High-efficiency intelligent sorting system for low-value recoverable living-source materials

Abstract

The invention relates to the technical field of recoverable material sorting, and discloses an efficient intelligent sorting system for life source low-value recoverable materials, which comprises an acquisition unit used for monitoring environmental parameters of an urban low-value recoverable material sorting center and the speed of a conveyor belt in real time, synchronously acquiring voiceprint data, three-dimensional point cloud data and hyperspectral data of the recoverable materials, and sending the voiceprint data, the three-dimensional point cloud data and the hyperspectral data to a cloud server; through close cooperation and data interaction among all the functional units, rapid and comprehensive analysis and accurate classification of low-value recoverable materials are achieved, under the complex working condition of a large-scale solid waste treatment scene, the system can recognize and judge the recoverable materials conveyed at a high speed and then classify and sort the recoverable materials, and in the process, the recoverable materials are classified and sorted, so that the recovery efficiency of the recoverable materials is improved. According to the method, the interference of external conveyor belt speed change and environmental factor fluctuation on identification can be reduced, so that the accuracy of overall classification is ensured, and the overall identification effect can completely meet the requirements of mixing and sorting of recyclable materials under complex working conditions.

Description

Technical Field

[0001] The present invention relates to the technical field of sorting recyclables, and particularly to an efficient and intelligent sorting system for low-value recyclables from domestic sources. Background Art

[0002] In the sorting process of low-value recyclables from domestic sources, in order to improve the recognition efficiency and accuracy, the conventional operation is to equip a resonance cavity directly above the conveyor belt. When various low-value recyclables move along the conveyor belt and hit the resonance cavity, a unique sound will be generated. Through a professional voiceprint analysis system, key features such as the frequency and amplitude of the sound wave are quickly analyzed, thereby accurately determining the material type of the recyclables and achieving efficient classification. For some leading sorting systems, in addition to relying on voiceprint recognition, three-dimensional contour scanning is also introduced to assist in shape recognition, and the appearance features such as the shape and volume of the recyclables are carefully scanned and analyzed from different angles to ensure the high quality and efficiency of the sorting operation.

[0003] However, in large-scale solid waste treatment scenarios, such as urban low-value recyclable sorting centers, low-value recyclables often exist in a mixed form, and need to be continuously and rapidly conveyed during the classification process. During this process, the voiceprint is easily interfered by external factors, resulting in inaccurate voiceprint analysis. Even if three-dimensional contour scanning is introduced to assist in shape recognition, the overall recognition effect is still difficult to meet the requirements of mixed sorting of recyclables under complex working conditions. Summary of the Invention

[0004] Aiming at the deficiencies of the prior art, the present invention provides an efficient and intelligent sorting system for low-value recyclables from domestic sources, which solves the above problems.

[0005] The above technical objectives of the present invention are achieved through the following technical solutions:

[0006] An efficient and intelligent sorting system for low-value recyclables from domestic sources, comprising:

[0007] An acquisition unit for real-time monitoring of the environmental parameters and conveyor belt speed of an urban low-value recyclable sorting center, and synchronously acquiring the voiceprint data, three-dimensional point cloud data, and hyperspectral data of recyclables;

[0008] A compensation unit for compensating the voiceprint data through environmental parameters to obtain compensated voiceprint data;

[0009] A correction unit for adjusting the linear array scanning sampling frequency in real time according to the conveyor belt speed and correcting the three-dimensional point cloud data to generate compensated three-dimensional shape data;

[0010] An analysis unit for analyzing the material composition of the compensated voiceprint data to obtain material attributes;

[0011] A fusion unit, configured to fuse the compensated three-dimensional morphological data with the material attributes to generate a primary classification result;

[0012] A matching unit, configured to perform material composition analysis on the hyperspectral data to generate a spectral matching label, and correct the primary classification result through the spectral matching label to obtain a secondary classification result;

[0013] A determination unit, configured to determine the secondary classification result to obtain a final classification result.

[0014] Further, synchronously collect the voiceprint data, three-dimensional point cloud data, and hyperspectral data of recyclables, including:

[0015] The voiceprint data refers to the voiceprint information generated and recorded when the recyclables collide in the resonance cavity;

[0016] The three-dimensional point cloud data refers to the data presenting the three-dimensional morphological characteristics of the recyclables through 4K line array scanning of the recyclables;

[0017] The hyperspectral data refers to the data covering the recyclables in different spectral bands obtained by photographing the recyclables with a hyperspectral camera.

[0018] Further, compensate the voiceprint data through environmental parameters to obtain compensated voiceprint data, including:

[0019] Extract the environmental parameters to obtain the acoustic attenuation coefficient;

[0020] Construct an adaptive filter based on the acoustic attenuation coefficient to generate a frequency-domain pure voiceprint;

[0021] Calculate the conveyor belt speed to obtain the acoustic wave propagation time delay;

[0022] Use the acoustic wave propagation time delay to correct the frequency-domain pure voiceprint to generate a speed-synchronized energy parameter;

[0023] Fuse the acoustic attenuation coefficient, the speed-synchronized energy parameter, and the spatial reflection characteristics of the three-dimensional point cloud data to generate anti-interference compensated voiceprint data.

[0024] Further, adjust the line array scanning sampling frequency in real time according to the conveyor belt speed, and correct the three-dimensional point cloud data to generate compensated three-dimensional morphological data, including:

[0025] Construct an acceleration-resonance coupling function based on the conveyor belt speed to generate a dynamic frequency parameter linked to the speed change;

[0026] Perform temporal resampling on the three-dimensional point cloud data through the dynamic frequency parameter to generate a geometric distortion factor representing speed-related deformation;

[0027] Spatially map and match the geometric distortion factor with the hyperspectral data to generate spatial calibration weights for eliminating the spatio-temporal offset of multi-dimensional data.

[0028] Perform weighted fusion of the dynamic frequency parameter, geometric distortion factor, and spatial calibration weights to generate compensated three-dimensional morphological data resistant to motion blur.

[0029] Furthermore, analyze the material composition of the compensated voiceprint data to obtain material properties, including:

[0030] Conduct time-frequency joint analysis on the compensated voiceprint data to generate multi-scale resonance factors that can characterize the material resonance characteristics.

[0031] Based on the multi-scale resonance factors, dynamically track the non-linear attenuation path of sound waves and materials, and extract dynamic transfer entropy values that can reflect the material density.

[0032] Perform feature fusion on the multi-scale resonance factors and dynamic transfer entropy values, construct a material response surface for sound wave energy attenuation, and generate a material attenuation fingerprint.

[0033] Match the material attenuation fingerprint with the attenuation characteristics in the material database to generate material properties including material category and thickness.

[0034] Furthermore, fuse the compensated three-dimensional morphological data with the material properties to generate a primary classification result, including:

[0035] Extract the dynamic geometric parameters of recyclables based on the compensated three-dimensional morphological data, and construct a morphology-material association matrix in combination with the material properties. Specifically, from the compensated three-dimensional morphological data, use a three-dimensional feature extraction algorithm to extract the three dynamic geometric parameters of the length, width, and height of the recyclables, combine the thickness parameter in the material properties with the three dynamic geometric parameters of the length, width, and height of the recyclables, construct a matrix with geometric parameters as rows and thickness parameters as columns, and each element represents the association relationship between the corresponding geometric parameter and thickness, thus forming a morphology-material association matrix.

[0036] Expand the morphology-material association matrix to generate multi-physical field coupling parameters including material density gradient and deformation coefficient.

[0037] Input the multi-physical-field coupling parameters into the deep residual network, and based on the fusion weights of three-dimensional morphology and acoustic features, output the primary classification results including probability confidence, specifically: After preprocessing the multi-physical-field coupling parameters by standardization, use them as the initial input features to access the first layer of the deep residual network. Through the residual blocks, perform cross-layer feature extraction and non-linear mapping. At the same time, after the three-dimensional morphological features pass through 3D convolution and the acoustic features pass through Mel-frequency cepstral coefficient extraction, generate feature vectors through independent sub-networks respectively. Use the attention mechanism to calculate the fusion weights of the two, perform weighted aggregation on the feature vectors, then input the fused features into the fully connected layer, and calculate the probability confidence of each classification through the Softmax function to generate the primary classification results including probability confidence.

[0038] Furthermore, perform material composition analysis on the hyperspectral data to generate spectral matching labels, including:

[0039] Dynamically compensate the optical path offset of the hyperspectral data based on the conveyor belt speed to generate a spectral-material association vector, specifically including: Calculate the spectral pixel displacement amount through the conveyor belt speed combined with the sampling frequency of the hyperspectral instrument, use the interpolation algorithm to perform displacement compensation on the original hyperspectral data to achieve dynamic correction of the optical path offset, and simultaneously match the corrected spectral data with the spectral features in the material sample library to generate a complete spectral-material association vector;

[0040] Perform cross-modal feature enhancement on the spectral-material association vector and the material attenuation fingerprint to obtain cross-modal fusion weights resistant to noise interference;

[0041] Perform spatio-temporal alignment on the cross-modal fusion weights to generate spectral matching labels, where the spectral matching labels contain gradient encodings of material component proportions and pollution degrees, specifically including: Decompose the cross-modal fusion weights through the matrix decomposition algorithm, extract their spatial and temporal feature components, multiply the spatial calibration weights element-wise with them, and use the alternating direction multiplier method for optimization and solution to align the fusion weights in the spatial and temporal dimensions. Use the decomposed weight matrix to perform weighted summation on the spectral-material association vector, and use the non-negative matrix factorization algorithm to decompose it into the product of a material component proportion matrix and a pollution degree gradient encoding matrix. Normalize the component proportion matrix column-wise and the pollution degree matrix row-wise, and finally generate spectral matching labels containing gradient encodings of material component proportions and pollution degrees.

[0042] Furthermore, correct the primary classification results through the spectral matching labels to obtain secondary classification results, including:

[0043] Extract the gradient encoding in the spectral matching labels, decompose it into two parameters, namely the material component proportion parameter and the pollution degree parameter, and construct a dynamic correction coefficient matrix based on the two parameters;

[0044] Non-linearly weightedly fuse the dynamic correction coefficient matrix with the probability confidence of the primary classification result to generate a cross-modal confidence compensation parameter;

[0045] Based on the spatial calibration weight, perform spatio-temporal consistency constraints on the cross-modal confidence compensation parameter, and output the secondary classification result.

[0046] Furthermore, judge the secondary classification result to obtain the final classification result, including:

[0047] Generate a dynamic confidence offset parameter based on the cross-modal confidence compensation parameter and the material attenuation fingerprint;

[0048] Generate a dynamic threshold matrix based on the physical relevance between the dynamic confidence offset parameter and the material properties.

[0049] Furthermore, judge the secondary classification result to obtain the final classification result, which also includes:

[0050] Non-linearly couple the dynamic threshold matrix with the gradient encoding of the spectral matching label to generate a classification optimization parameter;

[0051] Perform multi-dimensional confidence calibration on the secondary classification result according to the classification optimization parameter, and generate the final classification result including the material type, morphological characteristics and pollution coefficient.

[0052] In summary, the present invention mainly has the following beneficial effects:

[0053] Through the acquisition and fusion of multi-modal data, the problem of accurate sorting of low-value recyclables under complex working conditions is effectively solved. By synchronously acquiring voiceprint, three-dimensional point cloud and hyperspectral data through the acquisition unit, a rich information basis is provided for subsequent analysis. The compensation unit and the correction unit respectively perform accurate correction on the voiceprint data and the three-dimensional point cloud data to eliminate the influence of environmental interference and conveyor belt speed changes, making the data more conform to the characteristics of actual recyclables. The analysis unit deeply analyzes the material composition of the voiceprint data, and the fusion unit then organically combines the three-dimensional morphological data with the material properties to generate a primary classification result including probability confidence, and its accuracy is greatly improved compared with the traditional single recognition method. Subsequently, the matching unit introduces the material composition analysis of the hyperspectral data to generate a spectral matching label to correct the primary result, further improving the classification accuracy. Finally, through the comprehensive judgment of the judgment unit, the final classification result including multi-dimensional information is obtained, which can accurately distinguish recyclables of different materials, thicknesses, morphologies and pollution degrees, improve the accuracy of classification and sorting, and effectively meet the requirements of fine classification and sorting of low-value recyclables in large-scale solid waste treatment scenarios.

[0054] In the complex and ever-changing environment of the urban low-value recyclable sorting center, the compensation unit accurately extracts the acoustic attenuation coefficient based on the real-time monitored environmental parameters and constructs an adaptive filter. At the same time, combined with the correction of the acoustic wave propagation time delay, it generates anti-interference compensated voiceprint data to ensure the reliability of voiceprint analysis in different environments. The correction unit dynamically adjusts the linear array scanning sampling frequency according to the conveyor belt speed. Through complex operations such as constructing an acceleration-resonance coupling function, it generates compensated three-dimensional shape data that resists motion blur, effectively coping with the problem of shape data distortion caused by the fluctuation of the conveyor belt speed, and ensuring the stable data acquisition and analysis ability of the system during high-speed continuous conveying. In addition, the units cooperate closely, and the real-time compensation and correction mechanism for environmental and motion factors is integrated into the entire process from data acquisition to final determination. In the solid waste treatment scenario with long-term, high-intensity, and complex working conditions, the system can still stably and continuously output accurate classification results, reducing the probability of sorting errors caused by environmental changes or equipment operation fluctuations, and improving the efficiency and stability of the entire solid waste treatment process. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 is a block diagram of the high-efficiency intelligent sorting system for low-value recyclables from domestic sources of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0056] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0057] Refer to Figure 1 , the high-efficiency intelligent sorting system for low-value recyclables from domestic sources includes:

[0058] An acquisition unit for real-time monitoring of the environmental parameters and conveyor belt speed of the urban low-value recyclable sorting center, and synchronously acquiring the voiceprint data, three-dimensional point cloud data, and hyperspectral data of recyclables;

[0059] A compensation unit for compensating the voiceprint data through environmental parameters to obtain compensated voiceprint data;

[0060] A correction unit for adjusting the linear array scanning sampling frequency in real time according to the conveyor belt speed and correcting the three-dimensional point cloud data to generate compensated three-dimensional shape data;

[0061] An analysis unit for analyzing the material composition of the compensated voiceprint data to obtain material properties;

[0062] A fusion unit for fusing the compensated three-dimensional morphological data with the material properties to generate a primary classification result;

[0063] A matching unit for performing material composition analysis on the hyperspectral data to generate a spectral matching label, and correcting the primary classification result through the spectral matching label to obtain a secondary classification result;

[0064] A determination unit for determining the secondary classification result to obtain a final classification result.

[0065] The acquisition unit monitors the environmental parameters and conveyor belt speed of the urban low-value recyclable sorting center in real time, and synchronously acquires the voiceprint, three-dimensional point cloud, and hyperspectral data of the recyclables, providing rich and real-time data support for subsequent processing. The compensation unit compensates the voiceprint data according to the environmental parameters, and the correction unit adjusts the linear array scanning sampling frequency and corrects the three-dimensional point cloud data according to the conveyor belt speed, improving the accuracy and reliability of the data and ensuring the accuracy of subsequent analysis. The analysis unit analyzes the compensated voiceprint data to obtain the material properties, and the fusion unit fuses the compensated three-dimensional morphological data with the material properties to generate a primary classification result, realizing a preliminary comprehensive judgment of the form and material of the recyclables. The matching unit performs material composition analysis on the hyperspectral data, generates a spectral matching label and corrects the primary classification result, further improving the classification accuracy. The determination unit determines the secondary classification result to obtain the final classification result. The entire process realizes the efficient and intelligent sorting of low-value recyclables from domestic sources through the acquisition, processing, analysis, and fusion of multi-dimensional data, improving the sorting efficiency and accuracy, and helping to improve the recycling rate of recyclables.

[0066] In one case of this embodiment, synchronously acquiring the voiceprint data, three-dimensional point cloud data, and hyperspectral data of the recyclables includes:

[0067] Voiceprint data refers to the voiceprint information generated and recorded when recyclables collide in the resonance cavity. Specifically, in the design of the resonance cavity structure, an annular cavity with sound-absorbing cotton and hard reflector plates arranged at intervals on the inner wall is adopted. An elastic buffer bracket is installed at the bottom of the cavity to fix the vibration sensor array (including 8 groups of piezoelectric ceramic sensors and 2 groups of MEMS microphones). When recyclables fall into the resonance cavity through the diversion groove, the infrared pair-sensor at the entrance is first triggered. During the falling process, the recyclables collide with the three-stage stepped impact plate on the side wall of the cavity. The first stage is a 45° inclined steel plate, the middle stage is an arc-shaped ABS plastic plate, and the last stage is a porous aluminum alloy mesh plate. The vibration signals generated by each impact are transmitted to the bottom sensor array through the cavity structure. The piezoelectric ceramic sensors collect the structural vibration signals in the 10 - 20 kHz frequency band in real time, and the MEMS microphones synchronously collect the air sound pressure signals in the 20 - 200 Hz frequency band. After the two types of signals are converted by a 24-bit ADC analog-to-digital converter, they are time-stamped and aligned through a synchronous clock module, and finally voiceprint data including time-domain waveforms and frequency-domain energy distributions is formed;

[0068] Three-dimensional point cloud data refers to the data that presents the three-dimensional morphological characteristics of recyclables through 4K linear array scanning. Specifically, the scanning device consists of two groups of orthogonally arranged 4K linear array lidars and an industrial-grade synchronous controller, which is installed 200 mm above the conveyor belt to form a scanning field of 150 mm × 150 mm. When recyclables enter the scanning area along with the conveyor belt, the optoelectronic sensor array is first triggered. The distance between every two optoelectronic sensors is 5 mm. The industrial-grade synchronous controller calculates the object movement speed in real time according to the conveyor belt speed sensor and dynamically adjusts the scanning frequency of the linear array lidar to ensure that the distance between adjacent scanning lines ≤ 0.8 mm. Each group of lidars emits 850 nm infrared laser beams, which form a fan-shaped light surface through a cylindrical lens and are projected onto the object surface to generate diffuse reflection. The receiver calculates the three-dimensional coordinates of each pixel point through the principle of triangulation ranging. During the scanning process, the system synchronously collects the instantaneous speed and acceleration of the conveyor belt, then converts the time stamp of each scanning point into the conveyor belt displacement, fits the motion trajectory through second-order Taylor expansion, and then maps the coordinates of each point to the coordinate system at the start time of scanning. The generated three-dimensional point cloud data is filtered by voxel to remove discrete noise points, and three-dimensional point cloud data including surface curvature and normal vector information is formed;

[0069] Hyperspectral data refers to the data obtained by using a hyperspectral camera to photograph recyclables, covering the data of recyclables in different spectral bands. Specifically, it includes: diffuse reflection uniform light sources are equipped on both sides of the conveyor belt. When the recyclables enter the shooting area along the conveyor belt, the hyperspectral camera starts to scan line by line, adjusts the exposure time of each line according to the object movement speed to ensure that adjacent line images overlap. During the scanning period, the system collects data of the standard white board and the dark current reference board every 5 minutes to correct the spectral response of the camera. For the collected original spectral data, first, defective pixels are repaired, and then environmental light interference is removed. After processing, cube data containing 224 spectral bands is obtained. For transparent or semi-transparent objects, the system turns on the back monochromatic light source, and by calculating the difference between the transmission spectrum and the reflection spectrum, the spectral characteristics of the material are enhanced, and finally, available hyperspectral data is formed;

[0070] In one case of this embodiment, the voiceprint data is compensated by environmental parameters to obtain the compensated voiceprint data, including:

[0071] Extract environmental parameters to obtain the acoustic attenuation coefficient, specifically including: using a high-precision temperature and humidity sensor to collect environmental temperature and humidity data, substituting the real-time temperature and humidity values into the Mattson formula to calculate the time delay deviation caused by the change in sound speed, and obtaining the temperature and humidity correction coefficient. At the same time, using a microphone array to collect environmental noise signals, performing spectral analysis through the fast Fourier transform algorithm, extracting the signal-to-noise ratio attenuation amount in a specific frequency band, calculating the noise attenuation factor, substituting the temperature and humidity correction coefficient and the noise attenuation factor into the Savitzky-Golay filtering algorithm for weighted fusion, and finally calculating the acoustic attenuation coefficient, which is used to compensate and correct the amplitude spectrum and phase spectrum of the original voiceprint data;

[0072] Among them, substituting the real-time temperature and humidity values into the Mattson formula to calculate the time delay deviation caused by the change in sound speed, and obtaining the temperature and humidity correction coefficient , the calculation process is: ;

[0073] In the formula, represents the sound wave propagation distance, represents the sound wave propagation time, represents the environmental temperature value, represents the environmental humidity value, represents the sound speed;

[0074] Construct an adaptive filter based on the acoustic attenuation coefficient to generate a pure voiceprint in the frequency domain, specifically including: constructing an adaptive filter based on the acoustic attenuation coefficient, using the ideal pure voiceprint spectrum as the guiding signal, dynamically estimating the acoustic attenuation coefficient through the Kalman filtering algorithm, updating the filter weights in real time, taking the difference between the original voiceprint data and the ideal voiceprint spectrum as the error signal, adjusting the filter coefficients through the least mean square error algorithm, after frequency domain adaptive filtering, using the short-time Fourier transform to convert the time-domain signal into frequency-domain features, and then reconstructing the pure voiceprint through the inverse transform to generate a pure voiceprint in the frequency domain;

[0075] Calculate the conveyor belt speed to obtain the acoustic wave propagation time delay, specifically including: installing a laser displacement sensor and an encoder on both sides of the conveyor belt. The laser displacement sensor emits laser light to measure the distance between the object surface and the sensor. As the conveyor belt moves and the position of the recyclable object changes, the sensor continuously samples to obtain a series of displacement data. The encoder monitors the rotation angle and speed of the motor or drum, and combines with the circumference of the conveyor belt to calculate the real-time speed of the conveyor belt;

[0076] At the same time, install a laser rangefinder at the sound source of the conveyor belt. The laser rangefinder emits a laser beam, receives the reflected light, calculates the round-trip time between the two, and calculates the straight-line distance based on the speed of light.

[0077] According to the speed-distance-time formula, that is, time distance / speed, calculate the theoretical propagation time delay of the acoustic wave in the conveyor belt environment through the conveyor belt speed and the straight-line distance, and use a high-precision clock system to record the accurate timestamps of the acoustic wave emission and reception respectively. By subtracting these two timestamps, the actual propagation time delay of the acoustic wave is obtained;

[0078] Compare the theoretical propagation time delay with the actual propagation time delay. Since environmental factors such as temperature, humidity, and air flow will affect the acoustic wave propagation time delay, there is a certain deviation. To correct this deviation, use the least squares method to fit the actual propagation time delay obtained from multiple measurements to obtain the environmental correction coefficient. Finally, apply the environmental correction coefficient to the actual propagation time delay data to compensate for the error caused by environmental factors, so as to obtain the accurate acoustic wave propagation time delay with speed compensation;

[0079] Use the acoustic wave propagation time delay to correct the pure voiceprint in the frequency domain to generate a speed-synchronized energy parameter, specifically including: based on the acoustic wave propagation time delay, use the interpolation compensation algorithm to correct the phase of the pure voiceprint in the frequency domain to eliminate the phase deviation caused by the time delay. Through energy normalization processing, use the energy integration algorithm to calculate the energy of the corrected voiceprint, and combine the acoustic wave propagation speed and distance to generate a speed-synchronized energy parameter, so as to improve the accuracy of the voiceprint data and ensure its accurate matching with the actual acoustic characteristics;

[0080] Fuse the acoustic attenuation coefficient, velocity synchronization energy parameter, and spatial reflection characteristics of the 3D point cloud data to generate anti-interference compensated voiceprint data, specifically including: performing weighted summation on the acoustic attenuation coefficient and velocity synchronization energy parameter through a weighted fusion algorithm. In the weighted summation, the parameter of the acoustic attenuation coefficient is 0.6, and the weight of the velocity synchronization energy parameter is 0.4. Then, utilize the spatial reflection characteristics of the 3D point cloud data and combine with the spatial filtering algorithm to perform spatial feature correction on the weighted fused data, thereby generating anti-interference compensated voiceprint data;

[0081] Extract environmental parameters in multiple dimensions and calculate through algorithms such as the fast Fourier transform algorithm to obtain the acoustic attenuation coefficient. This coefficient can compensate and correct the amplitude spectrum and phase spectrum of the original voiceprint data, effectively eliminating the interference of environmental factors on the voiceprint data. At the same time, by constructing an adaptive filter, using the ideal pure voiceprint spectrum as the guiding signal, combining the Kalman filtering algorithm and the least mean square error algorithm to further purify the voiceprint data. Through the accurate calculation of the conveyor belt speed, combine the least squares method to compensate and correct the acoustic wave propagation delay, and then use the interpolation compensation algorithm and energy normalization processing to generate the velocity synchronization energy parameter, improving the accuracy of the voiceprint data from multiple aspects and laying a solid foundation for the accurate sorting of subsequent recyclables;

[0082] By fusing the acoustic attenuation coefficient, velocity synchronization energy parameter, and spatial reflection characteristics of the 3D point cloud data, anti-interference compensated voiceprint data is generated. Through the weighted fusion algorithm, the acoustic attenuation coefficient and velocity synchronization energy parameter are scientifically proportioned and summed, and then combined with the spatial reflection characteristics of the 3D point cloud data, and the spatial filtering algorithm is used for spatial feature correction, effectively suppressing the interference of factors such as environmental noise and conveyor belt movement on the voiceprint data. This multi-data fusion method not only improves the stability and reliability of the voiceprint data, but also enhances the adaptability of the entire sorting system to complex environments, making the analysis of the material composition of recyclables based on voiceprint data more accurate, ensuring the efficient and stable operation of the sorting system, and improving the sorting efficiency and quality of low-value recyclables from domestic sources.

[0083] In one case of this embodiment, the linear array scanning sampling frequency is adjusted in real time according to the conveyor belt speed, and the 3D point cloud data is corrected to generate compensated 3D morphological data, including:

[0084] Construct an acceleration-resonance coupling function based on the conveyor belt speed to generate dynamic frequency parameters linked to speed changes, where the conveyor belt acceleration is set as , and the resonance fundamental frequency is , calculate the dynamic frequency parameter , and the calculation formula is as follows: ;

[0085] In the formula, represents the dynamic frequency parameter at time t, represents the historical acceleration integration time window, represents the resonant gain coefficient of the material refractive index in the hyperspectral data, represents the acceleration gradient tensor, represents the acceleration mutation compensation factor, represents the second derivative of the acceleration, represents the differential symbol in the integration, represents in the time window , the acceleration and the resonant gain coefficient of the material refractive index as well as the acceleration gradient tensor the cumulative effect of the relationship between them;

[0086] Perform temporal resampling on the three-dimensional point cloud data through the dynamic frequency parameter to generate a geometric distortion factor representing velocity-associated deformation, specifically including: According to the dynamic frequency parameter, use the linear interpolation algorithm to perform temporal resampling on the original three-dimensional point cloud data to obtain a new point cloud sequence, and use the least squares method to fit the spatial distribution difference of the new point cloud sequence to generate a geometric distortion factor representing velocity-associated deformation;

[0087] Perform spatial mapping and matching on the geometric distortion factor and the hyperspectral data to generate a spatial calibration weight for eliminating the spatio-temporal offset of multi-dimensional data, specifically including: Through the affine transformation algorithm, preliminarily align the spatial coordinates of the geometric distortion factor and the hyperspectral data to construct a spatial mapping relationship matrix, and through the least squares method, fit the deformation parameters of the geometric distortion factor and the spatial positioning information of the hyperspectral data. The fitting process is as follows: Select evenly distributed ground control points, obtain the accurate geographical coordinates of the ground control points, find the coordinates of the distorted pixels corresponding in the hyperspectral data, use at least 3 ground control points (usually select 10 - 20 to improve accuracy), correspond the real coordinates with the distorted pixel coordinates, and continuously adjust the parameters in the affine transformation with the least squares method to make the error between the real coordinates of the control points and the transformed coordinates as small as possible, so as to obtain a spatial calibration weight for eliminating the spatio-temporal offset of multi-dimensional data;

[0088] The dynamic frequency parameter, geometric distortion factor, and spatial calibration weight are weighted and fused to generate compensated three-dimensional morphological data that resists motion blur. Specifically, it includes: normalizing the dynamic frequency parameter, geometric distortion factor, and spatial calibration weight in a unified coordinate system, using a weighted average algorithm, and allocating weights to each parameter with a preset weight coefficient. The weight of the dynamic frequency parameter is set to 0.4, the weight of the geometric distortion factor is set to 0.3, and the weight of the spatial calibration weight is set to 0.3. Using the polynomial fitting method, the weighted data is subjected to a fitting operation, and the fitted data is substituted into the three-dimensional point cloud data to directly generate compensated three-dimensional morphological data that resists motion blur;

[0089] The dynamic frequency parameter is calculated by constructing an acceleration-harmonic coupling function, which can accurately capture the cumulative effect and mutation characteristics of the conveyor belt speed change. When the conveyor belt acceleration fluctuates with time, the dynamic frequency parameter responds to the speed change trend in real time through the exponential integral term, and at the same time uses the second derivative term of acceleration to compensate for instantaneous mutations, enabling the scanning sampling frequency to form a dynamic linkage with the conveyor belt motion state. This mechanism effectively avoids the spatio-temporal misalignment problem of traditional fixed-frequency sampling in variable-speed scenarios, ensuring that the time series of three-dimensional point cloud data is strictly aligned with the actual motion trajectory of the conveyor belt. In the time-series resampling link, the linear interpolation algorithm based on the dynamic frequency parameter can adaptively adjust the time interval of the point cloud data, perform density compensation and phase calibration on the original point cloud sequence, and combine the geometric distortion factor generated by fitting the spatial distribution difference using the least squares method, which can accurately describe the stretching or compression deformation of the object shape caused by speed changes. Compared with the motion blur or data redundancy caused by traditional fixed-frequency sampling, it can reduce the synchronization error of the point cloud time during the conveyor belt acceleration and deceleration process, significantly improving the spatio-temporal consistency of the three-dimensional point cloud data;

[0090] By performing affine transformation and fitting with ground control points, an accurate spatial mapping relationship between the geometric distortion factor and hyperspectral data is established. Using at least 3 ground control points for least squares optimization can control the spatio-temporal offset error of multi-dimensional data at the sub-pixel level, solving the problem of insufficient spatial registration accuracy of different sensor data in traditional methods. This calibration mechanism not only ensures the unity of the coordinate system of the three-dimensional point cloud and hyperspectral data, but also eliminates the spatio-temporal misalignment of cross-modal data caused by conveyor belt motion through the dynamic allocation of spatial calibration weights. The weighted fusion strategy realizes the complementary advantages of multi-source compensation parameters through normalization processing and preset weight allocation. The dynamic frequency parameter dominates the adaptive sampling frequency in the time dimension, the geometric distortion factor focuses on spatial shape correction, and the spatial calibration weight ensures cross-modal data alignment. The three form a collaborative compensation through polynomial fitting, reducing the motion blur error of the three-dimensional morphological data in the conveyor belt speed fluctuation scenario, and is applicable to the sorting operation of recyclables.

[0091] In a case of this embodiment, the material composition of the compensated voiceprint data is analyzed to obtain material properties, including:

[0092] Perform time-frequency joint analysis on the compensated voiceprint data to generate multi-scale resonance factors that can characterize the resonance characteristics of the material. Among them, for the multi-scale resonance factors Perform calculations, and the calculation formula is as follows: ;

[0093] In the formula, where K represents the total number of scales, k represents the index of K, Represents the weight of the k-th scale, Represents the voiceprint signal Of the wavelet transform, Represents the tensor convolution operation, Represents the derivative of the frequency-domain energy density with respect to frequency, Represents the Gaussian kernel related to the acoustic attenuation coefficient, Represents the analysis of the attenuation characteristics in the frequency domain Mapped back to the time-frequency joint space and coupled with the wavelet transform result;

[0094] Based on the multi-scale resonance factors, dynamically track the non-linear attenuation path of the sound wave and the material, and extract the dynamic transfer entropy value that can reflect the material density, specifically including: based on the multi-scale resonance factors, use the Gaussian kernel function and the non-linear regression algorithm to dynamically track the non-linear attenuation path of the sound wave in the material, and by calculating the change rate of the acoustic parameters at each point on the attenuation path, combined with the acoustic attenuation coefficient, use the information entropy calculation method to quantify the uncertainty in the sound wave attenuation process, so as to extract the dynamic transfer entropy value that can reflect the material density;

[0095] Perform feature fusion on the multi-scale resonance factors and the dynamic transfer entropy value, construct a material response surface for the sound wave energy attenuation, and generate a material attenuation fingerprint, specifically including: perform normalization processing on the multi-scale resonance factors and the dynamic transfer entropy value, determine the weight coefficients of the two based on the Gaussian kernel function, and fuse the two groups of weighted data into new feature data through matrix operations; use the non-linear regression algorithm, with the fused feature data as the input and the sound wave energy attenuation parameter as the output, to fit the material attenuation fingerprint;

[0096] By matching the material attenuation fingerprint with the attenuation characteristics of the material database, the material attributes including material category and thickness are generated. Specifically, the material database can be formed by statistically analyzing the material data of recyclables in the city's low-value recyclables sorting center, comparing the material attenuation fingerprint with the attenuation characteristics of each sample in the material database, and using the cosine similarity algorithm to calculate the similarity between the two, and find the sample with the highest similarity. The corresponding material category is the target material category. For material thickness, a linear regression algorithm is used to fit the similarity and sample thickness information in the database, and finally a material attribute including material category and thickness is generated;

[0097] Through the tensor convolution of wavelet transform and frequency domain energy derivative, the multi-dimensional characteristics of material resonance in the voiceprint signal can be deeply mined. The setting of different scale weights enables the algorithm to capture the material resonance characteristics from macro to micro, and the Gaussian kernel related to the acoustic attenuation coefficient effectively suppresses the interference of environmental noise on material characteristics. When dynamically tracking the sound wave attenuation path, the synergy of the Gaussian kernel function and the nonlinear regression algorithm can accurately characterize the nonlinear attenuation process of sound waves propagating inside the material. By quantifying the change rate and uncertainty of acoustic parameters on the attenuation path, the extracted dynamic migration entropy value can directly reflect the material density characteristics. Compared with the single frequency analysis method, it reduces the error of density recognition of complex materials, significantly improves the analysis accuracy of material acoustic characteristics, and provides a more comprehensive and accurate data basis for material attribute judgment.

[0098] Through normalization processing and Gaussian kernel weight distribution, the material resonance characteristics and density information are effectively integrated, avoiding the limitations of a single feature in material identification. The material attenuation fingerprint generated by nonlinear regression fitting converts complex acoustic data into a unique feature identifier, greatly enhancing the distinction of material properties. In practical applications, the material database built based on the city's low-value recyclables sorting center enables the algorithm to quickly adapt to recyclable scenarios, and realizes the simultaneous identification of material category and thickness through the cosine similarity algorithm. In the mixed material detection scenario, the accuracy of material category identification is improved, which greatly improves the efficiency of automated processing in garbage sorting, waste recycling and other fields, and provides reliable technical support for resource recycling.

[0099] In one case of this embodiment, the compensated three-dimensional morphological data is fused with the material attributes to generate a primary classification result, including:

[0100] Extract the dynamic geometric parameters of recyclables based on the compensated three-dimensional shape data, and construct a shape-material correlation matrix in combination with the material properties. Specifically, from the compensated three-dimensional shape data, use a three-dimensional feature extraction algorithm to extract the three dynamic geometric parameters of the length, width, and height of the recyclables. Combine the thickness parameter in the material properties with the three dynamic geometric parameters of the length, width, and height of the recyclables. Take the geometric parameters as rows and the thickness parameter as columns to construct a matrix, and each element represents the correlation relationship between the corresponding geometric parameter and the thickness, thus forming a shape-material correlation matrix;

[0101] Expand the shape-material correlation matrix to generate multi-physical field coupling parameters including material density gradient and deformation coefficient. Specifically, based on the shape-material correlation matrix, use the radial basis function interpolation algorithm to construct an interpolation function with the original geometric parameters and thickness parameters in the matrix elements as nodes. Perform non-linear interpolation in the interpolation space through this function to fill the blank areas of the matrix and increase the data density. Then, calculate the material density gradient of the interpolated matrix data through the non-linear finite element analysis algorithm. For recyclables, according to their own material properties and geometric shape characteristics, adopt the large deformation analysis method in geometric non-linear theory, and combine the relationships between strain and displacement, stress and strain in material mechanics to directly solve the deformation coefficient. Finally, obtain the deformation characteristic parameters of each region, and spatially fuse the material density gradient and the deformation coefficient through tensor field superposition to generate multi-physical field coupling parameters including material density gradient and deformation coefficient;

[0102] Input the multi-physical field coupling parameters into a deep residual network, and output a primary classification result including probability confidence based on the fusion weight of three-dimensional shape and acoustic features. Specifically, after standardizing and preprocessing the multi-physical field coupling parameters, use them as the initial input features to access the first layer of the deep residual network. Perform cross-layer feature extraction and non-linear mapping through residual blocks. At the same time, after the three-dimensional shape features pass through 3D convolution and the acoustic features pass through Mel cepstral coefficient extraction, generate feature vectors through independent sub-networks respectively. Use the attention mechanism to calculate the fusion weight of the two, perform weighted aggregation on the feature vectors, and then input the fused features into the fully connected layer. Calculate the probability confidence of each classification through the Softmax function to generate a primary classification result including probability confidence;

[0103] By constructing a morphology-material correlation matrix and introducing multi-physical-field coupling parameters, the deep fusion of geometric features and material properties is achieved. On the one hand, the extraction of dynamic geometric parameters combined with three-dimensional feature algorithms accurately captures the dynamic changes of recyclables in the spatial dimension, solving the problem that static size parameters in traditional classification cannot reflect the actual morphological differences of objects. For example, for plastic bottles with different degrees of flattening or cardboard boxes in a folded state, the dynamic data of their length, width, and height can more realistically describe the spatial characteristics of the objects. On the other hand, by incorporating the thickness parameter into the matrix construction, the physical properties of the material are correlated with the geometric form, breaking through the limitations of single geometric or material feature analysis. Further, through radial basis function interpolation and non-linear finite element analysis, multi-physical-field parameters such as density gradient and deformation coefficient generated can characterize the mechanical properties and internal density distribution of the material when it is deformed under force, providing a three-dimensional feature set including shape, material, and mechanical behavior. This multi-dimensional data fusion not only enhances the richness of the feature space but also achieves the spatial alignment of physical-field information through tensor field superposition, providing high-precision and high-discrimination input data for subsequent deep network processing;

[0104] Through the classification architecture of the deep residual network combined with the multi-modal feature fusion strategy, among which, the cross-layer connection mechanism of the residual block effectively alleviates the problem of gradient disappearance in the training of deep networks and is suitable for dealing with the morphological variability of recyclables caused by material differences (such as the curved surface reflection of metal cans and the sharp edges of cardboard boxes). At the same time, the three-dimensional morphological features retain the spatial structure information through 3D convolution, and the acoustic features capture auditory correlation features such as the surface texture of the material after being extracted by Mel cepstral coefficients. The two modalities dynamically allocate fusion weights through the attention mechanism, which can adaptively focus on key features. For example, for transparent plastic bottles, it gives priority to the refraction light patterns of the three-dimensional contour, and for metal products, it focuses on the material resonance characteristics of the acoustic echo, avoiding the information redundancy or omission problems of traditional fixed-weight fusion. The finally output probability confidence index provides a decision-making fault tolerance basis for the automated sorting system. By setting a confidence threshold, low-certainty classifications can be reduced, significantly reducing the misjudgment rate caused by similar shapes or material confusion while improving the classification efficiency, especially suitable for the accurate classification requirements in complex recycling scenarios.

[0105] In one case of this embodiment, the hyperspectral data is analyzed for material composition to generate spectral matching labels, including:

[0106] Dynamically compensate for the optical path offset of hyperspectral data based on the conveyor belt speed to generate a spectral-material correlation vector, which specifically includes: calculating the spectral pixel displacement amount through the conveyor belt speed combined with the sampling frequency of the hyperspectral instrument, using an interpolation algorithm to perform displacement compensation on the original hyperspectral data to achieve dynamic correction of the optical path offset, and synchronously matching the corrected spectral data with the spectral features in the material sample library to generate a correlation vector. The dimension of the correlation vector covers the corresponding relationship between the spectral values of each band and the material properties, thereby constructing a complete spectral-material correlation vector;

[0107] Perform cross-modal feature enhancement on the spectral-material correlation vector and the material attenuation fingerprint to obtain a cross-modal fusion weight resistant to noise interference. Among them, calculate the cross-modal fusion weight using the following calculation formula: ;

[0108] In the formula, represents the spectral-material correlation vector, represents the material attenuation fingerprint, d represents the feature dimension, is the spatio-temporal gradient constraint term of the spatial calibration weight, represents the Hadamar product, represents a non-linear function, , T represents the transpose operation of the matrix;

[0109] Perform spatio-temporal alignment on the cross-modal fusion weight to generate a spectral matching label. Among them, the spectral matching label includes the gradient encoding of the material component proportion and the pollution degree, which specifically includes: decomposing the cross-modal fusion weight through a matrix decomposition algorithm, extracting its spatial and temporal feature components, multiplying the spatial calibration weight element-wise with them, and using the alternating direction multiplier method for optimization to achieve alignment of the fusion weight in the spatial and temporal dimensions. Use the decomposed weight matrix to perform weighted summation on the spectral-material correlation vector, and use the non-negative matrix factorization algorithm to decompose it into the product of the material component proportion matrix and the pollution degree gradient encoding matrix. The component proportion matrix is normalized by column, and the pollution degree matrix is normalized by row, and finally generate a spectral matching label containing the gradient encoding of the material component proportion and the pollution degree;

[0110] The dynamic optical path offset compensation mechanism based on the conveyor belt speed effectively solves the spatio-temporal misalignment problem in the process of hyperspectral data acquisition. In the scenario of high-speed operation of the conveyor belt, traditional static sampling easily causes spectral pixel position offset, resulting in spectral feature distortion. In this invention, the displacement is accurately calculated by combining the conveyor belt speed and the sampling frequency, and the interpolation algorithm is used to dynamically correct the original data, ensuring the precise correspondence between the spectral information in the hyperspectral data and the actual position of the measured object. At the same time, the corrected spectral data is matched with the spectral features of the material sample library to construct a spectral-material correlation vector, covering the correspondence between multi-band spectral values and material characteristics, and capable of comprehensively capturing the reflection and absorption characteristics of the material in different spectral bands. This data processing method not only eliminates the data error caused by the movement of the conveyor belt, but also provides an accurate and complete data basis for subsequent material component analysis through the construction of a multi-dimensional correlation vector, greatly improving the reliability of material identification;

[0111] Through the cross-modal feature enhancement and spatio-temporal alignment strategy, the analysis stability and anti-interference ability of the spectral matching label are significantly improved. By cross-modal fusion of the spectral-material correlation vector and the material attenuation fingerprint, a fusion weight resistant to noise interference is generated using a specific calculation formula, fully exploring the potential connection between the spectral data and the inherent characteristics of the material. At the same time, a spatio-temporal gradient constraint term of the spatial calibration weight is introduced to enable the fusion weight to effectively suppress the interference of noise on the features. In the spatio-temporal alignment link, the application of the matrix decomposition algorithm and the alternating direction multiplier method realizes the precise alignment of the fusion weight in the spatial and temporal dimensions, ensuring the consistency of spectral feature analysis. The finally generated spectral matching label not only includes the proportion of material components, but also reflects the degree of contamination through gradient coding. Using non-negative matrix factorization for normalization processing enables the label information to clearly and accurately reflect the true composition and contamination status of the material. This analysis method of multi-algorithm fusion effectively overcomes the problem that spectral data is vulnerable to noise in complex environments, providing a reliable guarantee for the accurate analysis of material components.

[0112] In one case of this embodiment, the primary classification result is corrected by the spectral matching label to obtain the secondary classification result, including:

[0113] Extract the gradient coding in the spectral matching label, decompose it into two parameters, namely the proportion parameter of material components and the contamination degree parameter, and construct a dynamic correction coefficient matrix based on the two parameters, specifically including:

[0114] When extracting the gradient encoding in the spectral matching tags, the differential spectroscopy method is used to calculate the first-order / second-order derivatives of the spectral curve to obtain the gradient encoding. After filtering and noise reduction by the Savitzky-Golay algorithm, the gradient encoding is decomposed into the proportion parameters of material components (linear combination coefficients based on the gradient characteristics of standard materials in the spectral library) and the pollution degree parameters (fitting the gradient offset of pollution characteristics using the locally weighted scatterplot smoothing method) through non-negative matrix factorization. Taking these two parameters as input variables, a dynamic correction coefficient matrix is constructed through multivariate adaptive regression splines to achieve parametric correction of the primary classification results;

[0115] The dynamic correction coefficient matrix and the probability confidence of the primary classification results are nonlinearly weighted and fused to generate cross-modal confidence compensation parameters, specifically including: using a radial basis function neural network, with the elements of the dynamic correction coefficient matrix as input and the primary classification probability confidence as the expected output. Calculate the distance between the input and the center vector of the hidden layer through the Gaussian radial basis function, and adjust the weights and the center vector using the gradient descent method. After network training, the dynamic correction coefficient matrix and the primary classification probability confidence are input into the network, and through nonlinear transformation and weighted fusion, cross-modal confidence compensation parameters are generated;

[0116] Based on the spatial calibration weights, spatio-temporal consistency constraints are imposed on the cross-modal confidence compensation parameters to output the secondary classification results, specifically including: using geographically weighted regression to calculate the spatial calibration weight matrix, defining the neighborhood spatial attenuation weights with the target pixel as the center through the Gaussian kernel function, constructing a spatio-temporal coupling constraint term by combining the covariance matrices of multi-temporal images, and then performing a tensor contraction operation on the cross-modal confidence compensation parameters and the spatial calibration weight matrix. Solve the spatio-temporal regularization objective function using the alternating direction method of multipliers to eliminate spatial heterogeneity and temporal mutation noise. Finally, through projection gradient descent iterative optimization, the corrected confidence parameters are made to satisfy local stationarity and temporal smoothness, and the secondary classification results are output;

[0117] Through the combination of differential spectroscopy method, Savitzky-Golay algorithm and non-negative matrix factorization, the noise reduction decomposition and parameter decoupling of gradient encoding are realized, enabling the proportion parameters of material components and the pollution degree parameters to accurately represent the material composition and environmental interference information of spectral characteristics. Based on multivariate adaptive regression splines, a dynamic correction coefficient matrix is constructed, which can achieve parametric correction for different spectral characteristics, effectively compensating for the classification bias caused by material mixing or pollution interference in the primary classification, improving the accuracy and robustness of the classification results in complex scenarios, and ensuring the sorting effect of recyclables;

[0118] Construct a spatial calibration weight matrix with spatio-temporal coupling constraints through geographically weighted regression and Gaussian kernel function, which can effectively eliminate the local classification fluctuations and temporal mutation noises caused by spatial heterogeneity. Combined with projection gradient descent optimization, it ensures that the cross-modal confidence compensation parameters meet the spatio-temporal consistency constraints, making the secondary classification results not only maintain the relevance of spectral features in the spatial neighborhood but also conform to the smoothness law of the time series, significantly enhancing the spatio-temporal stability and semantic consistency of the classification results, and thus facilitating the effective classification of recyclables.

[0119] In one case of this embodiment, the secondary classification results are judged to obtain the final classification results, including:

[0120] Generate a dynamic confidence offset parameter based on the cross-modal confidence compensation parameter and the material attenuation fingerprint, specifically including: By the weighted average algorithm, add the cross-modal confidence compensation parameter and the material attenuation fingerprint with specific weights. First, normalize the cross-modal confidence compensation parameter to make it within the range of 0-1, and the weight is taken as 0.6. Then, enhance the contrast of the material attenuation fingerprint through exponential transformation, and the weight is taken as 0.4. The calculation formula is: dynamic confidence offset parameter = 0.6×(normalized cross-modal confidence compensation parameter)+0.4×(material attenuation fingerprint after exponential transformation);

[0121] Generate a dynamic threshold matrix based on the physical correlation between the dynamic confidence offset parameter and the material attributes, specifically including: Based on the correlation between the dynamic confidence offset parameter and the material attributes, construct a set of physical parameters of the material attributes (such as reflectivity, transmittance, roughness, etc.). Align the dimensions of the set of physical parameters of the material attributes with the dynamic confidence offset parameter, adopt the matrix mapping algorithm, establish a parameter correlation matrix according to the material categories, and through element-wise dot multiplication operation, make each physical parameter of the material attribute form a mapping relationship with the corresponding dynamic confidence offset parameter. After normalization, generate a dynamic threshold matrix by category combination to ensure that the matrix elements correspond one by one to the material classification dimensions;

[0122] Through the innovative construction of dynamic confidence offset parameters and dynamic threshold matrices, the accuracy and robustness of material classification in complex scenarios are significantly improved. On the one hand, the generation mechanism of dynamic confidence offset parameters effectively integrates cross-modal information and material attenuation characteristics: through the normalization of cross-modal confidence compensation parameters and the exponential enhancement of material attenuation fingerprints, while retaining the complementarity of multi-modal data, the material feature differences are amplified. The 0.6:0.4 weight allocation strategy realizes the organic balance of modal information and material characteristics, avoiding the one-sided defects of single-modal information. On the other hand, the construction of the dynamic threshold matrix establishes a direct mapping between confidence offset and material physical properties, incorporates physical parameters such as reflectivity and transmittance into the classification decision system, and makes the threshold setting shift from empirical judgment to scientific derivation based on the essential characteristics of materials. The matrix mapping algorithm and element-wise dot product operation ensure the accurate alignment of parameter dimensions. The normalized dynamic threshold matrix corresponds one-to-one with the material classification dimension, retaining the unique physical properties of different materials and forming a unified classification decision standard. Through the deep integration of data-driven and physical mechanisms, the classification system is endowed with the ability to dynamically adapt to complex lighting and multi-modal input scenarios, effectively solving the problem of insufficient adaptability in traditional fixed-threshold classification.

[0123] In one case of this embodiment, to determine the secondary classification result and obtain the final classification result, it further includes:

[0124] Nonlinearly couple the dynamic threshold matrix with the gradient encoding of the spectral matching label to generate classification optimization parameters, specifically including: normalizing the gradient encoding of the spectral matching label to make it consistent with the dimension of the dynamic threshold matrix, multiplying the corresponding elements of the dynamic threshold matrix and the normalized gradient encoding in an element-wise manner, using the Sigmoid algorithm to perform a non-linear transformation on the multiplication result, compressing the numerical value to the range of 0-1, and then summing the transformed results by column. The obtained numerical value is the classification optimization parameter;

[0125] Perform multi-dimensional confidence calibration on the secondary classification result according to the classification optimization parameter to generate the final classification result including material type, morphological characteristics, and pollution coefficient, specifically including: multiplying the secondary classification result and the classification optimization parameter element-wise to generate preliminary calibration data, such as material type, and using multi-dimensional matrix operations combined with the classification optimization parameter and the Bayesian algorithm to perform probability correction on the confidence of material types such as metal, plastic, and glass;

[0126] For morphological characteristics, such as flat, rough, porous, etc., use the Bayesian algorithm to determine their probabilities;

[0127] For the mild, moderate, and severe pollution coefficients, the corresponding probabilities are calculated using the Bayesian algorithm. After obtaining the confidence intervals of material type, morphological characteristics, and pollution coefficient, the category corresponding to the maximum value of the probability density function is selected to determine the final classification result. The final classification result is composed of material type + morphological characteristics + pollution coefficient.

[0128] According to the classification results, set the sorting decision priority as follows:

[0129] Material type: composite material (low value) > plastic / glass (medium value) > metal (high value);

[0130] Morphological characteristics: blocky> flake-like> flocculent;

[0131] Pollution coefficient: high pollution > medium pollution > low pollution.

[0132] By nonlinearly coupling the dynamic threshold matrix with the spectral matching label gradient encoding to generate classification optimization parameters, the potential correlation between spectral information and material characteristics is fully explored. The introduction of the Sigmoid algorithm effectively compresses the numerical range, avoids abnormal data fluctuations, and ensures parameter stability and reliability. Multi-dimensional confidence calibration is combined with the Bayesian algorithm to perform probability correction on material type, morphological characteristics and pollution coefficient, breaking through the limitations of traditional single-dimensional classification, accurately characterizing sample characteristics from multiple angles, and greatly improving the accuracy and comprehensiveness of classification results. The setting of sorting decision priority closely combines the classification results with the city's low-value recyclables sorting center, clarifies the sorting order according to the material value, morphology and pollution degree, realizes efficient utilization and reasonable disposal of resources, and improves the intelligence and adaptability of the sorting system.

[0133] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An efficient and intelligent sorting system for low-value recyclables from domestic sources, characterized in that, Including: A collection unit, which is used to monitor the environmental parameters and conveyor belt speed of the urban low-value recyclable sorting center in real time, and synchronously collect the voiceprint data, three-dimensional point cloud data, and hyperspectral data of recyclables; A compensation unit, which is used to compensate the voiceprint data through environmental parameters to obtain compensated voiceprint data; A correction unit, which is used to adjust the linear array scanning sampling frequency in real time according to the conveyor belt speed, and correct the three-dimensional point cloud data to generate compensated three-dimensional morphological data; An analysis unit, which is used to analyze the material composition of the compensated voiceprint data to obtain material attributes; A fusion unit, which is used to fuse the compensated three-dimensional morphological data with the material attributes to generate a primary classification result; A matching unit, which is used to analyze the material composition of the hyperspectral data to generate a spectral matching label, and correct the primary classification result through the spectral matching label to obtain a secondary classification result; A determination unit, which is used to determine the secondary classification result to obtain the final classification result.

2. The efficient and intelligent sorting system for low-value recyclables from domestic sources according to claim 1, characterized in that, Synchronously collecting the voiceprint data, three-dimensional point cloud data, and hyperspectral data of recyclables, including: The voiceprint data refers to the voiceprint information generated and recorded when the recyclables collide in the resonance cavity; The three-dimensional point cloud data refers to the data that presents the three-dimensional morphological characteristics of the recyclables through 4K linear array scanning of the recyclables; The hyperspectral data refers to the data covering the recyclables in different spectral bands obtained by photographing the recyclables with a hyperspectral camera.

3. The efficient and intelligent sorting system for low-value recyclables from domestic sources according to claim 1, characterized in that, Compensating the voiceprint data through environmental parameters to obtain compensated voiceprint data, including: Extracting the environmental parameters to obtain the acoustic attenuation coefficient; Constructing an adaptive filter based on the acoustic attenuation coefficient to generate a frequency-domain pure voiceprint; Calculating the conveyor belt speed to obtain the acoustic wave propagation time delay; Using the acoustic wave propagation time delay to correct the frequency-domain pure voiceprint to generate a speed-synchronized energy parameter; Fusing the acoustic attenuation coefficient, speed-synchronized energy parameter, and the spatial reflection characteristics of the three-dimensional point cloud data to generate anti-interference compensated voiceprint data.

4. The efficient and intelligent sorting system for low-value recyclables from domestic sources according to claim 1, characterized in that, Adjusting the linear array scanning sampling frequency in real time according to the conveyor belt speed, and correcting the three-dimensional point cloud data to generate compensated three-dimensional morphological data, including: Constructing an acceleration-resonance coupling function based on the conveyor belt speed to generate a dynamic frequency parameter linked to the speed change; Performing time-series resampling on the three-dimensional point cloud data through the dynamic frequency parameter to generate a geometric distortion factor representing speed-related deformation; Performing spatial mapping matching on the geometric distortion factor and the hyperspectral data to generate a spatial calibration weight for eliminating the spatio-temporal offset of multi-dimensional data; Performing weighted fusion on the dynamic frequency parameter, geometric distortion factor, and spatial calibration weight to generate anti-motion-blur compensated three-dimensional morphological data.

5. The efficient and intelligent sorting system for low-value recyclables from domestic sources according to claim 3, characterized in that, Analyzing the material composition of the compensated voiceprint data to obtain material attributes, including: Performing time-frequency joint analysis on the compensated voiceprint data to generate a multi-scale resonance factor capable of characterizing the material resonance characteristics; Based on the multi-scale resonance factor, dynamically tracking the non-linear attenuation path of the acoustic wave and the material, and extracting a dynamic migration entropy value capable of reflecting the material density; Feature fusion of multi-scale resonance factors and dynamic transfer entropy values is performed to construct a material response surface for acoustic wave energy attenuation and generate a material attenuation fingerprint. By matching the material attenuation fingerprint with the attenuation characteristics in the material database, material attributes including material category and thickness are generated.

6. The efficient and intelligent sorting system for low-value recyclables from domestic sources according to claim 5, characterized in that, The compensated three-dimensional shape data is fused with the material attributes to generate a primary classification result, including: Dynamic geometric parameters of recyclables are extracted based on the compensated three-dimensional shape data, and a shape-material association matrix is constructed in combination with the material attributes, specifically including: from the compensated three-dimensional shape data, three dynamic geometric parameters of length, width, and height of recyclables are extracted using a three-dimensional feature extraction algorithm, and the thickness parameter in the material attributes is combined with the three dynamic geometric parameters of length, width, and height of recyclables. A matrix is constructed with geometric parameters as rows and thickness parameters as columns, and each element represents the association relationship between the corresponding geometric parameter and thickness, thus forming a shape-material association matrix. The shape-material association matrix is extended to generate multi-physical field coupling parameters including material density gradient and deformation coefficient. The multi-physical field coupling parameters are input into a deep residual network, and based on the fusion weights of three-dimensional shape and acoustic features, a primary classification result including probability confidence is output, specifically including: after the multi-physical field coupling parameters are preprocessed by standardization, they are used as initial input features to access the first layer of the deep residual network. Cross-layer feature extraction and non-linear mapping are performed through residual blocks. At the same time, three-dimensional shape features are convolved by 3D convolution, and acoustic features are extracted by Mel cepstral coefficients. Feature vectors are generated through independent sub-networks respectively. The fusion weights of the two are calculated using the attention mechanism. After weighted aggregation of the feature vectors, the fused features are input into a fully connected layer, and the probability confidence of each classification is calculated through the Softmax function to generate a primary classification result including probability confidence.

7. The high-efficiency intelligent sorting system for low-value recyclables from domestic sources according to claim 6, wherein Material composition analysis is performed on the hyperspectral data to generate spectral matching labels, including: Based on the conveyor belt speed, the optical path offset of the hyperspectral data is dynamically compensated to generate a spectral-material association vector, specifically including: the spectral pixel displacement amount is calculated by combining the conveyor belt speed and the sampling frequency of the hyperspectral instrument, and the original hyperspectral data is displacement-compensated using an interpolation algorithm to achieve dynamic correction of the optical path offset. At the same time, the corrected spectral data is matched with the spectral features in the material sample library to generate a complete spectral-material association vector. Cross-modal feature enhancement is performed on the spectral-material association vector and the material attenuation fingerprint to obtain cross-modal fusion weights resistant to noise interference. Perform spatio-temporal alignment on the cross-modal fusion weights to generate spectral matching labels, where the spectral matching labels include gradient encodings of the material component ratio and the pollution degree, specifically including: decomposing the cross-modal fusion weights through a matrix decomposition algorithm, extracting their spatial and temporal feature components, multiplying the spatial calibration weights element-wise with them, and using the alternating direction method of multipliers for optimization and solution to align the fusion weights in the spatial and temporal dimensions. Use the decomposed weight matrix to perform a weighted sum on the spectral-material correlation vectors, and use the non-negative matrix factorization algorithm to decompose it into the product of a material component ratio matrix and a pollution degree gradient encoding matrix. The component ratio matrix is normalized column-wise, and the pollution degree matrix is normalized row-wise, and finally generate spectral matching labels containing the material component ratio and the pollution degree gradient encoding.

8. The high-efficiency intelligent sorting system for low-value recyclables from domestic sources according to claim 7, wherein Correct the primary classification results through the spectral matching labels to obtain secondary classification results, including: Extract the gradient encoding in the spectral matching labels, decompose it into two parameters, namely the material component ratio parameter and the pollution degree parameter, and construct a dynamic correction coefficient matrix based on the two parameters; Perform non-linear weighted fusion on the dynamic correction coefficient matrix and the probability confidence of the primary classification results to generate a cross-modal confidence compensation parameter; Perform spatio-temporal consistency constraints on the cross-modal confidence compensation parameter based on the spatial calibration weights and output the secondary classification results.

9. The high-efficiency intelligent sorting system for low-value recyclables from domestic sources according to claim 8, wherein Judge the secondary classification results to obtain the final classification results, including: Generate a dynamic confidence offset parameter based on the cross-modal confidence compensation parameter and the material attenuation fingerprint; Generate a dynamic threshold matrix based on the physical correlation between the dynamic confidence offset parameter and the material properties.

10. The high-efficiency intelligent sorting system for low-value recyclables from domestic sources according to claim 9, wherein Judge the secondary classification results to obtain the final classification results, which also include: Perform non-linear coupling on the dynamic threshold matrix and the gradient encoding of the spectral matching labels to generate a classification optimization parameter; Perform multi-dimensional confidence calibration on the secondary classification results according to the classification optimization parameter to generate the final classification results including the material type, morphological features, and pollution coefficient.

Images