A waste intelligent sorting method based on multi-sensor fusion

Abstract

This invention relates to the field of solid waste treatment technology, and particularly to a smart waste sorting method based on multi-sensor fusion, comprising: dynamic registration of multiple sensors; multi-level fusion decision-making: suppressing wastewater reflection in visible light images; identifying and eliminating liquid contamination areas based on near-infrared spectral matching; generating spatial grid feature vectors by fusing visible light texture, spectral features, and electromagnetic signals; identifying material categories through a cascaded classifier and dynamically adjusting the weights of each sensor in the decision-making process based on environmental parameters; adaptive sorting execution: triggering the corresponding sorting execution mechanism according to the identification results; calculating the advance compensation amount based on conveyor belt speed, execution mechanism response delay, and material physical properties; adaptively extending the execution mechanism's action time according to the stacking level for stacked materials; and online feedback optimization. Through dynamic compensation and timing adjustment, the temporal consistency of multi-sensor data is ensured, overcoming the influence of belt speed fluctuations.

Description

Technical Field

[0001] This invention relates to the field of solid waste treatment technology, and in particular to a smart waste sorting method based on multi-sensor fusion. Background Technology

[0002] In the field of municipal solid waste recycling, automated sorting systems based on sensor technology have become core equipment for improving waste recycling efficiency, especially for sorting mixed plastics, metals, and paper. Current mainstream technologies have the following limitations:

[0003] Visible light imaging technology fails severely when identifying surface contamination: when plastic bottles are covered with oil stains or kitchen residue, their RGB texture features are distorted, resulting in a misclassification rate of 30-40% for PET and PVC.

[0004] Near-infrared spectroscopy is almost ineffective in material stacking scenarios: when multiple layers of plastic films overlap on a conveyor belt, the signal collected by the spectrometer is actually a mixed absorption spectrum of multiple materials, which cannot resolve the true molecular bond information of the underlying materials.

[0005] Electromagnetic induction technology only responds to metallic materials: it has no ability to identify composite contaminants such as plastic sheets with metal coatings on their surfaces or paper containing metallic ink, resulting in a decrease in sorting purity.

[0006] In addition, the speed fluctuation of the conveyor belt (typically 0.5-2m / s) causes spatiotemporal misalignment of multi-sensor data: for example, when running at high speed, the spectral acquisition point of the same material lags behind the image acquisition point by 5-15cm, resulting in a deviation in the feature fusion coordinates.

[0007] The random tumbling posture of materials causes distortion in feature extraction: especially for irregular objects such as aluminum cans and wrinkled plastic bags, a single-view sensor cannot obtain complete surface information, and the signal attenuation error of metal detectors reaches ±25%.

[0008] Therefore, there is an urgent need for a waste intelligent sorting method based on multi-sensor fusion to solve the above problems. Summary of the Invention

[0009] To achieve the above objectives, the present invention provides a waste intelligent sorting method based on multi-sensor fusion, comprising:

[0010] Multi-sensor dynamic registration:

[0011] Simultaneously acquire visible light images, short-wave infrared spectra, and electromagnetic induction signals of the materials on the conveyor belt;

[0012] Real-time measurement of conveyor belt speed and calculation of timing compensation offset; dynamic adjustment of sampling timing of each sensor to eliminate the influence of speed fluctuation.

[0013] Establish a three-dimensional spatial projection model and calculate the reliability weighting factor of spectral data based on the occlusion relationship between materials;

[0014] Multi-level fusion decision-making:

[0015] Suppressing wastewater reflection in visible light images: Identifying and removing liquid contamination areas based on near-infrared spectral matching;

[0016] Spatial grid feature vectors are generated by fusing visible light textures, spectral features, and electromagnetic signals;

[0017] Material categories are identified through a cascaded classifier, and the weight of each sensor in the decision-making process is dynamically adjusted based on environmental parameters.

[0018] Adaptive sorting execution:

[0019] The corresponding sorting execution mechanism is triggered based on the identification result;

[0020] Calculate the advance compensation amount based on conveyor belt speed, actuator response delay, and material physical properties;

[0021] The actuator's operating time is adaptively extended according to the stacking level for stacked materials;

[0022] Online feedback optimization:

[0023] The sorting results are verified by end-point detection equipment, and the classifier training weights are updated based on the features of erroneous samples.

[0024] Preferably, the generation of the timing compensation offset includes:

[0025] The displacement data of the conveyor belt surface is collected at a fixed frequency using a laser rangefinder, and the real-time linear velocity and acceleration are calculated based on the displacement change rate.

[0026] For shortwave infrared spectrometers, the sampling trigger advance is increased proportionally during acceleration in the transport band:

[0027] The response delay curve of the spectrometer was obtained in a calibration experiment, in which a marker at a known position was launched under different accelerations of a conveyor belt, and the time difference between the spectrometer being triggered and capturing the marker was measured.

[0028] For electromagnetic induction coils, shorten the signal integration period when the conveyor belt decelerates:

[0029] The integration time is dynamically adjusted based on the mapping relationship between the deceleration amplitude and the coil saturation threshold. This mapping relationship is determined by measuring the signal-to-noise ratio of the coil output signal at different speeds.

[0030] Preferably, the generation of the credibility weighting factor includes:

[0031] The three-dimensional point cloud model of the material is reconstructed using a binocular vision camera, and the projected area blocked by adjacent materials within the field of view of the short-wave infrared spectrometer is calculated.

[0032] Obtain the orthogonal projected area of ​​the material on the conveyor belt plane;

[0033] Calculate the ratio of the obstructed area to the total projected area, and determine the weight attenuation coefficient based on this ratio:

[0034] The growth curve of the error rate of spectral data identification under different occlusion ratios was calibrated by experiments, and the error rate increment was converted into a weight decay coefficient.

[0035] Final credibility weight factor = 1 - weight decay coefficient.

[0036] Preferably, the wastewater reflectivity suppression includes:

[0037] Extract pixel regions in a visible light image whose saturation is higher than a dynamic threshold, wherein the dynamic threshold is adaptively adjusted according to the ambient light intensity;

[0038] Obtain the near-infrared reflectance of the corresponding region and match it with a preset wastewater spectral library:

[0039] The wastewater spectral library is constructed by collecting the absorption peak depth ratio of sludge samples of different concentrations / components in a specific wavelength band. The absorption peak depth ratio is the reflectance ratio of the characteristic valley at 1450 nm to that at 1950 nm.

[0040] If the match is successful, it is identified as a reflective area of ​​sewage and pixel replacement is performed:

[0041] Replacement pixels are generated by interpolating the texture features of adjacent uncontaminated areas.

[0042] Preferably, the step of dynamically adjusting the weights of each sensor in the decision-making process based on environmental parameters includes:

[0043] Real-time acquisition of data from dust concentration sensors and material surface humidity sensors;

[0044] Establish a negative correlation function between visible light camera weights and dust concentration:

[0045] By testing the camera recognition error rate under different dust concentrations in a dust environment, the weight decay slope was fitted.

[0046] Establish a positive correlation function between infrared spectrometer weights and humidity:

[0047] By spraying a fixed amount of water mist onto the material surface, the humidity measurement increases the improvement in spectral recognition accuracy, and the fitting weight gain coefficient is used.

[0048] The final decision score of each classifier = Σ(sensor output × dynamic weight).

[0049] Preferably, the calculation of the advance compensation amount includes:

[0050] The height of the material's center of gravity is calculated using three-dimensional point computing.

[0051] Based on the installation height of the high-pressure gas nozzle and the airflow diffusion angle, calculate the theoretical time for the airflow to reach the material surface;

[0052] The lead distance of the air injection point of action is calculated by considering the following factors:

[0053] Real-time speed of conveyor belt, electromagnetic response delay of air valve, theoretical airflow time, and center of gravity height compensation item;

[0054] The center of gravity height compensation item is calibrated by comparing the actual sorting deviation of materials of different thicknesses.

[0055] Preferably, the adaptive extension of the actuator's operating time according to the stacking level of the stacked materials includes:

[0056] Identify the number of stacked material layers using a depth camera;

[0057] Increase the duration of the actuator based on the number of layers:

[0058] Based on the baseline duration of a single-layer material, the action time is linearly extended according to the proportion of the number of layers, and the proportionality coefficient is determined by measuring the minimum air flow rate required for the separation of multi-layer materials.

[0059] Spatial positioning of the target layer in the stacked material:

[0060] Based on the projected coordinates of the materials in each layer in the forward direction of the conveyor belt, the triggering sequence of the actuators in each layer is calculated independently.

[0061] Preferably, the updated classifier training weights include:

[0062] When the end-of-line X-ray fluorescence detector identifies a sorting error, the fusion feature vector of that material is extracted;

[0063] Retrieve the K most similar samples from the historical sample database;

[0064] Increase the weighting coefficients of the retrieved samples:

[0065] The higher the similarity, the greater the increase in weight. Similarity is measured by the Euclidean distance between feature vectors.

[0066] The weighted sample set is then input into the classifier for incremental training.

[0067] Preferably, it also includes online calibration of the spectrometer:

[0068] When the ambient temperature changes beyond a threshold, the built-in blackbody radiation source is triggered to perform wavelength calibration.

[0069] The calibration process includes:

[0070] Collect the blackbody radiation spectrum and extract the measured positions of characteristic absorption peaks;

[0071] Wavelength offset is generated by comparing with a standard peak position database;

[0072] An offset-temperature compensation mapping table is established, which is obtained through step temperature increase calibration in the temperature-controlled chamber.

[0073] Preferably, the method for acquiring the electromagnetic induction signal includes:

[0074] Three sets of Helmholtz coils are used to cover the width of the transmission band;

[0075] Set the excitation frequency independently for each group of coils:

[0076] The low-frequency group is used to detect eddy current losses in ferromagnetic metals, the medium-frequency group is used to detect non-ferrous metals, and the high-frequency group is used to detect alloy materials.

[0077] Separate signal components for each metal type using spectral analysis:

[0078] After bandpass filtering the original signal, the ratio of the signal attenuation slope of each frequency band is calculated to distinguish the metal category.

[0079] The beneficial effects of this invention are:

[0080] 1. This invention suppresses sewage reflection before image acquisition and uses the matching of near-infrared spectrum with sewage spectral library to effectively identify and eliminate liquid contamination areas, thereby avoiding imaging failure caused by surface stains and significantly reducing the misjudgment rate.

[0081] 2. This invention improves the recognition accuracy of near-infrared spectrometers in complex stacking scenarios by establishing a three-dimensional spatial projection model and calculating the reliability weight factor of spectral data in combination with the occlusion relationship of materials, thereby ensuring the sorting accuracy of bottom materials.

[0082] 3. This invention combines electromagnetic induction signals with visible light images and spectral data through multi-sensor fusion technology to generate a comprehensive feature vector, ensuring effective identification of metal and non-metal composite materials, improving the electromagnetic induction technology's ability to identify complex materials, and enhancing the purity of the sorting system.

[0083] 4. This invention eliminates the impact of speed fluctuations on data acquisition by measuring the speed of the conveyor belt in real time, calculating the timing compensation offset, and dynamically adjusting the sampling timing of the sensors according to the acceleration or deceleration of the conveyor belt. This achieves accurate registration of multi-sensor data and improves the accuracy of sorting results.

[0084] 5. This invention utilizes a depth camera to identify the number of material stacking layers and adaptively extends the action time of the actuator according to the number of material layers, ensuring that accurate feature information can be obtained for sorting regardless of whether the material is tumbling or not. Attached Figure Description

[0085] To more clearly illustrate the technical solutions in this invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0086] Figure 1 This is a flowchart of the steps of the method of the present invention;

[0087] Figure 2 This is a flowchart illustrating the steps involved in generating the timing compensation offset in the method of the present invention.

[0088] Figure 3 This is a flowchart illustrating the steps involved in calculating the advance compensation amount in the method of the present invention. Detailed Implementation

[0089] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. It should also be noted that, to make the embodiments more comprehensive, the following embodiments are the best and preferred embodiments, and those skilled in the art can use other alternative methods to implement some well-known technologies; moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0090] Please see Figures 1-3This invention provides a waste intelligent sorting method based on multi-sensor fusion. First, visible light images, short-wave infrared spectra, and electromagnetic induction signals of the materials on the conveyor belt are simultaneously acquired. Visible light images provide texture information of the material surface, short-wave infrared spectra reveal the chemical composition of the material, and electromagnetic induction signals are mainly used for metal detection. By measuring the conveyor belt speed in real time, the temporal compensation offset of each sensor can be calculated, thereby dynamically adjusting the sampling timing of the sensors to eliminate time misalignment caused by conveyor belt speed fluctuations. This step ensures the temporal synchronization of multi-sensor data and improves the accuracy of feature extraction. Furthermore, a three-dimensional spatial projection model is established, and by analyzing the occlusion relationships between materials, a confidence weight factor is assigned to the spectral data, thereby optimizing the accuracy and effectiveness of the spectral data. This process eliminates material overlap and occlusion problems, improving the sorting effect in complex stacking environments.

[0091] To overcome the effects of stains or liquid contamination, wastewater reflection suppression is applied to visible light images. Near-infrared spectral matching technology is used to identify and remove liquid-contaminated areas, avoiding image distortion caused by surface stains. Furthermore, texture features, spectral features, and electromagnetic induction signals from the visible light images are fused to generate a spatial grid feature vector. This multi-dimensional feature vector provides more comprehensive information for subsequent classification decisions. A cascaded classifier is used to identify material categories, and the weights of each sensor in the classification decision are dynamically adjusted according to different environmental conditions, thereby improving classification accuracy and adapting to different working environments and material characteristics.

[0092] After material category identification, the corresponding sorting actuators are triggered based on the material category. To address conveyor belt speed fluctuations and actuator response delays, advance compensation is calculated to ensure the sorting action matches the material arrival time. For stacked materials, the system adaptively extends the actuator's action time based on the stacking level, ensuring each layer of material receives thorough sorting. This process effectively avoids incomplete sorting caused by material stacking.

[0093] Finally, the sorting results are verified in real time by end-point detection equipment, identifying erroneous samples and feeding them back to the system to update the classifier's training weights. This optimization mechanism allows the sorting system to be continuously adjusted and optimized during actual operation, thereby improving sorting accuracy and efficiency.

[0094] By implementing the above-mentioned technical features, the present invention effectively solves the limitations of traditional technologies in the waste sorting process in terms of material mixing, surface contamination, and stacking of materials, thereby improving sorting efficiency and accuracy and having strong practical application prospects.

[0095] In one possible implementation, firstly, a laser rangefinder is used to collect real-time displacement data of the conveyor belt surface at a fixed frequency. The laser rangefinder obtains the rate of displacement change by accurately measuring the positional changes on the conveyor belt surface, and then calculates the real-time linear velocity and acceleration. The real-time calculation of linear velocity and acceleration provides an accurate reference for subsequent sensor timing adjustments, effectively solving the timing misalignment problem caused by changes in conveyor belt speed.

[0096] For short-wave infrared spectrometers, the sampling trigger advance needs to be increased proportionally during conveyor belt acceleration. To achieve this adjustment, the response delay curves of the spectrometer under different accelerations are first obtained through calibration experiments. Specifically, in the experiment, markers at known locations are launched on the conveyor belt, and the time difference between the spectrometer's trigger signal and the acquisition of the markers is measured. The spectrometer's response delay is determined based on different accelerations, and the sampling advance is dynamically adjusted according to the real-time acceleration of the conveyor belt in practical applications. This ensures that the short-wave infrared spectrometer acquires the spectral information of the material in a timely manner during acceleration, reducing timing deviations caused by acceleration and improving sorting accuracy.

[0097] For electromagnetic induction coils, the system adjusts the signal acquisition timing by shortening the signal integration period when the conveyor belt decelerates. Specifically, the signal-to-noise ratio of the coil output signal is first measured at different speeds to determine the mapping relationship between the deceleration amplitude and the coil saturation threshold. Based on this, the signal integration time is dynamically adjusted to ensure that the coil can accurately capture a sufficiently strong signal during conveyor belt deceleration, avoiding signal distortion or loss due to rapid speed changes. This adjustment effectively optimizes the acquisition of electromagnetic induction signals and improves the accuracy of metal material identification.

[0098] Through the aforementioned timing compensation measures, the system can adjust the sampling timing of each sensor in real time according to the speed changes of the conveyor belt, thereby ensuring precise synchronization of data from multiple sensors in time. Real-time velocity and acceleration data provided by the laser rangefinder provide the basis for compensation calculations, while the adaptive adjustments of the short-wave infrared spectrometer and electromagnetic induction coils resolve the impact of speed fluctuations on their respective signals, improving the stability and accuracy of the sorting system. These technical measures effectively solve the spatiotemporal misalignment problem caused by conveyor belt fluctuations in traditional sorting systems, thereby improving sorting accuracy and processing efficiency.

[0099] In one possible implementation, images of the materials on the conveyor belt are first captured using a binocular vision camera, and a 3D point cloud model is reconstructed using stereo vision principles. This model can accurately describe the spatial position and shape of the materials, helping the system understand the relative positions and occlusion between materials. By calculating the projected area of ​​adjacent parts of the materials within the field of view of a short-wave infrared spectrometer, the area occluded by adjacent materials can be obtained. In the 3D point cloud model, the projected area of ​​the materials is converted into an orthogonal projected area on a two-dimensional plane, thus clearly identifying the impact of occlusion on signal acquisition.

[0100] Next, the system calculates the ratio of the material obstruction area to the total projected area. This ratio represents the degree of material obstruction and reflects the visibility and accuracy of the spectral signal acquisition. Based on the experimental calibration results, the study found that when the obstruction ratio is large, the recognition error rate of the spectral data increases significantly. To quantify this change, the system experimentally determined the growth curve of the spectral data recognition error rate under different obstruction ratios. By converting the relationship between the error rate increment and the obstruction ratio into a weighted attenuation coefficient, the system can effectively adjust the reliability of the signal data and reduce the interference of the obstructed portion on the results.

[0101] Finally, based on the calculated weight decay coefficient, the confidence weight factor is determined to be 1 − weight decay coefficient. This formula ensures that as the occlusion ratio increases, the weight of the material's spectral data in the decision-making process will gradually decrease, thereby avoiding the impact of inaccurate identification caused by occlusion on the sorting results.

[0102] By dynamically adjusting the confidence weighting factor based on material occlusion, the system can adjust the influence of each sensor's data in real time according to the degree of material occlusion, enhancing the robustness and accuracy of the waste sorting process. Especially in cases of material stacking or overlap, proper compensation for the occlusion effect can effectively prevent erroneous spectral identification, improving the system's adaptability to complex environments and thus achieving more accurate and efficient intelligent waste sorting.

[0103] In one possible implementation, the system first acquires visible light images of the material on the conveyor belt and performs saturation analysis on the pixels in the images. When the saturation of certain pixel areas exceeds a set dynamic threshold, these areas are considered to potentially have reflective issues. To ensure that this threshold is adaptive under different ambient lighting conditions, the dynamic threshold is adjusted according to the real-time ambient light intensity. The threshold increases when the light intensity is high to avoid false identification; and decreases accordingly when the light intensity is low, thereby improving the system's adaptability to different lighting conditions.

[0104] Next, the system acquires the near-infrared reflectance of suspected areas. These areas have been marked as potentially reflective in the visible light image. By measuring the near-infrared reflectance of these areas, the system matches this data with a pre-defined wastewater spectral library. This library is constructed by collecting the absorption peak depth ratios of sludge samples with different concentrations and compositions at specific wavelengths, with particular focus on the characteristic valleys at 1450 nm and 1950 nm. The absorption peak depth ratio is calculated as the ratio of the reflectance of these two characteristic valleys; using this ratio, the system can accurately identify the reflective characteristics of wastewater.

[0105] Once reflectivity matching is successful and an image region is confirmed to be a wastewater reflective area, the system will replace the pixels in that region. The replacement process uses texture features from adjacent uncontaminated areas for interpolation to generate replacement pixels. In this way, the influence of wastewater reflective areas is effectively suppressed, avoiding interference from these reflective areas on waste sorting results, thus ensuring image quality and the accuracy of subsequent data processing.

[0106] Through the aforementioned technical steps, reflective areas in wastewater can be accurately identified and processed, effectively mitigating the impact of ambient light and wastewater reflection on image quality. Dynamically adjusted thresholds ensure robustness under varying lighting conditions, while matching near-infrared reflectivity with the wastewater spectral library guarantees high-precision determination of wastewater reflection. Finally, the pixel replacement method preserves the image's texture features, resulting in more natural and fluid image processing. Overall, wastewater reflection suppression improves the accuracy and reliability of images during waste sorting, further optimizing sorting results and enhancing the stability and adaptability of the sorting system.

[0107] In one possible implementation, the dust concentration and surface humidity of the material are monitored in real time. A dust concentration sensor detects the concentration of airborne dust, while a humidity sensor monitors changes in the humidity of the material surface. These environmental parameters directly affect the sensors' recognition capabilities; therefore, the weight of each sensor in the decision-making process needs to be dynamically adjusted based on their changes.

[0108] The presence of dust can affect the recognition accuracy of visible light cameras, especially in high-dust environments, where the camera may fail to accurately capture material details. To quantify this impact, tests were first conducted in environments with varying dust concentrations, recording the camera's recognition error rate under these conditions. Through data fitting, a negative correlation function between dust concentration and camera weight was established, resulting in an attenuation coefficient. As dust concentration increases, the camera weight is correspondingly reduced to mitigate the interference of dust on the recognition results.

[0109] Humidity significantly affects the recognition accuracy of infrared spectrometers. By spraying a measured amount of water mist onto the material surface to simulate recognition environments under different humidity conditions, the impact of humidity changes on spectral recognition accuracy was measured. As humidity increases, the spectrometer's recognition accuracy typically improves. Therefore, through fitting experimental data, a positive correlation function between humidity and the infrared spectrometer's weights was established, deriving the weight gain coefficient. When humidity increases, the infrared spectrometer's weights increase accordingly to improve recognition accuracy.

[0110] After all sensor data is collected and weights are adjusted, the final classifier decision score is obtained by weighted summation, that is, the output of each sensor is multiplied by its dynamic weight and then summed. The formula is:

[0111] The final decision score of each classifier = Σ(sensor output × dynamic weight);

[0112] This calculation method allows the final decision to combine the outputs of multiple sensors and adjust the importance of each sensor according to environmental changes, thereby making the sorting results more accurate.

[0113] By dynamically adjusting the weights of the sensors, this method can adapt to different environmental conditions and improve the accuracy of waste sorting. In environments with high dust concentrations, the weight of the visible light camera is reduced to avoid false identification; while in environments with high humidity, the weight of the infrared spectrometer is increased, optimizing its identification performance. Finally, by integrating and weighting the outputs of different sensors, the system can make more accurate decisions under various environmental changes, improving the intelligence and reliability of waste sorting.

[0114] In one possible implementation, firstly, three-dimensional spatial data of the material on the conveyor belt is acquired using 3D point cloud technology. From this data, the system extracts the material's center of gravity height. Center of gravity height refers to the vertical position of the material in three-dimensional space, and it is related to factors such as the material's shape, thickness, and density. The material's center of gravity height affects the airflow jet effect, therefore requiring accurate calculation. Through 3D point cloud computing, the material's center of gravity information can be obtained with high precision, thus providing necessary data support for subsequent motion compensation.

[0115] The process of airflow being ejected from the high-pressure nozzle and reaching the material surface requires a certain amount of time. This time is calculated taking into account the nozzle's installation height and the airflow's diffusion angle. Based on the vertical distance from the nozzle to the material surface and the airflow's diffusion angle, the time required for the airflow to reach the target position can be estimated. This time is crucial for calculating the advance compensation amount, as the injection action must be precisely matched to the material's movement.

[0116] Taking into account multiple factors, the system calculates the lead distance of the gas injection point. This distance is mainly determined by the following factors:

[0117] Real-time speed of the conveyor belt: The speed of the material on the conveyor belt directly affects the lead time of the airflow injection point, because the position of the material is constantly changing.

[0118] Electromagnetic response delay of the air valve: The response speed of the air valve is another important factor. When the air valve is electromagnetically controlled, its response delay will affect the timing of airflow injection, and therefore needs to be compensated for.

[0119] Theoretical airflow time: The theoretical time for airflow to reach the material surface is also an important component of advance compensation.

[0120] Center of gravity height compensation: Calculate the corresponding compensation amount based on the material's center of gravity height. A higher center of gravity may mean that the airflow needs to be injected earlier to ensure that the airflow accurately reaches the material surface.

[0121] To improve the accuracy of the compensation, the center of gravity height compensation term needs to be calibrated by comparing the actual sorting deviations of materials with different thicknesses. Through experimental testing, the sorting deviations at different material thicknesses are measured, and the actual results are compared with the expected targets to ultimately obtain a compensation coefficient related to the center of gravity height. This compensation term helps adjust the timing of the airflow injection, making it more precisely matched to the characteristics of materials with different thicknesses.

[0122] By precisely calculating the advance compensation amount, this method ensures that the point of application of the air jet is precisely when the material passes through the nozzle. This compensation mechanism takes into account factors such as the material's movement speed, jet delay, and center of gravity height, ensuring that the jetting action is synchronized with the material sorting, thereby improving the accuracy and efficiency of sorting. Especially when processing materials of different thicknesses or shapes, the center of gravity height compensation item, through calibration, effectively reduces the actual sorting deviation, guaranteeing sorting accuracy. This advance compensation method greatly optimizes the intelligent waste sorting process, improving the system's flexibility and reliability.

[0123] In one possible implementation, a depth camera is first used to perform a 3D scan of the material on the conveyor belt. By acquiring the depth information of the material, it is possible to identify whether the material is stacked and to calculate the number of stacked layers. The depth camera can provide the spatial location of each material layer, accurately determining the stacking situation. This process provides material layer information for subsequent sorting work, enabling more detailed sorting.

[0124] When materials are stacked, the separation of each layer requires a corresponding amount of time. If the material is in a single layer, the actuator's action time is the baseline duration. If the material is stacked in multiple layers, the actuator's action time will increase according to the number of stacked layers. By extending the baseline duration proportionally to the number of layers, it can be ensured that each layer of material is fully separated. The more layers there are, the longer the action time should be, thus avoiding incomplete sorting due to insufficient action time.

[0125] The extended processing time is proportional to the number of layers; that is, the processing time increases linearly based on the baseline time for a single layer of material. The proportionality coefficient is determined experimentally by measuring the minimum airflow rate required for separating multiple layers of material. Specifically, by testing the relationship between airflow rate and separation efficiency on stacked materials with different numbers of layers, a suitable proportionality coefficient is determined to ensure that each layer receives an appropriate processing time.

[0126] To more precisely control the sorting of each layer of material, the target layer within the stacked materials must be spatially located. By using 3D coordinate data of the material layers provided by a depth camera, combined with the projected coordinates of each layer in the conveyor belt's forward direction, the precise position of each layer can be determined. Based on this data, the triggering sequence of the actuators can be calculated independently; that is, the processing time and order of each layer can be precisely arranged according to its position on the conveyor belt. In this way, the actuators can act on the target layer at the exact time, ensuring effective material separation.

[0127] By identifying the number of stacked material layers and extending the action time, precise sorting of multi-layered materials is achieved. Dynamically adjusting the actuator's action time adapts to the processing needs of different material layers, avoiding sorting errors caused by insufficient action time. Simultaneously, precise control of the position of each layer ensures that the target layer is processed within the correct time sequence, thereby improving sorting efficiency and accuracy. This technology makes the waste sorting process more intelligent, better able to handle the processing requirements of different stacked materials, and improves sorting reliability and system adaptability.

[0128] In one possible implementation, firstly, when the end-of-line X-ray fluorescence spectrometer (used to detect the chemical composition and material properties of waste) identifies a sorting error, the system extracts a fused feature vector of the erroneous material. These feature vectors contain the results of fusion data from multiple sensors, such as the material's morphology, spectral information, and color. After extracting these feature vectors, the system can accurately record the specific details of the missorting and provide necessary data support for subsequent classifier updates.

[0129] Next, the system will search the historical sample database to find the K samples most similar to the current material feature vector. These historical samples serve as a reference for the classifier, helping the system better understand the reasons for current misclassification. By comparing historical data, the system can infer which features might cause misclassification in the classifier.

[0130] To optimize the classifier's training performance, the system weights the retrieved K samples. The weighting coefficients are proportional to the sample similarity. Specifically, samples with higher similarity receive a larger increase in their weighting coefficient. Similarity is measured by the Euclidean distance between feature vectors. A smaller Euclidean distance indicates higher similarity between samples, thus granting that sample greater weight during training and providing more reference data. This weighting mechanism helps the classifier prioritize historical samples most similar to the features of incorrectly sorted materials, thereby improving the relevance and effectiveness of the training.

[0131] Finally, the weighted set of K samples is input into the classifier for incremental training. Incremental training refers to fine-tuning the existing classifier model using new data, rather than retraining the entire model. Through incremental training, the classifier can quickly adapt to new data features while maintaining its original training performance. After each update, the classifier's weights are optimized to improve the accuracy of identifying similar misclassification cases.

[0132] This method effectively addresses potential identification errors during waste sorting. By extracting the feature vectors of erroneously sorted materials in real time and retrieving the most similar samples from a historical database, the system dynamically adjusts the classifier's training data, thereby improving its identification accuracy. The introduction of a weighting mechanism allows the classifier to place greater emphasis on historical data with features similar to the currently erroneous materials, thus enhancing its ability to identify similar situations in the future. This method not only optimizes the classifier's performance but also improves the robustness and adaptability of the sorting system in practical applications.

[0133] In one possible implementation, changes in ambient temperature can affect the performance of the spectrometer, particularly its wavelength measurements. To maintain measurement accuracy, the system automatically triggers a built-in blackbody radiation source for wavelength calibration when the ambient temperature change exceeds a preset threshold. A blackbody radiation source is a device that emits broad-spectrum radiation and is widely used in spectrometer calibration because it provides stable spectral characteristics, avoiding measurement errors caused by external environmental factors.

[0134] During calibration, spectral data from the blackbody radiation source is first acquired using a spectrometer. This spectral data contains multiple characteristic absorption peaks, each corresponding to a specific wavelength position. By analyzing the actual positions of these absorption peaks, the system can determine whether the currently measured wavelength has shifted. The positions of these absorption peaks serve as standard features, providing a reference for wavelength calibration.

[0135] By comparing the measured absorption peak positions with predefined peak positions in a standard peak position database, the offset of each absorption peak can be determined. This offset represents the wavelength change caused by variations in ambient temperature. In this way, the wavelength deviation of the spectrometer under current environmental conditions can be accurately calculated, providing data support for subsequent compensation.

[0136] Based on the offset data at different temperatures, the system establishes a compensation mapping table between temperature and wavelength offset. This mapping table is obtained through step-by-step temperature increase calibration within the temperature-controlled chamber. During this process, the system controls the temperature changes within the chamber, measures the wavelength offset at each temperature point, and progressively records the relationship between temperature changes and wavelength offset. This calibration method yields accurate compensation data, providing the system with a reference for compensating for temperature changes.

[0137] This online calibration method for the spectrometer effectively eliminates spectral measurement errors caused by changes in ambient temperature, ensuring the sorting system maintains high-precision performance under various environmental conditions. Through automatic wavelength calibration and real-time temperature compensation, the system can continuously provide stable spectral data under changing environmental conditions. This not only improves the accuracy of waste sorting but also enhances the system's reliability and adaptability. Especially in practical applications where temperature fluctuations are unavoidable, this calibration method ensures the sorting system can accurately identify material characteristics, further improving the overall effectiveness of the intelligent sorting process.

[0138] In one possible implementation, a Helmholtz coil, a device commonly used to generate a uniform magnetic field, is used. By covering the width of the conveyor belt with three different coil groups, extensive electromagnetic induction signal acquisition can be performed during waste sorting. These coil groups form a uniform magnetic field, thereby effectively capturing the electromagnetic responses generated by metallic substances in the waste. The design of each coil group ensures that the signals generated by different metals as they pass through can be accurately detected, thus improving sorting accuracy.

[0139] To effectively distinguish between different types of metals, different excitation frequencies are set for the coils in the low-frequency, medium-frequency, and high-frequency groups, respectively, for ferromagnetic metals, non-ferrous metals, and alloy materials. The low-frequency group is specifically used to detect eddy current losses in ferromagnetic metals, as these metals exhibit a more significant eddy current effect under low-frequency excitation. The medium-frequency group is used to detect non-ferrous metals, as their electromagnetic response is more pronounced under medium-frequency excitation. The high-frequency group is mainly used for alloy materials, whose electromagnetic properties are more evident under high-frequency excitation. By independently setting the excitation frequency for each group of coils, high efficiency and accuracy in identifying different types of metals can be ensured.

[0140] After acquiring the electromagnetic induction signal, spectral analysis is performed to separate the signal components of different metal types. Spectral analysis extracts different frequency components from the signal, allowing the electromagnetic response of each metal to be clearly distinguished within a specific frequency range. This process is crucial, enabling the system to accurately differentiate various metal materials from complex electromagnetic signals, thus providing precise data support for subsequent sorting processes.

[0141] Based on spectral analysis, the original signal undergoes bandpass filtering to remove unnecessary frequency components, focusing on signals within relevant frequency bands. Subsequently, the ratio of signal attenuation slopes for each frequency band is calculated, a crucial step in distinguishing different metal types. Ferromagnetic metals, non-ferrous metals, and alloys exhibit different signal attenuation slopes at different frequencies; ratio analysis effectively differentiates these metal types. This method provides an efficient metal classification tool through quantitative signal analysis.

[0142] This method of acquiring and analyzing electromagnetic induction signals can effectively improve the metal sorting accuracy of intelligent waste sorting systems. By employing three sets of coils with different excitation frequencies, and combining spectral analysis with bandpass filtering, different types of metal materials can be accurately identified against complex signal backgrounds. This not only improves sorting efficiency but also enhances the system's robustness, enabling it to accurately identify various metal types during actual waste treatment, thereby significantly improving the overall performance and resource recovery efficiency of the waste sorting system.

[0143] The following examples will illustrate this in detail:

[0144] This embodiment uses metal sorting in a scrap metal recycling plant as a scenario, specifically applied to the metal sorting process in waste electronic equipment (such as electrical appliances, home appliances, mobile phones, etc.). The equipment contains various metal materials, and the multi-sensor fusion method based on electromagnetic induction signals of this invention is used to achieve accurate sorting. The waste equipment is transported to the sorting area via a conveyor belt, and the electromagnetic induction signal acquisition system of Helmholtz coils accurately identifies and classifies different types of metals.

[0145] Three sets of Helmholtz coils are used, each covering the width of the transmission band, with each set of coils evenly distributed perpendicular to the bandwidth. The radius of each set of coils is 20cm, and the spacing is 15cm. The length of each coil is 30cm.

[0146] The operating frequency is set as follows: Low frequency group (ferromagnetic metal eddy current loss): The excitation frequency is set to the range of 1kHz to 10kHz, mainly for ferromagnetic metals such as iron and steel.

[0147] Intermediate frequency group (non-ferrous metals): The excitation frequency is set to 20kHz to 50kHz, mainly for non-ferrous metals such as aluminum and copper.

[0148] High-frequency group (alloy materials): The excitation frequency is set to 100kHz to 200kHz, mainly for alloy materials such as stainless steel and nickel alloys.

[0149] As discarded electronic equipment passes through the conveyor belt, the electromagnetic induction signals captured by the Helmholtz coil are sent to the signal processing system for real-time analysis. The signal acquisition process includes the following steps:

[0150] Each set of coils receives electromagnetic signals reflected or transmitted from the scrap equipment, and the original signals include the electromagnetic response of the target metal material.

[0151] The acquired raw signals are bandpass filtered to remove noise in irrelevant frequency bands. For example, low-frequency signals are filtered through a bandpass filter from 1kHz to 10kHz, mid-frequency signals through a bandpass filter from 20kHz to 50kHz, and high-frequency signals through a bandpass filter from 100kHz to 200kHz.

[0152] For each signal group, calculate its spectral attenuation slope. The attenuation slope is expressed by the formula:

[0153] ;

[0154] in, and The signals at different frequencies f 1 and f The amplitude at point 2. The attenuation slope reflects the attenuation characteristics of the signal in the frequency band. Different metal materials have different attenuation slopes, and this feature can be used to distinguish metal types.

[0155] Different metals are distinguished by using spectral analysis results combined with attenuation slope ratios:

[0156] Ferromagnetic metals: The attenuation slope is relatively large in the low frequency group (e.g., attenuation slope S=5dB / Hz).

[0157] Non-ferrous metals: The attenuation slope of the mid-frequency group is relatively small (e.g., S=2dB / Hz).

[0158] Alloy materials: The attenuation slope of the high-frequency group has a special variation (e.g., S=3dB / Hz).

[0159] By comparing the attenuation slopes of different metals, ferromagnetic metals, non-ferrous metals, and alloy materials can be accurately identified. The specific sorting process classifies materials by setting thresholds: if the attenuation slope is greater than a certain set value, it is judged as a ferromagnetic metal; if it is less than that value, it is a non-ferrous metal or alloy.

[0160] Considering the impact of external environmental factors (such as temperature changes) on the signal, this embodiment also includes a temperature compensation mechanism. By comparing standard test data with actual test data, a compensation mapping table for temperature and wavelength offset is constructed, and adjustments are made according to the real-time ambient temperature to ensure the accuracy of signal measurement.

[0161] Specifically, waste electronic equipment enters the sorting area via a conveyor belt. Helmholtz coils, based on the characteristics of different metals, collect reflected or transmitted electromagnetic signals through a set excitation frequency. The collected raw signals are passed through a bandpass filter to remove noise from irrelevant frequency bands, extracting signals within a specific frequency range. Spectral analysis is performed on each frequency band to calculate the signal attenuation slope. Based on the calculated attenuation slope, different types of metals are classified using set thresholds, precisely sorting ferromagnetic metals, non-ferrous metals, and alloy materials. The system automatically adjusts the sorting machinery based on the classification results to ensure efficient waste metal recycling and sorting.

[0162] To verify the effectiveness of the present invention, a set of comparative experiments were conducted, comparing the traditional magnetic induction method with the electromagnetic induction signal processing method of the present invention.

[0163] Experiment 1: Traditional Magnetic Induction Method: This method was used for sorting scrap metal using traditional magnetic induction. This method can only distinguish between ferromagnetic and non-ferrous metals, cannot effectively identify alloy materials, and is easily affected by noise during signal processing.

[0164] Results: The sorting accuracy of the traditional method was 85%, and the recognition rate for alloy materials was less than 50%.

[0165] Experiment 2: Electromagnetic induction signal processing method of the present invention: The multi-band Helmholtz coil and spectrum analysis method of the present invention are used to sort the signals by combining the attenuation slope of different metals.

[0166] Results: The sorting accuracy of the method of the present invention is 98%, and the identification rate of alloy materials exceeds 90%.

[0167] This invention uses multi-frequency excitation with three sets of Helmholtz coils to enable clear differentiation of signals from different metals, greatly improving the accuracy of metal material identification.

[0168] This invention can effectively identify alloy materials, solving the problem that traditional magnetic induction methods cannot handle alloy metals, and has strong adaptability to environmental changes (such as temperature).

[0169] By implementing this invention, efficient and precise metal sorting can be achieved in the waste metal recycling process. Compared with traditional methods, this invention excels in the following aspects:

[0170] This invention can accurately distinguish between ferromagnetic metals, non-ferrous metals, and alloy materials, with a significant advantage, especially when processing alloy metals. Employing multi-band signal processing and spectral analysis, it effectively filters out noise, ensuring efficient sorting in complex environments. It can automatically adjust to different operating environments (such as temperature variations) to guarantee signal accuracy and sorting performance.

[0171] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0172] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for intelligent waste sorting based on multi-sensor fusion, characterized in that, The application relates to a multi-sensor dynamic registration system for online material sorting. The system comprises: A multi-sensor dynamic registration system for online material sorting. Synchronously collecting visible light images, short-wave infrared spectra and electromagnetic induction signals of materials on a conveying belt; Real-time measuring the conveying belt speed and calculating a time sequence compensation offset, and dynamically adjusting the sampling time sequence of each sensor to eliminate the influence of speed fluctuation; Establishing a three-dimensional space projection model, and calculating a reliability weight factor of spectral data according to the shielding relationship between materials; Multi-level fusion decision: Inhibiting sewage reflection on the visible light image: identifying and removing liquid contaminated areas based on near-infrared spectrum matching; Fusing visible light texture, spectral characteristics and electromagnetic signals to generate a spatial grid feature vector; Identifying the material category through a cascade classifier, and dynamically adjusting the weight of each sensor in decision-making according to environmental parameters; Self-adaptive sorting execution: Triggering the corresponding sorting execution mechanism according to the identification result; Based on the conveying belt speed, the response delay of the execution mechanism and the physical properties of the materials, the action advance compensation amount is calculated; For stacked materials, the execution mechanism action time is self-adaptively prolonged according to the stacking level; Online feedback optimization:

2. The method as claimed in claim 1, wherein Verifying the sorting result through an end detection device, and updating the classifier training weight based on the error sample features. The generation of the time sequence compensation offset comprises: Collecting the displacement data of the conveying belt surface at a fixed frequency through a laser range finder, and calculating the real-time linear speed and acceleration according to the displacement rate; For the short-wave infrared spectrometer, the sampling trigger advance amount is increased in proportion when the conveying belt accelerates: Obtaining the response delay curve of the spectrometer in the calibration experiment, wherein the calibration experiment is performed by emitting a known position marker under different acceleration of the conveying belt, and the time difference from triggering to capturing the marker by the spectrometer is measured; For the electromagnetic induction coil, the signal integration period is shortened when the conveying belt decelerates:

3. The method as claimed in claim 1, wherein, According to the mapping relationship between the deceleration amplitude and the coil saturation threshold, the integration time is dynamically adjusted, and the mapping relationship is determined by measuring the signal-to-noise ratio of the coil output signal under different speeds. The generation of the reliability weight factor comprises: Reconstructing the three-dimensional point cloud model of the material through a binocular vision camera, calculating the projection area of the short-wave infrared spectrometer field of view which is shielded by adjacent materials, and calculating the orthogonal projection area of the material on the conveying belt plane; The ratio of the shielding area to the total projection area is calculated, and the weight attenuation coefficient is determined according to the ratio: Through the experiment, the growth curve of the error rate of spectral data under different shielding ratios is calibrated, the error rate increment is converted into the weight attenuation coefficient, and finally the reliability weight factor = 1-weight attenuation coefficient. The sewage reflection inhibition comprises:

4. The method of claim 1, wherein Extracting the pixel area with a saturation higher than a dynamic threshold in the visible light image, and the dynamic threshold is self-adaptively adjusted according to the environmental light intensity; Obtaining the near-infrared band reflectivity of the corresponding area, and matching with a preset sewage spectrum library: The sewage spectrum library is constructed by collecting the absorption peak depth ratio of sludge samples with different concentrations / ingredients in a specific wave band, and the absorption peak depth ratio is the reflectivity ratio of the characteristic valleys at 1450 nm and 1950 nm; If the matching is successful, the sewage reflection area is determined and the pixel replacement is performed: The texture features of the adjacent non-contaminated area are used for interpolation to generate the replacement pixels. The dynamic adjustment of the weight of each sensor in decision-making according to the environmental parameters comprises:

5. The method as claimed in claim 1, wherein, ​ Real-time acquisition of dust concentration sensor and material surface humidity sensor data; Establish a negative correlation function between visible light camera weight and dust concentration: By testing the camera recognition error rate under different concentrations in a dust environment, the weight attenuation slope is fitted; Establish a positive correlation function between infrared spectrometer weight and humidity: By spraying a certain amount of water mist on the material surface, measure the improvement of spectral recognition accuracy with increasing humidity, and fit the weight gain coefficient; The final decision score of each classifier = Σ (sensor output × dynamic weight).

6. The method as claimed in claim 1, wherein The calculation of the action advance compensation amount includes: Calculate the material gravity center height by three-dimensional point cloud; According to the installation height of high-pressure air nozzle and the air flow diffusion angle, calculate the theoretical time of air flow reaching the material surface; Calculate the advance distance of air jet action point by considering the following factors: Real-time speed of the conveying belt, electromagnetic response delay of the air valve, theoretical time of air flow, gravity center height compensation term; The gravity center height compensation term is calibrated by comparing the actual separation deviation of materials with different thicknesses.

7. The method as claimed in claim 1, wherein, The adaptive extension of the actuator action time for stacked materials according to the stacking level includes: Identify the number of material stacking layers by a depth camera; Increase the actuator action time according to the number of layers: On the basis of the single-layer material reference time, extend the action time in proportion to the number of layers, and the proportion coefficient is determined by measuring the minimum air flow required for multi-layer material separation; Spatially locate the target layer in the stacked material: According to the projection coordinates of each layer of material in the forward direction of the conveying belt, independently calculate the actuator trigger timing of each layer.

8. The method of claim 1, wherein, The update of the classifier training weight includes: When the end X fluorescence detector identifies a separation error, extract the fusion feature vector of the material; Retrieve the K samples with the highest similarity in the historical sample library; Increase the weight allocation coefficient of the retrieved samples: The higher the similarity, the greater the weight increase, and the similarity is measured by the Euclidean distance between feature vectors; Input the weighted sample set into the classifier for incremental training.

9. The method as claimed in claim 1, wherein It also includes online calibration of the spectrometer: When the ambient temperature changes by more than a threshold value, trigger the built-in blackbody radiation source for wavelength calibration; The calibration process includes: Collect blackbody radiation spectrum and extract the measured position of the characteristic absorption peak; Compare with the standard peak position database to generate wavelength offset; Establish a compensation mapping table of offset and temperature, which is obtained by stepwise temperature calibration in the temperature control chamber.

10. The method of claim 1, wherein, The collection method of electromagnetic induction signals includes: Use three groups of Helmholtz coils to cover the width of the conveying belt; Set an independent excitation frequency for each group of coils: Low-frequency group detects ferromagnetic metal eddy current loss, medium-frequency group detects non-ferrous metal, and high-frequency group detects alloy material; Separate the signal components of each metal type through spectral analysis: After band-pass filtering the original signal, calculate the ratio of the attenuation slope of each frequency band to distinguish the metal category.

Images