Method and equipment for dynamically weighing live poultry

Abstract

The embodiment of the invention aims to provide a method, equipment and a computer program product for dynamically weighing live poultry. The method comprises the following steps: when living poultry is put into a weighing box, carrying out continuous multi-modal data acquisition on the living poultry according to a preset acquisition window to obtain a plurality of continuous data packets; performing feature extraction from the corresponding data packet according to the feature extraction window to obtain a plurality of continuous feature vectors; calculating real-time exercise intensity according to the continuous data packet corresponding to the recent reference time window; determining a corresponding main motion state according to a dynamic time window corresponding to the real-time motion intensity and a corresponding continuous feature vector, and determining a corresponding sub-motion state according to the main motion state and a real-time feature vector; and according to the weight estimation model corresponding to the main motion state, carrying out preliminary weight estimation on the living poultry, and according to the sub-motion state, determining a dynamic compensation value of the preliminary weight estimation so as to obtain a final weight value.

Description

Technical Field

[0001] The various embodiments of the present disclosure relate to the field of intelligent farming technology, and in particular to a method for dynamically weighing live poultry. Background Art

[0002] In modern smart farming, the demand for dynamic weighing of live poultry is inextricably linked to the trend toward larger-scale, more refined management. Traditional manual weighing methods are inefficient and prone to causing stress in birds. Dynamic weighing technology, however, allows for frequent weight collection without disrupting the birds' normal movements. This provides a key basis for accurate feed rationing, early disease warning, and optimal marketing timing.

[0003] The core components of an intelligent conveyor belt system include a conveyor mechanism, a weighing module, and a data processing unit. The conveyor mechanism typically utilizes an endless belt made of non-slip material, transporting live birds at a steady speed through a weighing area. The weighing module, embedded in a specific section of the conveyor belt, contains a highly sensitive strain gauge or piezoelectric sensor that captures the weight of the birds as they pass. The data processing unit uses a filtering algorithm to eliminate noise interference caused by the birds' flapping and shaking, ultimately outputting the weight value.

[0004] However, these systems still have limitations. Uncontrollable behaviors of live poultry (such as violent struggling and crowding) can distort weighing signals, and even with algorithmic corrections, the errors can still be significant. Furthermore, the system's installation and operation rely on standardized farming processes, such as the need for bird guideways, making farm renovations difficult. Summary of the Invention

[0005] The purpose of various embodiments of the present disclosure is to provide a method, apparatus, computer program product, and computer program storage medium for dynamic weighing of live poultry.

[0006] According to one aspect of the present disclosure, a method for dynamically weighing live poultry is provided, wherein the method comprises the following steps:

[0007] When the live poultry is placed in the weighing box, continuous multimodal data collection is performed on the poultry according to a predetermined collection window to obtain multiple continuous data packets, each of which includes pressure data, motion measurement data, and optical ranging data of the predetermined collection window;

[0008] Extracting features from the corresponding data packet according to the feature extraction window to obtain a plurality of continuous feature vectors, each of the feature vectors corresponding to one of the feature extraction windows;

[0009] According to the reference time window, the real-time motion intensity is calculated based on the continuous data packets corresponding to the most recent reference time window;

[0010] According to the dynamic time window corresponding to the real-time motion intensity, determining a corresponding main motion state according to the continuous feature vector corresponding to the dynamic time window, and determining a corresponding sub-motion state according to the main motion state and the real-time feature vector;

[0011] A preliminary weight estimate is performed on the live poultry according to the weight estimation model corresponding to the main motion state, and a dynamic compensation value for the preliminary weight estimate is determined according to the sub-motion state, thereby obtaining a final weight value.

[0012] According to one aspect of the present disclosure, there is also provided an apparatus for dynamic weighing of live poultry, wherein the apparatus comprises a memory and a processor, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the apparatus is configured to perform the following operations:

[0013] When the live poultry is placed in the weighing box, continuous multimodal data collection is performed on the poultry according to a predetermined collection window to obtain multiple continuous data packets, each of which includes pressure data, motion measurement data, and optical ranging data of the predetermined collection window;

[0014] Extracting features from the corresponding data packet according to the feature extraction window to obtain a plurality of continuous feature vectors, each of the feature vectors corresponding to one of the feature extraction windows;

[0015] According to the reference time window, the real-time motion intensity is calculated based on the continuous data packets corresponding to the most recent reference time window;

[0016] According to the dynamic time window corresponding to the real-time motion intensity, determining a corresponding main motion state according to the continuous feature vector corresponding to the dynamic time window, and determining a corresponding sub-motion state according to the main motion state and the real-time feature vector;

[0017] A preliminary weight estimate is performed on the live poultry according to the weight estimation model corresponding to the main motion state, and a dynamic compensation value for the preliminary weight estimate is determined according to the sub-motion state, thereby obtaining a final weight value.

[0018] According to one aspect of the present disclosure, a computer program product is further provided, comprising computer program instructions, wherein when the computer program instructions are executed by a computer device, the computer device is configured to perform a method for dynamic weighing of live poultry, wherein the method comprises the following steps:

[0019] When the live poultry is placed in the weighing box, continuous multimodal data collection is performed on the poultry according to a predetermined collection window to obtain multiple continuous data packets, each of which includes pressure data, motion measurement data, and optical ranging data of the predetermined collection window;

[0020] Extracting features from the corresponding data packet according to the feature extraction window to obtain a plurality of continuous feature vectors, each of the feature vectors corresponding to one of the feature extraction windows;

[0021] According to the reference time window, the real-time motion intensity is calculated based on the continuous data packets corresponding to the most recent reference time window;

[0022] According to the dynamic time window corresponding to the real-time motion intensity, determining a corresponding main motion state according to the continuous feature vector corresponding to the dynamic time window, and determining a corresponding sub-motion state according to the main motion state and the real-time feature vector;

[0023] A preliminary weight estimate is performed on the live poultry according to the weight estimation model corresponding to the main motion state, and a dynamic compensation value for the preliminary weight estimate is determined according to the sub-motion state, thereby obtaining a final weight value.

[0024] The embodiments of the present disclosure achieve high-precision weight estimation in complex biological motion scenarios through multimodal data fusion and layered state perception mechanism.

[0025] Among them, the dynamic time window adjustment mechanism adaptively matches the analysis granularity according to the real-time motion intensity, adopts a long window to achieve noise suppression in the static state, switches to a short window to ensure response speed during intense exercise, and achieves the optimal balance between accuracy and efficiency on resource-constrained embedded platforms. The hierarchical state classification model maintains the stability of macroscopic motion intensity recognition and analyzes microscopic behavioral differences (such as distinguishing pecking from pacing) through a collaborative architecture of main state judgment and sub-state segmentation, providing an accurate decision-making basis for differentiated compensation. The dynamic compensation mechanism combines physical modeling with data-driven methods, adopts phase-synchronized vibration compensation for high-frequency flapping, implements inertial advance compensation for impact behavior, and forms a multi-dimensional interference suppression system. The solution disclosed in this disclosure breaks through the limitations of the adaptability of traditional weighing equipment to dynamic interference of living organisms, and still maintains stable output in complex scenarios such as instantaneous strong vibrations, providing high-reliability technical support for automated management of farms. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Other features, objects and advantages of the present disclosure will become more apparent from a reading of the detailed description of non-limiting embodiments made with reference to the following drawings:

[0027] Figure 1A flow chart of a method for dynamic weighing of live poultry according to an exemplary embodiment of the present disclosure is shown.

[0028] The same or similar reference numerals in the drawings represent the same or similar components. DETAILED DESCRIPTION

[0029] The specific embodiments of the present disclosure will be further described below with reference to the accompanying drawings.

[0030] Before discussing the exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments of the present disclosure are described as devices represented by block diagrams and processes or methods represented by flow charts. Although the flow charts describe the operating processes of the various embodiments of the present disclosure as sequential processing, many of the operations therein can be implemented in parallel, concurrently or simultaneously. In addition, the order of the operations can be rearranged. The processes of the various embodiments of the present disclosure can be terminated when their operations are completed, but can also include additional steps not shown in the flow charts. The processes of the various embodiments of the present disclosure can correspond to methods, functions, procedures, subroutines, subprograms, etc.

[0031] The methods illustrated by the flowcharts and the devices illustrated by the block diagrams discussed below may be implemented by hardware, software, firmware, middleware, microcode, hardware description languages, or any combination thereof. When implemented in software, firmware, middleware, or microcode, the program code or code segments that perform the necessary tasks may be stored in a machine or computer-readable medium such as a storage medium. (One or more) processors may perform the necessary tasks.

[0032] Similarly, it will also be understood that any flow charts, flow diagrams, state transition diagrams, and the like represent various processes that can be fully described as program code stored in a computer-readable medium and thereby executed by a computer device or processor, whether or not such computer device or processor is explicitly shown.

[0033] As used herein, the term "storage medium" may refer to one or more devices for storing data, including read-only memory (ROM), random access memory (RAM), magnetic RAM, kernel memory, magnetic disk storage media, optical storage media, flash memory devices, and / or other machine-readable media for storing information. The term "computer-readable medium" may include, but is not limited to, portable or fixed storage devices, optical storage devices, and various other media capable of storing and / or containing instructions and / or data.

[0034] A code segment can represent a procedure, function, subroutine, program, routine, subroutine, module, software package, class, or any combination of instructions, data structures, or program descriptions. A code segment can be coupled to another code segment or hardware circuit by passing and / or receiving information, data, arguments, parameters, or memory contents. Information, arguments, parameters, data, etc. can be passed, forwarded, or transmitted via any suitable means, including shared memory, message passing, token passing, network transmission, etc.

[0035] In this context, "computer device" refers to an electronic device that can perform predetermined processing procedures such as numerical calculations and / or logical calculations by running predetermined programs or instructions. It may include at least a processor and a memory, wherein the processor executes program instructions pre-stored in the memory to perform the predetermined processing procedure, or the predetermined processing procedure is performed by hardware such as ASIC, FPGA, DSP, or a combination of the above two.

[0036] The above-mentioned "computer device" is usually expressed in the form of a general-purpose computer device, and its components may include but are not limited to: one or more processors or processing units, system memory. The system memory may include a computer-readable medium in the form of a volatile memory, such as a random access memory (RAM) and / or a cache memory. The "computer device" may further include other removable / non-removable, volatile / non-volatile computer-readable storage media. The memory may include at least one computer program product having a set (e.g., at least one) program modules that are configured to perform the functions and / or methods of the various embodiments of the present disclosure. The processor executes various functional applications and data processing by running the programs stored in the memory.

[0037] For example, a computer program for executing various functions and processes of the various embodiments of the present disclosure is stored in the memory. When the processor executes the corresponding computer program, the various embodiments of the present disclosure are implemented.

[0038] Typically, a computer device may be, for example, a user device or a network device, or even a combination of the two. The user device includes, but is not limited to, a personal computer (PC), a laptop computer, a mobile terminal, etc., and the mobile terminal includes, but is not limited to, a smartphone, a tablet computer, etc.; the network device includes, but is not limited to, a single network server, a server group consisting of multiple network servers, or a cloud based on cloud computing (Cloud Computing) consisting of a large number of computers or network servers, wherein cloud computing is a type of distributed computing, a super virtual computer consisting of a group of loosely coupled computer sets. The computer device may be run independently to implement the various embodiments of the present disclosure, or may be connected to a network and implement the various embodiments of the present disclosure through interactive operations with other computer devices in the network. The network in which the computer device is located includes, but is not limited to, the Internet, a wide area network, a metropolitan area network, a local area network, a VPN network, etc.

[0039] It should be noted that the user devices, network devices and networks are merely examples. Other existing or future computing devices or networks that are applicable to the embodiments of the present disclosure should also be included in the scope of protection of the present disclosure and are incorporated herein by reference.

[0040] The specific structural and functional details disclosed herein are merely representative and are for the purpose of describing exemplary embodiments of the present disclosure. However, the various embodiments of the present disclosure may be implemented in many alternative forms and should not be construed as being limited to only the embodiments set forth herein.

[0041] It should be understood that although the terms "first," "second," and the like may be used herein to describe various elements, these elements should not be limited by these terms. These terms are used solely to distinguish one element from another. For example, a first element may be referred to as a second element, and similarly, a second element may be referred to as a first element, without departing from the scope of the exemplary embodiments. The term "and / or" as used herein includes any and all combinations of one or more of the listed associated items.

[0042] The terms used herein are intended only to describe specific embodiments and are not intended to limit exemplary embodiments. Unless the context clearly indicates otherwise, the singular forms "a", "an", "an item" used herein are also intended to include the plural. It should also be understood that the terms "comprise" and / or "include" used herein specify the presence of stated features, integers, steps, operations, units and / or components, and do not preclude the presence or addition of one or more other features, integers, steps, operations, units, components and / or combinations thereof.

[0043] It should also be noted that, in some alternative implementations, the functions / actions mentioned may occur in a different order than that indicated in the accompanying drawings. For example, two figures shown in succession may actually be performed substantially simultaneously or may sometimes be performed in the reverse order, depending on the functions / actions involved.

[0044] Figure 1 A flow chart of a method for dynamically weighing live poultry according to an embodiment of the present disclosure is shown.

[0045] See Figure 1 In step S1, when the live poultry is placed in the weighing box, the weighing device collects continuous multimodal data according to the predetermined collection window to obtain multiple continuous data packets; in step S2, the weighing device extracts features from the corresponding data packets according to the feature extraction window to obtain multiple continuous feature vectors; in step S3, the weighing device calculates the real-time motion intensity according to the continuous data packets corresponding to the most recent reference time window according to the reference time window; in step S4, the weighing device determines the corresponding main motion state according to the dynamic time window corresponding to the real-time motion intensity and the continuous feature vector corresponding to the dynamic time window, and determines the corresponding sub-motion state according to the main motion state and the real-time feature vector; in step S5, the weighing device performs a preliminary weight estimation on the live poultry according to the weight estimation model corresponding to the main motion state, and determines the dynamic compensation value for the preliminary weight estimation according to the sub-motion state, thereby obtaining the final weight value.

[0046] Specifically, if Figure 1 As shown, in step S1, when a live poultry is placed in a weighing box, the weighing equipment continuously collects multimodal data according to a predetermined collection window to obtain multiple continuous data packets. Each data packet corresponds to a predetermined collection window and includes pressure data, motion measurement data, and optical ranging data.

[0047] Here, live poultry such as chickens, ducks, geese, pigeons and other farmed poultry. For wild poultry, such as various rare and protected poultry, the dynamic weighing solution disclosed in the present invention is also applicable.

[0048] According to one example, the weighing equipment is configured with a pressure sensor array, a motion measurement unit and an optical ranging unit around the weighing box to collect pressure data, motion measurement data (such as IMU data), and optical ranging data (such as depth map data) after the live poultry enters the weighing box.

[0049] The pressure sensor array is integrated into the load-bearing platform at the bottom of the weighing box and is evenly distributed across the effective weighing area. The array consists of multiple pressure sensing units arranged in a grid format, such as a 4×4 matrix. Each sensing unit independently measures local pressure, collectively forming a complete plantar pressure distribution map. The pressure sensors are covered with a non-slip, wear-resistant coating that is flush with the bottom of the weighing box to prevent discomfort or measurement errors when standing on live birds.

[0050] A motion measurement unit, typically an inertial measurement unit (IMU), is mounted at the center of a rigid frame within the load-bearing structure at the bottom of the weighing box, precisely aligned with the weighing box's geometric coordinate system. This unit is rigidly connected to the weighing box via mechanical fasteners, ensuring accurate sensing of the three-dimensional acceleration and angular velocity changes caused by the motion of the live bird. The unit must be mounted away from the deformation zone of the pressure sensor array to prevent interference between pressure measurements and vibration data.

[0051] Optical ranging units, typically Time of Flight (ToF) cameras, are symmetrically arranged on both sides and in the center of the top of the weighing box to form a multi-view three-dimensional perception network. The top central camera is aimed vertically downward at the center of the weighing area, and the cameras on both sides are tilted at a 45-degree angle toward the inside of the box, together covering the three-dimensional posture information of live poultry. An anti-fouling and light-transmitting film is set on the surface of the camera lens, and it works in conjunction with the lighting module on the top of the box to ensure the stability of the depth map data under different lighting conditions. The optical axis of the camera is parallel to the long axis of the weighing box to ensure the uniformity of the three-dimensional coordinate system.

[0052] The spatial arrangement of each sensor must meet the requirements for synchronous data collection. The pressure sensor array and the inertial measurement unit are physically coupled via a rigid structure to ensure phase consistency of the vibration signal. The installation height of the time-of-flight camera is determined based on the field of view. A typical configuration is 30-50 cm from the weighing platform at the bottom of the weighing box to ensure the complete outline of the live bird is captured while avoiding perspective distortion. Metal reflective surfaces should be avoided inside the weighing box to prevent interference with the accuracy of time-of-flight ranging.

[0053] During the high-precision synchronous data acquisition process, the weighing equipment coordinates the data capture timing of multiple sensors through a unified clock source and adopts a high-precision clock synchronization mechanism based on the IEEE 1588 protocol to achieve μs-level clock synchronization among the pressure sensor array, motion measurement unit, and optical ranging equipment.

[0054] The data acquisition process uses a fixed 100ms cycle, known as the predefined acquisition window, during which each sensor operates in parallel at a preset sampling rate. For example, the pressure sensor array scans pressure data from 16 measurement points at a rate of 100Hz, with each measurement point supporting dynamic measurements of 0-20kg. The six-axis motion sensor captures acceleration and angular velocity information along three axes at a frequency of 200Hz, with an acceleration measurement range of ±16g and an angular velocity detection range of ±2000dps. The ToF camera outputs three 640×480 pixel depth images (at a frame rate of 30fps) per cycle, with each pixel recording three-dimensional distance data with millimeter-level accuracy. The image acquisition time is strictly aligned with the master clock via a hardware trigger signal.

[0055] Multimodal data is structured and encapsulated into standardized data packets. Each packet header embeds a unified time reference marker accurate to the μs level, serving as the time reference for subsequent data processing. Pressure data is stored in a 16-dimensional array with numerical accuracy reaching g-level units. Motion sensor data includes instantaneous measurements of 3D acceleration and 3D angular velocity, stored in floating-point format. Depth image data is processed using a lossless compression algorithm to eliminate spatial redundancy and then converted into a binary data stream.

[0056] According to an example, a data packet is generated every 100ms, containing:

[0057] Time base mark (Unix timestamp, accurate to μs);

[0058] Pressure matrix (16×1 array, unit: g);

[0059] IMU six-dimensional data (Ax, Ay, Az, ω x ,ω y ,ω z );

[0060] Depth map compressed data (binary data stream).

[0061] The weighing equipment completes the collection, verification and packaging of all data sources within a 100ms cycle, ensuring the temporal and spatial consistency of multi-dimensional data.

[0062] This synchronous acquisition mechanism solves the challenge of integrating multi-source heterogeneous data in live dynamic weighing scenarios through hardware-level clock alignment and periodic data packaging strategies. Compared to traditional time-sharing sampling systems, synchronous acquisition achieves millisecond-level synchronous capture of pressure distribution, posture movement, and three-dimensional morphological data, enabling precise correlation and analysis of subtle movements and mass changes in live poultry. The collaborative working mode of multiple sensors under a unified time and space reference provides live dynamic weighing solutions with dynamic error compensation capabilities that cannot be achieved by traditional single sensors, maintaining stable measurement accuracy under complex working conditions such as animal agitation and environmental vibration.

[0063] In step S2, the weighing device performs feature extraction from the data packets corresponding to each feature extraction window according to the feature extraction window to obtain multiple continuous feature vectors.

[0064] In some embodiments, the feature extraction window in step S2 is different from the predetermined acquisition window in step S1. For example, the predetermined acquisition window is 100ms and the feature extraction window is 1s, that is, the weighing device acquires 10 consecutive data packets to perform a feature extraction, and each feature vector corresponds to a consecutive 1s time window (10 100ms data packets).

[0065] According to an example, the feature vector corresponding to a feature extraction window is an 8-dimensional feature vector, including pressure distribution entropy, main vibration frequency, center of mass X offset, center of mass Y offset, motion trajectory curvature radius, angular velocity energy integral, pressure distribution symmetry, and depth mutation variance.

[0066] For the pressure distribution characteristics, 10 pressure matrix data sets were first integrated, and the 160 pressure measurement values were normalized and converted into a probability distribution model. The pressure distribution entropy value was calculated using the information entropy algorithm, which reflects the degree of balance of the pressure distribution. The entropy value calculation result was mapped to a quantitative range of 0 to 4.09, where an entropy value above 3.5 indicates that the living body has drastic posture changes. When living poultry and livestock move frequently, the plantar pressure shows a high entropy characteristic, while the entropy value decreases significantly when standing steadily. For example, the entropy value is usually less than 1.5 in a static state, but it can reach above 3.8 when the wings flap. The pressure distribution entropy value can effectively distinguish between static loads and dynamic interference.

[0067] Furthermore, the pressure matrix in each data packet is divided into left and right symmetrical halves along the central axis. The total pressure values in each half of the pressure matrix are accumulated within a 1-second window, and the ratio is calculated. The absolute value of the logarithm of the ratio is taken to obtain the pressure distribution symmetry, which is used as a quantitative indicator of standing symmetry in living subjects. When this metric approaches zero, it indicates a symmetrical standing posture. Deviations from zero reflect the degree of center of gravity shift. This deviation is mapped to a normalized range of 0-1 using a hyperbolic tangent function. A sliding average filter is used in this calculation to eliminate interference from instantaneous pressure fluctuations, ensuring the stability of the symmetry indicator.

[0068] When processing IMU time series data, digital filtering is used to remove high-frequency noise from the Z-axis acceleration signal, and then a fast Fourier transform is used to extract the primary vibration frequency. A detection range of 2-15 Hz is set to cover the typical motion frequency range of live livestock, including biomechanical characteristics such as walking cadence and wing flapping. After spectral analysis, the vibration frequency component with the highest energy is selected as the primary vibration frequency, with accuracy controlled to 0.1 Hz, ensuring accurate capture of periodic motion.

[0069] Angular velocity energy integration is achieved by summing the squares of the angular velocity components of all sampling points within each data packet. Specifically, the squares of the three-axis angular velocity data in each data packet are summed, integrated over the time step, and then accumulated over 10 data packets to generate an energy parameter that reflects the intensity of the rotational motion of the living organism. This parameter effectively characterizes three-dimensional rotational behaviors such as head turning and rolling in birds.

[0070] During the construction of three-dimensional spatial features, the real-time center of mass coordinates of live poultry and livestock are calculated through 3D point cloud processing. The center of mass offset calculation is based on a dynamic reference system establishment mechanism and utilizes differential processing. First, the center of mass coordinates of the pressure distribution in the first data packet are calculated as the reference point. For each subsequent data packet, the displacement of the current center of mass relative to the reference point in the X and Y axes is calculated in real time. By calculating the root mean square value / maximum fluctuation range of the X and Y displacements in 10 data packets, the center of mass X and Y offset values are generated to accurately capture the lateral movement and forward and backward tilting movements of the poultry. A coordinate rotation correction algorithm is embedded in the calculation process to eliminate measurement errors introduced by the inclination angle of the weighing platform installation.

[0071] The center of mass coordinates are recorded at 100-millisecond intervals to generate a motion trajectory sequence. The curvature radius parameter (i.e., the curvature radius of the motion trajectory) is derived based on the geometric characteristics of the trajectory curve. This curvature feature can identify subtle posture changes such as turning or raising the head. For example, a sudden increase in the curvature of the center of mass trajectory indicates that the living person is adjusting the body's orientation.

[0072] Depth mutation variance is achieved through joint analysis of multiple point clouds. Pixel-level distance statistics are collected for the depth map in each data packet, and the standard deviation of all valid pixels in a single frame is calculated. The instantaneous standard deviation values of 30 depth map frames are collected within a 1-second window. By calculating the variance of these standard deviations, the intensity of dynamic changes in the surface contour of a living organism is quantified. This feature is highly sensitive to surface deformations such as feather erection and limb extension and retraction.

[0073] Through multimodal feature fusion and adaptive window design, this technology addresses the industry challenge of motion interference and noise coupling in in vivo dynamic weighing. Traditional methods rely on single sensor data and struggle to distinguish true weight signals from biological motion noise. By introducing pressure distribution entropy to quantify plantar dynamic characteristics, combined with vibration frequency domain analysis and 3D center of mass trajectory modeling, this technology enables accurate analysis of complex motion patterns. The weighing device maintains stable feature extraction capabilities, particularly in scenarios such as high-frequency flapping and rapid movement. While ensuring interpretability of biological motion features, it effectively extracts weak dynamic interference. Furthermore, the combined use of pressure symmetry metrics and deep mutation variance distinguishes static load offsets from true biological motion, resolving the miscompensation issue caused by center of gravity shift in traditional methods. Angular velocity energy integration overcomes the limitations of planar motion detection and provides a quantitative basis for 3D posture analysis. The structured design of feature vectors significantly improves the generalization performance of subsequent state classification models, making weight compensation calculations biomechanically interpretable.

[0074] In step S3, the weighing device calculates the real-time exercise intensity according to the reference time window and the continuous data packets corresponding to the most recent reference time window.

[0075] Here, the weighing equipment calculates the real-time movement intensity of the birds through a multi-stage process, ensuring that the dynamic weighing process responds accurately to the biological movement.

[0076] According to an example, the reference time window for calculating the real-time motion intensity may be 500 ms, and the update frequency of the reference time window is once every 100 ms, that is, the data overlap rate of adjacent windows is 80%.

[0077] First, obtain the continuous data packets corresponding to the latest reference time window.

[0078] Step S2 caches the latest 10 consecutive data packets through the ring buffer. When step S3 calculates the real-time motion intensity, the continuous data packets corresponding to the latest reference time window can be extracted from the ring buffer, such as the data packets of the latest 500ms (the latest 5 data packets).

[0079] By elastically splitting subwindows, the real-time requirements of a short-term benchmark window (500ms benchmark time window) are met while maintaining the accuracy of long-term feature extraction (1s feature extraction window). This elastic splitting avoids the degradation of feature quality caused by shortening the window, especially in high-frequency motion scenarios, preserving the spectral resolution advantage of the 1s feature extraction window. Although real-time calculation of short-term dynamic features requires additional computing resources, it prevents the smoothing interference of long windows on short-term features.

[0080] Subsequently, a weighted sum of each motion component is calculated and dynamically normalized to obtain a real-time motion intensity value. The motion components include Z-axis acceleration variance, pressure distribution entropy, and center of mass displacement, capturing the motion characteristics of live birds from three dimensions: vertical vibration, pressure distribution, and horizontal motion.

[0081] For the Z-axis acceleration data set, the numerical fluctuation index for the last 500ms is calculated. This index is generated by statistically analyzing the deviation of each Z-axis acceleration data point from the average value, effectively reflecting the intensity of the subject's vertical movement. The pressure distribution entropy of the last five data packets is also calculated to characterize the disordered state of plantar pressure changes at the current moment. For the center of mass movement distance, the linear distance change between the center of mass positions at adjacent time points is accumulated to quantify the amplitude of the subject's horizontal movement.

[0082] The above motion components are linearly weighted and superimposed to generate an intermediate value for comprehensive motion intensity. The weighting emphasizes the influence of vertical vibration while also taking into account the contributions of pressure distribution and center of mass shift. For example, the weights are 0.5 for Z-axis acceleration variance, 0.2 for pressure distribution entropy, and 0.3 for center of mass shift. Dynamic normalization is used to fuse the multi-dimensional motion characteristic components into a unified motion intensity index. The weighing device maintains a continuously updated historical database, storing the fused motion component values for all calculation cycles within the last 30 seconds. Scaling is performed based on the historical intensity extremes within the past 30 seconds. This dynamic scaling interval is constructed by identifying the maximum and minimum values in the historical data. The current fused motion component value is then mapped to a standard range of 0-1 to obtain the final real-time motion intensity value. This mechanism enables the weighing device to automatically adapt to the differences in motion intensity among individual live animals, eliminating the interference of individual mass differences in status assessment. The dynamic normalization strategy automatically establishes an individualized assessment benchmark through long-term data learning, making the motion intensity of juveniles and adults comparable. This feature significantly improves classification accuracy in continuous weighing scenarios for group-reared live poultry.

[0083] In step S4, the weighing device determines the corresponding main motion state according to the dynamic time window corresponding to the real-time motion intensity and the continuous feature vector corresponding to the dynamic time window, and determines the corresponding sub-motion state according to the main motion state and the real-time feature vector.

[0084] Here, motion state adaptive classification data preparation is achieved through a dynamic time window adjustment mechanism.

[0085] The system uses a range of real-time exercise intensity values to categorize exercise intensity. When the normalized real-time exercise intensity reaches or exceeds 0.7, vigorous exercise mode is activated. Moderate exercise mode is enabled when the intensity is between 0.3 and 0.7, and the system switches to static monitoring mode when the intensity falls below 0.3. These grading thresholds were calibrated through biomechanical experiments to match common behavioral characteristics of live poultry.

[0086] Differentiated data window management strategies are implemented for different motion modes (exercise intensity intervals). In intense motion mode, a short-term analysis window containing the most recent 1s of data is used. This window consists of one eigenvector to ensure that transient features of fast movements are captured. This window length can quickly capture state changes in high-frequency motion scenarios, such as flapping wings or rapid turning behaviors. In moderate motion mode, it is extended to a standard 3s window, covering 3 consecutive eigenvectors to balance motion details and pattern stability. The 3s window balances motion tracking accuracy and historical data continuity, and is suitable for routine activities such as walking and pecking. In static mode, it is further extended to a 5s long window, and the wide time span of 5 eigenvectors is used to suppress environmental noise interference, which can eliminate weak interference such as respiratory tremors.

[0087] The weighing equipment performs multi-level motion state analysis based on the feature vector sequence within the dynamic time window.

[0088] The main motion state determination module receives a continuous sequence of feature vectors within a dynamic time window and extracts temporal features using a two-layer LSTM network. The first layer analyzes short-term correlation patterns between feature vectors, such as capturing the periodic pressure entropy fluctuations associated with wing flapping behavior. The second layer identifies long-term evolutionary trends across the time window, such as the transition from a static state to vigorous motion. The network output layer generates a main state probability distribution encompassing three categories: static, micro-motion, and vigorous motion. When the probability of a category exceeds 0.7, the corresponding main state flag is triggered.

[0089] When the short-term analysis window determined by real-time motion intensity values includes only one eigenvector, a sliding window expansion scheme is used to input the previous eigenvector along with the eigenvectors within the window into a two-layer LSTM network to extract temporal features and perform primary state identification. This allows the model to maintain high-precision state classification even in intense motion scenarios. This sliding window expansion scheme preserves the LSTM's temporal modeling capabilities, breaking through the bottleneck of short-window temporal analysis and providing a universal state identification framework for dynamic weighing systems. It demonstrates exceptional technical adaptability, particularly in high-frequency motion interference scenarios.

[0090] In one example, the model's predicted probability of the primary state must exceed a threshold to ensure high confidence in the judgment result and reduce the risk of misjudgment. Furthermore, real-time motion intensity is used as a physical verification condition to prevent misjudgments caused by bias in training data, such as mistaking a brief pause for stillness.

[0091] Examples of mixed judgment conditions for the main state are shown in Table 1 below:

[0092]

[0093] Table 1

[0094] The real-time feature vector is the last feature vector within the dynamic time window, representing the instantaneous motion characteristics at the current moment. During the sub-motion state determination phase, this real-time feature vector is input into the rule engine along with the determined main state label. For example, when the main state is inching, the engine detects the periodic reciprocating changes in the center of mass trajectory in the real-time feature vector (alternating increases and decreases in the X / Y offset) and fluctuations in the pressure entropy value between 1.8 and 2.2, thus determining the sub-state as "pacing." This determination process simultaneously references the evolution trend of historical feature vectors to avoid transient noise interference.

[0095] The state judgment rule library dynamically loads differentiated analysis strategies according to the main state category. For the main state of intense motion, the rule library enables high-frequency feature monitoring mode, focusing on tracking acceleration peaks and center of mass mutation parameters; the main state of micro-motion activates periodic pattern recognition to detect rhythmic changes in pressure entropy and angular velocity energy. The rule threshold is dynamically adjusted according to the main state probability. For example, when the probability of intense motion is 0.8, the acceleration threshold for collision judgment is increased from 5m / s to 10m / s. 2 Down to 4.2m / s 2 To improve sensitivity.

[0096] The mapping relationship between the main state and the sub-state is shown in Table 2 below:

[0097]

[0098] Table 2

[0099] The judgment logic for the main motion state and sub-motion state is based on the collaborative analysis of multimodal sensor data. The standing sub-state in the static state is determined by the dual conditions of the fluctuation range threshold of the three-dimensional coordinate of the center of mass and the pressure distribution entropy value. When the standard deviation of the center of mass displacement is less than the preset limit and the pressure entropy is continuously below the critical level, the living body is judged to be in a static standing posture. The micro-motion state is further refined into pecking and pacing sub-states. The core characteristic of pecking behavior is the periodic fluctuation of the pressure distribution entropy value, whose fluctuation frequency is consistent with the extension and retraction movement of the living body's neck, and the acceleration spectrum shows a significant energy peak in the range of 2-4Hz. Pacing is identified by the regular sinusoidal oscillation of the center of mass coordinates in the horizontal plane, accompanied by periodic changes in the pressure distribution symmetry index, and its period matches the gait frequency.

[0100] During intense exercise, the high-frequency flapping sub-state is determined by the acceleration signal's dominant frequency exceeding the normal threshold for biological motion, such as a bird's wing flapping frequency exceeding 5Hz. Simultaneously, the pressure distribution entropy exceeds the critical threshold for disturbance, indicating a dramatic fluctuation in plantar pressure. The collision box sub-state is determined through a combination of center of mass coordinate mutation detection and acceleration peak exceeding the limit. A collision is detected when the center of mass position offset exceeds the dynamic threshold within 100ms and the instantaneous value of the three-axis acceleration reaches the impact threshold. A hysteresis debounce mechanism is introduced into the judgment logic for each sub-state to prevent state transitions caused by transient noise interference. For example, high-frequency flapping must continuously meet the frequency condition for more than 300ms to trigger a state switch.

[0101] This hierarchical judgment system achieves multi-granular analysis of the motion characteristics of live poultry by integrating time, frequency, and spatial domain features. Main-state classification focuses on macroscopic activity intensity, while sub-state analysis captures microscopic behavioral differences. The two work together to enhance classification robustness in complex motion scenarios. The precise distinction between stillness and micro-motion prevents respiratory tremors from being misidentified as valid movement, while sub-state identification during intense movement provides the basis for differentiated weight compensation strategies. For example, high-frequency flapping triggers phase synchronization compensation, while collisions with the enclosure activate shock filtering algorithms, significantly improving dynamic weighing accuracy. Compared to traditional single-layer classification models, the sub-state rule engine is customized for the main state, significantly improving the recognition accuracy of high-frequency flapping while reducing the false positive rate. The introduction of real-time feature vectors reduces response latency to transient movements, meeting the real-time requirements of dynamic weighing.

[0102] In step S5, the weighing device performs a preliminary weight estimation on the live poultry according to the weight estimation model corresponding to the main motion state, and determines a dynamic compensation value for the preliminary weight estimation according to the sub-motion state, thereby obtaining a final weight value.

[0103] Here, the weighing device performs dynamic weight compensation calculation based on the layer state recognition result.

[0104] In one example, a preset weight estimation model is loaded based on the primary motion state tag. For example, a sliding window mean model is used in static conditions, a Kalman filter (UKF) model is used in micro-motion conditions, and an enhanced UKF model is used in intense motion conditions. When the model is loaded, a parameter set is initialized simultaneously, including core parameters such as noise covariance and convergence rate, to ensure that it matches the current motion intensity.

[0105] The sliding window mean model performs multi-period averaging on static data to suppress minor fluctuations caused by respiratory tremors. The Kalman filter model addresses periodic vibrations during micro-motion, separating true weight signals from motion noise through state prediction and observation correction. The anti-interference-enhanced UKF model utilizes a combined frequency-domain filtering and inertia compensation algorithm to offset nonlinear interference caused by intense movement.

[0106] Dynamic weighing at rest

[0107] In the static state, the sliding window mean model reads 50 consecutive pressure matrices from the ring buffer within the last 5 seconds, accumulates all pressure measurements at the same moment, and generates a total pressure value for that moment. A weighted average algorithm is used to calculate a preliminary weight estimate for these 50 consecutive total pressure values. Weights are assigned using a linearly decreasing strategy, with data closer to the current moment receiving higher weights. For example, the most recent data point receives a weight of 0.3, the previous point receives a weight of 0.27, and so on, until the weight of the furthest point is reduced to 0.01. This weighting mechanism enhances the algorithm's sensitivity to recent pressure changes while retaining the smoothing effect of historical data. For sudden pressure outliers within the window, such as sudden pressure drops caused by a discharge event, the model pre-processes them using a median filter to remove them, preventing abnormal data from contaminating the mean calculation.

[0108] The sliding window mean model outputs a noise-suppressed static weight estimate, reflecting the baseline weight of live poultry at rest. When the estimated fluctuations for three consecutive windows are less than a preset threshold, the weight data is considered stable, triggering a rapid output mechanism. The sliding window mean model effectively eliminates physiological interferences such as respiratory tremors and muscle micro-vibrations through time-domain smoothing, providing a highly stable baseline for dynamic weighing.

[0109] In the static state, no dynamic compensation based on the sub-state is required, such as vibration compensation and center of mass compensation. The weight estimate output by the sliding window mean model is the final weight value of the current live poultry.

[0110] Dynamic weighing in micro-motion state

[0111] During micro-motion, the input data for the unscented Kalman filter includes 30 packets from the last three seconds, including the pressure matrix from the pressure sensor array, the three-axis acceleration and angular velocity signals from the IMU, and the depth image provided by the ToF camera. The pressure sensor data is averaged to suppress high-frequency noise, while the acceleration signal is bandpass filtered to extract 2-5Hz bio-motion features and suppress high-frequency mechanical vibration interference. The center of mass coordinates are smoothed using a trajectory algorithm to eliminate positioning jitter.

[0112] An unscented Kalman filter establishes a state-space model of the dynamic weighing system, using the baseline weight, weight slow-change rate, and motion disturbance amplitude of live poultry as the state variables to be estimated. During the prediction phase, a state transition equation is constructed based on biomechanical properties to simulate the combined effects of periodic vibration and slow center of mass change on the pressure sensor. During the observation phase, the total pressure value and vibration energy value are integrated to calculate the predicted pressure value through nonlinear mapping. After iterative correction, a preliminary weight estimate is output. This value reflects the stable mass baseline of live poultry in a micro-motion state and provides reliable input for subsequent compensation calculations.

[0113] Based on the pecking state, the 2-3Hz characteristic frequency component is extracted from the vibration spectrum, and the pressure fluctuation amplitude caused by the vertical periodic vibration is calculated. The vertical vibration compensation coefficient is obtained based on the pre-built frequency-amplitude lookup table, and an inverted waveform is generated to offset the periodic interference in the sensor reading. For example, when a pecking frequency of 2.5Hz is detected, a negative compensation pulse of the corresponding amplitude is applied. For unforeseen frequencies (such as 3.2Hz), the compensation coefficient is calculated by interpolation of adjacent table entries. For example, if the table only contains coefficients for 3.0Hz and 3.5Hz, the compensation amount for 3.2Hz is generated by linear interpolation. Based on the table lookup method, lightweight online learning (such as LMS filtering) is introduced to fine-tune the compensation coefficient. For example, the table lookup provides a baseline value, and the amplitude of ±10% is adaptively adjusted based on the real-time residual.

[0114] Based on the pacing sub-state and the horizontal offset pattern of the center of mass trajectory, the time-domain correlation between the displacement rate in the X-axis and the symmetry of the pressure distribution is analyzed. The horizontal inertia compensation weight is dynamically adjusted according to the center of mass movement rate, with the compensation strength proportional to the square of the displacement rate to match the physical characteristics of the inertial force. For regular oscillation behavior, the pacing main frequency is extracted through fast Fourier transform and a phase prediction model is constructed to generate a 1 / 4-cycle-ahead compensation waveform to offset system response delays. For sudden non-periodic offsets, a second-order kinematic integral prediction is performed in conjunction with the real-time acceleration data of the inertial measurement unit. The displacement increment within the next 50ms is calculated in advance and pre-compensation is applied. A sliding mean filter is introduced to eliminate high-frequency noise interference and ensure the smoothness and stability of dynamic corrections.

[0115] The initial weight estimate and sub-state compensation values are combined using a multi-dimensional weight allocation strategy. The contribution of compensation items is dynamically adjusted based on the intensity of the main motion state and the type of sub-state. For example, vibration compensation for pecking accounts for 340%, while center of mass compensation for pacing accounts for 30%. The result of this combination is the final weight value.

[0116] Furthermore, the rationality of the synthesis result is verified by the density-volume relationship verification module. The biological volume is calculated based on the three-dimensional model reconstructed by the ToF camera, and the compensated weight estimate is divided by the volume value to obtain the real-time density. If the density value exceeds the reasonable physiological range of the living body (for example, 0.95-1.05 kg / L), the adaptive correction mechanism is triggered. This mechanism extracts the sliding window mean of the historical density data in the last 30 seconds, and combines it with the current volume value to reversely correct the weight estimate. For example, when an abnormal increase in density is detected, the density mean in the window is multiplied by the current volume to generate the calibrated weight to ensure that the result conforms to the laws of biophysics.

[0117] Dynamic weighing during strenuous exercise

[0118] During intense motion, the enhanced UKF model extracts the 10 data packets from the last second. These data packets include the pressure matrix from the pressure sensor array, the three-axis acceleration and angular velocity signals from the IMU, and the depth image provided by the ToF camera. The pressure data undergoes anti-aliasing filtering, retaining high-frequency components above 20 Hz to capture the transient impact of wing flapping. The acceleration signal undergoes bandpass filtering to extract the 5-15 Hz frequency band characteristic of intense biological motion and suppress mechanical noise. The center of mass trajectory is smoothed and extrapolated using a motion prediction algorithm to eliminate point cloud positioning jitter caused by rapid movement.

[0119] An enhanced unscented Kalman filter constructs an extended state-space model, adding a vibration phase tracking term and a center of mass shift coupling coefficient. During the prediction phase, the model simulates the combined disturbances of high-frequency vibration and sudden center of mass changes, such as periodic pressure fluctuations caused by wing flapping and transient inertial shifts caused by impacts. During the observation phase, the high-frequency components of the pressure sensor are integrated with vibration phase information, separating the true weight signal from motion artifacts through a nonlinear mapping. After the observation phase, the filter enters an iterative correction process. The residual between the predicted weight and the sensor measured value is calculated and combined with the covariance matrix to dynamically adjust the confidence weights of the state estimate. With each iteration, the model adaptively shrinks or expands the parameter search range based on the residual, gradually approaching the optimal solution. When the fluctuation of the weight estimate for three consecutive iterations is less than a preset convergence threshold, the state estimate is considered stable and a preliminary weight estimate stripped of motion disturbances is output. This process ensures that the weight estimate quickly converges to the theoretical benchmark under the combined disturbances of wing flapping and impacts, avoiding system oscillations caused by over- or under-compensation.

[0120] For high-frequency flapping sub-states, the main frequency components exceeding 8Hz are extracted from the vibration spectrum, and a reverse compensation waveform is generated based on a pre-calibrated frequency-amplitude relationship table. The compensation module introduces a phase synchronization mechanism, predicting the interference trend of the next cycle through historical vibration phase data and applying anti-phase cancellation pulses in advance. For example, when a flapping frequency of 10Hz is detected, a reverse vibration compensation amount with a phase lead of 25ms is applied to eliminate the periodic fluctuations of the sensor reading. For unforeseen frequencies (such as 10.8Hz), the compensation coefficient is calculated by interpolation of adjacent table entries. For example, if the table only contains coefficients for 10.4Hz and 11Hz, the compensation amount for 10.8Hz is generated by linear interpolation. Based on the table lookup method, lightweight online learning (such as LMS filtering) is introduced to fine-tune the compensation coefficient. For example, the table lookup provides a baseline value, and the amplitude is adaptively adjusted by ±10% based on the real-time residual.

[0121] For the collision box sub-state, the impact peak of the acceleration signal and the rate of change in the center of mass coordinate are analyzed. When the X / Y axis acceleration exceeds the biological motion limit threshold, the impact filtering algorithm is triggered. This algorithm uses the real-time acceleration integral of the inertial measurement unit to predict the center of mass displacement increment within the next 30ms. It then calculates the horizontal inertial force compensation based on the rate of change in the pressure distribution symmetry. For example, if a lateral collision causes a sudden change in the center of mass velocity of 0.2m / s, the compensation weight is dynamically increased based on the square of the displacement, generating a force correction term that acts in the opposite direction of motion.

[0122] The initial weight estimate and sub-state compensation values are combined using a dynamic weight allocation strategy. In the high-frequency flapping sub-state, the vibration compensation weight is increased to 70%, while the center of mass compensation weight is set to 30%. In the box impact sub-state, the center of mass compensation weight increases to 65%, while the vibration compensation weight decreases to 35%. The weight coefficients are dynamically adjusted via a fuzzy logic controller. Input parameters include the main vibration frequency amplitude, the center of mass mutation rate, and the pressure distribution entropy. The output is a smooth weight curve to avoid abrupt changes in the compensation amount.

[0123] Furthermore, the rationality of the synthesis result is verified by the density-volume relationship verification module. The biological volume is calculated based on the three-dimensional model reconstructed by the ToF camera, and the compensated weight estimate is divided by the volume value to obtain the real-time density. If the density value exceeds the reasonable physiological range of the living body (for example, 0.95-1.05 kg / L), the adaptive correction mechanism is triggered. This mechanism extracts the sliding window mean of the historical density data in the last 30 seconds, and combines it with the current volume value to reversely correct the weight estimate. For example, when an abnormal increase in density is detected, the density mean in the window is multiplied by the current volume to generate the calibrated weight to ensure that the result conforms to the laws of biophysics.

[0124] In the dynamic weighing solution, a dynamic compensation mechanism based on multi-state layering achieves high-precision mass estimation in all motion scenarios by stripping common interference and adapting and correcting individual behaviors. Differentiated modeling and compensation strategies are adopted for different states of live poultry and livestock, such as stillness, micro-motion, and intense movement, breaking through the adaptability limitations of traditional solutions to complex biological motion characteristics. In the micro-motion state, Kalman filtering is used to separate periodic vibrations and slowly varying inertial interference, effectively reducing the high-frequency vibration compensation error of pecking behavior, improving the accuracy of center of mass offset compensation for pacing behavior, and significantly suppressing asymmetric fluctuations in plantar pressure distribution. In intense motion scenarios, the enhanced filtering model combines phase prediction with inertial advance compensation to improve the suppression rate of periodic interference from high-frequency flapping and reduce the instantaneous impact error of colliding with the box, thereby significantly reducing weighing fluctuations under extreme motion.

[0125] It should be noted that the various embodiments of the present disclosure may be implemented in software and / or a combination of software and hardware, for example, may be implemented using an application specific integrated circuit (ASIC), a general purpose computer, or any other similar hardware device. In one embodiment, the software programs of the various embodiments of the present disclosure may be executed by a processor to implement the steps or functions described above. Similarly, the software programs of the various embodiments of the present disclosure (including related data structures) may be stored in a computer-readable recording medium, for example, a RAM memory, a magnetic or optical drive, or a floppy disk and the like. In addition, some steps or functions of the various embodiments of the present disclosure may be implemented using hardware, for example, as a circuit that cooperates with a processor to perform various steps or functions.

[0126] In addition, at least a portion of the various embodiments of the present disclosure may be implemented as a computer program product, such as computer program instructions, which, when executed by a computing device, can invoke or provide the methods and / or technical solutions according to the various embodiments of the present disclosure through the operation of the computing device. The program instructions for invoking / providing the methods of the various embodiments of the present disclosure may be stored in a fixed or removable recording medium, and / or transmitted via a data stream in a broadcast or other signal-carrying medium, and / or stored in the working memory of a computing device that operates according to the program instructions.

[0127] For those skilled in the art, it is obvious that the embodiments of the present disclosure are not limited to the details of the above-mentioned exemplary embodiments, and the embodiments of the present disclosure can be implemented in other specific forms without departing from the spirit or basic characteristics of the embodiments of the present disclosure. Therefore, no matter from which point of view, the embodiments should be regarded as exemplary and non-restrictive, and the scope of the embodiments of the present disclosure is limited by the appended claims rather than the above description, and it is intended that all changes that fall within the meaning and scope of the equivalent elements of the claims are included in the embodiments of the present disclosure. Any figure mark in the claims should not be regarded as limiting the claims involved. In addition, it is obvious that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units or devices stated in the system claim can also be implemented by one unit or device through software or hardware. Words such as first and second are used to indicate names and do not indicate any specific order.

Claims

1. A method for dynamic weighing of live poultry, wherein: The method comprises the following steps: When the live poultry is placed in the weighing box, continuous multimodal data collection is performed on the poultry according to a predetermined collection window to obtain multiple continuous data packets, each of which includes pressure data, motion measurement data, and optical ranging data of the predetermined collection window; Extracting features from the corresponding data packet according to the feature extraction window to obtain a plurality of continuous feature vectors, each of the feature vectors corresponding to one of the feature extraction windows; According to the reference time window, the real-time motion intensity is calculated based on the continuous data packets corresponding to the most recent reference time window; According to the dynamic time window corresponding to the real-time motion intensity, determining a corresponding main motion state according to the continuous feature vector corresponding to the dynamic time window, and determining a corresponding sub-motion state according to the main motion state and the real-time feature vector; A preliminary weight estimate is performed on the live poultry according to the weight estimation model corresponding to the main motion state, and a dynamic compensation value for the preliminary weight estimate is determined according to the sub-motion state, thereby obtaining a final weight value.

2. The method according to claim 1, wherein The step of calculating the real-time exercise intensity specifically includes: Obtaining continuous data packets corresponding to the latest reference time window; According to the continuous data packets, a weighted sum of multi-dimensional motion components is calculated and dynamically normalized to obtain the real-time motion intensity, wherein the motion components include Z-axis acceleration variance, pressure distribution entropy value and center of mass movement distance.

3. The method according to claim 1, wherein The step of determining the dynamic time window specifically includes: Determining the exercise intensity interval based on the numerical range of the real-time exercise intensity; Based on the motion intensity interval, the dynamic time window for determining the main motion state is determined.

4. The method according to claim 1 or 3, wherein According to the continuous feature vector corresponding to the dynamic time window, the corresponding main motion state is determined through a double-layer LSTM network.

5. The method according to claim 4, wherein When the continuous feature vectors corresponding to the dynamic time window include only one feature vector, the previous feature vector is input into the double-layer LSTM network together with the feature vector in the dynamic time window through a sliding window expansion scheme.

6. The method according to claim 1, wherein The corresponding relationship between the main motion state and the weight estimation model includes any one of the following: -The static state uses a sliding window mean model; -The Kalman filter model is used for micro-motion state; -Using enhanced UKF model for intense exercise.

7. The method according to claim 6, wherein: The dynamic compensation of the sub-motion state includes horizontal center of mass compensation and / or vertical vibration compensation.

8. The method according to claim 1, wherein Each of the characteristic vectors includes at least: a pressure distribution entropy value, a main vibration frequency, a center of mass offset value, and a motion trajectory curvature radius.

9. An apparatus for dynamic weighing of live poultry, wherein: The device includes a memory and a processor, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the device is configured to perform the following operations: When the live poultry is placed in the weighing box, continuous multimodal data collection is performed on the poultry according to a predetermined collection window to obtain multiple continuous data packets, each of which includes pressure data, motion measurement data, and optical ranging data of the predetermined collection window; Extracting features from the corresponding data packet according to the feature extraction window to obtain a plurality of continuous feature vectors, each of the feature vectors corresponding to one of the feature extraction windows; According to the reference time window, the real-time motion intensity is calculated based on the continuous data packets corresponding to the most recent reference time window; According to the dynamic time window corresponding to the real-time motion intensity, determining a corresponding main motion state according to the continuous feature vector corresponding to the dynamic time window, and determining a corresponding sub-motion state according to the main motion state and the real-time feature vector; A preliminary weight estimate is performed on the live poultry according to the weight estimation model corresponding to the main motion state, and a dynamic compensation value for the preliminary weight estimate is determined according to the sub-motion state, thereby obtaining a final weight value.

10. A computer program product comprising computer program instructions, wherein: When the computer program instructions are executed by a computer device, the computer device is configured to perform the method according to any one of claims 1 to 8.

Images