A multi-category plastic automatic backflow sorting method and system and storage medium

Abstract

The application provides a multi-category plastic automatic reflow sorting method and system and a storage medium. The method comprises the following steps: obtaining point cloud data of a bottle body, processing the point cloud data by using a laser triangulation method to obtain bottle body shape characteristics of the bottle body; obtaining initial spectral information of the bottle body and working temperature and humidity of a working environment, determining whether compensation of the spectrum is needed based on the working temperature and humidity and the initial spectral information, if needed, substituting the working temperature and humidity and / or the initial spectral information into a preset compensation model to obtain spectral compensation information, and determining actual spectral information of the bottle body based on the initial spectral information and the spectral compensation information; substituting the bottle body shape characteristics and the actual spectral information into a hybrid neural network model to obtain a plastic type of the bottle body and a corresponding confidence degree, determining a confidence level corresponding to the confidence degree, and generating a corresponding sorting instruction based on the confidence level to realize sorting of the bottle body. The application can efficiently and accurately realize multi-category plastic sorting work.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of sorting, in particular to a multi-category plastic automatic backflow sorting method and system and a storage medium. BACKGROUND

[0002] With the wide application of plastic products in daily life and industrial production, the recycling and reuse of waste plastics have become a key link to solve resource waste and environmental pollution. However, there are many types of plastics, which have significant differences in physical and chemical properties, and the recyclable waste plastics need to be sorted and recycled.

[0003] At present, due to the low efficiency and high cost of manual sorting, it has been gradually replaced by equipment to carry out automatic sorting. When using equipment to carry out automatic sorting, the conventional infrared spectrum recognition method or near-infrared sorting method is often used for plastic sorting work. Compared with manual sorting, although the sorting efficiency can be improved, the accuracy of plastic sorting is low due to the influence of factors such as plastic color and surface contamination. Therefore, the existing method cannot efficiently and accurately realize plastic sorting work. SUMMARY

[0004] In order to efficiently and accurately realize multi-category plastic sorting work, the embodiments of the present application provide a multi-category plastic automatic backflow sorting method, system and storage medium.

[0005] In a first aspect, the embodiments provide a multi-category plastic automatic backflow sorting method, which comprises:

[0006] Obtaining point cloud data of a bottle body, and processing the point cloud data using a laser triangulation method to obtain bottle body morphology characteristics of the bottle body;

[0007] Obtaining initial spectral information of the bottle body and working temperature and humidity of a working environment, and determining whether compensation of the spectrum is needed based on the working temperature and humidity and the initial spectral information,

[0008] If needed, substituting the working temperature and humidity and / or the initial spectral information into a preset compensation model to obtain spectral compensation information, and determining actual spectral information of the bottle body based on the initial spectral information and the spectral compensation information;

[0009] Substituting the bottle body morphology characteristics and the actual spectral information into a hybrid neural network model to obtain a plastic type of the bottle body and a corresponding confidence, determining a confidence level corresponding to the confidence, and generating a corresponding sorting instruction based on the confidence level to realize sorting of the bottle body.

[0010] In some embodiments, the method is applied to a sorting device comprising three lasers, the three lasers are uniformly distributed in a ring shape on the sorting device, and the obtaining of the point cloud data of the bottle body comprises:

[0011] The point cloud data of the bottle body is synchronously obtained based on a trigger signal sent by the sorting device.

[0012] In some embodiments, the preset compensation model comprises a temperature and humidity layer compensation model and a pollution layer compensation model, the spectral compensation information comprises temperature and humidity layer spectral compensation information and / or pollution layer spectral compensation information, and the substituting of the working temperature and humidity and / or the initial spectral information into the preset compensation model to obtain the spectral compensation information comprises:

[0013] Substituting the working temperature and humidity into the temperature and humidity layer compensation model to obtain the temperature and humidity layer spectral compensation information;

[0014] And / or, substituting the initial spectral information into the pollution layer compensation model to obtain the pollution layer spectral compensation information.

[0015] In some embodiments, the sorting device comprises a conveying belt, and a set of standard reflectors is installed on each side of the conveying belt, each set of standard reflectors comprising a white plate, and the method further comprises:

[0016] The actual reflectivity of the white plate is obtained every preset time interval, and it is determined whether the actual reflectivity deviation between the actual reflectivity and the set reflectivity of the white plate is greater than a trigger compensation reflectivity deviation, if not, a reflector no deviation instruction is generated;

[0017] If greater, reflector spectral compensation information is generated based on the actual reflectivity deviation, and the actual spectral information is updated using the spectral compensation information.

[0018] In some embodiments, the hybrid neural network model is a CNN-Transformer hybrid neural network, and the substituting of the bottle body morphological features and the actual spectral information into the hybrid neural network model to obtain the plastic type of the bottle body and the corresponding confidence comprises:

[0019] The bottle body morphological features are processed by a three-layer CNN to obtain first intermediate output information;

[0020] The actual spectral information is attention weighted by a Transformer to obtain second intermediate output information;

[0021] The first intermediate output information and the second intermediate output information are tensor spliced to obtain to-be-processed information;

[0022] The information to be processed is processed using a first loss function to obtain the plastic type of the bottle and the corresponding confidence level.

[0023] In some embodiments, the method further includes:

[0024] The information to be processed is processed using a second loss function to obtain the contamination type and degree value of the bottle.

[0025] In some embodiments, generating corresponding sorting instructions based on the confidence level to sort the bottles includes:

[0026] If the confidence level is a direct sorting level, the working information of the nozzle is determined based on the bottle shape characteristics, and a sorting instruction corresponding to the working information is generated to sort the bottle.

[0027] If the confidence level is a reflux sorting level, a reflux sorting instruction corresponding to the reflux channel is generated to open the reflux channel and realize the sorting of the bottle.

[0028] If the confidence level is an intervention sorting level, an intervention sorting instruction corresponding to an unknown type is generated to sort the bottles.

[0029] In some embodiments, the sorting device includes a hyperspectral camera, and acquiring the initial spectral information of the bottle includes:

[0030] The hyperspectral camera synchronously acquires the initial spectral information of the bottle based on the trigger signal sent by the sorting device.

[0031] Secondly, this embodiment provides a multi-category automated return sorting system for plastics, the system comprising: an information acquisition module, an information processing module, and a plastic sorting module; wherein,

[0032] The information acquisition module is used to acquire point cloud data of the bottle, process the point cloud data using laser triangulation to obtain the bottle shape characteristics, and acquire the initial spectral information of the bottle and the working temperature and humidity of the working environment.

[0033] The information processing module is used to determine whether spectral compensation is needed based on the working temperature and humidity and the initial spectral information. If so, the working temperature and humidity and / or the initial spectral information are substituted into a preset compensation model to obtain spectral compensation information. The actual spectral information of the bottle is determined based on the initial spectral information and the spectral compensation information.

[0034] The plastic sorting module is used to input the bottle shape characteristics and the actual spectral information into a hybrid neural network model to obtain the plastic type and corresponding confidence level of the bottle, determine the confidence level corresponding to the confidence level, and generate corresponding sorting instructions based on the confidence level to achieve the sorting of the bottle.

[0035] Thirdly, this embodiment provides a computer-readable storage medium having a computer program stored thereon that can run on a processor, wherein when the computer program is executed by the processor, it implements a multi-category automated return sorting method for plastics as described in the first aspect.

[0036] By employing the above method, this application first acquires point cloud data of the bottle body, and then processes the point cloud data using laser triangulation to obtain the bottle body's morphological characteristics. It then acquires the initial spectral information of the bottle body and the operating temperature and humidity of the working environment. Based on the operating temperature and humidity and the initial spectral information, it determines whether spectral compensation is needed. If so, it substitutes the operating temperature and humidity and / or the initial spectral information into a preset compensation model to obtain spectral compensation information. Based on the initial spectral information and the spectral compensation information, it determines the actual spectral information of the bottle body. By considering contamination of the bottle body to compensate for the detected information, it obtains more accurate spectral information that reflects the actual spectral information of the bottle body material, providing accurate information guidance for subsequent bottle sorting.

[0037] Then, the bottle's morphological features and actual spectral information are substituted into a hybrid neural network model to obtain the bottle's plastic type and corresponding confidence level.

[0038] Finally, the confidence level corresponding to the confidence score is determined, and corresponding sorting instructions are generated based on the confidence level to achieve the sorting of the bottles. By using a hybrid neural network to specifically process the actual spectral information, and by superimposing the high accuracy of the actual spectral information, the hybrid neural network model can output more accurate information about the plastic type and corresponding confidence score of the bottle. Furthermore, the sorting device can automatically complete the sorting process, achieving high efficiency. Ultimately, this enables efficient and accurate sorting of multiple types of plastics. Attached Figure Description

[0039] Figure 1 This is a block diagram of an automated return sorting method for multiple types of plastics provided in this application.

[0040] Figure 2 This is a block diagram of a method provided in this application for substituting working temperature and humidity and / or initial spectral information into a preset compensation model to obtain spectral compensation information.

[0041] Figure 3This application provides a method block diagram for inputting bottle morphological features and actual spectral information into a hybrid neural network model to obtain the plastic type of the bottle and the corresponding confidence level.

[0042] Figure 4 This is a block diagram of a method for sorting bottles by generating corresponding sorting instructions based on confidence levels, as disclosed in this application.

[0043] Figure 5 This is a connection diagram of an automated return sorting system for multiple types of plastics provided in this application. Detailed Implementation

[0044] To better understand the purpose, technical solutions, and advantages of this application, the application has been described and illustrated below with reference to the accompanying drawings and embodiments. However, those skilled in the art should understand that this application can be implemented without these details. It will be apparent to those skilled in the art that various modifications can be made to the embodiments disclosed in this application, and the general principles defined in this application can be applied to other embodiments and application scenarios without departing from the principles and scope of this application. Therefore, this application is not limited to the illustrated embodiments, but conforms to the broadest scope consistent with the scope of protection claimed in this application.

[0045] The embodiments of this application will now be described in further detail with reference to the accompanying drawings.

[0046] The specific application scenario of this application is a sorting device. This sorting device includes a main conveyor belt, a laser, a hyperspectral camera, a microwave sensor, a valve array, a screw conveyor, an environmental sensor, a secondary conveyor belt, and a sorting control terminal. The main conveyor belt is 800mm wide with an adjustable speed range of 0.5-4m / s. The laser is a 3D linear array camera with a resolution of 2048*2048 pixels and a frame rate of 120fps. Three lasers are mounted at a 30° angle, 1.2m above the main conveyor belt, at a 75° angle to the material movement direction, ensuring that each bottle is scanned with at least 200 spectral lines. The hyperspectral camera is a six-band InGaAs linear array camera with a wavelength range of 900-1700nm, a spectral resolution of 10nm, and a scanning line frequency of 200Hz, and is equipped with a narrowband filter. The light-reflecting wheel enables rapid switching between six characteristic wavelength bands: 905nm, 1070nm, 1215nm, 1300nm, 1450nm, and 1550nm. The air valve array consists of five groups spaced 30cm apart. The nozzles employ a fan-shaped diffusion design with a diameter of 8mm, operating at a pressure of 0.6MPa and a coverage width of 1.2m, ensuring effective sorting of large bottles. The screw conveyor has a lifting height of 3m and a processing capacity of 8 tons / hour. Environmental sensors include temperature and humidity sensors. A secondary conveyor belt transports objects brought by the screw conveyor. The sorting control unit includes a data processing platform, such as an NVIDIA Jetson AGX Xavier edge computing unit, which incorporates relevant sorting algorithms and control protocols, such as EtherCAT real-time communication with a 1ms cycle.

[0047] The system employs a spatially coordinated layout of lasers and hyperspectral cameras for information acquisition. Additionally, a set of standard reflectors, including white and black boards, are placed on each side of the main conveyor belt. When a bottle falls onto the main conveyor belt, the laser, hyperspectral camera, environmental sensors, and standard reflectors acquire information about the bottle and perform compensation to obtain more accurate bottle information. Then, relevant algorithms process the bottle information to determine where the bottle should be sorted and control the corresponding air valves to implement this, ultimately achieving efficient and accurate bottle sorting.

[0048] Figure 1 This is a block diagram of an automated return sorting method for multiple categories of plastics provided in this application. Figure 1 As shown, an automated recycle sorting method for multiple types of plastics includes the following steps:

[0049] Step S100: Obtain point cloud data of the bottle body, and process the point cloud data using laser triangulation to obtain the bottle body shape features.

[0050] This application describes the process from the perspective of the sorting control end of a sorting device. A multi-laser collaborative triangulation scheme is employed to achieve sub-millimeter-level 3D reconstruction. Three lasers emit laser light to form parallel stripes with a spacing of 15 nm on the bottle surface. The deformed laser stripes are then captured at 120 fps to obtain the laser data of the bottle. The laser exposure time can be automatically adjusted according to the speed of the main conveyor belt to ensure that the deformed laser stripes are captured at 120 fps, thus eliminating motion blur.

[0051] After the sorting control terminal acquires the laser data, it can establish the laser plane equation through a standard spherical target, determine the center line of the bottle using the Steger algorithm, and then obtain the point cloud data of the bottle based on the camera optical center, image point, and laser plane equation.

[0052] The sorting control unit synchronously acquires point cloud data of the bottles based on trigger signals sent by the sorting device. Specifically, when bottles need to be sorted, the sorting device sends a trigger signal, activating a laser to obtain laser data about the bottle. This data is then sent to the sorting control unit, which processes the laser data to obtain point cloud data. This provides guidance on when to acquire the initial spectral information of the bottle, ensuring that both point cloud data and initial spectral information are available for the same bottle.

[0053] After acquiring point cloud data, the sorting control unit can perform spiral fitting on the point cloud data of the bottle body area to obtain targeted point cloud data for the bottle opening. This bottle opening point cloud data is then used to replace the corresponding point cloud data, resulting in more accurate point cloud data for the bottle body. Finally, RANSAC is used to process the newly obtained point cloud data to obtain the bottle's morphological characteristics. This approach, by utilizing three lasers and processing the laser data using the Steger algorithm, as well as reprocessing specific locations, yields more accurate point cloud data, ensuring the accuracy of the subsequently obtained bottle morphological characteristics. Furthermore, trigger signals from the sorting device are used to synchronously acquire the bottle's point cloud data, providing guidance for subsequently acquiring appropriate initial spectral information for the same bottle.

[0054] Step S200: Obtain the initial spectral information of the bottle and the working temperature and humidity of the working environment, and determine whether spectral compensation is needed based on the working temperature and humidity and the initial spectral information.

[0055] The hyperspectral camera synchronously acquires the initial spectral information of the bottle based on the trigger signal sent by the sorting device. Specifically, the sorting device simultaneously sends a trigger signal to the hyperspectral camera, which, upon receiving the trigger signal, immediately engages with the narrowband filter wheel to rapidly switch between six characteristic bands: 905nm, 1070nm, 1215nm, 1300nm, 1450nm, and 1550nm, to acquire the initial spectral information of the bottle. This information is then sent to the sorting control unit, ensuring that the sorting control is almost synchronized in obtaining the initial spectral information and point cloud data. This approach ensures that both point cloud data and initial spectral information are obtained for the same bottle, reducing the likelihood of mismatches between the point cloud data and the initial spectral information for the same bottle.

[0056] The aforementioned working temperature and humidity specifically refer to the temperature and humidity of the bottle's environment. This can be obtained using environmental sensors and sent to the sorting control terminal so that the terminal can acquire the working environment's temperature and humidity. Specifically, the working temperature and humidity can be acquired each time point cloud data or initial spectral information is obtained. The difference between the current working temperature and humidity and the previous working temperature and humidity is then calculated and compared to a preset temperature and humidity level. If the working temperature and humidity exceed the preset level, spectral compensation is needed; if they do not exceed the preset level, no compensation is required. The preset temperature and humidity level can be determined based on the material's coefficient of thermal expansion, which is obtained from historical experiments and stored in the sorting control terminal.

[0057] Among the six characteristic wavelength bands—905nm, 1070nm, 1215nm, 1300nm, 1450nm, and 1550nm—the 1300nm band is specifically used to detect the hydroxyl absorption peak of PET bottles, the 1550nm band identifies the CH bond stretching vibration characteristics of HDPE, and the 1450nm band detects the third harmonic absorption peak of CH bonding in PP materials. Without spectral compensation, there will be high reflectivity in a certain characteristic wavelength band; with spectral compensation, there will be low reflectivity in another characteristic wavelength band. For example, if the bottle is contaminated with oil or labels, the reflectivity of the bottle in the characteristic wavelength band corresponding to that material will decrease. The bottle will not have reflectivity in characteristic wavelength bands not corresponding to its material. Therefore, by examining the initial optical distribution information, it can be determined whether spectral compensation is needed at the level of contamination.

[0058] Step S300: If necessary, substitute the working temperature and humidity and / or the initial spectral information and / or the bottle motion information into the preset compensation model to obtain spectral compensation information, and determine the actual spectral information of the bottle based on the initial spectral information and the spectral compensation information.

[0059] When determining which spectrum needs compensation, whether it is a spectrum that needs compensation based on temperature and humidity or a spectrum that needs compensation based on pollution, it is determined that a spectrum needs compensation.

[0060] The preset compensation models include a temperature and humidity layer compensation model and a pollution layer compensation model. When only temperature and humidity layer spectral compensation is needed, the spectral compensation information includes temperature and humidity layer spectral compensation information. When only pollution layer spectral compensation is needed, the spectral compensation information includes pollution layer spectral compensation information. When both temperature and humidity layer spectral compensation and pollution layer spectral compensation are needed, the spectral compensation information includes both temperature and humidity layer spectral compensation information and pollution layer spectral compensation information. Figure 2 This is a block diagram of the method provided in this application for obtaining spectral compensation information by substituting operating temperature and humidity and / or initial spectral information into a preset compensation model. For example... Figure 2 As shown, the process of substituting operating temperature and humidity and / or initial spectral information into a preset compensation model to obtain spectral compensation information includes the following steps:

[0061] Step S301: Substitute the working temperature and humidity into the temperature and humidity layer compensation model to obtain the temperature and humidity layer spectral compensation information.

[0062] Step S302, and / or, substitute the initial spectral information into the contamination layer compensation model to obtain contamination layer spectral compensation information.

[0063] The aforementioned temperature and humidity layer compensation model refers to the correspondence between temperature / humidity and spectral shift. Specifically, a material property database is provided. For temperature compensation, the thermal expansion coefficients of each material are pre-stored. The information requiring compensation due to temperature can be obtained by using the temperature and humidity layer spectral compensation information = thermal expansion coefficient * (operating temperature and humidity - 25).

[0064] Furthermore, the sorting control terminal also stores optical path attenuation curves for different wavelengths under different humidity levels. For humidity compensation, information on the need for compensation due to humidity can be obtained by using the corresponding optical path attenuation curve.

[0065] When determining the required temperature and humidity layer compensation spectrum, the temperature and humidity are substituted into the aforementioned temperature and humidity layer compensation model to obtain the temperature and humidity layer spectral compensation information.

[0066] The aforementioned contamination layer compensation model is specifically an anti-interference spectral database. The establishment of this database comprises three parts: laboratory simulation, data acquisition, and model building. Laboratory simulation involves using accelerated aging experiments to simulate real-world contamination scenarios, including 12 contamination states such as gradient contamination and labels made of different materials. Data acquisition uses a hyperspectral camera to record the spectral attenuation curves under each contamination state. Model building employs Gaussian process regression to establish a contamination type-waveform attenuation compensation model, dynamically adjusting parameters through Bayesian optimization to ensure model accuracy. Specifically, after obtaining the initial spectral information, a Markov decision process is used to process the initial spectral information to determine the contamination type, which is then substituted into the contamination type-waveform attenuation compensation model to obtain the contamination layer spectral compensation information.

[0067] Next, the spectral compensation information of the temperature and humidity layer and the spectral compensation information of the contamination layer are superimposed on different wavelengths to obtain the spectral compensation information. Finally, the initial spectral information and the spectral compensation information are superimposed on different waveforms to obtain the actual spectral information of the bottle after compensation.

[0068] Preferably, the sorting control terminal also acquires the actual reflectivity of the whiteboard at preset time intervals, determines whether the actual reflectivity deviation between the actual reflectivity and the set reflectivity of the whiteboard is greater than the trigger compensation reflectivity deviation, if it is not greater than the set reflectivity deviation, generates a reflector no deviation command; if it is greater than the set reflectivity deviation, generates reflector spectral compensation information based on the actual reflectivity deviation, and uses the spectral compensation information to update the actual spectral information.

[0069] A set of standard reflectors is installed on each side of the main conveyor belt. Each set includes a white board and a black board, with the white board set to have a reflectivity of 98%. A hyperspectral camera is used at preset time intervals to acquire the actual reflectivity of the white board. The difference between the actual reflectivity and the set reflectivity is then used to obtain the actual reflection deviation. Next, the actual reflection deviation is compared with a set trigger compensation reflection deviation. If the driver's reflectivity deviation is not greater than the trigger compensation reflection deviation, it indicates that the standard reflector has minimal influence on the initial spectral information and can be ignored, requiring no further compensation. At this point, the next step can be performed: generating a reflector no-deviation command to further determine the plastic type of the bottle.

[0070] If the actual reflectance deviation is greater than the triggered compensation reflectance deviation, it indicates that the standard reflector will affect the initial spectral information. In this case, a trigger compensation mechanism is needed. This involves substituting the actual reflectance deviation into a preset trigger compensation algorithm to obtain reflector spectral compensation information. Then, this reflector spectral compensation information is superimposed on the obtained actual spectral information for different waveforms to update the actual spectral information and obtain new actual spectral information, thus completing the update of the actual spectral information. The preset trigger compensation algorithm can be obtained through extensive offline experiments, and the preset time can be determined according to actual conditions; this application preferentially sets it to thirty minutes. By further considering the state of the standard reflector and further correcting the actual spectral information of the bottle, the obtained actual spectral information can be made more accurate. This provides accurate bottle information for subsequent plastic sorting, thereby improving sorting efficiency when using a sorting device and indirectly improving the accuracy of subsequent sorting work.

[0071] Step S400: Substitute the bottle shape features and actual spectral information into the hybrid neural network model to obtain the plastic type of the bottle and the corresponding confidence level, determine the confidence level corresponding to the confidence level, and generate corresponding sorting instructions based on the confidence level to achieve the sorting of the bottle.

[0072] The hybrid neural network model described above is a CNN-Transformer hybrid neural network. Figure 3 This is a flowchart illustrating the method provided in this application for inputting bottle morphological features and actual spectral information into a hybrid neural network model to obtain the plastic type of the bottle and its corresponding confidence level. For example... Figure 3 As shown, the steps to input bottle morphology features and actual spectral information into a hybrid neural network model to obtain the bottle's plastic type and corresponding confidence level include:

[0073] Step S401: A three-layer CNN is used to process the bottle shape features to obtain the first intermediate output information.

[0074] Step S402: Attention weighting is applied to the actual spectral information using a Transformer to obtain the second intermediate output information.

[0075] Step S403: Tensor concatenation is performed on the first intermediate output information and the second intermediate output information to obtain the information to be processed.

[0076] Step S404: Use the first loss function to process the information to be processed to obtain the plastic type of the bottle and the corresponding confidence level.

[0077] The hybrid neural network consists of an input and an output. The input requires two data points: a 64x64x64 voxelized 3D matrix, including morphological features such as the bottle neck's threads and the bottle bottom's unevenness; and a 256x256 reflectance matrix composed of actual 6-band spectral information. The output is the material classification probability, i.e., the confidence score, specifically the plastic type and its corresponding confidence score.

[0078] For the CNN branch, which is a 3D type, it uses three layers of 3*3*3 convolutional kernels to process the bottle shape features. Each layer is combined with ReLU activation and BatchNorm, and then max pooling (stride=2) is used to gradually reduce the feature map to 8*8*8, thus obtaining the first intermediate output information. In particular, deformable convolution is introduced in the third layer to adaptively capture the bottle deformation features, which can quickly and accurately obtain the first intermediate output information.

[0079] For the Transformer branch, the actual spectral information of the 6 bands is converted into a 256-dimensional embedding vector, and a 4-head attention mechanism is used to focus on learning the correlation of certain key bands. Environmental parameters such as temperature and humidity are added to the position encoding to obtain the second intermediate output information.

[0080] Next, bilinear interpolation is used to tensor-concatenate the first and second intermediate outputs, and a cross-attention mechanism is employed to achieve spatiotemporal alignment, thus obtaining the information to be processed. Finally, FocalLoss, the first loss function, is used to process the information, resulting in the output of the hybrid neural network showing the plastic type and corresponding confidence level of the bottle. By specifically processing the actual spectral information using a hybrid neural network, and by combining this with the high accuracy of the actual spectral information, the hybrid neural network model can output a more accurate information about the plastic type and corresponding confidence level of the bottle.

[0081] Preferably, a second loss function can also be used to process the information to be processed in order to obtain the pollution type and pollution degree value of the bottle.

[0082] Specifically, the hybrid neural network model also incorporates a second loss function, Smooth L1 Loss. This second loss function is used to process the information to be processed, yielding a more accurate value for the type and degree of contamination of the bottle. This further verifies the accuracy of the contamination type obtained solely from the initial spectral information, thus providing a more accurate contamination type and its corresponding degree of contamination.

[0083] The sorting control terminal pre-stores a confidence level and confidence grade relationship table. This table was obtained through extensive offline experiments. The corresponding confidence grade can be obtained by substituting the obtained confidence levels into the table. The table includes three levels: direct sorting grade, return sorting grade, and intervention sorting grade. For example, when the confidence level is greater than 90%, the corresponding confidence grade is direct sorting; when the confidence level is not less than 70% and not greater than 90%, the corresponding confidence grade is return sorting; and when the confidence level is less than 70%, the corresponding confidence grade is intervention sorting.

[0084] Once the confidence level of the bottle is determined, it is determined where in the sorting device the bottle needs to be sorted, and then the sorting action needs to be carried out. Figure 4 This is a block diagram of a method for sorting bottles based on confidence levels, as disclosed in this application. Figure 4 As shown, generating corresponding sorting instructions based on confidence levels to achieve bottle sorting includes the following steps:

[0085] Step S405: If the confidence level is direct sorting level, determine the working information of the nozzle based on the bottle shape characteristics, and generate the sorting instruction corresponding to the working information to achieve the sorting of the bottle.

[0086] Step S406: If the confidence level is the reflux sorting level, generate a reflux sorting instruction corresponding to the reflux channel to open the reflux channel and realize the sorting of the bottle.

[0087] Step S407: If the confidence level is the intervention sorting level, generate the intervention sorting instruction corresponding to the unknown type to achieve the sorting of the bottle.

[0088] When the confidence level is direct sorting level, it indicates that a nozzle is needed to provide driving force to the bottle so that the bottle moves to the sorting position corresponding to its material. The bottle shape characteristics obtained in step S100 above can be used to determine the orientation of the platform on the main conveyor belt, thereby further determining the working information required for the nozzle to work. This working information includes the working angle, working nozzle, and working air valve volume to ensure that the bottle can move accurately to the corresponding position.

[0089] Specifically, the formula for calculating the working air valve volume is: Working air valve volume = Nozzle coefficient * (Cylinder material density * Main conveyor belt speed + Bottle deformation compensation item). (1 / 2)The nozzle coefficient, obtained through offline testing and calibrated, reflects the influence of nozzle shape on airflow efficiency. The calibration method involves measuring the airflow coverage of different nozzles under standard operating conditions and fitting the result to obtain the nozzle coefficient. This application preferentially selects a nozzle coefficient of 1.05. The bottle material density determines the airflow impact momentum requirement. After determining the bottle material, the density can be obtained by referring to a table. The main conveyor belt speed affects the bottle displacement compensation amount, which can be obtained by reading the motor's operating speed. The bottle deformation compensation term is calculated from the volume compression ratio, i.e., bottle deformation compensation term = compensation coefficient * volume compression ratio. This compensation coefficient is determined through impact mechanics experiments, and the volume compression ratio can be obtained simultaneously from the initial spectral information obtained in step S200 above.

[0090] For nozzle selection, the coordinates of the bottle's center of gravity can be detected by a laser. Then, based on the preset nozzle selection, the coordinates of the bottle's center of gravity are substituted into the valve number = x-coordinate of the bottle's center of gravity / y-coordinate of the bottle's center of gravity + 1. Next, taking into account the position prediction error caused by the deformation of the bottle, the valve number is corrected by a motion tracking algorithm to determine which nozzle needs to be used.

[0091] To determine the working angle, the offset of the bottle's center of gravity (vertical coordinate) is substituted into the formula: nozzle angle = arctan(offset of bottle's center of gravity (vertical coordinate) / (main conveyor belt speed * airflow duration)) to obtain the initial working angle. The airflow duration can be determined based on actual conditions; in this application, it is preferably 50 ms. Next, considering the torque effect of the airflow on the flattened bottle, the initial working angle needs to be substituted into the angle correction algorithm to obtain the final working angle. By considering the bottle's state to determine the necessary working information for the nozzle, the bottle can be more accurately sorted into its corresponding position, thus achieving fast and accurate bottle sorting.

[0092] When the confidence level is reflux sorting level, it indicates that the label area of ​​the bottle is too large. In this case, a screw conveyor is needed to transport the bottle to the buffer bin, that is, to generate a reflux sorting instruction corresponding to the reflux channel to open the reflux channel and realize the sorting of the bottle.

[0093] In addition, during the process of transporting the bottle to the buffer chamber, microwave sensors detect the water content of the bottle and X-rays detect metal residues, so as to prepare for the subsequent sorting of the bottle and facilitate the sorting of the bottle later.

[0094] When the confidence level is at the manual sorting level, it indicates that an unknown material or a material that matches the specifications has been detected. At this point, X-rays can be activated to detect the metal composition, and high-resolution images and spectral data can be uploaded to the cloud. After a certain number of images have been accumulated, manual sorting can be performed, generating intervention sorting instructions corresponding to the unknown type to achieve the sorting of the bottles. In this way, by obtaining the plastic type of the bottle and its corresponding confidence level, different sorting operations can be performed based on different situations, enabling fast and accurate sorting of the bottles.

[0095] Figure 5 This is a connection diagram of a multi-category automated return sorting system for plastics provided in this application. Figure 5 As shown, a multi-category automated plastic return sorting system includes: an information acquisition module, an information processing module, and a plastic sorting module.

[0096] The system comprises several modules: The information acquisition module acquires point cloud data of the bottle, processes it using laser triangulation to determine its morphological features, and obtains initial spectral information and the operating temperature and humidity of the working environment. The information processing module determines whether spectral compensation is needed based on the operating temperature and humidity and initial spectral information. If so, it substitutes the operating temperature and humidity and / or initial spectral information into a preset compensation model to obtain spectral compensation information. Based on the initial spectral information and the spectral compensation information, it determines the actual spectral information of the bottle. The plastic sorting module inputs the bottle's morphological features and actual spectral information into a hybrid neural network model to obtain the bottle's plastic type and corresponding confidence level. It then determines the confidence level corresponding to the confidence level and generates corresponding sorting instructions based on the confidence level to achieve bottle sorting.

[0097] The other functions performed by the information acquisition module, information processing module, and plastic sorting module, as well as the technical details of each function, are the same as or similar to the corresponding features in the multi-category automated return sorting method for plastics described above, so they will not be repeated here.

[0098] This application also provides a computer storage medium storing a computer program that, when run on a computer, enables the computer to execute the steps in the previously described automated return sorting method for multiple types of plastics.

[0099] It should be understood that although the steps in the flowcharts in the accompanying figures are shown sequentially as indicated by the arrows, these steps are not necessarily performed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order requirement for the execution of these steps, and they can be performed in other orders.

[0100] The above description is only a partial embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A multi-category automated return sorting method for plastics, characterized in that, The method is applied to a sorting device comprising three uniformly distributed ring lasers, a hyperspectral camera, and standard reflectors on both sides of a conveyor belt. The method includes: The point cloud data of the bottle is acquired, and the point cloud data is processed using laser triangulation to obtain the bottle shape characteristics of the bottle. The system acquires the initial spectral information of the bottle and the operating temperature and humidity of the working environment. Based on the operating temperature and humidity and the initial spectral information, it determines whether spectral compensation is needed. If necessary, the working temperature and humidity and / or the initial spectral information are substituted into the preset compensation model to obtain spectral compensation information, and the actual spectral information of the bottle is determined based on the initial spectral information and the spectral compensation information. The bottle's morphological features and the actual spectral information are substituted into a hybrid neural network model to obtain the bottle's plastic type and corresponding confidence level. The confidence level corresponding to the confidence level is determined, and a corresponding sorting instruction is generated based on the confidence level to achieve the sorting of the bottle. The preset compensation model includes a temperature and humidity layer compensation model and a pollution layer compensation model. The spectral compensation information includes temperature and humidity layer spectral compensation information and / or pollution layer spectral compensation information. Substituting the working temperature and humidity and / or the initial spectral information into the preset compensation model to obtain the spectral compensation information includes: The operating temperature and humidity are substituted into the temperature and humidity layer compensation model to obtain the temperature and humidity layer spectral compensation information. And / or, substitute the initial spectral information into the contamination layer compensation model to obtain contamination layer spectral compensation information; Based on the trigger signal sent by the sorting device, the laser data of the bottle is acquired synchronously by three lasers. The Steger algorithm is used to determine the center line of the bottle. After the point cloud of the bottle area is fitted with a spiral line to correct the point cloud data of the bottle mouth, the point cloud data is processed by the laser triangulation method to obtain the bottle shape characteristics. Based on the trigger signal sent by the sorting device, the initial spectral information of the bottle is synchronously acquired by the hyperspectral camera, the working temperature and humidity of the working environment are acquired by the environmental sensor, and the actual reflectivity of the white board in the standard reflector is acquired at preset intervals. If the deviation between the actual reflectivity and the set reflectivity of the white board is greater than the trigger compensation reflectivity deviation, reflector spectral compensation information is generated based on the deviation.

2. The method according to claim 1, characterized in that, The method is applied in a sorting device containing three lasers, which are uniformly distributed in a ring on the sorting device. The acquisition of point cloud data of the bottle includes: Point cloud data of the bottle is acquired synchronously based on the trigger signal sent by the sorting device.

3. The method according to claim 2, characterized in that, The sorting device includes a conveyor belt, and a set of standard reflective plates are installed on each side of the conveyor belt. Each set of standard reflective plates includes a white board. The method further includes: At preset time intervals, the actual reflectivity of the whiteboard is obtained, and it is determined whether the actual reflectivity deviation between the actual reflectivity and the set reflectivity of the whiteboard is greater than the trigger compensation for reflectivity deviation. If it is not greater, a reflector no-deviation command is generated. If the deviation is greater than the actual reflectivity, reflector spectral compensation information is generated based on the actual reflectivity deviation, and the actual spectral information is updated using the spectral compensation information.

4. The method according to claim 1, characterized in that, The hybrid neural network model is CNN- The Transformer hybrid neural network, wherein the step of substituting the bottle morphology features and the actual spectral information into the hybrid neural network model to obtain the plastic type of the bottle and the corresponding confidence level includes: The bottle shape features are processed using a three-layer CNN to obtain the first intermediate output information; The actual spectral information is attention-weighted using a Transformer to obtain the second intermediate output information. Tensor concatenation is performed on the first intermediate output information and the second intermediate output information to obtain the information to be processed; The information to be processed is processed using a first loss function to obtain the plastic type of the bottle and the corresponding confidence level.

5. The method according to claim 4, characterized in that, The method further includes: The information to be processed is processed using a second loss function to obtain the contamination type and degree value of the bottle.

6. The method according to claim 1, characterized in that, The step of generating corresponding sorting instructions based on the confidence level to sort the bottles includes: If the confidence level is a direct sorting level, the working information of the nozzle is determined based on the bottle shape characteristics, and a sorting instruction corresponding to the working information is generated to sort the bottle. If the confidence level is a reflux sorting level, a reflux sorting instruction corresponding to the reflux channel is generated to open the reflux channel and realize the sorting of the bottle. If the confidence level is an intervention sorting level, an intervention sorting instruction corresponding to an unknown type is generated to sort the bottles.

7. The method according to claim 2, characterized in that, The sorting device includes a hyperspectral camera, and acquiring the initial spectral information of the bottle includes: The hyperspectral camera synchronously acquires the initial spectral information of the bottle based on the trigger signal sent by the sorting device.

8. A multi-category automated return sorting system for plastics, characterized in that, The system is applied to a sorting device comprising three uniformly distributed ring lasers, a hyperspectral camera, and standard reflectors on both sides of a conveyor belt. The system includes: an information acquisition module, an information processing module, and a plastic sorting module; wherein... The information acquisition module is used to acquire point cloud data of the bottle, process the point cloud data using laser triangulation to obtain the bottle shape characteristics, and acquire the initial spectral information of the bottle and the working temperature and humidity of the working environment. The information processing module is used to determine whether spectral compensation is needed based on the working temperature and humidity and the initial spectral information. If so, the working temperature and humidity and / or the initial spectral information are substituted into a preset compensation model to obtain spectral compensation information. The actual spectral information of the bottle is determined based on the initial spectral information and the spectral compensation information. The plastic sorting module is used to input the bottle shape characteristics and the actual spectral information into a hybrid neural network model to obtain the plastic type and corresponding confidence level of the bottle, determine the confidence level corresponding to the confidence level, and generate corresponding sorting instructions based on the confidence level to realize the sorting of the bottle. The preset compensation model includes a temperature and humidity layer compensation model and a pollution layer compensation model. The spectral compensation information includes temperature and humidity layer spectral compensation information and / or pollution layer spectral compensation information. Substituting the working temperature and humidity and / or the initial spectral information into the preset compensation model to obtain the spectral compensation information includes: The operating temperature and humidity are substituted into the temperature and humidity layer compensation model to obtain the temperature and humidity layer spectral compensation information. And / or, substitute the initial spectral information into the contamination layer compensation model to obtain contamination layer spectral compensation information; Based on the trigger signal sent by the sorting device, the laser data of the bottle is acquired synchronously by three lasers. The Steger algorithm is used to determine the center line of the bottle. After the point cloud of the bottle area is fitted with a spiral line to correct the point cloud data of the bottle mouth, the point cloud data is processed by the laser triangulation method to obtain the bottle shape characteristics. Based on the trigger signal sent by the sorting device, the initial spectral information of the bottle is synchronously acquired by the hyperspectral camera, the working temperature and humidity of the working environment are acquired by the environmental sensor, and the actual reflectivity of the white board in the standard reflector is acquired at preset intervals. If the deviation between the actual reflectivity and the set reflectivity of the white board is greater than the trigger compensation reflectivity deviation, reflector spectral compensation information is generated based on the deviation.

9. A computer-readable storage medium having a computer program stored thereon that can run on a processor, characterized in that, When the computer program is executed by the processor, it implements a multi-category automated return sorting method for plastics as described in any one of claims 1 to 7.

Images