Method for controlling timing of injection valve for multi-channel PET bottle flake sorting

Abstract

The present application relates to the technical field of artificial intelligence pruning training, and discloses a spray valve timing control method for multi-channel PET bottle chip sorting, comprising: first, establishing a bottle chip cross-nozzle sequence according to the transverse coverage range of the PET bottle chip and the transverse sequence of the nozzle; then, establishing a spray valve timing control neural network that can be pruned, and obtaining the spray valve action output after removing the candidate network unit when evaluating the candidate network unit. If the spray valve action label in the spray valve action output indicates that all the nozzles of the action are located within the bottle chip cross-nozzle sequence and the nozzle numbers are continuous, the candidate network unit is allowed to be deleted, and a pruned model is obtained. The pruned model is used to process the PET bottle chips to be sorted, generate a spray valve action label, a spray valve opening time and a spray valve closing time, and control the opening and closing of the corresponding spray valve, thereby completing the pneumatic sorting of the PET bottle chips.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence pruning training technology, and more specifically, to a spray valve timing control method for multi-channel PET flake sorting. Background Technology

[0002] Multi-channel PET flake sorting lines typically include conveying, imaging and recognition, control calculation, and spray valve array execution. PET flakes are discretely distributed on the conveyor belt, with variations in lateral position and width. The control side needs to match the flakes to be rejected to one or more spray nozzles based on the recognition results, and determine the opening and closing times of each spray valve. To meet high-speed sorting cycle times, the production line controller often uses neural networks to process images and generate spray valve timing outputs, while pruning is used to reduce model size and operating overhead.

[0003] In existing implementations, pruning is often based on weights, channel importance, or runtime overhead, which can weaken the lateral sequence constraints of the nozzle array. For PET flakes spanning multiple nozzles, if the nozzles triggered after model compression exceed the flake coverage area, or if nozzles that should be adjacent fail to trigger consecutively, the spray valve's spray area may not match the actual area the flake passes through, affecting pneumatic sorting execution. Therefore, in multi-channel spray valve control scenarios, it is necessary to ensure that model compression constraints and spray valve array execution constraints are consistent, avoiding decisions based solely on internal model metrics while ignoring the correspondence between the flake coverage area and the spray valve channels. Summary of the Invention

[0004] This invention provides a spray valve timing control method for multi-channel PET bottle flake sorting, which solves the technical problems mentioned in the background art.

[0005] To achieve the above-mentioned objectives, the present invention provides the following technical solution.

[0006] A method for timing control of spray valves in multi-channel PET bottle flake sorting includes: Based on the correspondence between the lateral coverage of PET bottle flakes and multiple nozzles, a bottle flake cross-nozzle order is established to characterize the continuous nozzle range of the PET bottle flakes. The multiple nozzles are arranged in lateral order, have nozzle numbers that characterize the lateral order, and have a definite correspondence with each spray valve. A valve timing control neural network is established that can be used for structured pruning and generate valve timing outputs. The valve timing outputs include valve action flags, valve opening times, and valve closing times for each valve. For a PET bottle sample with the bottle sheet cross-nozzle sequence and a candidate network structure unit of the spray valve timing control neural network, obtain the spray valve action output after removing the candidate network structure unit. The nozzles whose actions are indicated by the spray valve action mark in the spray valve action output are taken as the action nozzles. It is determined whether all action nozzles are located within the bottle piece cross-spray nozzle sequence and whether the nozzle numbers are consecutive. If both are true, it is determined that the condition of maintaining the bottle piece cross-spray nozzle sequence is met. Structure deletion is performed only on candidate network structural units that satisfy the condition of maintaining the order of bottle flakes across nozzles, resulting in a structured pruning model.

[0007] Compared with the prior art, the present invention has the following substantial features and significant progress: This invention links the lateral coverage of PET bottle flakes, the continuity of nozzle numbering, and the deletion criteria of neural network pruning, so that the model compression process is constrained by the sequence of nozzle channels. The pruned model still outputs nozzle action markers, nozzle opening times, and nozzle closing times according to the bottle flake cross-nozzle sequence, and drives the corresponding nozzles to open and close. In this way, the nozzle selection is kept in line with the bottle flake coverage in multi-channel pneumatic sorting of PET bottle flakes, reducing the channel out-of-bounds and nozzle discontinuity problems caused by the timing output of nozzles during pruning. Attached Figure Description

[0008] Figure 1 This is a diagram of the spray valve timing control architecture for multi-channel PET bottle flake sorting according to the present invention. Detailed Implementation

[0009] The following description is provided in conjunction with the accompanying drawings. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the invention; those skilled in the art can make equivalent substitutions or combinations for the specific implementations.

[0010] This embodiment relates to a valve timing control method for multi-channel PET flake sorting, applicable to scenarios where PET flakes move along a PET flake conveyor belt and undergo pneumatic sorting via a multi-channel valve array. In this scenario, after the PET flakes pass through the image acquisition area, the control processor generates a valve timing output based on the PET flake image, causing the corresponding valves to perform opening and closing actions according to their opening and closing times. Multiple nozzles in the multi-channel valve array are arranged in a horizontal sequence, each nozzle having a nozzle number representing the horizontal sequence, and the nozzle number has a definite correspondence with each valve.

[0011] Traditional neural network compression processes typically focus on network structure size, channel importance, training loss, or validation loss. However, in the pneumatic sorting of PET flakes, there is a spatial relationship between nozzles. PET flakes have a practical coverage area in the lateral direction, and the nozzles that can effectively act on a PET flake should fall within the range of consecutive nozzles corresponding to that lateral coverage area, and the nozzle numbers should remain consecutive. If structured pruning only performs structure deletion based on channel weights or other conventional channel importance, the structured pruning model may still output valve action markers, valve opening times, and valve closing times, but the spatial correspondence between the action nozzles and the lateral coverage area of ​​the PET flake may be disrupted.

[0012] In this implementation, when establishing the valve timing control neural network and implementing structured pruning, the order of bottle flakes across nozzles is used as the admission criterion for whether candidate network structural units can be deleted. The order of bottle flakes across nozzles represents the range of consecutive nozzles corresponding to the PET bottle flakes, derived from the correspondence between the lateral coverage of the PET bottle flakes and multiple nozzles. After a candidate network structural unit is removed, the control processor obtains the valve action output and takes the nozzles marked with the valve action flag in the valve action output as the action nozzles; only when all action nozzles are within the bottle flakes across nozzle order and the nozzle numbers are consecutive, is it determined that the condition of maintaining the bottle flakes across nozzle order is met, and structural deletion of the candidate network structural unit is allowed.

[0013] See Figure 1 This embodiment can be implemented collaboratively by a PET flake conveyor belt 100, PET flakes 110, an image acquisition unit 200, a control processor 300, a flake cross-nozzle sequence generation module 310, a spray valve timing control neural network 320, a candidate network structure unit evaluation module 330, a pruning admission judgment module 340, a spray valve driver 400, and a multi-channel spray valve array 500. The PET flake conveyor belt 100 carries the PET flakes 110, the image acquisition unit 200 acquires images of the PET flakes, the control processor 300 executes flake cross-nozzle sequence generation, spray valve timing control neural network processing, candidate network structure unit evaluation, and pruning admission judgment, the spray valve driver 400 generates a spray valve drive signal based on the spray valve timing output, and the multi-channel spray valve array 500 performs pneumatic sorting of the PET flakes 110 based on the spray valve drive signal.

[0014] Furthermore, a nozzle-crossing order for PET bottle flakes is established to characterize the continuous nozzle range corresponding to the PET bottle flakes, as follows: In one embodiment, the control processor 300 first establishes a bottle sheet cross-nozzle order representing the range of consecutive nozzles corresponding to the PET bottle sheet, based on the correspondence between the lateral coverage area of ​​the PET bottle sheet and multiple nozzles. The lateral coverage area of ​​the PET bottle sheet refers to the coverage region of the PET bottle sheet in the direction of the multiple nozzles; the multiple nozzles are arranged in lateral order and have nozzle numbers representing the lateral order. The bottle sheet cross-nozzle order is not a general image recognition result, but a sequence of consecutive nozzle numbers obtained by mapping the lateral coverage area of ​​the PET bottle sheet to the nozzle number order.

[0015] During implementation, the image acquisition device 200 acquires images of PET bottle flakes, and the control processor 300 determines the left and right positions of the PET bottle flakes within the images. The left position of the bottle flake is determined using... This indicates that the right side of the bottle is used... Indicates. The first of multiple nozzles Each nozzle has a lateral nozzle position The lateral position of the nozzles is determined according to the lateral order of the nozzle numbers. The control processor 300 determines the left and right nozzles based on the left and right positions of the bottle piece and the lateral position of each nozzle, and uses the nozzle numbers arranged consecutively from the left to the right as the bottle piece's nozzle crossing order.

[0016] ,in, Indicates the order in which the bottle flakes cross the nozzle. Indicates the nozzle number, Indicates the lateral position of the nozzle. Indicates the left side of the bottle. This indicates the position on the right side of the bottle label. This formula describes an feasible method for establishing the bottle label's cross-nozzle sequence, where the left-end nozzle is determined by a nozzle number that is no more than or close to the left side of the bottle label, and the right-end nozzle is determined by a nozzle number that is no less than or close to the right side of the bottle label. The nozzle numbers between the left and right ends form a continuous nozzle range in a horizontal order.

[0017] When generating the bottle flake cross-nozzle sequence, the bottle flake cross-nozzle sequence generation module 310 can determine the lateral coverage range of the PET bottle flake based on the contour, boundary, or mask in the PET bottle flake image, or it can determine it based on the correspondence between the calibrated lateral coordinates of the image and the lateral position of the nozzle. During calibration, the lateral arrangement direction of multiple nozzles in the multi-channel spray valve array 500 is used as the unified lateral coordinate direction. Within the acquisition field of view of the image acquisition unit 200, the lateral calibration position corresponding to the lateral position of the nozzle is selected, so that the lateral coordinates of the image and the lateral position of the nozzle are in the same lateral coordinate. The lateral position of each nozzle is the position of the corresponding nozzle center in the unified lateral coordinate, and is recorded with the nozzle number. After obtaining the PET bottle flake contour, boundary, or mask in the PET bottle flake image, the control processor 300 reads the left and right positions of the PET bottle flake in the unified lateral coordinate, and then determines the bottle flake cross-nozzle sequence based on the left and right positions of the bottle flake and the lateral position of the nozzle.

[0018] A correspondence is maintained between the order of bottle flakes across the nozzles and the PET bottle flake samples. PET bottle flake samples used for subsequent evaluation of candidate network structural units can be generated by associating the order of bottle flakes across the nozzles with the corresponding PET bottle flake images and nozzle valve action records. The nozzle valve action records include the nozzle valve action markers, valve opening times, and valve closing times for each nozzle. Thus, when nozzle valve action output is obtained under the influence of removing candidate network structural units, the pruning admission judgment module 340 can directly obtain the order of bottle flakes across the nozzles corresponding to the PET bottle flake sample and determine whether the action nozzle is still within the range of the corresponding continuous nozzles of the PET bottle flake.

[0019] In terms of training sample organization, two consecutive adjacent frames of images can be used as a model input sample. The valve action markers, valve opening times, and valve closing times of all valves in the same PET flake removal event are used as the valve timing labels for the model input sample. The valve opening and closing times can use a unified image timestamp as the time reference. This organization method establishes a correspondence between PET flake images, valve action records, and the order of flakes across nozzles within the same PET flake sample, facilitating the subsequent acquisition of valve action outputs under the influence of candidate network structural units being removed.

[0020] Furthermore, a valve timing control neural network is established that allows for structured pruning and generates valve timing outputs, as detailed below: The control processor 300 establishes a valve timing control neural network that allows for structured pruning and generates valve timing outputs. The valve timing control neural network 320 processes PET bottle flake samples and generates valve timing outputs based on these samples. The valve timing outputs include valve action flags, valve opening times, and valve closing times for each valve. The valve action flags are binary; a flag of 1 indicates the corresponding valve is performing an action, while a flag of 0 indicates it is not performing an action. The valve opening time indicates when the corresponding valve begins to open, and the valve closing time indicates when it closes. After the valve timing control neural network 320 outputs the valve action flags, the control processor 300 directly identifies the nozzles with a valve action flag of 1 as the operating nozzles.

[0021] The valve timing control neural network 320 can employ a neural network structure capable of structured pruning. In one specific implementation, the valve timing control neural network 320 includes an input layer that receives two consecutive frames of images, several convolutional layers for extracting information about the appearance and motion of the PET bottle flakes, and an output layer with a defined output relationship to the valve channel. This structure enables the valve timing control neural network 320 to extract features related to the position and motion of the PET bottle flakes from the PET bottle flake images, and to generate valve action markers, valve opening times, and valve closing times for the valve itself.

[0022] When establishing the valve timing control neural network 320, the initial state can be set as the unpruned state. In the unpruned state, the valve timing control neural network 320 retains all candidate network structure units to be evaluated, facilitating the subsequent acquisition of the valve action output under the influence of removing each candidate network structure unit. At this stage, the valve timing control neural network 320 is used for the valve timing output of multi-channel valve control.

[0023] When extracting features from PET bottle flake samples, the valve timing control neural network 320 can obtain appearance and motion information from the PET bottle flake images through convolutional layers. Appearance information reflects the region, shape, and lateral coverage of the PET bottle flake in the image, while motion information reflects the impact of changes in the PET bottle flake position between adjacent images on the valve opening and closing times. Based on the feature extraction results, the valve timing control neural network 320 generates valve action markers, valve opening times, and valve closing times for each valve, thus forming a complete valve timing output.

[0024] The training of the valve timing control neural network 320 can be supervised learning using PET bottle samples and their valve timing labels. During training, the valve action classification error is used to constrain the valve action label, the valve opening time error is used to constrain the valve opening time, and the valve closing time error is used to constrain the valve closing time. For valves whose actions are represented by the valve action label, the valve opening time and valve closing time have practical control significance. Therefore, the training process can calculate the valve opening time error and valve closing time error for valves whose actions are represented by the valve action label.

[0025] Furthermore, the valve action output after removing candidate network structure units is obtained, as follows: For PET bottle flake samples with a cross-nozzle sequence and candidate network structure units of the valve timing control neural network 320, the candidate network structure unit evaluation module 330 obtains the valve action output after removing the candidate network structure unit. The candidate network structure unit is a structure unit in the valve timing control neural network 320 that can be structurally pruned. Its removal will change the output state of the valve timing control neural network 320 for the PET bottle flake samples.

[0026] In one specific embodiment, the convolutional output channels in the valve timing control neural network 320 are used as candidate network structure units. Each convolutional output channel has a corresponding channel weight. The candidate network structure unit evaluation module 330 determines the candidate removal order based on the channel importance corresponding to the channel weights of the candidate network structure units, and sequentially obtains the valve action output under the influence of the removal of each candidate network structure unit according to the candidate removal order. In this embodiment, channel importance can be represented by channel scores, which are determined based on the L1 norm of the channel weights.

[0027] ,in, Indicates channel score. Indicates channel weight, Indicates the output channel number of the convolution. The L1 norm represents the channel weights. Channel scores are used to determine the candidate deletion order, which in turn determines the order in which the candidate network structure unit evaluation module 330 evaluates the candidate network structure units.

[0028] Obtaining the valve action output under the influence of removing a candidate network structural unit can be achieved by temporarily silencing the candidate network structural unit. Temporary silencing involves suppressing the output of the candidate network structural unit while maintaining network connections, thus obtaining the corresponding valve action output under temporary silencing. Specifically, when the candidate network structural unit is a convolutional output channel, the control processor 300 sets the output of that convolutional output channel to zero, and the network connections and model weights in the valve timing control neural network 320 remain intact. Forward inference is then performed on the PET bottle sample in this state to obtain the valve action output. Because the network connections are not deleted during the temporary silencing phase, the candidate network structural unit evaluation module 330 can observe the impact of removing the candidate network structural unit.

[0029] The valve action output indicates the state of the nozzle pointed to by the valve action marker after a candidate network structural unit is removed. The valve action output differs from the complete valve timing output, which includes the valve action marker, valve opening time, and valve closing time. In the pruning admission judgment, the valve action output is primarily used to determine which nozzles are marked as action nozzles by the valve action marker. Through this distinction, the pruning admission judgment module 340 can judge whether the action nozzles conform to the bottle-plate cross-nozzle order, while the valve timing control neural network 320 retains the ability to generate valve timing output.

[0030] The candidate network structure unit evaluation module 330 can perform the above processing on each of the PET bottle flake samples used to evaluate the candidate network structure units. Each PET bottle flake sample has a corresponding bottle flake cross-nozzle order. After the candidate network structure unit is temporarily silenced, the valve timing control neural network 320 outputs the valve action output for that PET bottle flake sample. Subsequently, the pruning admission judgment module 340 uses the bottle flake cross-nozzle order of the PET bottle flake sample as a spatial constraint to judge the action nozzle in the valve action output.

[0031] Furthermore, it is determined whether the action nozzle meets the condition of maintaining the order of bottle flakes across the nozzles, as follows: The pruning admission judgment module 340 takes the nozzles marked with the spray valve action output as the action nozzles and judges whether all action nozzles are located within the bottle piece cross-nozzle sequence and whether the nozzle numbers are consecutive. If both are true, the condition of maintaining the bottle piece cross-nozzle sequence is satisfied. This judgment includes two parallel and must be true simultaneously: first, all action nozzles are located within the bottle piece cross-nozzle sequence; second, the nozzle numbers of all action nozzles are consecutive.

[0032] All actuated nozzles are located within the bottle-piece cross-nozzle sequence. This means that the nozzles whose actions are indicated by the spray valve action marker in the spray valve action output all have nozzle numbers belonging to the bottle-piece cross-nozzle sequence of the corresponding PET bottle-piece sample. Since the bottle-piece cross-nozzle sequence represents the range of continuous nozzles corresponding to the PET bottle-piece, this determination can limit the spray valve action output after structured pruning from exceeding the nozzle range corresponding to the lateral coverage area of ​​the PET bottle-piece.

[0033] Continuous nozzle numbering means that the nozzle numbers of all actuated nozzles in the valve's output sequence are in lateral order without any gaps. For pneumatic sorting of PET flakes, the lateral coverage area of ​​the PET flakes corresponds to the continuous nozzle range. If the actuated nozzles are scattered across discontinuous nozzle numbers, even if some actuated nozzles are located within the flake spanning nozzle sequence, it does not indicate that the valve's output maintains the correspondence between the lateral coverage area of ​​the PET flakes and the lateral nozzle sequence. Therefore, nozzle number continuity and range inclusion together constitute the conditions for maintaining the flake spanning nozzle sequence.

[0034] When maintaining the order of bottle flakes across the nozzle is used as an admission criterion for candidate network structural units, the pruning admission judgment module 340 makes a judgment based on the PET bottle flake sample set used to evaluate the candidate network structural units. If all PET bottle flake samples in the PET bottle flake sample set satisfy the condition of maintaining the order of bottle flakes across the nozzle, the candidate network structural unit is determined as a candidate network structural unit that can be deleted. This admission criterion directly links the structured pruning decision with the spatial order of the nozzles corresponding to the lateral coverage area of ​​the PET bottle flakes.

[0035] The condition of maintaining the order of bottle flakes across the nozzle acts as an admission condition for candidate network structural units. In other words, the channel importance of candidate network structural units can be used to determine the candidate deletion order, but whether a candidate network structural unit performs structural deletion must also be subject to the constraint of maintaining the order of bottle flakes across the nozzle. Therefore, the structural compression process of the valve timing control neural network 320 is not solely determined by channel importance, but also by maintaining the spatial order of the action nozzles relative to the lateral coverage of the PET bottle flakes.

[0036] Furthermore, structural deletion is performed on candidate network structural units that satisfy the condition of maintaining the order of bottle flakes across the nozzle, as follows: Once the pruning admission judgment module 340 determines that the candidate network structure unit meets the condition of maintaining the bottle piece cross-nozzle order, the control processor 300 performs structure deletion only on the candidate network structure units that meet the condition of maintaining the bottle piece cross-nozzle order, thus obtaining a structured pruning model. Here, structure deletion actually changes the network structure of the valve timing control neural network 320, so that the candidate network structure units that meet the condition of maintaining the bottle piece cross-nozzle order no longer participate in the subsequent valve timing output generation.

[0037] In the implementation where the candidate network structural unit is a convolutional output channel, structural deletion of the candidate network structural unit includes: deleting the convolutional kernel corresponding to the convolutional output channel, and, if the convolutional output channel has a subsequent layer connection, deleting the input convolutional kernel in the next layer that takes the convolutional output channel as input. Through this process, the deleted convolutional output channel and its subsequent connection relationships are removed synchronously, ensuring that the structured pruning model remains executable in terms of network structure.

[0038] The structured pruning model is the valve timing control neural network 320 obtained after performing the aforementioned constrained structure deletion. Since each candidate network structural unit that is actually deleted has already passed the condition of maintaining the order of the bottle flakes across the nozzle, the structured pruning model preserves the spatial order constraint between the lateral coverage of the PET bottle flakes and the action nozzle.

[0039] After obtaining the structured pruning model, it can be retrained based on PET bottle flake samples to obtain the valve action output of the retrained model. The retrained model that maintains the bottle flake sequence across the nozzles is retained. Retraining uses PET bottle flake samples and valve timing labels to update the retained model weights, enabling the structured pruning model to continue generating valve action markers, valve opening times, and valve closing times.

[0040] The retrained structured pruning model still accepts the constraint of maintaining the order of bottle flakes across nozzles. After candidate network structural units are removed, the structured pruning model's output of valve actions for PET bottle flake samples should continue to ensure that all action nozzles are within the order of bottle flakes across nozzles and that the nozzle numbers of the action nozzles are consecutive.

[0041] Furthermore, a structured pruning model is used to generate the timing output of the spray valves and control the opening and closing of the corresponding spray valves, as detailed below: During the online sorting stage, the control processor 300 uses a structured pruning model to process the PET bottle flakes to be sorted and generates a spray valve timing output. It then controls the opening and closing of the corresponding spray valves according to the spray valve timing output to complete the pneumatic sorting of the PET bottle flakes. When the PET bottle flakes to be sorted pass through the acquisition area of ​​the image acquisition unit 200, the image acquisition unit 200 acquires an image of the PET bottle flake and inputs it into the control processor 300. The structured pruning model generates a spray valve timing output based on the PET bottle flake image.

[0042] The valve timing output includes the valve action flag, valve opening time, and valve closing time for each valve. The valve action flag uses a binary flag; a flag of 1 indicates that the corresponding valve performs an action, while a flag of 0 indicates that the corresponding valve does not perform an action. When the valve action flag is 1, the valve opening time and valve closing time determine the opening and closing sequence of the corresponding valve. The control processor 300 can send the valve timing output to the valve driver 400, which generates valve driving signals based on the valve timing output.

[0043] After receiving the spray valve drive signal, the multi-channel spray valve array 500 causes the corresponding spray valve to open when the spray valve opens and close when the spray valve closes. Since the formation process of the structured pruning model has taken the order of bottle flakes across the spray nozzles as the admission constraint, the action nozzles in the spray valve timing output can be consistent with the continuous nozzle range corresponding to the lateral coverage of the PET bottle flakes, so that the pneumatic sorting of PET bottle flakes can still be completed after structured pruning.

[0044] In this embodiment, the spray valve timing output is the final output used to control the opening and closing of the corresponding spray valve. The spray valve action output is used to determine the order of bottle flakes across the spray nozzle during the structured pruning stage, and the spray valve timing output is used for spray valve control during the online sorting stage.

[0045] In a complete numerical embodiment, the multi-channel nozzle array 500 includes 12 nozzles arranged in a laterally ordered sequence, numbered 1 to 12. Each nozzle has a defined correspondence with a nozzle valve, and the lateral positions of the nozzles are as follows: , , , , , , , , , , , The PET bottle flakes 110 move along the PET bottle flake conveyor belt 100 to the acquisition area of ​​the image acquisition device 200, where the image acquisition device 200 acquires the image of the PET bottle flakes and inputs it into the control processor 300.

[0046] In the At the start of each control cycle, the control processor 300 determines the left-hand position of the PET bottle flake from the PET bottle flake image. The right side of the bottle is The bottle flake cross-nozzle sequence generation module 310 determines the left and right nozzles based on the left and right positions of the bottle flake and the lateral position of the nozzle. Following the method for establishing the bottle flake cross-nozzle sequence, the left nozzle is designated as nozzle number 3, and the right nozzle as nozzle number 7. The bottle flake cross-nozzle sequence is as follows: The sequence of nozzles across the PET bottle indicates the range of consecutive nozzles corresponding to that PET bottle.

[0047] The control processor 300 sets the PET bottle sample and the bottle cross-nozzle sequence. The data is linked together, and the image and valve action record of the corresponding PET bottle flake are used as the data content of the PET bottle flake sample. The valve action record includes the valve action mark, valve opening time, and valve closing time of each valve. For example, in the valve timing label of this PET bottle flake sample, the valve action mark of the valve corresponding to nozzle numbers 3, 4, 5, 6, and 7 indicates that it is in action, while the valve action mark of the valve corresponding to other nozzle numbers indicates that it is not in action; the valve opening times of the valves corresponding to nozzle numbers 3, 4, 5, 6, and 7 are respectively... , , , , The valve closing times are respectively , , , , .

[0048] The valve timing control neural network 320 takes the PET bottle sample as input to generate valve action markers, valve opening times, and valve closing times for each valve. Before structured pruning, the candidate network structural unit evaluation module 330 uses the convolutional output channels as candidate network structural units and reads the channel weights of the candidate network structural units. For example, the convolutional output channel number... The channel weight is Channel scores are obtained based on the L1 norm of the channel weights. Convolution output channel number The channel weight is Channel scores are obtained based on the L1 norm of the channel weights. When the candidate network structure unit evaluation module 330 determines the candidate deletion order, the convolution output channel number is... Prior to convolution output channel numbering Undergo evaluation.

[0049] The candidate network structure unit evaluation module 330 first numbers the convolution output channels. The corresponding candidate network structure unit is temporarily silenced. Temporary silence suppresses the output of the candidate network structure unit while maintaining network connections. Under this temporary silence, the valve timing control neural network 320 obtains the valve action output for the aforementioned PET bottle sample. The valve action output is marked with valve action markers indicating that the nozzles for the action are nozzle numbers 3, 4, 5, and 6. The pruning admission judgment module 340 uses nozzle numbers 3, 4, 5, and 6 as the action nozzles and correlates them with the bottle sample's cross-nozzle sequence. A comparison was made. All the nozzles were located within the nozzle sequence of the bottle flakes, and the nozzle numbers 3, 4, 5, and 6 were consecutive. Therefore, it was determined that the PET bottle flake sample met the condition of maintaining the nozzle sequence of the bottle flakes.

[0050] Subsequently, the candidate network structure unit evaluation module 330 numbers the convolution output channels. The corresponding candidate network structure unit is temporarily silenced. Under this temporary silence, the valve action output indicates that the valve action is performed by nozzle numbers 3, 5, and 7. Although nozzle numbers 3, 5, and 7 are all located in the bottle flake cross-nozzle sequence. However, the nozzle numbers of the action nozzles are not consecutive, so the pruning admission judgment module 340 determines that the PET bottle sample does not meet the condition of maintaining the order of bottle pieces across nozzles. This candidate network structure unit is not determined as a candidate network structure unit that can be deleted.

[0051] For convolution output channel numbering For the corresponding candidate network structure unit, if all PET bottle samples in the PET bottle sample set satisfy the condition of maintaining the order of bottle sheets across the nozzle, the control processor 300 determines the candidate network structure unit as a candidate network structure unit that can be deleted, and performs structure deletion on the candidate network structure unit. During structure deletion, the convolution kernel corresponding to the convolution output channel is deleted, and if there is a next-layer connection for the convolution output channel, the input convolution kernel in the next layer that uses the convolution output channel as input is deleted. After the structure deletion is completed, the structured pruning model is obtained.

[0052] After obtaining the structured pruning model, the control processor 300 retrains the model based on PET bottle flake samples. The retrained structured pruning model obtains valve action outputs for the same PET bottle flake sample. In the valve action output, the valve action markers indicate that the nozzles involved in the action are nozzle numbers 3, 4, 5, 6, and 7. All action nozzles are located in the bottle flake cross-nozzle sequence. The nozzles are numbered 3, 4, 5, 6, and 7 consecutively, so the retrained structured pruning model is preserved.

[0053] During online sorting, another PET bottle piece enters the image acquisition area. The control processor 300 processes this PET bottle piece using a structured pruning model and generates a spray valve timing output. In the spray valve timing output, the spray valve action markers corresponding to nozzle numbers 4, 5, and 6 indicate the action, and the spray valve opening time corresponding to nozzle number 4 is... The valve closing time is The nozzle number 5 corresponds to the nozzle valve opening time is The valve closing time is The nozzle number 6 corresponds to the nozzle valve opening time is The valve closing time is The control processor 300 outputs control signals to open and close the corresponding spray valves according to the spray valve timing sequence. The spray valve driver 400 generates spray valve drive signals. The multi-channel spray valve array 500 causes the corresponding spray valve to open when the spray valve is open and close when the spray valve is closed, thus completing the pneumatic sorting of the PET bottle flakes to be sorted.

[0054] Although embodiments of this application have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of this application, and all such modifications and variations fall within the scope defined by the appended claims.

Claims

1. A method for timing control of spray valves in multi-channel PET flake sorting, characterized in that, include: Based on the correspondence between the lateral coverage of PET bottle flakes and multiple nozzles, a bottle flake cross-nozzle order is established to characterize the continuous nozzle range of the PET bottle flakes. The multiple nozzles are arranged in lateral order, have nozzle numbers that characterize the lateral order, and have a definite correspondence with each spray valve. A valve timing control neural network is established that can be structured and pruned to generate valve timing outputs. The valve timing outputs include valve action flags, valve opening times, and valve closing times for each valve. For a PET bottle sample with the bottle sheet cross-nozzle sequence and a candidate network structure unit of the spray valve timing control neural network, obtain the spray valve action output after removing the candidate network structure unit. The nozzles whose actions are indicated by the spray valve action mark in the spray valve action output are taken as the action nozzles. It is determined whether all action nozzles are located within the bottle piece cross-spray nozzle sequence and whether the nozzle numbers are consecutive. If both are true, it is determined that the condition of maintaining the bottle piece cross-spray nozzle sequence is met. Structural deletion is performed only on candidate network structural units that satisfy the condition of maintaining the order of bottle flakes across the nozzle, resulting in a structured pruning model; The structured pruning model is used to process the PET bottle flakes to be sorted and generate the timing output of the spray valves. The corresponding spray valves are opened and closed according to the timing output of the spray valves to complete the pneumatic sorting of PET bottle flakes.

2. The spray valve timing control method for multi-channel PET flake sorting according to claim 1, characterized in that, Establishing the bottle flake cross-nozzle sequence includes: determining the left and right positions of the PET bottle flake; determining the left and right end nozzles based on the left and right positions of the bottle flake, the lateral position of each nozzle; and using the nozzle numbers arranged consecutively from the left to the right end nozzle as the bottle flake cross-nozzle sequence.

3. The spray valve timing control method for multi-channel PET flake sorting according to claim 1, characterized in that, The PET bottle sample used to obtain the output of the spray valve action is formed by associating the order of the bottle across the nozzle with the image of the corresponding PET bottle and the spray valve action record. The spray valve action record includes the spray valve action mark of each spray valve, the spray valve opening time, and the spray valve closing time.

4. The spray valve timing control method for multi-channel PET flake sorting according to claim 1, characterized in that, The establishment of the timing control neural network for the spray valves includes: extracting features from PET bottle flake samples, and generating spray valve action markers, spray valve opening times, and spray valve closing times for each spray valve based on the feature extraction results.

5. The spray valve timing control method for multi-channel PET flake sorting according to claim 1, characterized in that, The convolutional output channels in the valve timing control neural network are used as candidate network structure units. The candidate deletion order is determined according to the channel importance corresponding to the channel weight of the candidate network structure unit. The valve action output under the influence of the removal of each candidate network structure unit is obtained in sequence according to the candidate deletion order.

6. The spray valve timing control method for multi-channel PET flake sorting according to claim 5, characterized in that, Obtaining the valve action output under the influence of the removal of the candidate network structure unit includes: performing temporary silencing on the candidate network structure unit, wherein the temporary silencing is to suppress the output of the candidate network structure unit and keep the network connection from being deleted, and obtaining the corresponding valve action output under the temporary silencing.

7. The spray valve timing control method for multi-channel PET bottle flake sorting according to claim 1, characterized in that, Using the condition of maintaining the order of bottle flakes across the nozzle as the admission condition for the candidate network structure unit includes: judging based on the PET bottle flake sample set used to evaluate the candidate network structure unit, and determining the candidate network structure unit as a candidate network structure unit that can be deleted when all PET bottle flake samples in the PET bottle flake sample set meet the condition of maintaining the order of bottle flakes across the nozzle.

8. The spray valve timing control method for multi-channel PET flake sorting according to claim 5, characterized in that, Performing structural deletion on the candidate network structural unit includes: deleting the convolution kernel corresponding to the convolution output channel, and, if the convolution output channel has a next-layer connection, deleting the input convolution kernel in the next layer that takes the convolution output channel as input.

9. The spray valve timing control method for multi-channel PET bottle flake sorting according to claim 3, characterized in that, After obtaining the structured pruning model, the structured pruning model is retrained based on the PET bottle sample to obtain the valve action output of the retrained structured pruning model, and the retrained structured pruning model that satisfies the condition of maintaining the order of bottle pieces across the nozzle is retained.

Images