Automatic feather sorting method based on form recognition and reinforcement learning reward function

Abstract

The invention discloses an automatic feather sorting method based on form recognition and a reinforcement learning reward function, and the method comprises the steps: collecting the related data of feathers through multi-source sensing fusion, constructing a dynamic algorithm system through the combination of feather type adaptation, environment pre-judgment and microstructure detection, and achieving the precise sorting of overlapped feathers through the linkage of the reinforcement learning reward function and sorting execution parameters. According to the scheme, the problem that in the prior art, the sorting success rate, the form adaptability and the environment adaptability cannot be considered at the same time is solved, and the method has the advantages of being accurate in feather type recognition, rapid in environment sudden change response and low in sorting damage and is suitable for efficient sorting scenes of badminton manufacturing production lines.

Description

Technical Field

[0001] This invention relates to the field of automatic sorting technology in badminton manufacturing, specifically to an automatic feather sorting method based on morphological recognition and reinforcement learning reward function. Background Technology

[0002] Current feather sorting technologies, whether traditional threshold segmentation or template matching schemes, or conventional deep learning classification schemes, do not fully consider the coupled effects of feather microscopic functional structure, feather shape differences, and environmental parameters. In actual production lines, the damage to the interlocking structure between feather shafts directly affects the impact resistance of subsequent shuttlecocks, while the differences in shaft stiffness among different feather shapes lead to persistently high breakage rates even with uniform sorting parameters. Simultaneously, changes in humidity, vibration interference, and light fluctuations in the workshop environment further exacerbate the identification errors of overlapping feathers, resulting in problems such as over-sampling, under-sampling, or irreversible feather deformation during sorting. These factors combine to make it difficult for existing technologies to simultaneously guarantee sorting speed while meeting the stringent requirements of production lines for sorting success rate, feather integrity, and product consistency.

[0003] Based on the above problems, there is an urgent need for an automated feather sorting technology that can coordinate microstructure detection, feather shape adaptation, and environmental adaptation. Summary of the Invention

[0004] This invention provides an automatic feather sorting method based on morphological recognition and reinforcement learning reward function, comprising: acquiring image data of feathers on a conveyor belt and performing preprocessing; characterized in that it further comprises: Multi-source sensors synchronously collect relevant environmental parameters and material property parameters of feathers; A coupling relationship model is constructed based on feather morphological characteristics, material properties, and environmental parameters; Dynamically adjust the weights of the reinforcement learning reward function; The sorting target is determined by combining the output of the coupling relationship model and the reward function. Dynamically adjust sorting execution parameters according to sorting objectives; Control the sorting mechanism to perform sorting actions according to the adjusted parameters; Real-time feedback of sorting results and iterative optimization of model parameters.

[0005] Preferably, the multi-source sensors include an industrial camera, a near-infrared spectral sensor, a humidity sensor, a triaxial accelerometer, and a feather density detection lens. The parameters acquired simultaneously include feather images, spectral data, ambient humidity, vibration data, feather density, and moment of inertia of the feather shaft cross section.

[0006] Further preferred, the preprocessing includes distortion correction, adaptive dynamic threshold segmentation and binary mask generation. The dynamic threshold is adjusted in real time based on the rate of change of environmental parameters to ensure that the feather edge features are completely preserved under different environments.

[0007] Further preferred features include morphological characteristics such as the coordinates of key points on the bristle, the distribution pattern of the barbs, and the wing symmetry; and material properties such as bristle stiffness, elastic modulus, and barb sliplock integrity, which are obtained through image texture analysis and spectral feature back deduction.

[0008] A further preferred coupling model is constructed using the feather-lock integrity and bending coupling coefficient, with the specific formula as follows: ; Where S is the barb sliplock integrity, ranging from 0 to 1, obtained by extracting the cascade structure between barbs through image texture features; ω is the coupling coefficient, used to balance the interaction between humidity and stiffness; H is the ambient humidity, in percentage; K is the barb stiffness, in Newtons per millimeter; A is the ambient vibration amplitude, in gravitational acceleration; μ is the illumination influence coefficient, used to correct the interference of illumination deviation on the coupling relationship; L is the real-time illumination intensity, in lux; L0 is the reference illumination intensity, in lux.

[0009] A further optimized dynamic reward weight correction function is: ; ; in The original tier reward weight, The original bar reward weight, ξ represents the original adaptability reward weight; C represents the wingside recognition confidence level, ranging from 0 to 1, determined by the fusion of spectral features and morphological symmetry; ξ represents the barb sliplock integrity and bending coupling coefficient. The effective signal-to-noise ratio is expressed in decibels; ζ is the vibration attenuation coefficient, used to quantify the weakening effect of vibration on the signal-to-noise ratio; θ is the confidence gain coefficient; erf is the error function; ε is the minimum constant, used to avoid the denominator being zero, and takes the value of 10 to the power of negative six.

[0010] Further optimized, the sorting execution parameters include gripping force and gripping acceleration, and the linkage formula is: ; ;

[0011] in The clamping force is measured in Newtons. is the reference force coefficient, with the dimension of Newtons per millimeter; K is the stiffness of the burr bar, with the dimension of Newtons per millimeter; ξ is the bending coefficient of the sprue, derived from ξ; The clamping acceleration is measured in meters per second squared. The reference clamping acceleration is measured in meters per second squared; φ is the feather shape adaptation coefficient. The value is the predicted elastic deformation of the bar, in millimeters. This is the total output value of the dynamic reward function.

[0012] Further preferred, the sorting target determination includes top-layer feather identification, isolation screening, and wing attribute confirmation. Top-layer identification is based on the output threshold of the reward function, isolation screening is achieved through dynamic radius circular domain detection, and wing attributes are confirmed by fusing near-infrared spectral features and morphological symmetry.

[0013] Further preferably, the sorting mechanism is a SCARA robot or a Delta robot, the robot drive module supports continuously adjustable gripping acceleration, the suction cup has a rotation function, and can perform negative pressure suction action after adjusting the posture according to the direction angle of the hair rod.

[0014] Further optimization involves iteratively updating the coefficient parameters in the coupling relationship model and reward function by statistically analyzing the sorting success rate, breakage rate, and wing side misjudgment rate, thus adapting to changes in feather characteristics and production line environment for different batches.

[0015] The technical effects achieved by the above embodiments include: The core inventive technology of this invention lies in constructing a multi-parameter coupling relationship model, a dynamic reward weight correction mechanism, and a sorting execution parameter linkage system. These three elements work synergistically to solve the core problems in the background technology, such as the lack of detection of feather microstructure, insufficient feather shape adaptation, and lagging response to environmental interference. The solution achieves accurate identification of overlapping feathers, ensuring feather integrity during sorting, improving sorting success rate and product consistency, adapting to different production line environments and feather batch differences, and providing efficient and stable sorting technology support for badminton shuttlecock manufacturing. Attached Figure Description

[0016] Figure 1 This is a flowchart of the automatic feather sorting method based on morphological recognition and reinforcement learning reward function proposed in this application. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0018] Traditional technical solutions have the following technical problems: existing feather sorting methods do not address the coupled effects of feather microstructure, feather shape characteristics and environmental parameters, resulting in frequent over-sampling and under-sampling during the sorting process, high feather breakage rate and poor product consistency.

[0019] Based on this, please refer to Figure 1 This embodiment provides an automatic feather sorting method based on morphological recognition and reinforcement learning reward function, including: acquiring image data of feathers on a conveyor belt and performing preprocessing; characterized in that it further includes: S1: Multi-source sensors simultaneously collect relevant environmental parameters and material properties of feathers; S2: Construct a coupling relationship model based on feather morphological characteristics, material properties, and environmental parameters; S3: Dynamically adjust the weights of the reinforcement learning reward function; S4: Determine the sorting target by combining the output results of the coupling relationship model and the reward function; S5: Dynamically adjust sorting execution parameters according to sorting targets; S6: Control the sorting mechanism to perform sorting actions according to the adjusted parameters; S7: Real-time feedback of sorting results and iterative optimization of model parameters.

[0020] It is worth mentioning that this solution uses multi-source sensor fusion as its data foundation, simultaneously acquiring key parameters such as feather images, spectral data, environmental humidity, and vibration data through various devices including industrial cameras and near-infrared spectral sensors. This enables comprehensive perception of the macroscopic morphology, microstructure, and environmental conditions of the feathers. Based on this multi-dimensional data, a model of the integrity and bending coupling coefficient of the feather barb lock is constructed to quantify the interactive influence of microstructure, material properties, and environmental parameters, providing data support for subsequent decision-making. The weights of the reinforcement learning reward function are dynamically adjusted, incorporating key indicators such as wing-side recognition confidence and effective signal-to-noise ratio into the weight calculation, making the reward function more accurately reflect the core needs of feather sorting. Combining the results of the coupling model and the reward function, sorting targets are determined from three dimensions: hierarchical recognition, isolation screening, and wing-side attributes, ensuring accurate positioning of target feathers. Execution parameters such as gripping force and acceleration are dynamically adjusted according to target characteristics to avoid feather damage during sorting. The sorting mechanism executes the adjusted actions to complete the sorting, and sorting-related indicators are statistically analyzed in real time, iteratively optimizing model parameters and continuously improving sorting performance. The entire technical process forms a closed-loop system from data acquisition and model building to decision-making and execution feedback, with each link working together to achieve efficient and accurate sorting.

[0021] The technical effects achieved by the above embodiments include: enabling collaborative detection of feather microstructure and macromorphology, improving the accuracy of sorting target identification, reducing the probability of over-sampling and under-sampling, reducing feather breakage, ensuring product consistency, and adapting to different environments and feather batches.

[0022] Traditional technical solutions have the following technical problems: existing sorting technologies use only single sensing devices that can collect feather image data but cannot obtain material characteristics and environmental parameters, resulting in a lack of data support for judging the integrity of feathers and their environmental adaptability.

[0023] Based on this, the multi-source sensors include an industrial camera, a near-infrared spectral sensor, a humidity sensor, a triaxial accelerometer, and a feather density detection lens. The parameters collected simultaneously include feather images, spectral data, ambient humidity, vibration data, feather density, and the moment of inertia of the feather shaft cross section.

[0024] It is worth mentioning that the multi-source sensing system integrates five core detection devices: an industrial camera to acquire 2448×2048 resolution feather images, providing a foundation for morphological feature extraction; a near-infrared spectral sensor with a sampling rate of 100fps, analyzing the spectral response characteristics of feathers to infer material parameters such as feather shaft stiffness and elastic modulus; a humidity sensor to monitor the workshop's environmental humidity in real time, ensuring accuracy in capturing the impact of subtle humidity changes on feather properties; a triaxial accelerometer with a range of ±2g to collect vibration data during conveyor belt operation, reflecting environmental interference; and a feather barb density detection lens specifically designed to capture feather barb distribution patterns and calculate feather barb density parameters. These devices achieve coordinated data acquisition through a synchronous triggering mechanism, ensuring consistency of different types of parameters over time. The acquired multi-dimensional data provides comprehensive and accurate data input for subsequent coupled model construction, reward function adjustment, and sorting decisions, solving the problem of insufficient data dimensionality from a single sensing device.

[0025] The technical effects achieved by the above embodiments include: realizing comprehensive collection of feather morphology, material and environmental parameters, providing sufficient data support for subsequent stages, improving the scientific nature and accuracy of sorting decisions, and enhancing the adaptability of the solution to environmental changes and differences in feather characteristics.

[0026] Traditional technical solutions have the following technical problems: the existing preprocessing process uses fixed threshold segmentation, which cannot cope with the fluctuations in feather image quality caused by changes in environmental parameters, resulting in the loss of feather edge features or increased noise interference, affecting the accuracy of subsequent recognition.

[0027] Based on this, the preprocessing includes distortion correction, adaptive dynamic thresholding and binary mask generation. The dynamic threshold is adjusted in real time based on the rate of change of environmental parameters to ensure that the feather edge features are completely preserved under different environments.

[0028] It's worth noting that the preprocessing workflow comprises three core steps in sequence. First, distortion correction is performed, correcting image distortion caused by lens optical characteristics using preset camera calibration parameters to restore the true shape of the feathers. Next, adaptive dynamic thresholding segmentation is executed. This step uses the rate of change of environmental parameters as the core adjustment criterion. By calculating the rate of change of environmental humidity, vibration data, and light intensity over the past five frames, a mapping relationship is established between the rate of change and the threshold adjustment. When environmental parameters fluctuate significantly, the threshold is dynamically adjusted accordingly, ensuring effective separation of feathers from the background even in complex environments such as high humidity, strong vibration, or sudden changes in light intensity. Finally, a binary mask is generated based on the segmentation results to highlight the feather area and suppress background noise. The entire preprocessing workflow requires no manual intervention, adapts to different environmental conditions, and consistently maintains the integrity and accuracy of feather edge features, laying a solid foundation for subsequent morphological feature extraction and recognition steps.

[0029] The technical effects achieved by the above embodiments include: realizing environmental adaptation in the preprocessing process, avoiding feature loss or noise interference caused by fixed thresholds, ensuring stable feather image quality, and improving the accuracy and reliability of subsequent recognition steps.

[0030] Traditional technical solutions have the following technical problems: existing technologies only focus on the macroscopic morphological characteristics of feathers and ignore the detection of material properties, which makes it impossible to judge the integrity of the feather structure and its use value. After sorting, some feathers are affected by microstructural damage or material incompatibility, affecting product quality.

[0031] Based on this, the morphological features include the coordinates of key points on the bristle, the distribution pattern of the barbs, and the wing symmetry. The material properties include bristle stiffness, elastic modulus, and barb slip lock integrity, which are obtained through image texture analysis and spectral feature inference.

[0032] It is worth mentioning that morphological feature extraction and material property detection are performed simultaneously. Regarding morphological features, edge detection and contour extraction are performed on preprocessed feather images to locate the coordinates of four key nodes of the barb, clarifying its length and direction. Image segmentation algorithms are used to separate individual barbs, statistically analyzing their number and distribution density to obtain their distribution patterns. A symmetry calculation model is used to analyze the symmetry of barb distribution on both sides of the feather, determining wing-side property tendencies. Regarding material property parameters, near-infrared spectral sensors are used to acquire feather spectral data. A model relating spectral features to material parameters is used to inversely deduce barb stiffness and elastic modulus. Image texture analysis algorithms are employed to magnify image details at barb connections, detecting the integrity of the interlocking structure between barbs and determining whether there is breakage, deformation, or other damage. The detection results of morphological features and material property parameters complement each other, comprehensively reflecting the structural state and usability of the feather, providing multi-dimensional basis for sorting decisions.

[0033] The technical effects achieved by the above embodiments include: enabling a comprehensive evaluation of feather morphology and material characteristics, screening out feathers with complete structure and qualified materials, improving product quality, and reducing subsequent rework caused by inconsistent feather characteristics.

[0034] Traditional technical solutions have the following technical problems: existing technologies have not established a correlation model between feather microstructure, material properties and environmental parameters, which makes it impossible to quantify the impact of multi-factor coupling on the sorting process, and sorting decisions lack a scientific basis.

[0035] Based on this, the coupling relationship model is constructed using the feather-lock integrity and bending coupling coefficient, with the specific formula as follows: ; Where S is the barb sliplock integrity, ranging from 0 to 1, obtained by extracting the cascade structure between barbs through image texture features; ω is the coupling coefficient, used to balance the interaction between humidity and stiffness; H is the ambient humidity, in percentage; K is the barb stiffness, in Newtons per millimeter; A is the ambient vibration amplitude, in gravitational acceleration; μ is the illumination influence coefficient, used to correct the interference of illumination deviation on the coupling relationship; L is the real-time illumination intensity, in lux; L0 is the reference illumination intensity, in lux.

[0036] It is worth noting that the core of the coupling relationship model is the coupling coefficient ξ between the integrity of the feather clip lock and its bending. This coefficient comprehensively quantifies the coupling effects of multiple factors, including the integrity of the feather clip lock (S), ambient humidity (H), feather stiffness (K), ambient vibration amplitude (A), and real-time illumination intensity (L). The integrity of the feather clip lock (S) is obtained through image texture analysis. When the cascaded structure between feathers is intact and undamaged, S approaches 1; when damage exists, the value of S decreases. Ambient humidity (H) directly affects the flexibility of the feathers; higher humidity makes the feathers more easily bent. This value is collected in real-time by a humidity sensor. Feather stiffness (K) reflects the feather's ability to resist deformation and is derived from spectral characteristics, with dimensions in Newtons per millimeter. Ambient vibration amplitude (A) reflects the stability of the conveyor belt operation; greater vibration causes more severe interference to the sorting process, with dimensions in the form of gravitational acceleration. The deviation between the real-time illumination intensity (L) and the reference illumination intensity (L0) affects the image recognition accuracy, thus indirectly affecting the determination of the coupling relationship. In the formula, the tanh function smooths the coupling effect of humidity, stiffness, and vibration, preventing large fluctuations in coefficients caused by abrupt changes in a single parameter; the exponential function corrects for the influence of illumination deviations, ensuring coefficient stability. A larger ξ value indicates a more complete feather microstructure, superior material properties, and less environmental interference, resulting in higher sorting priority. This model integrates multi-dimensional, dispersed parameters into a single coupling coefficient, providing a concise and effective quantitative basis for subsequent reward function adjustments and sorting decisions.

[0037] The core design purpose of this formula is to quantify the coupling effect of feather sliplock integrity, feather material properties, and environmental parameters, overcoming the technical bottleneck of existing technologies that cannot comprehensively assess the impact of multiple factors on sorting. This provides a unified quantitative basis for subsequent reward function adjustments and sorting execution parameter optimization. The formula adopts a multi-factor product structure, with each factor corresponding to a type of key influencing factor. Furthermore, the selection of specific mathematical functions achieves a non-linear, smooth fusion of the influences of each factor, ensuring a high coupling coefficient. Its output is stable and has clear technical significance.

[0038] The first factor S in the formula serves as a fundamental core parameter, directly reflecting the integrity of the microscopic functional structure of the barb lock. Its value is limited to between 0 and 1, and is a normalized result calculated after extracting the cascade structure features between barbs using image texture analysis technology. As a crucial connecting part of the feather composition, the integrity of the lock structure directly determines the impact resistance of the subsequently manufactured shuttlecock. Therefore, using S as the fundamental factor in the formula reflects the core influence of the microstructure on sorting priority.

[0039] Part Two Factors This is the core element in achieving the coupling between environmental parameters and material properties. The ratio of H to K is designed based on the inverse relationship between humidity and feather stiffness on feather bending characteristics: higher humidity increases the flexibility of feather fibers after absorbing water, making them more prone to bending; conversely, greater feather stiffness increases the feather's resistance to bending. The ratio accurately reflects the strength of this interaction. The 0.3 power design of A is based on experimental data. Extensive testing revealed that the effect of vibration on feather bending is not linear; the impact is smaller at low amplitude vibrations and slows down at high amplitude vibrations. The 0.3 power nonlinear transformation accurately fits this characteristic, avoiding excessive amplification of interference from low-amplitude vibrations. The selection of a function mainly utilizes its saturation property, when If the calculation result is too large or too small, The function output tends to stabilize, thus avoiding the coupling coefficient being affected by a sudden change in a single parameter. In the event of significant fluctuations, ensure the robustness of the assessment results; As a coupling coefficient, its value is obtained through multiple sets of orthogonal experiments. It is used to balance the influence weights among humidity, stiffness and vibration, so that the output range of this factor is stable between 0 and 1, consistent with the value range of S, which facilitates subsequent multi-factor product calculations.

[0040] Part Three Factors Used to correct for interference from light intensity deviations in coupling relationship assessment. This represents the relative deviation between real-time illumination intensity and reference illumination intensity. This design eliminates the influence of absolute illumination values ​​and focuses on the interference caused by relative changes. The exponential function is chosen to utilize its monotonically decreasing characteristic; when the relative illumination deviation is 0, the exponential function output is 1, affecting the coupling coefficient. No correction effect; as the relative deviation increases, the output of the exponential function gradually decreases, affecting... The gradual attenuation correction conforms to the actual law that the greater the illumination deviation, the lower the image recognition accuracy and the worse the reliability of coupling relationship evaluation. As the illumination influence coefficient, its value is determined through sorting experiments under different illumination environments to ensure that the correction range matches the actual impact of illumination deviation on recognition accuracy.

[0041] The output of the entire formula The value range is stable between 0 and 1. The closer the value is to 1, the more complete the feather barb lock structure, the better the adaptability of the feather shaft stiffness, and the less environmental interference, resulting in a higher sorting priority. Conversely, a value closer to 1 indicates that the feather has structural damage, incompatible materials, or serious environmental interference, resulting in a lower sorting priority or direct classification as unqualified.

[0042] The technical effects achieved by the above embodiments include: quantifying the impact of multi-factor coupling on the sorting process, providing a scientific quantitative basis for sorting decisions, improving the accuracy and rationality of decisions, and avoiding the limitations of single-parameter judgment.

[0043] Traditional technical solutions have the following technical problems: existing reinforcement learning reward functions use fixed weights, which cannot dynamically adjust the reward focus according to feather characteristics and environmental conditions, resulting in a disconnect between reward output and actual sorting needs, affecting recognition accuracy.

[0044] Based on this, the dynamic reward weight adjustment function is: ; ; in The original tier reward weight, The original bar reward weight, ξ represents the original adaptability reward weight; C represents the wingside recognition confidence level, ranging from 0 to 1, determined by the fusion of spectral features and morphological symmetry; ξ represents the barb sliplock integrity and bending coupling coefficient. The effective signal-to-noise ratio is expressed in decibels; ζ is the vibration attenuation coefficient, used to quantify the weakening effect of vibration on the signal-to-noise ratio; θ is the confidence gain coefficient; erf is the error function; ε is the minimum constant, used to avoid the denominator being zero, and takes the value of 10 to the power of negative six.

[0045] It is worth mentioning that the dynamic reward weight correction function includes two correlation formulas, the core of which is based on the wingside identification confidence C, the feather sliplock integrity and bending coupling coefficient ξ, and the effective signal-to-noise ratio. Three key parameters dynamically adjust the original reward weights. The wingside recognition confidence score C is determined by fusing spectral features and morphological symmetry; a C value closer to 1 indicates more reliable wingside attribute recognition. Effective signal-to-noise ratio... The second formula, calculated to comprehensively reflect the impact of light intensity L and vibration amplitude A on image quality, indicates that higher light intensity and lower vibration amplitude are associated with better image quality. The larger the value, the better the image quality; ζ is the vibration attenuation coefficient, used to quantify the weakening effect of vibration on the signal-to-noise ratio; the error function erf is used to enhance the gain effect of the wing-side recognition confidence C and coupling coefficient ξ on the effective signal-to-noise ratio. In the first formula, the original hierarchical reward weight... With (1+0.25C) correction, the more reliable the wing-side recognition, the higher the weight of the hierarchical reward; the original hair shaft reward weight... By correcting the coupling coefficient ξ, the better the feather's microstructure and material properties, the higher the weight of the feather shaft reward; the original fit reward weight... pass The ratio of ξ is corrected; the better the image quality and the higher the coupling coefficient, the higher the weight of the fit reward. ε is a local constant used to avoid the denominator being zero and to ensure the stability of the formula calculation. The corrected total weight. It can dynamically adapt to feather characteristics and environmental conditions, making the reward function output more in line with actual sorting needs.

[0046] The first formula is the core optimization module of the reinforcement learning reward function. It aims to address the problem of reward output being disconnected from actual sorting needs due to the use of fixed reward weights in existing technologies. By incorporating the quantified results of feather characteristics and environmental parameters into the weight adjustment, the reward function can accurately focus on key evaluation indicators in the sorting process, improving the recognition accuracy and decision-making rationality of the reinforcement learning model. The formula adopts a weighted summation structure, specifically modifying the original reward weights. The design of each modification term is directly related to the core requirements of sorting.

[0047] First correction The optimization of hierarchical reward weights involves correlating the reliability of wing attribute recognition with the reward weight of the highest-level feather recognition. Hierarchical recognition is a fundamental step in sorting, and the confidence level C of wing attribute recognition reflects the overall reliability of morphological feature extraction. The closer C is to 1, the more accurate the feather morphological feature recognition, and the higher the reliability of the hierarchical recognition results. The linear correction can appropriately increase the proportion of hierarchical reward weights, making the reward function more focused on incentivizing the accuracy of hierarchical judgments. The coefficient design of 0.25 is based on the results of experimental data calibration, ensuring that the correction range of the wing confidence to hierarchical weights is within a reasonable range, reflecting the correlation without excessively amplifying the influence of a single factor.

[0048] Second revision The reward weight for the feather bar is optimized by combining the integrity of the feather bar slip lock with the bending coupling coefficient. It is directly integrated into the correction process. The core value of feather shaft identification lies in ensuring the structural integrity and usability of the feathers, while As a key indicator for comprehensively quantifying the microstructure, material properties, and environmental adaptability of feathers, its value directly reflects the quality level of the feather shaft and the overall feather. The closer the value is to 1, the more intact the feather structure and the better the material compatibility. Increasing the weight of the feather shaft reward at this point incentivizes the reinforcement learning model to prioritize identifying these high-quality feathers, aligning with the core requirements for feather quality during sorting. This direct product-based correction method establishes a strong correlation between the feather shaft reward weight and feather quality, ensuring that the reward output is consistent with the quality orientation of sorting.

[0049] The third revision The weighting of the suitability reward was optimized, and a collaborative correction mechanism for image quality and feather quality was constructed. The core of suitability evaluation is determining whether the feathers are suitable for the current sorting environment and subsequent processing requirements, which requires simultaneous consideration of the reliability of image recognition (as determined by...). (Quantification) and the quality level of the feathers themselves (by Quantification). The larger the value, the more reliable the image recognition and the more sufficient the basis for the suitability judgment; The larger the feather, the better its quality and the more solid its fit. The ratio of these two factors can comprehensively reflect the overall reliability of the fit assessment. (Introduced into the denominator) This can avoid An error occurred when the denominator was zero, and then... The existence of this constraint prevents an excessive increase in the fit weight simply because the image quality is good, even when the feather quality is poor. This collaborative correction method allows the fit reward weight to dynamically balance the dual impact of image quality and feather quality, ensuring that the reward function output better matches the actual fit requirements of sorting.

[0050] The output of the entire formula As a dynamically adjusted total reward weight, it can flexibly adjust the weight ratio of each sub-reward according to the specific characteristics of each feather and the real-time environmental conditions, making the evaluation criteria of the reinforcement learning reward function more targeted and reasonable.

[0051] The second formula is the foundational formula for dynamic reward weight correction. Its core function is to quantify the effective signal-to-noise ratio (SNR) of feather images, addressing the inaccurate image quality assessment caused by the use of fixed SNR thresholds in existing technologies. It provides a quantitative basis for the dynamic adjustment of reward weights in terms of image quality. By integrating two key environmental parameters—illuminance and vibration amplitude—the formula constructs an adaptive SNR assessment model, whose design logic closely aligns with the actual image acquisition scenarios in feather sorting.

[0052] The numerator of the formula This demonstrates the amplifying effect of light intensity on signal strength and the attenuating effect of vibration on signal strength. Light intensity L is the direct determinant of image signal strength; the larger L is, the higher the contrast between the feathers and the background, and the stronger the signal strength. This relationship is directly reflected through linear multiplication. This is used to quantify the nonlinear attenuation effect of vibration on signals. Based on experimental observations, it was found that the interference intensity of vibration on image signals increases with the square of the vibration amplitude. The exponential decay function can accurately fit this law: when the vibration amplitude A is 0, the output of the exponential function is 1, with no attenuation effect; as A increases, the output of the exponential function decays rapidly, and the signal strength is significantly weakened. As the vibration attenuation coefficient, its value is determined by image quality testing under different vibration amplitudes to ensure that the attenuation amplitude is consistent with the actual interference level, so that the numerator can accurately reflect the effective signal strength under the influence of environmental factors.

[0053] denominator of the formula The design is mainly to avoid the calculation error of zero denominator when the vibration amplitude A is 0. At the same time, by introducing a linear term of A, the overall impact of vibration on signal-to-noise ratio is further balanced. As a local constant, it takes the value of 10 to the power of negative six. Its value is far smaller than the vibration amplitude that might occur in real-world scenarios, and therefore will not have a substantial impact on the calculation results; it only serves to ensure computational stability. The linear term of A is retained in the denominator because vibration not only affects the signal-to-noise ratio by attenuating the signal strength, but also increases noise by causing image blurring. This dual effect is comprehensively quantified by combining the nonlinear attenuation of the numerator with the linear growth of the denominator. The calculation results are closer to the actual image quality.

[0054] The output of the formula The unit of measurement is decibels, and its value directly reflects the clarity and noise level of a feather image. The larger the value, the better the image quality, and the higher the reliability of subsequent morphology recognition and feature extraction.

[0055] The technical effects achieved by the above embodiments include: realizing dynamic adaptive adjustment of the reward function weights, making the reward output accurately match the feather characteristics and environmental conditions, and improving the recognition accuracy and decision rationality of the reinforcement learning model.

[0056] Traditional technical solutions have the following technical problems: the existing sorting execution parameters use fixed values ​​and are not dynamically adjusted according to the characteristics of the feathers and the recognition results, resulting in improper gripping force or acceleration, causing feather breakage or loss.

[0057] Based on this, the sorting execution parameters include gripping force and gripping acceleration, and the linkage formula is: ; ;

[0058] in The clamping force is measured in Newtons. is the reference force coefficient, with the dimension of Newtons per millimeter; K is the stiffness of the burr bar, with the dimension of Newtons per millimeter; ξ is the bending coefficient of the sprue, derived from ξ; The clamping acceleration is measured in meters per second squared. The reference clamping acceleration is measured in meters per second squared; φ is the feather shape adaptation coefficient. The value is the predicted elastic deformation of the bar, in millimeters. This is the total output value of the dynamic reward function.

[0059] It is worth mentioning that the sorting execution parameters include the gripping force. With clamping acceleration The two are dynamically calculated through a correlation formula, achieving deep linkage with feather characteristics and recognition results. (Grip force) In the calculation, K0 is the reference force coefficient, with the dimension of Newtons per millimeter, used to convert the product of stiffness and coupling coefficient into force units; K is the barb stiffness, with the dimension of Newtons per millimeter, reflecting the feather's ability to resist deformation; ξ is the barb lock integrity and bending coupling coefficient, comprehensively reflecting the feather structure and environmental adaptability. ξ is the bending coefficient of the feather shaft, which is derived from ξ. The larger the bending coefficient, the easier the feather is to bend. The effective signal-to-noise ratio reflects the reliability of image recognition; C is the confidence level of wing-side recognition. The formula reflects the force requirements of feather material and structure through the product of K and ξ, and corrects the influence of bending characteristics through (1+α). The influence of +0.3C) fusion recognition accuracy is taken into account to obtain the appropriate gripping force. Gripping acceleration In the calculation, a0 is the reference clamping acceleration, with the dimension of meters per second squared; φ is the feather type adaptation coefficient, which reflects the degree of adaptation between the feather type and the sorting mechanism. The value is the predicted elastic deformation of the feather shaft, with the dimension of millimeters. The larger the predicted deformation value, the more easily the feather is subjected to irreversible deformation. This is the total output value of the dynamic reward function, reflecting the overall quality of the identified target. The formula is derived through... and The ratio balances feather conformation and deformation risk, through The impact of target quality on acceleration is corrected to ensure that acceleration meets sorting efficiency requirements while avoiding feather damage. The two formulas are interrelated and work together to achieve precise dynamic adjustment of sorting execution parameters.

[0060] The first formula is the core control formula for the sorting execution process. It aims to solve the problems of feather breakage or unstable gripping caused by the fixed gripping force in existing technologies. By incorporating multiple parameters such as feather material properties, structural integrity, image quality, and recognition reliability into the force calculation, it achieves dynamic adaptive adjustment of the gripping force, minimizing the risk of feather damage while ensuring gripping stability. The formula design follows the principle of "quality-oriented + reliability-constrained," and each part of the structure is closely related to the actual needs of sorting execution.

[0061] Formula Part 1 It is the basic calculation module for clamping force, and its core is to determine the force benchmark value based on the characteristics of the feather itself. As a benchmark force coefficient, its core function is to achieve the conversion of physical dimensions, linking the stiffness K of the burr bar with the coupling coefficient. The product of K and K is converted into a force unit (Newton) that meets actual sorting needs. Its value is determined through extensive calibration experiments to ensure that the baseline force can meet the grasping requirements of regular feathers. The product of K and K directly reflects the quality characteristics of the feather: the larger K is, the higher the stiffness of the feather shaft, and the greater the clamping force it can withstand. The larger the value, the more complete the feather structure, the better the environmental adaptability, and the higher the safety margin during gripping. The product of these two factors accurately reflects the feather's ability to withstand gripping force. (Denominator) This is used to introduce constraints on the bending characteristics of the sprue. As by The derived feather shaft bending coefficient indicates that the larger the value, the more easily the feather is bent and deformed. By linearly increasing the denominator, the clamping force can be appropriately reduced to avoid feather bending damage due to excessive force, thus forming a reasonable logic that "the stronger the load-bearing capacity, the lower the bending risk, and the greater the force."

[0062] Formula Part 2 It is a reliability correction module for clamping force, used to adjust the force based on image recognition quality and feature recognition reliability. As the effective signal-to-noise ratio, a higher value indicates better image quality and higher recognition accuracy of features such as feather shaft key points and barb distribution. This also leads to higher positioning accuracy of the gripping position, and appropriately increasing the gripping force can enhance grasping stability. C, representing the wingside recognition confidence level, indicates higher overall reliability of morphological feature recognition and more sufficient basis for gripping decisions. A linear correction of 0.3C can further optimize the accuracy of force adjustment. The coefficient of 0.3 is designed based on experimental data calibration results, ensuring that the correction range of the wingside confidence level to the force is moderate, reflecting the impact of recognition reliability without causing the force to deviate from the feather's load-bearing capacity due to excessive correction.

[0063] The output of the formula The final clamping force can be dynamically adapted to the specific condition of each feather, achieving precise control by "applying more force to high-quality feathers to ensure stability, and reducing force to low-quality or easily bent feathers to avoid damage".

[0064] The second formula is another core control formula for sorting execution parameters. Working in conjunction with the gripping force formula, it aims to solve the imbalance between sorting efficiency and feather protection caused by the fixed gripping acceleration in existing technologies. By integrating factors such as feather shape adaptability, deformation risk, and target quality identification, it achieves dynamic optimization of acceleration, minimizing the risk of irreversible feather deformation while meeting the production line's sorting cycle requirements. The formula design follows the principle of "efficiency and protection balance," with each correction module cooperating to form a complete acceleration control logic.

[0065] Formula Part 1 This is the baseline value for the clamping acceleration, determined based on the production line's sorting cycle time requirements. It ensures that under normal conditions, a sorting speed of at least "1 piece / second" can be met, providing a basic reference for dynamic adjustment of the acceleration. The baseline value is not arbitrarily selected but determined through multiple speed tests and cycle time calibrations, taking into account the motion characteristics of the sorting mechanism and actual operating parameters such as the production line conveyor belt speed. This ensures both sorting efficiency and allows for sufficient adjustment flexibility.

[0066] Formula Part 2 It is an accelerated feather shape adaptation and deformation risk constraint module, the core of which is to balance the needs of feather shape adaptability and feather deformation protection. As a feather type adaptation coefficient, the larger the value, the better the feather type is adapted to the gripping method of the sorting mechanism, and the greater the acceleration it can withstand. Through the linear growth of molecules, the acceleration can be appropriately increased, thereby improving sorting efficiency. As a predictive value for the elastic deformation of the feather shaft, a higher value indicates a greater risk of irreversible deformation during the clamping process. Linearly increasing the denominator can reduce acceleration and avoid deformation damage. The constant term 0.1 in the denominator is mainly to avoid... When the denominator is 0, the denominator becomes too small, leading to excessive acceleration. At the same time, it is important to ensure that... Even at a smaller denominator, reasonable constraints are still provided, making acceleration adjustments more robust. The design of this module follows the logic that "the better the feather shape adaptability and the lower the deformation risk, the greater the acceleration," achieving a preliminary balance between efficiency and protection.

[0067] Formula Part 3 It is the target mass correction module for acceleration, used to adjust acceleration based on the overall mass of the identified target. As the total output value of the dynamic reward function, a larger value indicates a higher overall quality of the identified feathers (including multiple dimensions such as grade, shaft, and fit), making them high-value targets that need to be protected during the sorting process. The linear correction appropriately reduces acceleration when the target quality is high, further reducing the risk of damage to high-value feathers; while for lower-quality feathers, it maintains a relatively high acceleration to ensure sorting efficiency. The coefficient design of 0.2 is based on the results of calibration using a large amount of experimental data, ensuring that the magnitude of the quality correction can effectively protect high-value feathers without excessively affecting the overall sorting rhythm.

[0068] The output of the formula The final clamping acceleration, whose value can be coordinated with the clamping force, enables precise sorting and control of feathers of different qualities and characteristics.

[0069] The technical effects achieved by the above embodiments include: achieving accurate matching between sorting execution parameters and feather characteristics and recognition results, avoiding feather damage caused by improper clamping force or acceleration, and improving sorting integrity and efficiency.

[0070] Traditional technical solutions have the following technical problems: existing sorting target determination only focuses on the identification of the highest layer and isolated screening, without confirming the wing attributes, resulting in mixed packing of feathers after sorting, which affects the subsequent badminton shuttlecock assembly quality and flight performance.

[0071] Based on this, the sorting target determination includes top-layer feather identification, isolation screening, and wing attribute confirmation. Top-layer identification is based on the reward function output threshold determination, isolation screening is achieved through dynamic radius circular domain detection, and wing attributes are confirmed by fusing near-infrared spectral features and morphological symmetry.

[0072] It's worth noting that the sorting target determination comprises three core steps, forming a comprehensive screening mechanism. The top-layer feather identification is based on the output of a dynamic reward function, with a set reasonable threshold. When the total output value of the reward function exceeds the threshold, it is identified as a top-layer feather, ensuring priority sorting of unobstructed and easily graspable feathers. Isolated feather screening employs dynamic radius circular domain detection. Centered on the key node of the feather shaft, the radius of the circular domain is dynamically adjusted according to the coupling coefficient ξ. A larger ξ value indicates a more complete feather structure and less environmental interference, resulting in a larger radius; conversely, a smaller ξ value results in a smaller radius. By detecting the presence of other feathers or interfering objects within the circular domain, it ensures that the sorting target is free from surrounding interference, avoiding over- or under-sampling. Wing attribute confirmation is achieved through the fusion of near-infrared spectral features and morphological symmetry. The spectral response of the left and right wing feathers differs in the spectral data collected by the near-infrared spectral sensor. Combined with the morphological symmetry analysis results, the wing attributes of the feathers are accurately determined, clearly distinguishing between left and right wing feathers. These three steps are executed sequentially and complement each other, ensuring that the sorting target is not only a top-layer, isolated feather but also has clearly defined wing attributes, providing a basis for subsequent box sorting.

[0073] The technical effects achieved by the above embodiments include: enabling multi-dimensional and accurate determination of sorting targets, avoiding mixed feather packing, improving subsequent assembly quality and product flight performance, and reducing rework rate.

[0074] Traditional technical solutions have the following technical problems: existing sorting mechanisms have limited functions, fixed gripping acceleration, and the suction cups cannot rotate, making them unable to adapt to feathers with different shank angles, resulting in unstable gripping or feather damage.

[0075] Based on this, the sorting mechanism is a SCARA robot or a Delta robot. The robot drive module supports continuously adjustable gripping acceleration, and the suction cup has a rotation function. It can adjust its posture according to the direction angle of the bristle and then perform a negative pressure suction action.

[0076] It's worth noting that the sorting mechanism uses either SCARA or Delta robots, both of which possess high-speed and precise positioning capabilities, adapting to the sorting cycle requirements of the production line. The robot drive module has undergone specialized optimization, supporting continuously adjustable gripping acceleration within the range of 0 to 1.5 meters per second squared. It can accurately respond based on dynamically calculated gripping acceleration a_clamp, achieving smooth acceleration adjustment and avoiding feather loss or damage caused by sudden acceleration changes. The suction cup assembly has a rotation function, with a rotation angle range of 0 to 360 degrees. The rotation position is fed back in real time through an angle sensor, ensuring angle control accuracy. During sorting, the robot first receives the feather stalk direction angle θ information, drives the suction cup to rotate to the corresponding angle, aligning the suction cup with the feather stalk direction, and then applies the calculated gripping force. With acceleration The robot performs a negative pressure suction action to ensure stable and reliable grasping. It communicates in real-time with the preceding data processing and decision-making modules via an industrial bus, enabling rapid transmission of instructions and data to guarantee the real-time performance and accuracy of the sorting operation.

[0077] The technical effects achieved by the above embodiments include: enabling the sorting mechanism to adapt to feathers with different stalk angles, improving gripping stability, reducing feather damage, and ensuring sorting efficiency and quality.

[0078] Traditional technical solutions have the following technical problems: existing technologies lack an iterative optimization mechanism and cannot adapt to the long-term changes in the characteristics of different batches of feathers and the production line environment, resulting in a gradual decline in sorting performance and an inability to maintain stable sorting results.

[0079] Based on this, iterative optimization dynamically updates the coefficient parameters in the coupling relationship model and reward function by statistically analyzing the sorting success rate, breakage rate, and wing side misjudgment rate, adapting to the characteristics of different batches of feathers and changes in the production line environment.

[0080] It's worth noting that the iterative optimization mechanism is based on the statistics of key sorting indicators, providing real-time statistics on three core metrics: sorting success rate, breakage rate, and wing-side misjudgment rate. The sorting success rate is the ratio of the number of feathers successfully grasped and correctly sorted to the total number of sorted feathers; the breakage rate is the ratio of the number of feathers that suffer irreversible deformation or damage during sorting to the total number of sorted feathers; and the wing-side misjudgment rate is the ratio of the number of feathers with incorrect wing-side attribute determinations to the total number of sorted feathers. Target thresholds are set for each indicator, and a parameter update process is triggered when a statistical indicator deviates from the target threshold. For coefficients such as ω and μ in the coupling relationship model, a mapping relationship between indicator deviation and coefficient adjustment is established. When the breakage rate is too high, the ω value is appropriately increased to strengthen the coupling effect of humidity and stiffness; when the recognition accuracy is insufficient, the μ value is adjusted to optimize the illumination correction effect. For coefficients such as ζ and θ in the reward function, adjustments are made based on the deviation between the sorting success rate and the wing-side misjudgment rate, improving the sensitivity of the reward function to key features. The parameter updates adopt a small-step iterative approach, continuously monitoring the changes in indicators after each adjustment until the indicators return to the target threshold range, ensuring the stability and reliability of the optimization process, and achieving dynamic adaptation to changes in the characteristics of different batches of feathers and the production line environment.

[0081] The technical effects achieved by the above embodiments include: establishing a continuously optimized closed-loop system, dynamically adapting to feather characteristics and environmental changes, maintaining stable sorting performance over a long period of time, and preventing the sorting effect from declining over time.

[0082] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. An automatic feather sorting method based on morphological recognition and reinforcement learning reward function, comprising: The method involves acquiring and preprocessing image data of feathers on a conveyor belt; its characteristic is that it further includes: Multi-source sensors synchronously collect relevant environmental parameters and material property parameters of feathers; A coupling relationship model is constructed based on feather morphological characteristics, material properties, and environmental parameters; Dynamically adjust the weights of the reinforcement learning reward function; The sorting target is determined by combining the output of the coupling relationship model and the reward function. Dynamically adjust sorting execution parameters according to sorting targets; Control the sorting mechanism to perform sorting actions according to the adjusted parameters; Real-time feedback of sorting results and iterative optimization of model parameters.

2. The automatic feather sorting method based on morphological recognition and reinforcement learning reward function according to claim 1, characterized in that, The multi-source sensors include an industrial camera, a near-infrared spectral sensor, a humidity sensor, a triaxial accelerometer, and a feather density detection lens. The parameters acquired simultaneously include feather images, spectral data, ambient humidity, vibration data, feather density, and the moment of inertia of the feather shaft cross section.

3. The automatic feather sorting method based on morphological recognition and reinforcement learning reward function according to claim 1, characterized in that, Preprocessing includes distortion correction, adaptive dynamic thresholding, and binary mask generation. The dynamic threshold is adjusted in real time based on the rate of change of environmental parameters to ensure that the feather edge features are completely preserved under different environments.

4. The automatic feather sorting method based on morphological recognition and reinforcement learning reward function according to claim 1, characterized in that, Morphological features include the coordinates of key points on the bristle, the distribution pattern of barbs, and wing symmetry. Material properties include bristle stiffness, elastic modulus, and barb sliplock integrity, which are obtained through image texture analysis and spectral feature inference.

5. The automatic feather sorting method based on morphological recognition and reinforcement learning reward function according to claim 1, characterized in that, The coupling relationship model is constructed using the feather-lock integrity and bending coupling coefficient, with the specific formula as follows: ； Where S is the barb sliplock integrity, ranging from 0 to 1, obtained by extracting the cascade structure between barbs through image texture features; ω is the coupling coefficient, used to balance the interaction between humidity and stiffness; H is the ambient humidity, in percentage; K is the barb stiffness, in Newtons per millimeter; A is the ambient vibration amplitude, in gravitational acceleration; μ is the illumination influence coefficient, used to correct the interference of illumination deviation on the coupling relationship; L is the real-time illumination intensity, in lux; L0 is the reference illumination intensity, in lux.

6. The automatic feather sorting method based on morphological recognition and reinforcement learning reward function according to claim 1, characterized in that, The dynamic reward weight adjustment function is: ； ； in The original tier reward weight, The original bar reward weight, The original adaptability reward weight; C represents the confidence level of wing side identification, ranging from 0 to 1, and is determined by fusing spectral features with morphological symmetry; ξ represents the barb sliplock integrity and bending coupling coefficient. The effective signal-to-noise ratio is expressed in decibels; ζ is the vibration attenuation coefficient, used to quantify the weakening effect of vibration on the signal-to-noise ratio; θ is the confidence gain coefficient; erf is the error function; ε is the minimum constant, used to avoid the denominator being zero, and takes the value of 10 to the power of negative six.

7. The automatic feather sorting method based on morphological recognition and reinforcement learning reward function according to claim 1, characterized in that, The sorting execution parameters include gripping force and gripping acceleration, and the linkage formula is: ； ； in The clamping force is measured in Newtons. is the reference force coefficient, with the dimension of Newtons per millimeter; K is the stiffness of the burr bar, with the dimension of Newtons per millimeter; ξ is the bending coefficient of the sprue, derived from ξ; The clamping acceleration is measured in meters per second squared. The reference clamping acceleration is measured in meters per second squared. φ is the feather type adaptation coefficient; The value is the predicted elastic deformation of the bar, in millimeters. This is the total output value of the dynamic reward function.

8. The automatic feather sorting method based on morphological recognition and reinforcement learning reward function according to claim 1, characterized in that, The sorting target determination includes top-layer feather identification, isolation screening, and wing attribute confirmation. Top-layer identification is based on the reward function output threshold determination, isolation screening is achieved through dynamic radius circular domain detection, and wing attributes are confirmed by fusing near-infrared spectral features and morphological symmetry.

9. The automatic feather sorting method based on morphological recognition and reinforcement learning reward function according to claim 1, characterized in that, The sorting mechanism is a SCARA robot or a Delta robot. The robot drive module supports continuously adjustable gripping acceleration, and the suction cup has a rotation function. It can adjust its posture according to the direction angle of the bristle and then perform a negative pressure suction action.

10. The automatic feather sorting method based on morphological recognition and reinforcement learning reward function according to claim 1, characterized in that, Iterative optimization dynamically updates the coefficient parameters in the coupling relationship model and reward function by statistically analyzing the sorting success rate, breakage rate, and wing side misjudgment rate, adapting to changes in feather characteristics and production line environment for different batches.

Images