proportion matrix is ​​generated according to the preset mapping rules.

[0100] In this embodiment, the component classification branch is the part of the deep learning model used to classify waste components. The soft classification probability refers to the probability of each material category being classified, with a value ranging from 0 to 1. The preset mapping rule is a predefined rule that converts the soft classification probability into a component proportion matrix.

[0101] In one embodiment, reference Figure 6 Step S32 may include steps S321-S324, which will be described in detail below: S321: Based on the fused feature representation, initial soft classification probability values ​​for organic matter, fiber, and oil are obtained through the component classification branch of the deep learning model.

[0102] In this embodiment, the component classification branch is the part of the deep learning model used to classify waste components. The initial soft classification probability value refers to the preliminary classification probability of organic matter, fibrous substances, and oily substances output by the component classification branch, and its value ranges from 0 to 1.

[0103] In one embodiment, the fused feature representation is input into the component classification branch of the deep learning model. This branch uses classification algorithms, such as fully connected layers and a softmax function, to process the fused feature representation and output initial soft classification probability values ​​for organic matter, fibrous materials, and oils. For example, after calculation by the component classification branch, the initial soft classification probability value is 0.6 for organic matter, 0.2 for fibrous materials, and 0.2 for oils. By obtaining the initial soft classification probability values, a preliminary understanding of the classification probability of different substances in the waste can be obtained.

[0104] S322: Normalize the initial soft classification probability value and use the normalized probability value as the mass proportion estimate of the corresponding material category to generate a preliminary component proportion vector.

[0105] In this embodiment, mass proportion estimation refers to using the normalized probability value as an estimate of the mass proportion of the corresponding substance category. The preliminary component proportion vector is a vector composed of normalized probability values, used to represent the preliminary proportion relationship between different substances.

[0106] In one embodiment, the initial soft classification probability values ​​are normalized. For example, the initial soft classification probability values ​​are 0.6 for organic matter, 0.2 for fiber, and 0.2 for oils. These three values ​​are summed to obtain a total of 1. Each value is then divided by the sum to obtain the normalized probability values: 0.6 for organic matter, 0.2 for fiber, and 0.2 for oils. These normalized probability values ​​are used as estimates of the mass proportion of the corresponding substance categories, generating a preliminary component proportion vector [0.6, 0.2, 0.2]. Through normalization and the generation of the preliminary component proportion vector, the initial soft classification probability values ​​can be converted into meaningful mass proportion estimates.

[0107] S323: Based on the actual detection deviation in historical feedback data of similar waste treatment, the preliminary component ratio vector is corrected.

[0108] In this embodiment, historical feedback data on similar waste treatment refers to the actual test results and related data recorded when similar kitchen waste was treated in the past. Actual test deviation refers to the difference between the preliminary component proportion vector and the actual test results. Correcting the preliminary component proportion vector is the process of adjusting the preliminary component proportion vector based on the actual test deviation.

[0109] In one embodiment, historical feedback data on similar waste treatment is collected, and the differences between the actual detection results and the preliminary component proportion vector are analyzed. For example, historical data shows that the actual proportion of organic matter in similar waste is usually 10% higher than estimated by the preliminary component proportion vector. Based on this actual detection deviation, the preliminary component proportion vector is corrected. If the preliminary component proportion vector is [0.6, 0.2, 0.2], the proportion of organic matter is adjusted to 0.7, and the proportions of fiber and oil are adjusted to 0.15 and 0.15, respectively. By correcting based on historical data, the accuracy of the component proportion vector can be improved.

[0110] S324: Encapsulate the corrected component proportions into a standardized triplet matrix and output it as the component proportion matrix.

[0111] In this embodiment, the standardized ternary matrix refers to encapsulating the modified component proportions into a matrix according to a certain format, usually a 1×3 matrix, which represents the proportions of organic matter, fiber, and oils respectively.

[0112] In one embodiment, the corrected component ratio is assumed to be 0.7 for organic matter, 0.15 for fiber, and 0.15 for oils. These ratios are encapsulated into a 1×3 matrix [0.7, 0.15, 0.15] in the order of organic matter, fiber, and oils, and output as the component ratio matrix. By encapsulating it into a standardized triplet matrix, subsequent processing and use can be facilitated, clearly representing the proportional relationship of different substances in the waste.

[0113] S33: Using the fusion feature representation, the absorption intensity of specific functional groups in the near-infrared spectrum and the image texture roughness index are analyzed through the degradation evaluation branch of the deep learning model to calculate the organic molecular chain breakage tendency score.

[0114] In this embodiment, the degradation assessment branch is the part of the deep learning model used to evaluate the degradation potential of waste. The specific functional group absorption intensity refers to the absorption intensity of functional groups related to organic matter degradation in the near-infrared spectrum. The image texture roughness index is an indicator reflecting the roughness of the waste surface texture. The organic matter molecular chain breakage tendency score is a score calculated based on the specific functional group absorption intensity and the image texture roughness index, reflecting the probability of organic matter molecular chain breakage.

[0115] In one embodiment, step S33 can be implemented as follows: G1: Based on the fusion feature representation, features directly related to degradation potential are extracted through degradation assessment branches, including the absorption intensity of easily degradable functional groups in the near-infrared spectrum and image texture roughness index.

[0116] In this embodiment, the absorption intensity of the recalcitrant functional groups refers to the absorption intensity of functional groups associated with recalcitrant and easily degradable organic matter in the near-infrared spectrum. The image texture roughness index is an indicator reflecting the roughness of the waste surface texture.

[0117] In one embodiment, the fused feature representation is input into the degradation evaluation branch of the deep learning model. This branch extracts the absorption intensity of easily degradable functional groups in the near-infrared spectrum from the fused feature representation through specific operations such as convolutional layers and pooling layers. For example, the absorption intensity of easily degradable carbohydrate functional groups and recalcitrant lignin functional groups in the near-infrared spectrum is extracted. Simultaneously, an image texture roughness index is extracted from the image data using an image feature extraction algorithm, such as the Local Binary Pattern (LBP) algorithm. Assuming the absorption intensity of easily degradable functional groups extracted from the near-infrared spectrum is 0.8, the absorption intensity of recalcitrant functional groups is 0.2, and the image texture roughness index extracted from the image is 0.6, extracting these features directly related to degradation potential can provide a basis for subsequent degradation potential assessment.

[0118] G2: Analyze the ratio of absorption intensity of readily degradable functional groups to recalcitrant functional groups in the near-infrared spectrum, and calculate the spectral degradation tendency score based on the ratio of absorption intensity.

[0119] In this embodiment, the ratio of absorption intensities of readily degradable functional groups to those of recalcitrant functional groups refers to the ratio of the absorption intensity of readily degradable functional groups to that of recalcitrant functional groups in the near-infrared spectrum. The spectral degradation tendency score is a score calculated based on this ratio and is used to assess the degradation tendency of waste at the spectral level.

[0120] In one embodiment, the ratio of absorption intensities of readily degradable functional groups to recalcitrant functional groups in the near-infrared spectrum is calculated. For example, the absorption intensity of readily degradable functional groups is 0.8, and the absorption intensity of recalcitrant functional groups is 0.2, resulting in an absorption intensity ratio of 4. A formula is established to calculate the spectral degradation tendency score, such as: Spectral degradation tendency score = Absorption intensity ratio / (Absorption intensity ratio + 1). Substituting the absorption intensity ratio of 4 into the formula, we obtain the spectral degradation tendency score = 4 / (4 + 1) = 0.8. By calculating the spectral degradation tendency score, the degradation potential of waste can be assessed from the perspective of near-infrared spectroscopy.

[0121] G3: Calculate the structural degradation difficulty score based on the image texture roughness index.

[0122] In one embodiment, a function related to an image texture roughness index is defined to calculate the structural degradation difficulty score. For example, assume the structural degradation difficulty score = 1 - image texture roughness index. Given an image texture roughness index of 0.6, then the structural degradation difficulty score = 1 - 0.6 = 0.4. By calculating the structural degradation difficulty score, the degradation difficulty of waste can be assessed from the perspective of image texture.

[0123] G4: The spectral degradation tendency score and the structural degradation difficulty score are weighted and fused to generate the final organic molecular chain breakage tendency score.

[0124] In one embodiment, the weight of the spectral degradation tendency score is set to 0.6, and the weight of the structural degradation difficulty score is set to 0.4. Given that the spectral degradation tendency score = 0.8 and the structural degradation difficulty score = 0.4, the final organic molecular chain breakage tendency score is calculated through weighted fusion: Organic molecular chain breakage tendency score = 0.6 × 0.8 + 0.4 × 0.4 = 0.64. Generating the final organic molecular chain breakage tendency score through weighted fusion allows for a more accurate assessment of the degradation potential of waste by comprehensively considering both spectral and structural factors.

[0125] S34: Based on the organic molecular chain breakage tendency score, environmental temperature and humidity parameters are introduced as adjustment factors for weighted correction to generate the normalized bioactivity index.

[0126] In this embodiment, the adjustment factor refers to the parameter used to adjust the score of the tendency of organic molecular chains to break, which is the environmental temperature and humidity parameter. The normalized bioactivity index refers to the bioactivity index after weighted correction and normalization, and its value ranges from 0 to 1.

[0127] In one embodiment, an organic molecular chain breakage tendency score is used as a basis, and then weighted and corrected according to environmental temperature and humidity parameters. For example, when the ambient temperature is high and the humidity is suitable, the activity of microorganisms is high, and the organic molecular chain breakage tendency score can be appropriately increased. A weighted correction formula is set to combine the organic molecular chain breakage tendency score and environmental temperature and humidity parameters. Then, the corrected score is normalized so that its value ranges between 0 and 1, generating a normalized bioactivity index. By introducing environmental temperature and humidity parameters for weighted correction, the degree of microbial availability of waste can be more accurately reflected.

[0128] S35: Verify the consistency between the degradation characteristics implied by the component ratio matrix and the degradation difficulty indicated by the bioactivity index; if the difference between the two exceeds a preset threshold, trigger the feature backtracking recalculation process to re-execute step S31; if they are consistent, output the final component ratio matrix and bioactivity index.

[0129] In this embodiment, consistency verification refers to checking whether the degradation characteristics implied by the component ratio matrix and the degradation difficulty indicated by the bioactivity index are consistent. The preset threshold is a pre-set standard for judging whether the difference between the two is too large. The feature backtracking recalculation process refers to the process of recalculating and processing the features.

[0130] In one embodiment, the proportion of different substances in the component ratio matrix is ​​analyzed to infer their implicit degradation characteristics. For example, a high proportion of fibrous materials may indicate greater degradation difficulty. This proportion is then compared with the degradation difficulty indicated by the bioactivity index. If the difference exceeds a preset threshold—for example, if the component ratio matrix shows greater degradation difficulty but the bioactivity index shows less, and the difference exceeds a preset 20%—then a feature backtracking recalculation process is triggered, and step S31 is re-executed to reweight and fuse the features. If the two are consistent, the final component ratio matrix and bioactivity index are output. Consistency verification ensures the accuracy and consistency of the component ratio matrix and bioactivity index.

[0131] The deep learning model in this application embodiment includes a multi-branch attention mechanism, a component classification branch, and a degradation evaluation branch.

[0132] The multi-branch attention mechanism serves as the front-end processing module of the model, receiving structured feature vectors as input. This mechanism contains multiple attention branches, each corresponding to a different data modality (such as image texture features, spectral grouping statistics, etc.). Each branch performs a linear transformation on the input features through a fully connected layer. Taking the image texture feature branch as an example, the input image texture feature F_image first undergoes a fully connected layer transformation, with transformation matrices W_q, W_k, and W_v, to obtain the query vector. key vector value vector Then, the attention weights are obtained by calculating the dot product of the query vector and the key vector and applying it through the softmax function, as shown in the formula. Where d_k is the dimension of the key vector. The value vector is then weighted and summed using attention weights to obtain the output features of this branch. The output features of each branch are fused through a concatenation operation to generate an optimized fused feature representation F_fused.

[0133] The component classification branch receives the optimized fused feature representation F_fused as input. This branch contains multiple fully connected layers and a softmax layer. The fused feature representation first undergoes feature transformation and non-linear activation through several fully connected layers, such as an FC_1 layer (with the ReLU activation function) to obtain... Then, after passing through the FC_2 layer, we get... The feature dimensions are gradually adjusted to match the number of output categories. Finally, a softmax layer converts the output into soft classification probabilities for each substance category. .

[0134] The degradation assessment branch also receives an optimized fused feature representation F_fused as input. This branch contains convolutional layers, pooling layers, and fully connected layers. The fused feature representation first passes through convolutional layers to extract local features related to degradation potential, for example, using a 3×3 convolutional kernel to obtain F_conv = Conv (F_fused, Kernel_3x3). Then, it undergoes feature dimensionality reduction through pooling layers (such as max pooling) to obtain F_pool = MaxPooling (F_conv). Next, the fully connected layers map the features to dimensions related to degradation assessment metrics, outputting an organic molecular chain breakage tendency score. .

[0135] The multi-branch attention mechanism is connected in series with the component classification branch and the degradation assessment branch. The output of the multi-branch attention mechanism serves as the input to the latter two branches. The component classification branch and the degradation assessment branch are processed in parallel to jointly complete the task of inferring the waste component distribution and assessing degradation potential.

[0136] In the deep learning model training process of this application embodiment, data preparation is first performed by collecting a large amount of multimodal sensing data of kitchen waste as a training dataset, and labeling the true component ratio and degradation potential index of each sample. The dataset is divided into a 70% training set, a 15% validation set, and a 15% test set. Then, the weights and biases of each layer of the model are randomly initialized, such as the weight matrix W and bias vector b of the fully connected layer.

[0137] During training, samples from the training set are input into the model, which sequentially passes through a multi-branch attention mechanism, a component classification branch, and a degradation evaluation branch to obtain prediction results. In the multi-branch attention mechanism, each data modality branch transforms the input features into query, key, and value vectors through a fully connected layer. For example, image texture features F_image are obtained after passing through a fully connected layer. , Then calculate the attention weights. The weighted summation yields the branch outputs, which are then concatenated to form an optimized fused feature representation, F_fused. The component classification branch receives F_fused and outputs a soft classification probability P_class after passing through multiple fully connected layers and a softmax layer. The degradation assessment branch receives F_fused and outputs an organic molecule chain breakage tendency score S_break after convolution, pooling, and fully connected layers. A loss function is defined; the component classification branch uses the cross-entropy loss function. The degradation evaluation branch uses the mean squared error loss function. Total loss function Based on the total loss function, the gradients of the parameters in each layer are calculated using the backpropagation algorithm. and The Adam optimizer (learning rate η = 0.001) is used to update parameters. During training, the weights of the component classification branch loss function can be adjusted according to the importance of different substance categories. For example, if the weight of oil is set to w_oil, the adjusted loss function is... Evaluate model performance on the validation set and adjust hyperparameters. Repeat the above steps until the model performance is satisfactory. Finally, calculate accuracy, recall, and mean absolute error on the test set. Indicators such as these are used to assess generalization ability.

[0138] In one embodiment, reference Figure 7 Step S40 may include steps S41-S45, which will be described in detail below: S41: Query the current motor load, chamber temperature and remaining processing capacity of the processing equipment to generate a snapshot of the equipment's operating status.

[0139] In this embodiment, the processing equipment refers to equipment for processing kitchen waste, such as a crusher or fermentation tank. Motor load refers to the working load of the motor in the processing equipment. Chamber temperature refers to the internal temperature of the processing equipment. Remaining processing capacity refers to the amount of waste the processing equipment can still hold. Equipment operating condition snapshot refers to a record of the operating status of the processing equipment at a certain moment.

[0140] In one embodiment, the system communicates with the control system of the processing equipment to query the current motor load, chamber temperature, and remaining processing capacity. For example, a query command is sent to the control system of the crusher to obtain the motor current value, and the motor load is calculated based on the current value. The chamber temperature inside the fermenter is obtained through a temperature sensor. The remaining space or amount of waste inside the processing equipment is obtained through a level sensor or weighing device, and the remaining processing capacity is calculated. This information is integrated to form a snapshot of the equipment's operating status.

[0141] S42: When the proportion of fiber in the component ratio matrix is ​​higher than the preset threshold and the bioactivity index is lower than the median level, set the coarse crushing mode and extend the premixing time.

[0142] In this embodiment, the preset threshold is a pre-set standard for judging whether the proportion of fibrous materials is too high. The median level is the median value of the bioactivity index. Coarse crushing mode refers to using a larger particle size when crushing waste. Premixing time refers to the time for mixing waste before fermentation.

[0143] In one embodiment, a preset threshold for the proportion of fibrous materials is set to 30%, and the median level of the bioactivity index is set to 0.5. When the proportion of fibrous materials in the component ratio matrix is ​​higher than 30% and the bioactivity index is lower than 0.5, a coarse crushing mode is set. For example, the crusher's particle size is set to a larger value, such as 5 cm. At the same time, the premixing time is extended, for example, from the original 10 minutes to 20 minutes. By setting a coarse crushing mode and extending the premixing time, the treatment effect of fibrous materials can be improved, making them easier for microorganisms to decompose.

[0144] S43: Dynamically compress the fermentation temperature rise slope according to the proportion of oily substances to avoid local anaerobic acidification and generate a stepped temperature control curve.

[0145] In this embodiment, the proportion of oily substances refers to the percentage of oily substances in the component proportion matrix. The fermentation temperature rise slope refers to the rate at which the temperature rises during fermentation. Localized anaerobic acidification refers to the acidification phenomenon that occurs in localized areas during fermentation due to insufficient oxygen. A stepped temperature control curve refers to dividing the fermentation process into multiple stages, each with a different temperature setpoint, forming a stepped temperature control curve.

[0146] In one embodiment, reference Figure 8 Step S43 may include steps S431-S435, which will be described in detail below: S431: Read the percentage of oily substances in the component ratio matrix and determine whether it is in a high-risk range.

[0147] In this embodiment, the component ratio matrix is ​​a matrix representing the proportions of different substances in the waste. The percentage of oily substances refers to the proportion of oily substances in the component ratio matrix. The high-risk zone is a pre-defined range of oily substance percentages; when the percentage of oily substances is within this range, problems such as localized anaerobic acidification may occur.

[0148] In one embodiment, the percentage of oily substances is read from a component ratio matrix. For example, if the component ratio matrix is ​​[0.6, 0.2, 0.2], then the percentage of oily substances is 0.2. A high-risk range is defined as an oily substance percentage greater than 0.15. It is then determined whether the read oily substance percentage falls within this high-risk range. In this example, 0.2 is greater than 0.15, so the oily substance percentage falls within the high-risk range. By determining whether the oily substance percentage falls within the high-risk range, potential problems can be identified promptly, providing a basis for adjusting subsequent processing strategies.

[0149] S432: If in a high-risk zone, multiply the standard heating slope by a scaling factor less than one to obtain the heating rate after compression.

[0150] In this embodiment, the standard temperature rise rate refers to the rate of temperature increase during fermentation under normal conditions. The scaling factor is a value less than 1 used to compress the standard temperature rise rate. The compressed temperature rise rate is the temperature rise rate obtained by multiplying the standard temperature rise rate by the scaling factor.

[0151] In one embodiment, when the proportion of oily substances is determined to be in a high-risk range, a scaling factor is set to 0.8. Assuming a standard heating rate of 5°C per hour, multiplying the standard heating rate by the scaling factor of 0.8 yields a compressed heating rate of 4°C per hour. By compressing the heating rate, localized anaerobic acidification caused by an excessively high proportion of oily substances can be avoided, making the fermentation process more stable.

[0152] S433: The entire fermentation cycle is divided into three stages: the initial adaptation stage, the main reaction stage, and the maturation stage, with different target temperature platforms assigned to each stage.

[0153] In this embodiment, the fermentation cycle refers to the entire time period from the start to the end of fermentation. The initial adaptation phase refers to the initial stage of fermentation, where microorganisms adapt to the environment. The main reaction phase refers to the main stage during fermentation where microorganisms multiply rapidly and decompose organic matter. The maturation phase refers to the later stage of fermentation, where the fermentation products are further processed and stabilized. The target temperature plateau refers to the temperature that needs to be reached and maintained at each stage.

[0154] In one embodiment, the entire fermentation cycle is divided into three stages: an initial adaptation stage, a main reaction stage, and a maturation stage. A target temperature plateau is assigned to the initial adaptation stage at 30°C, the main reaction stage at 50°C, and the maturation stage at 40°C. In the initial adaptation stage, the lower temperature allows the microorganisms to gradually adapt to the environment; in the main reaction stage, the higher temperature promotes microbial growth and metabolism; and in the maturation stage, a moderate temperature stabilizes the fermentation products. By dividing the fermentation cycle into stages and assigning target temperature plateaus, the fermentation process can be optimized, and the fermentation effect improved.

[0155] S434: Insert a constant temperature maintenance period between adjacent temperature platforms. The duration of this period is proportional to the proportion of fat to ensure that lipase can function fully.

[0156] In this embodiment, the isothermal maintenance period refers to the time during which the temperature remains constant between adjacent temperature plateaus. Lipase is an enzyme capable of breaking down fats and oils, and it can function effectively at a suitable temperature.

[0157] In one embodiment, when the temperature is increased from 30°C in the initial adaptation phase to 50°C in the main reaction phase, a constant temperature maintenance period is inserted. Assuming the proportion of lipids is 0.2%, the duration of the constant temperature maintenance period is calculated to be 4 hours based on a preset ratio. During these 4 hours, the temperature is kept constant, allowing lipase sufficient time to decompose lipids. By inserting a constant temperature maintenance period, the lipase can be ensured to function fully, thereby improving the decomposition efficiency of lipids.

[0158] S435: Combines the heating rate, target temperature and constant temperature duration of each segment into a sequence of temperature setpoints at discrete time points, forming a stepped temperature control curve.

[0159] In this embodiment, the heating rate refers to the speed at which the temperature rises at each stage. The target temperature refers to the temperature that needs to be reached at each stage. The isothermal duration refers to the time during which the temperature is kept constant between adjacent temperature plateaus. The discrete time point temperature setpoint sequence is a series of temperature setpoints arranged in chronological order, consisting of the heating rate, target temperature, and isothermal duration. The stepped temperature control curve is a curve composed of these temperature setpoints, exhibiting a stepped shape.

[0160] In one embodiment, the heating rate and target temperature (30°C) of the initial adaptation stage, the heating rate and target temperature (50°C) of the main reaction stage, the heating rate and target temperature (40°C) of the maturation stage, and the isothermal duration between adjacent temperature plateaus are combined. For example, the heating rate of the initial adaptation stage is 2°C per hour, requiring 5 hours to rise from 20°C to 30°C; the isothermal maintenance period is 2 hours; the heating rate of the main reaction stage is 4°C per hour, requiring 5 hours to rise from 30°C to 50°C; the isothermal maintenance period is 4 hours; and the heating rate of the maturation stage is 1°C per hour, requiring 10 hours to cool from 50°C to 40°C. These time points and their corresponding temperature setpoints are combined to form a discrete time-point temperature setpoint sequence, such as [20℃ (0 hours), 30℃ (5 hours), 30℃ (7 hours), 50℃ (12 hours), 50℃ (16 hours), 40℃ (26 hours)], thus creating a stepped temperature control curve. By constructing this stepped temperature control curve, temperature changes during fermentation can be precisely controlled, ensuring optimal fermentation results.

[0161] S44: The stirring frequency setting is matched according to the bioactivity index. High activity corresponds to high frequency intermittent stirring, and low activity corresponds to low frequency continuous stirring.

[0162] In this embodiment, the bioactivity index reflects the activity level of microorganisms in the waste. The stirring frequency setting refers to different stirring frequency settings of the stirring device in the treatment equipment. High-frequency intermittent stirring means that the stirring device performs intermittent stirring at a high frequency. Low-frequency continuous stirring means that the stirring device performs continuous stirring at a low frequency.

[0163] In one embodiment, the bioactivity index is divided into three ranges: high, medium, and low, each corresponding to a different stirring frequency level. When the bioactivity index is in the high range (e.g., greater than 0.8), the stirring device is set to a high-frequency intermittent stirring mode, for example, stirring for 1 minute every 5 minutes. When the bioactivity index is in the low range (e.g., less than 0.2), the stirring device is set to a low-frequency continuous stirring mode, for example, stirring for 1 minute every 30 minutes. By matching the stirring frequency level according to the bioactivity index, the stirring strategy can be adjusted according to the actual degradation of the waste, improving the contact efficiency between microorganisms and waste and promoting the degradation process.

[0164] S45: Arrange the crushing, temperature control and stirring strategies in chronological order into an irreversible execution queue, add safety interlock conditions, and form a phased processing instruction sequence.

[0165] In this embodiment, the irreversible execution queue refers to a series of processing strategies arranged in chronological order; once execution begins, the order cannot be arbitrarily changed. Safety interlock conditions are conditions set to ensure the safety of the processing process, such as stopping stirring when the temperature is too high. The staged processing instruction sequence is an instruction sequence that integrates crushing, temperature control, and stirring strategies to guide the operation of the processing equipment.

[0166] In one embodiment, the crushing, temperature control, and stirring strategies are sequentially arranged into an irreversible execution queue according to the time sequence of the processing flow. For example, coarse crushing is performed first, followed by fermentation according to a stepped temperature control curve, while the stirring frequency is adjusted based on the bioactivity index. Safety interlocking conditions are set between each operation step; for example, if the motor load is too high during crushing, the crushing operation is automatically stopped. These operation steps and safety interlocking conditions are combined to form a phased processing instruction sequence. By forming a phased processing instruction sequence, the orderly execution of the processing can be ensured, improving the safety and efficiency of the processing.

[0167] Accordingly, to better implement the above methods, this application also provides a dynamic food waste treatment system that integrates deep learning. For example... Figure 9 As shown, the deep learning-integrated dynamic food waste treatment system 80 includes:

[0168] The acquisition module 801 is used to collect multimodal sensing data during the disposal of kitchen waste. The multimodal sensing data includes visible light image sequences, near-infrared spectral signals and weight change time series, environmental temperature and humidity parameters, and the current filling status of the waste container. The vector generation module 802 is used to perform spatiotemporal alignment and noise suppression on the multimodal sensing data to generate a structured feature vector set, wherein the structured feature vector set includes a fusion representation after timestamp registration, spatial coordinate normalization and outlier removal. The deep learning module 803 is used to input the structured feature vector set into a preset deep learning model, perform waste component distribution inference and degradation potential assessment, and output a component ratio matrix and a bioactivity index, wherein the component ratio matrix represents the relative proportion of organic matter, fibrous and oily substances, and the bioactivity index reflects the degree of microbial availability. The instruction generation module 804 is used to dynamically generate a phased processing instruction sequence based on the component ratio matrix and bioactivity index, combined with the current operating conditions of the processing equipment. The phased processing instruction sequence includes a crushing particle size setting, a fermentation temperature range, and a stirring frequency adjustment strategy. The instruction sending module 805 is used to send the phased processing instruction sequence to the kitchen waste treatment terminal to complete the dynamic processing decision identifier for the current waste disposal, wherein the dynamic processing decision identifier is an executable process control signal.

[0169] The implementation details of each module are provided in the preceding method embodiments and will not be repeated here. The technical effects achieved by each module and device are described in the foregoing method embodiments.

[0170] like Figure 10 As shown, this application embodiment also provides a computer device 90, which includes a processor 901 and a memory 902, wherein the memory 902 stores a computer program, and when the computer program is executed by the processor 901, the processor 901 performs the steps of any of the methods described above.

[0171] The above-disclosed embodiments are merely preferred embodiments of this application and should not be construed as limiting the scope of this application. Therefore, any equivalent variations made in accordance with the claims of this application are still within the scope of this application.

Claims

1. A method for dynamic treatment of kitchen waste integrating deep learning, characterized in that, include: Collect multimodal sensing data during the disposal of kitchen waste, wherein the multimodal sensing data includes visible light image sequences, near-infrared spectral signals and weight change time series, environmental temperature and humidity parameters, and the current filling status of the waste container; The multimodal sensing data is spatiotemporally aligned and noise suppressed to generate a structured feature vector set, wherein the structured feature vector set includes a fusion representation after timestamp registration, spatial coordinate normalization and outlier removal. The structured feature vector set is input into a pre-set deep learning model to perform waste component distribution inference and degradation potential assessment, and output a component ratio matrix and a bioactivity index, wherein the component ratio matrix represents the relative proportion of organic matter, fibrous and oily substances, and the bioactivity index reflects the degree of microbial availability. Based on the component ratio matrix and bioactivity index, and combined with the current operating conditions of the processing equipment, a phased processing instruction sequence is dynamically generated, wherein the phased processing instruction sequence includes crushing particle size setting, fermentation temperature range and stirring frequency adjustment strategy. The phased processing instruction sequence is sent to the food waste treatment terminal to complete the dynamic processing decision identification for the current waste disposal, wherein the dynamic processing decision identification is an executable process control signal.

2. The method according to claim 1, characterized in that, The multimodal sensing data collected during the food waste disposal process includes: Trigger continuous frame image capture at the inlet of the waste disposal container to obtain a sequence of visible light images covering the complete falling trajectory; Simultaneously activate the near-infrared sensor array to scan the reflectance spectrum of the waste surface, collect the original spectral response values ​​according to the preset wavelength bands, and construct a spectral-spatial mapping table; The changes in mass increment are monitored in real time by the weighing unit at the bottom of the container, and the weight time series curve from the initial static state to the steady state is recorded, and the timestamps corresponding to the abrupt change points are marked. Read the filling height data output by the liquid level or ultrasonic ranging module built into the container and convert it into a volume occupancy rate value; The environmental monitoring unit is invoked to obtain the current air temperature and relative humidity readings, which are used as the environmental temperature and humidity parameters. Using the reference timestamp as a unified time reference, the visible light image sequence, near-infrared spectral response value, weight time series curve, volume occupancy rate value, and temperature and humidity readings are bound together to form a synchronous sensing data package with environmental context.

3. The method according to claim 2, characterized in that, The process involves spatiotemporal alignment and noise suppression of the multimodal sensing data to generate a structured feature vector set. This structured feature vector set includes a fused representation after timestamp registration, spatial coordinate normalization, and outlier removal, comprising: Using the weight mutation start time in the multimodal sensing data as the reference time origin, the image frames in the visible light image sequence and the spectral sampling points in the near-infrared spectral signal are resampled and aligned with millisecond-level precision to establish a spatiotemporal synchronized data block. Background segmentation is performed on the visible light image sequence in the spatiotemporal synchronization data block, and the viewpoint distortion is corrected by using a fixed reference object on the inner wall of the container, and a background-removed and geometrically normalized garbage area mask is output. The near-infrared spectral signals in the spatiotemporal synchronization data block are grouped according to the material absorption characteristic bands, the mean and variance of each group are calculated, and abnormal channels that deviate from the global statistical distribution by more than three standard deviations are removed. A sliding window smoothing filter is applied to the weight time-series curve in the multimodal sensing data to identify and remove high-frequency jitter components caused by vibration or airflow disturbance, and to extract the smoothed mass accumulation trend. Image texture features are extracted from the background-removed and geometrically normalized garbage area mask, spectral grouping statistics are extracted from the processed near-infrared spectral signal, the instantaneous slope of the mass accumulation trend is extracted from the smoothed weight time series curve, and the environmental temperature and humidity parameters and volume occupancy rate values ​​contained in the multimodal sensing data are obtained. The extracted image texture features, spectral grouping statistics, instantaneous slope of the cumulative mass trend, volume occupancy rate, and environmental temperature and humidity parameters are concatenated into a fixed-dimensional feature vector to form the structured feature vector set.

4. The method according to claim 1, characterized in that, The process of inputting the structured feature vector set into a pre-set deep learning model to perform waste component distribution inference and degradation potential assessment, and outputting a component proportion matrix and a bioactivity index, includes: The structured feature vectors are input into the multi-branch attention mechanism of the deep learning model, and feature weighting and fusion are performed based on the correlation between different data modalities and waste components and degradation characteristics to generate an optimized fused feature representation; Using the fused feature representation, the soft classification probability of each substance category is output through the component classification branch of the deep learning model, and the component proportion matrix is ​​generated according to the preset mapping rules. Using the fused feature representation, the degradation evaluation branch of the deep learning model is used to analyze the absorption intensity of specific functional groups in the near-infrared spectrum and the image texture roughness index, and to calculate the organic molecular chain breakage tendency score. Based on the organic molecular chain breakage tendency score, environmental temperature and humidity parameters are introduced as adjustment factors for weighted correction to generate the normalized bioactivity index. The consistency between the degradation characteristics implied by the component ratio matrix and the degradation difficulty indicated by the bioactivity index is verified. If the difference between the two exceeds a preset threshold, a feature backtracking recalculation process is triggered to re-execute the step of inputting the structured feature vector into the multi-branch attention mechanism of the deep learning model. If they are consistent, the final component ratio matrix and bioactivity index are output.

5. The method according to claim 1, characterized in that, The process of dynamically generating a phased processing instruction sequence based on the component ratio matrix and bioactivity index, combined with the current operating conditions of the processing equipment, includes: Query the current motor load, chamber temperature and remaining processing capacity of the processing equipment to generate a snapshot of the equipment's operating status; When the proportion of fiber in the component ratio matrix is ​​higher than the preset threshold and the bioactivity index is lower than the median level, the coarse crushing mode is set and the premixing time is extended. The temperature rise slope of the dynamic compression fermentation is determined by the proportion of oily substances to avoid local anaerobic acidification and generate a stepped temperature control curve. The stirring frequency is matched according to the level of the bioactivity index: high activity corresponds to high frequency intermittent stirring, and low activity corresponds to low frequency continuous stirring. The crushing, temperature control, and stirring strategies are arranged in chronological order into an irreversible execution queue, with added safety interlock conditions, forming a phased processing instruction sequence.

6. The method according to claim 3, characterized in that, The near-infrared spectral signals in the spatiotemporal synchronization data block are grouped according to the material absorption characteristic bands, the mean and variance of each group are calculated, and abnormal channels that deviate from the global statistical distribution by more than three standard deviations are removed, including: The full-band near-infrared signal in the spatiotemporal synchronization data block is divided into three logical groups: water molecule absorption region, carbon-hydrogen bond vibration region, and stray light interference region. The arithmetic mean of all channels in each group is calculated. A local Gaussian distribution model is constructed based on the average values ​​of each group to determine the absolute deviation of each channel from the mean of its group. Calculate the standard deviation of the deviation of all channels globally, and set the elimination threshold to three times the standard deviation. Channels whose deviation exceeds the removal threshold are marked as abnormal channels, and an index list of abnormal channels is generated. Remove the band data corresponding to the abnormal channel index from the original spectral signal, and retain the valid channels.

7. The method according to claim 4, characterized in that, The use of the fusion feature representation The component classification branch of the deep learning model outputs the soft classification probability of each substance category, and generates the component proportion matrix according to a preset mapping rule, including: Based on the fused feature representation, initial soft classification probability values ​​for organic matter, fiber, and oil are obtained through the component classification branch of the deep learning model; The initial soft classification probability value is normalized, and the normalized probability value is used as the mass proportion estimate of the corresponding material category to generate a preliminary component proportion vector. The preliminary component ratio vector is corrected based on actual detection deviations in historical feedback data of similar waste treatment. The corrected component proportions are encapsulated into a standardized triplet matrix and output as the component proportion matrix.

8. The method according to claim 5, characterized in that, The process of dynamically compressing the fermentation temperature rise slope according to the proportion of oily substances to avoid local anaerobic acidification and generate a step-by-step temperature control curve includes: Read the percentage of oily substances in the component ratio matrix to determine whether it is in a high-risk range; If it is in a high-risk zone, multiply the standard heating slope by a scaling factor of less than one to obtain the heating rate after compression. The entire fermentation cycle is divided into three stages: the initial adaptation stage, the main reaction stage, and the maturation stage, with different target temperature platforms assigned to each stage. A constant temperature maintenance period is inserted between adjacent temperature platforms, the duration of which is proportional to the proportion of oil, to ensure that lipase can fully function. The heating rate, target temperature, and constant temperature duration of each segment are combined into a sequence of temperature setpoints at discrete time points, forming a stepped temperature control curve.

9. The method according to claim 4, characterized in that, The degradation potential assessment is performed in parallel based on the fused feature representation. Through the degradation assessment branch of the model, the organic molecular chain breakage tendency score is calculated by analyzing the absorption intensity of specific functional groups in the near-infrared spectrum and the image texture roughness index, including: Based on the fusion feature representation, features directly related to degradation potential are extracted through degradation assessment branches, including the absorption intensity of easily degradable functional groups in the near-infrared spectrum and image texture roughness index. The ratio of absorption intensity of readily degradable functional groups to recalcitrant functional groups in the near-infrared spectrum was analyzed, and a spectral degradation tendency score was calculated based on the ratio of absorption intensity. The structural degradation difficulty score is calculated based on the image texture roughness index. The spectral degradation tendency score and the structural degradation difficulty score are weighted and fused to generate the final organic molecular chain breakage tendency score.

10. A dynamic kitchen waste treatment system integrating deep learning, characterized in that, The system includes: The acquisition module is used to collect multimodal sensing data during the disposal of kitchen waste. The multimodal sensing data includes visible light image sequences, near-infrared spectral signals and weight change time series, environmental temperature and humidity parameters, and the current filling status of the waste container. The vector generation module is used to perform spatiotemporal alignment and noise suppression on the multimodal sensing data to generate a structured feature vector set, wherein the structured feature vector set includes a fusion representation after timestamp registration, spatial coordinate normalization and outlier removal. The deep learning module is used to input the structured feature vector set into a preset deep learning model, perform waste component distribution inference and degradation potential assessment, and output a component ratio matrix and a bioactivity index, wherein the component ratio matrix represents the relative proportion of organic matter, fibrous and oily substances, and the bioactivity index reflects the degree of microbial availability. The instruction generation module is used to dynamically generate a phased processing instruction sequence based on the component ratio matrix and bioactivity index, combined with the current operating conditions of the processing equipment. The phased processing instruction sequence includes a crushing particle size setting, fermentation temperature range, and stirring frequency adjustment strategy. The instruction sending module is used to send the phased processing instruction sequence to the kitchen waste treatment terminal to complete the dynamic processing decision identifier for the current waste disposal, wherein the dynamic processing decision identifier is an executable process control signal.

Images