A Method and System for Poultry Weight Data Analysis Based on Multi-stage Cleaning and Distribution Correction

Abstract

This invention discloses a method and system for analyzing poultry weight data based on multi-level cleaning and distribution correction, belonging to the field of smart farming technology. The method includes: collecting raw weight data through intelligent weighing sensors, dynamically thresholding and compressing the data via an edge gateway, and then uploading it to a cloud platform; the cloud platform performs three-level cleaning—dynamic fine-tuning based on a large language model, time-aware deduplication, and intelligent slicing of stable intervals—to construct a high-quality, effective sample set; analyzing the statistical characteristics of the samples using a deep belief network, adaptively selecting the optimal correction strategy, and outputting representative weights; constructing spatiotemporal graph data based on historical and current weights, and predicting future trends in group weight and evenness through a spatiotemporal graph convolutional network; and automatically generating feed adjustment, health warning, and slaughter planning decision suggestions based on the prediction results. This invention effectively improves the quality and accuracy of poultry weight data analysis, realizing intelligent closed-loop management from data collection to decision support.

Description

Technical Field

[0001] This invention relates to the field of smart farming technology, and more specifically, to a method and system for analyzing poultry weight data based on multi-level cleaning and distribution correction. Background Technology

[0002] The modernization and intelligent transformation of poultry farming relies heavily on accurate and reliable data analysis and decision support. Accurate measurement of poultry weight and assessment of herd evenness are core aspects throughout the entire farming process, directly impacting feed efficiency, health levels, and ultimately, economic benefits. Traditional weight management primarily relies on manual, periodic sampling and weighing. This method is not only inefficient and prone to human error, but also lacks real-time data, failing to reflect the dynamic changes in the herd. In recent years, the deployment of intelligent weighing sensors and automated data acquisition equipment has enabled the automated acquisition and uploading of poultry weight data, laying the foundation for data-driven farming. However, in practical applications, the quality of the raw data collected is often questionable. Electromagnetic interference, mechanical vibrations, poultry stress behaviors, and fluctuations in data transmission links at the farming site all contribute to data containing significant noise, outliers, and duplicate records. Using this "dirty data" directly without processing will severely distort subsequent analysis results.

[0003] Existing data preprocessing methods are mostly simple threshold filtering or moving averages, lacking multi-level, interconnected cleaning mechanisms tailored to the characteristics of poultry weight data. This results in insufficient ability to identify deeply hidden outliers and unstable sequences, limiting the thoroughness and reliability of data cleaning. More importantly, even after data cleaning, the analytical capabilities of most current systems remain at the descriptive statistical level. Traditional mean methods are sensitive to skewed distribution data and easily affected by extreme values, causing representative weights to deviate from reality. Furthermore, most systems have not achieved a closed-loop process from data collection, cleaning, and correction to decision support, resulting in a disconnect between data processing and business applications. Therefore, there is an urgent need for a poultry weight data analysis method that integrates multi-level data cleaning mechanisms and intelligent distribution correction strategies to improve data quality and statistical accuracy, supporting scientific decision-making in poultry management. Summary of the Invention

[0004] To address the aforementioned technical issues, this invention proposes a method and system for analyzing poultry weight data based on multi-level cleaning and distribution correction. By integrating multi-level data cleaning, intelligent distribution correction, and deep learning prediction models, the entire process from data acquisition to decision support is optimized.

[0005] The first aspect of this invention provides a method for analyzing poultry weight data based on multi-level cleaning and distribution correction, comprising the following steps:

[0006] The raw weight data of poultry is collected in real time by an intelligent weighing sensor array. The edge computing gateway performs preliminary filtering and data compression using a preset dynamic threshold model, and then sends the data to the Internet of Things cloud platform through an adaptive transmission protocol.

[0007] After receiving the data, the cloud platform performs a three-level progressive cleaning process, including dynamic fine filtering based on large language models and business rules, time-series-aware deduplication of adjacent data, and intelligent slicing of stable intervals to extract the most densely distributed stable intervals of data and form a high-quality and effective sample set.

[0008] Calculate the statistical features of the high-quality effective sample set, input the statistical features into a pre-trained deep belief network, output the optimal correction strategy for the current data distribution, and transform and calculate the sample data according to the optimal correction strategy to output a representative weight value.

[0009] Using historical and current corrected representative weight data, a spatiotemporal map of the growth status of chicken flocks in various areas of the farm is constructed. A spatiotemporal map convolutional network model is then used to predict the trend of flock weight and evenness at a specified future time point.

[0010] Based on the predicted trend results, combined with preset breeding goals and business data, the system automatically generates breeding decision-making suggestions, including feed feeding adjustments, health management early warnings, and slaughter time planning.

[0011] In this solution, the edge computing gateway uses a pre-set dynamic threshold model to perform preliminary filtering and data compression on the raw weight data, and then sends it to the IoT cloud platform via an adaptive transmission protocol, including:

[0012] The raw weight data of poultry is collected in real time using an intelligent weighing sensor array deployed in the farm. A dynamic threshold model is pre-set in the edge computing gateway. The dynamic threshold model is embedded with a standard growth model that matches the poultry breed and age. The baseline expected weight is obtained according to the standard growth model.

[0013] A physiological fluctuation coefficient is set based on the daily weight growth rate and physiological fluctuation pattern of poultry at different ages, and historical data within a preset period is obtained. The dispersion of the historical data is analyzed to generate a historical stability coefficient. An adaptive fluctuation range of the baseline expected weight is set according to the physiological fluctuation coefficient and the historical stability coefficient.

[0014] Obtain context information from other sensors in the farm, perform context compensation on the threshold of the adaptive floating range based on the correlation between the context information and poultry weight, and use the finally determined adaptive floating range to initially filter the original weight data to obtain valid candidate data;

[0015] The valid candidate data is lightweight compressed using differential coding, and an adaptive transmission protocol is determined based on the network signal strength. The data is then sent to the IoT cloud platform using the adaptive transmission protocol.

[0016] In this solution, the first stage of cloud-based data cleaning involves dynamic and refined filtering of the received weight data based on a large language model and business rules, including:

[0017] The static attributes, dynamic health status, real-time environmental data, and historical weight data of the current batch of poultry are extracted from the poultry management system of the farm, and the extracted contextual information is used to construct a structured scenario description.

[0018] The scenario description is constructed as prompt words, input into a large language model trained and fine-tuned using aquaculture knowledge, the large language model is used to understand the context information, and a collaborative analysis is performed on the weight influencing factors in the scenario. The prediction range obtained by the standard growth model is asymmetrically adjusted to conform to the current scenario's dynamic weight threshold range.

[0019] The valid candidate data uploaded from the edge are compared with the dynamic weight threshold range, and data points that meet the threshold range requirements are retained.

[0020] In this solution, the second stage of cloud data cleaning involves introducing time-series behavior analysis to perform time-aware deduplication of adjacent data, including:

[0021] Maintain a time-series data buffer for each weighing unit. After matching the timestamps with the data filtered in the first stage, insert the data into the time-series data buffer in chronological order.

[0022] The fluctuation value of the sensor output within a preset time range when poultry is standing calmly on the smart weighing sensor is defined as the stable state. The fluctuation value of the output value when poultry walks on or off the smart weighing sensor or moves on the smart weighing sensor is defined as the dynamic state.

[0023] When it is detected that the similarity between the continuous fluctuation of weight data and the stable state within a preset time range meets the preset requirements, it is determined to be a stable state fluctuation. The average value of all data within the preset time range is calculated, and the midpoint of the preset time range is matched with the average value as a timestamp and recorded as a valid weight record.

[0024] If the similarity between the continuous fluctuation of weight data and the dynamic state within a preset time range meets the preset requirements, it is determined to be a dynamic state jitter. The numerical range with the longest duration and the smallest fluctuation within the preset time range is obtained and recorded as a valid weight record, and intermediate data is cleared.

[0025] After time-aware processing, the output is unique and valid weight data in the time dimension, and the deduplicated weight data is obtained.

[0026] In this solution, in the third stage of cloud data cleaning, a density-aware clustering algorithm is introduced to intelligently slice stable regions, including:

[0027] The deduplicated weight data is sorted in ascending order of numerical values. The absolute difference between adjacent data points in the sequence is calculated to generate a difference sequence. The difference threshold is subjected to distribution statistical analysis to obtain the median and dispersion for distribution modeling. Abnormal intervals greater than the preset difference threshold are identified through statistical anomaly retrieval.

[0028] The abnormal interval is used to initially divide the data region. A density-based clustering algorithm is introduced. The neighborhood radius and the minimum number of data points in the cluster are predefined according to the size of the data region and the data range. The dense data point region is filtered according to the neighborhood radius and the minimum number of data points to identify the core cluster.

[0029] The largest and densest cluster in the core cluster is selected as the main cluster. The minimum and maximum data values ​​in the main cluster are defined as the lower and upper limits of the smart slice, respectively. A stable interval is obtained based on the lower and upper limits of the cluster, and a high-quality and effective sample set is constructed based on the data in the stable interval.

[0030] In this scheme, statistical features of the high-quality, effective sample set are calculated, and these statistical features are input into a pre-trained deep belief network. The network outputs an optimal correction strategy for the current data distribution and transforms and calculates the sample data according to this optimal correction strategy, outputting representative weight values, including:

[0031] Calculate the statistical features of the high-quality effective sample set to obtain a multi-dimensional feature vector that comprehensively describes the distribution pattern;

[0032] Policy decision-making is performed using an improved deep belief network. In the training of the deep belief network at the bottom layer, intrinsic features are learned from the distribution of unlabeled weight data. A supervised learning mechanism is introduced at the top layer. The network is trained using an optimal correction policy dataset labeled with aquaculture knowledge to learn the mapping from statistical features to the optimal correction policy.

[0033] Multidimensional feature vectors are imported into a pre-trained deep belief network, an attention mechanism is introduced, different weights are assigned to different features, and the deep nonlinear structure of the deep belief network is activated using weighted features. An understanding vector of the distribution pattern of high-quality effective sample set data is generated in the highest hidden layer, and the probability distribution of strategy selection is obtained using the understanding vector.

[0034] The strategy with the highest screening probability is selected as the optimal correction strategy. Based on the optimal correction strategy, the sample data is transformed and calculated accordingly to finally obtain the representative weight value.

[0035] This solution utilizes historical and current corrected representative weight data to construct a spatiotemporal map of chicken growth trends in different areas of the farm. A spatiotemporal graph convolutional network model is then used to predict the trends in group weight and evenness at a specified future time point, including:

[0036] The farm is divided into several regions, and each region is defined as a graph node. Each node has a feature vector at each time step, which includes the region's representative body weight, evenness, and environmental factors.

[0037] Based on the distance between graph nodes and environmental flow, an edge structure and edge weights are constructed to build a multidimensional spatiotemporal graph of the growth status of chicken flocks in various areas of the farm.

[0038] A gated spatiotemporal convolution module is constructed based on graph convolutional networks and gated one-dimensional convolutions. This module learns and performs spatial correlation and temporal pattern perception in parallel on the multidimensional spatiotemporal graph data. In spatial dependency perception, graph convolutions are used to propagate information on the graph structure of time slices. Multi-hop spatial correlations are captured by stacked graph convolutions. In temporal dynamics capture, one-dimensional causal convolutions with gated mechanisms are used to extract temporal patterns from the historical feature vectors of each node based on gated one-dimensional convolutions.

[0039] By stacking gated spatiotemporal convolutional modules, shallow and deep spatiotemporal features at different scales are obtained. Multi-scale spatiotemporal features are fused, and a multi-task prediction output head is used to simultaneously predict the group weight and group evenness at future time points, thereby obtaining the future trends of weight and evenness in each region.

[0040] This solution, based on predicted trends and combined with pre-set breeding targets and business data, automatically generates breeding decision-making suggestions, including feed adjustment, health management early warning, and slaughter time planning.

[0041] The system obtains the future trends of body weight and uniformity in various areas of the farm, and constructs a decision context by combining the farm's production goals, real-time status and market data. The system then imports a large language model that has been fine-tuned with knowledge of the aquaculture field.

[0042] In the large language model, correlation analysis is used to diagnose changes in predicted weight and predicted uniformity, distinguishing between nutritional, health, and management issues. Based on the difference between predicted and target weight, feed adjustment suggestions are generated, including the types of nutrients to be adjusted, the adjustment range, and the execution time window. Based on the trend of predicted uniformity, health management early warning instructions are generated, including group management, health inspections, and equipment checks. Combining the predicted slaughter weight, market conditions, and cost data, the optimal slaughter time window suggestion and batch slaughter strategy are output.

[0043] Output a set of decision recommendations with natural language explanations, and push the set of decision recommendations directly to the administrators through the user terminal.

[0044] The second aspect of this invention provides a poultry weight data analysis system based on multi-level cleaning and distribution correction. The system includes: a data acquisition and edge preprocessing module, a cloud-based multi-level data cleaning module, an intelligent statistical correction and calculation module, a growth prediction and intelligent decision-making module, and a data service and user interaction module.

[0045] The data acquisition and edge preprocessing module collects the raw weight data of poultry in real time. The edge computing gateway performs preliminary filtering and data compression using a preset dynamic threshold model, and sends the data to the Internet of Things cloud platform through an adaptive transmission protocol.

[0046] The cloud-based multi-level data cleaning module is responsible for performing a three-level progressive cleaning process, including dynamic fine filtering based on large language models and business rules, time-series-aware deduplication of adjacent data, and intelligent slicing of stable intervals to extract the most densely distributed stable intervals of data and form a high-quality and effective sample set.

[0047] The intelligent statistical correction and calculation module calculates the statistical features of the high-quality effective sample set, inputs the statistical features into the pre-trained deep belief network, outputs the optimal correction strategy for the current data distribution, and transforms and calculates the sample data according to the optimal correction strategy to output representative weight values.

[0048] The growth prediction and intelligent decision-making module uses historical and current corrected representative weight data to construct a spatiotemporal map of the growth status of chicken flocks in various areas of the farm. Through a spatiotemporal map convolutional network model, it predicts the trend of flock weight and evenness at a specified future time point. Based on the predicted trend results, combined with preset breeding goals and business data, it automatically generates breeding decision suggestions, including feed adjustment, health management early warning, and slaughter time planning.

[0049] The data service and user interaction module visualizes the data analysis results, predicted trends, and decision-making suggestions, and pushes aquaculture decision-making suggestions to the user terminal through a standardized interface.

[0050] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0051] This invention employs a three-tiered, progressive data cleaning mechanism—edge preprocessing and cloud-based—to effectively identify and eliminate invalid and abnormal data caused by sensor noise, environmental interference, and abnormal poultry behavior, significantly improving the purity and reliability of the original data. An adaptive correction strategy based on deep belief networks is introduced to intelligently select the optimal processing method according to the actual data distribution, overcoming the estimation bias of traditional mean methods under skewed distributions and improving the accuracy and robustness of representative weight calculation. A spatiotemporal graph convolutional network prediction model is constructed to predict trends in population weight and evenness. Furthermore, a large language model decision engine incorporating expert knowledge generates highly targeted suggestions for feed adjustment, health warnings, and slaughter planning, achieving closed-loop management. The collaborative processing architecture between the edge and cloud reduces data transmission pressure and cloud computing load, ensuring real-time performance and stability in large-scale deployment scenarios. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in the embodiments or examples of the present invention, the drawings used in the embodiments or examples will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained according to these drawings without creative effort.

[0053] Figure 1 A flowchart is shown for a poultry weight data analysis method based on multi-stage cleaning and distribution correction;

[0054] Figure 2 A flowchart is shown to obtain the optimal correction strategy for the current data distribution pattern;

[0055] Figure 3 A flowchart is shown to predict the trend of group weight and evenness at a specified future time point;

[0056] Figure 4 A block diagram of a poultry weight data analysis system based on multi-level cleaning and distribution correction is shown. Detailed Implementation

[0057] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0058] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0059] like Figure 1 As shown, this embodiment provides a method for analyzing poultry weight data based on multi-level cleaning and distribution correction, including:

[0060] The raw weight data of poultry is collected in real time by an intelligent weighing sensor array. The edge computing gateway performs preliminary filtering and data compression using a preset dynamic threshold model, and then sends the data to the Internet of Things cloud platform through an adaptive transmission protocol.

[0061] After receiving the data, the cloud platform performs a three-level progressive cleaning process, including dynamic fine filtering based on large language models and business rules, time-series-aware deduplication of adjacent data, and intelligent slicing of stable intervals to extract the most densely distributed stable intervals of data and form a high-quality and effective sample set.

[0062] Calculate the statistical features of the high-quality effective sample set, input the statistical features into a pre-trained deep belief network, output the optimal correction strategy for the current data distribution, and transform and calculate the sample data according to the optimal correction strategy to output a representative weight value.

[0063] Using historical and current corrected representative weight data, a spatiotemporal map of the growth status of chicken flocks in various areas of the farm is constructed. A spatiotemporal map convolutional network model is then used to predict the trend of flock weight and evenness at a specified future time point.

[0064] Based on the predicted trend results, combined with preset breeding goals and business data, the system automatically generates breeding decision-making suggestions, including feed feeding adjustments, health management early warnings, and slaughter time planning.

[0065] It should be noted that the system uses an array of intelligent weighing sensors deployed in the farm to collect the poultry's raw weight data in real time. The sensors continuously collect raw weight signals at a high frequency to ensure that the system can capture the stable moments when the poultry stay on the scale. A dynamic threshold model is pre-configured in the edge computing gateway. This model embeds a standard growth model that matches the poultry breed and age, and a baseline expected weight is obtained based on the standard growth model. A physiological fluctuation coefficient is set based on the daily weight growth rate and physiological fluctuation pattern of poultry at different ages. The younger the age, the larger the relative fluctuation range. Historical data within a preset period is acquired, and the dispersion of the historical data is analyzed to generate a historical stability coefficient. If recent data fluctuations are large, the threshold is automatically widened to avoid false filtering. An adaptive floating range for the baseline expected weight is set based on the physiological fluctuation coefficient and the historical stability coefficient. Context information from other sensors in the farm is acquired, and the threshold of the adaptive floating range is context-compensated based on the correlation between the context information and the poultry weight. For example, after a set feeding time, the upper limit of the threshold is temporarily and appropriately increased. The finally determined adaptive floating range is used to initially filter the original weight data and obtain valid candidate data. Obvious abnormal values ​​caused by transient equipment failures, small foreign objects falling, or extreme poultry distress are filtered out. The valid candidate data is compressed using differential coding or similar lightweight methods. An adaptive transmission protocol is then determined based on network signal strength, and the data is sent to the IoT cloud platform using this protocol. When the network signal is good, data is uploaded in real-time using a small-batch, low-latency mode. When network instability or interruption is detected, the system automatically switches to a caching and batch resumption mode, packaging the data locally and prioritizing the transmission of time-sensitive critical data packets after network recovery.

[0066] It should be noted that in the first stage of cloud data cleaning, the received weight data is dynamically and finely filtered based on a large language model and business rules. Static attributes (poultry breed, current age), dynamic health status (text descriptions obtained from farmer reports or automatic monitoring systems), real-time environmental data (average ambient temperature and humidity, ventilation status), and historical weight data of the current batch of poultry are extracted from the farm's breeding management system. The extracted contextual information is used to construct a structured scenario description. This scenario description is then used as prompt words, input into a large language model trained and finely tuned using breeding knowledge. The large language model is used to understand the contextual information and perform collaborative analysis on weight-influencing factors in the scenario. The prediction range obtained from the standard growth model is asymmetrically adjusted to conform to the current scenario's dynamic weight threshold range. For groups with health risks, abnormally high weight data points may be unreliable, so the upper limit is appropriately lowered; if growth slows, the lower limit is slightly lowered to avoid misjudging individuals who are genuine but growing slowly as abnormal. Valid candidate data uploaded from the edge is compared with the dynamic weight threshold range, and data points that meet the threshold requirements are retained. It achieves filtering that incorporates expert experience, removes obvious noise, and identifies pseudo-normal data under specific physiological conditions.

[0067] In the second stage of cloud data cleaning, time-series behavioral analysis is introduced to distinguish between genuine poultry weight changes and invalid data redundancy caused by sensor jitter. Adjacent time-sensing data is deduplicated to retain valuable information. A time-series data buffer is maintained for each weighing unit. Data filtered in the first stage is matched with timestamps and inserted into the time-series data buffer in chronological order, retaining data within short time windows to ensure that the analysis focuses on recently occurring consecutive events.

[0068] A behavior-sensor coupling model based on the physiological and physical characteristics of poultry is introduced. The fluctuation value of the sensor output within a preset time range when the poultry is standing calmly on the smart weighing sensor is defined as the stable state. The fluctuation value of the output value when the poultry walks onto or off the smart weighing sensor or moves on it is defined as the dynamic state. When the similarity between the continuous fluctuation of weight data and the stable state within the preset time range meets preset requirements, it is determined to be a stable state fluctuation. The average value of all data within the preset time range is calculated, and the midpoint of the preset time range is matched with the average value as a timestamp, recorded as a valid weight record. When the similarity between the continuous fluctuation of weight data and the dynamic state within the preset time range meets preset requirements, it is determined to be a dynamic state fluctuation. The numerical interval with the longest duration and smallest fluctuation within the preset time range is obtained and recorded as a valid weight record, and intermediate data is cleared. For example, a reading changing from 1500 grams to 800 grams and then to 0 grams within 1 second corresponds to the process of the poultry leaving the weighing pan. After time-aware processing, the output is unique and valid weight data in the time dimension. The deduplicated weight data is obtained to eliminate duplication and redundancy caused by device vibration and physical processes.

[0069] In the third stage of cloud data cleaning, a density-aware clustering algorithm is introduced to intelligently slice stable intervals and extract the core data clusters that best represent the true weight of poultry in a resting state. The deduplicated weight data is sorted in ascending order of value, and the absolute difference between adjacent data points in the sequence is calculated to generate a difference sequence. Distribution statistical analysis is performed on the difference threshold to obtain the median and dispersion for distribution modeling. Statistical anomaly retrieval identifies abnormal intervals greater than the preset difference threshold. These abnormal intervals are used to initially divide the data region. A density-based clustering algorithm is then introduced, pre-defining the neighborhood radius and the minimum number of data points per cluster based on the data region size and data range. Dense regions of data points are filtered based on the neighborhood radius and the minimum number of data points. The sorted data is scanned; if a data point's neighboring data points within the neighborhood radius meet the minimum number of data points per cluster, they are considered to belong to the same dense cluster. By expanding the points that connect all densities, core clusters are identified; the largest and densest cluster in the core clusters is selected as the main cluster, and the minimum and maximum data values ​​in the main clusters are defined as the lower and upper limits of the smart slice, respectively. A stable interval is obtained based on the lower and upper limits of the cluster, and a high-quality and effective sample set is constructed based on the data in the stable interval.

[0070] It should be noted that, as Figure 2As shown, the statistical features of the high-quality, effective sample set are calculated, including basic features such as mean, median, and standard deviation, skewness and kurtosis to describe the asymmetry and sharpness of the distribution, and other robust statistics based on percentiles. A multi-dimensional feature vector comprehensively describing the distribution pattern is obtained based on the acquired features. A massive historical poultry weight sample set and a dataset of optimal correction strategies labeled by breeding experts based on the data distribution pattern are collected to calculate multi-dimensional statistical features and construct the feature vector. Features include, but are not limited to: mean, median, standard deviation, skewness, kurtosis, interquartile range, percentiles, etc. The optimal correction strategies labeled by experts are used as labels for supervised learning. An improved deep belief network is used for policy decision-making. Unsupervised pre-training is performed layer by layer in the bottom training of the deep belief network to learn intrinsic features from the unlabeled weight data distribution, obtain initial weight values, and avoid getting trapped in local optima. A final strategy selection output layer is added to the pre-trained network, and a supervised learning mechanism is introduced at the top layer. The network is trained using an optimal correction policy dataset labeled with aquaculture knowledge. The training objective is to minimize the difference between the network's predicted policy and the expert-labeled real policy, learning the mapping from statistical features to the optimal correction policy. An attention mechanism is introduced near the input layer. During training, different importance weights are assigned to different features in the multidimensional feature vector.

[0071] Multidimensional feature vectors are imported into a pre-trained deep belief network. An attention mechanism is introduced to assign different weights to different features. Weighted features are used to activate the deep nonlinear structure of the deep belief network for forward propagation. An understanding vector representing the distribution of high-quality, effective sample data is generated at the highest hidden layer. This understanding vector is used to obtain the probability distribution of policy selection, and the output node corresponds to a set of correction policies. For example: Policy 0: directly uses the arithmetic mean; Policy 1: uses the median; Policy 2: performs a square root transformation, calculates the mean, and then performs inverse correction; Policy 3: performs a logarithmic transformation, calculates the mean, and then performs inverse correction; Policy 4: uses a pruned mean. A probability score is output for each policy, and the policy with the highest probability is selected as the optimal correction policy. Based on the optimal correction policy, corresponding transformations and calculations are performed on the sample data to finally obtain representative weight values.

[0072] It should be noted that, using historical and current corrected representative weight data, a spatiotemporal graph of the growth status of chicken flocks in different areas of the farm is constructed. An improved spatiotemporal graph convolutional network model is then used to predict the trend of flock weight and evenness at a specified future time point. For example... Figure 3As shown, the physical farm is transformed into a dynamic graph structure representing its spatial layout and temporal evolution. The farm is divided into several areas, such as feeding areas and resting areas, and each area is defined as a graph node. Each node has a feature vector at each time step, which includes the representative weight, uniformity, and environmental factors of the area, such as the area's average daily temperature, humidity concentration, and ammonia concentration. Based on the distance between graph nodes and environmental circulation, an edge structure and edge weights are constructed to build a multidimensional spatiotemporal graph of the growth status of chickens in each area of ​​the farm. The environmental circulation is determined by the direction and intensity of the main ventilation paths.

[0073] A gated spatiotemporal convolution module is constructed based on graph convolutional networks and gated one-dimensional convolutions to learn the multidimensional spatiotemporal graph data and perform spatial correlation and temporal pattern perception in parallel. In spatial dependency perception, graph convolutions are used to propagate information on the graph structure of time slices, and each node can aggregate the feature information of its neighboring nodes. Multi-hop spatial correlations are captured by stacked graph convolutions. In temporal dynamic capture, one-dimensional causal convolutions with gating mechanisms are used to extract temporal patterns from the historical feature vectors of each node based on gated one-dimensional convolutions. The gating mechanism can retain important information in the time series while forgetting irrelevant information, and causal convolutions ensure that predictions do not use future information. By stacking gated spatiotemporal convolutional modules, shallow and deep spatiotemporal features at different scales are obtained. Shallow modules capture short-term local effects, such as the impact of yesterday's neighbor's weight on today's weight, while deep modules capture long-term global trends, such as the cumulative effect of environmental changes in the breeding area over the past week on growth. The multi-scale spatiotemporal features are fused, and a multi-task prediction output head is used to simultaneously predict the population weight and population evenness at future time points, obtaining the future trends of weight and evenness in each region. Multi-task synchronous prediction ensures the consistency of the predicted results for weight and evenness indicators.

[0074] It should be noted that, based on the predicted trend results and combined with preset breeding targets and business data, the system automatically generates breeding decision-making suggestions, including feed adjustment, health management early warning, and slaughter time planning, which are executed through an expert knowledge decision engine based on a large language model. The system obtains the future weight and uniformity trends of various areas of the farm, and constructs a decision context by combining the farm's production targets (target slaughter weight and target slaughter age, etc.), real-time status (current feed formula and inventory, poultry house environmental parameters, etc.), and market data (pork price trends and feed raw material costs, etc.). This context is then imported into a large language model fine-tuned with knowledge from the breeding field. Within this large language model, correlation analysis is performed on changes in predicted weight and predicted uniformity for diagnosis, distinguishing between nutritional, health, and management issues. By analyzing the correlation between weight and uniformity trends, the potential root causes of the problems are identified. Based on the difference between predicted and target body weight, and taking into account current feed inventory and raw material market prices, feed adjustment suggestions are generated, including the types of nutrients to be adjusted, the adjustment range, and the execution time window. Based on the predicted evenness trend, health management early warning instructions are generated, including group management, health inspections, and equipment checks. For example, for areas where evenness is predicted to deteriorate significantly, a suggestion is generated to manually group areas B within 48 hours, transferring individuals with weights below the lower limit to recovery pens. Combining predicted slaughter weight, market conditions, and cost data, revenue simulation is performed, outputting optimal slaughter time window suggestions and batch slaughter strategies. A set of decision suggestions with natural language explanations is output and directly pushed to management personnel via user terminals. After confirmation, these suggestions are automatically distributed to execution units such as automated feeding systems and environmental control systems. Feedback information obtained through data collection, cleaning, and prediction is fed to an expert knowledge decision engine based on a large language model, generating a continuous optimization closed loop.

[0075] like Figure 4 As shown, the second embodiment of the present invention provides a poultry weight data analysis system based on multi-level cleaning and distribution correction. The system includes: a data acquisition and edge preprocessing module, a cloud-based multi-level data cleaning module, an intelligent statistical correction and calculation module, a growth prediction and intelligent decision-making module, and a data service and user interaction module.

[0076] The data acquisition and edge preprocessing module includes an intelligent weighing sensor array and an edge computing gateway. The intelligent weighing sensor array collects the raw weight data of poultry in real time. The edge computing gateway performs preliminary filtering and data compression using a preset dynamic threshold model and sends the data to the Internet of Things cloud platform through an adaptive transmission protocol.

[0077] The cloud-based multi-level data cleaning module is responsible for performing a three-level progressive cleaning process, including: an LLM intelligent filtering unit, a time-aware deduplication unit, and a stable interval slicing unit. The LLM intelligent filtering unit receives data reported from the edge, calls a finely tuned large language model, and implements context-aware dynamic fine-grained filtering based on the large language model and business rules. The time-aware deduplication unit analyzes the data that has passed the fine-grained filtering and uses time-awareness to deduplicate adjacent data. The stable interval slicing unit performs in-depth analysis on the deduplicated data, applies a density clustering algorithm, and extracts the most densely distributed stable intervals to form a high-quality and effective sample set.

[0078] The intelligent statistical correction and calculation module includes a distribution feature calculation unit, an improved deep belief network, and a representative weight calculation unit. The distribution feature calculation unit calculates the statistical features of the high-quality and effective sample set. The improved deep belief network uses the statistical features to obtain the optimal correction strategy for the current data distribution. The representative weight calculation unit transforms and calculates the sample data according to the optimal correction strategy and outputs a representative weight value.

[0079] The growth prediction and intelligent decision-making module includes a spatiotemporal graph builder, a spatiotemporal graph convolutional network model, and an LLM expert decision engine. The spatiotemporal graph builder uses historical and currently corrected representative weight data to construct spatiotemporal graph data of the growth status of chicken flocks in various areas of the farm. Through the spatiotemporal graph convolutional network model, it predicts the trend of flock weight and evenness at a specified future time point. Based on the predicted trend results, the LLM expert decision engine automatically generates breeding decision suggestions, including feed adjustment, health management early warning, and slaughter time planning, in combination with preset breeding goals and business data.

[0080] The data service and user interaction module includes a data visualization interface and a decision command interface. The data visualization interface is used to visualize data analysis results, predicted trends and decision suggestions, which greatly reduces the threshold for use. The decision command interface is used to push aquaculture decision suggestions to user terminals through a standardized interface, or directly to the automated execution system, so as to realize the implementation of data-driven decision-making.

[0081] A third embodiment of the present invention provides a computer-readable storage medium, which includes a poultry weight data analysis method program based on multi-level cleaning and distribution correction. When the poultry weight data analysis method program based on multi-level cleaning and distribution correction is executed by a processor, it implements the steps of the poultry weight data analysis method based on multi-level cleaning and distribution correction.

[0082] In the several embodiments provided in this application, it should be understood that the disclosed methods and systems can be implemented in other ways. The system embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, and can be electrical, mechanical, or other forms. Furthermore, in the various embodiments of the present invention, all functional units can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0083] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0084] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for analyzing poultry weight data based on multi-level cleaning and distribution correction, characterized in that, Includes the following steps: The raw weight data of poultry is collected in real time by an intelligent weighing sensor array. The edge computing gateway performs preliminary filtering and data compression using a preset dynamic threshold model, and then sends the data to the Internet of Things cloud platform through an adaptive transmission protocol. After receiving the data, the cloud platform performs a three-level progressive cleaning process, including dynamic fine filtering based on large language models and business rules, time-series-aware deduplication of adjacent data, and intelligent slicing of stable intervals to extract the most densely distributed stable intervals of data and form a high-quality and effective sample set. Calculate the statistical features of the high-quality effective sample set, input the statistical features into a pre-trained deep belief network, output the optimal correction strategy for the current data distribution, and transform and calculate the sample data according to the optimal correction strategy to output a representative weight value. Using historical and current corrected representative weight data, a spatiotemporal map of the growth status of chicken flocks in various areas of the farm is constructed. A spatiotemporal map convolutional network model is then used to predict the trend of flock weight and evenness at a specified future time point. Based on the predicted trend results, combined with preset breeding goals and business data, the system automatically generates breeding decision-making suggestions, including feed feeding adjustments, health management early warnings, and slaughter time planning.

2. The poultry weight data analysis method based on multi-level cleaning and distribution correction according to claim 1, characterized in that, The edge computing gateway performs preliminary filtering and data compression on the raw weight data using a pre-set dynamic threshold model, and then sends it to the IoT cloud platform via an adaptive transmission protocol, including: The raw weight data of poultry is collected in real time using an intelligent weighing sensor array deployed in the farm. A dynamic threshold model is pre-set in the edge computing gateway. The dynamic threshold model is embedded with a standard growth model that matches the poultry breed and age. The baseline expected weight is obtained according to the standard growth model. A physiological fluctuation coefficient is set based on the daily weight growth rate and physiological fluctuation pattern of poultry at different ages, and historical data within a preset period is obtained. The dispersion of the historical data is analyzed to generate a historical stability coefficient. An adaptive fluctuation range of the baseline expected weight is set according to the physiological fluctuation coefficient and the historical stability coefficient. Obtain context information from other sensors in the farm, perform context compensation on the threshold of the adaptive floating range based on the correlation between the context information and poultry weight, and use the finally determined adaptive floating range to initially filter the original weight data to obtain valid candidate data; The valid candidate data is lightweight compressed using differential coding, and an adaptive transmission protocol is determined based on the network signal strength. The data is then sent to the IoT cloud platform using the adaptive transmission protocol.

3. The poultry weight data analysis method based on multi-level cleaning and distribution correction according to claim 1, characterized in that, In the first stage of cloud-based data cleaning, the received weight data is dynamically and finely filtered based on a large language model and business rules, including: The static attributes, dynamic health status, real-time environmental data, and historical weight data of the current batch of poultry are extracted from the poultry management system of the farm, and the extracted contextual information is used to construct a structured scenario description. The scenario description is constructed as prompt words, input into a large language model trained and fine-tuned using aquaculture knowledge, the large language model is used to understand the context information, and a collaborative analysis is performed on the weight influencing factors in the scenario. The prediction range obtained by the standard growth model is asymmetrically adjusted to conform to the current scenario's dynamic weight threshold range. The valid candidate data uploaded from the edge are compared with the dynamic weight threshold range, and data points that meet the threshold range requirements are retained.

4. The poultry weight data analysis method based on multi-level cleaning and distribution correction according to claim 1, characterized in that, In the second stage of cloud data cleaning, time-series behavior analysis is introduced to perform time-aware deduplication of adjacent data, including: Maintain a time-series data buffer for each weighing unit. After matching the timestamps with the data filtered in the first stage, insert the data into the time-series data buffer in chronological order. The fluctuation value of the sensor output within a preset time range when poultry is standing calmly on the smart weighing sensor is defined as the stable state. The fluctuation value of the output value when poultry walks on or off the smart weighing sensor or moves on the smart weighing sensor is defined as the dynamic state. When it is detected that the similarity between the continuous fluctuation of weight data and the stable state within a preset time range meets the preset requirements, it is determined to be a stable state fluctuation. The average value of all data within the preset time range is calculated, and the midpoint of the preset time range is matched with the average value as a timestamp and recorded as a valid weight record. If the similarity between the continuous fluctuation of weight data and the dynamic state within a preset time range meets the preset requirements, it is determined to be a dynamic state jitter. The numerical range with the longest duration and the smallest fluctuation within the preset time range is obtained and recorded as a valid weight record, and intermediate data is cleared. After time-aware processing, the output is unique and valid weight data in the time dimension, and the deduplicated weight data is obtained.

5. The poultry weight data analysis method based on multi-level cleaning and distribution correction according to claim 1, characterized in that, In the third stage of cloud data cleaning, density-aware clustering algorithms are introduced for intelligent slicing of stable regions, including: The deduplicated weight data is sorted in ascending order of numerical values. The absolute difference between adjacent data points in the sequence is calculated to generate a difference sequence. The difference threshold is subjected to distribution statistical analysis to obtain the median and dispersion for distribution modeling. Abnormal intervals greater than the preset difference threshold are identified through statistical anomaly retrieval. The abnormal interval is used to initially divide the data region. A density-based clustering algorithm is introduced. The neighborhood radius and the minimum number of data points in the cluster are predefined according to the size of the data region and the data range. The dense data point region is filtered according to the neighborhood radius and the minimum number of data points to identify the core cluster. The largest and densest cluster in the core cluster is selected as the main cluster. The minimum and maximum data values ​​in the main cluster are defined as the lower and upper limits of the smart slice, respectively. A stable interval is obtained based on the lower and upper limits of the cluster, and a high-quality and effective sample set is constructed based on the data in the stable interval.

6. The poultry weight data analysis method based on multi-level cleaning and distribution correction according to claim 1, characterized in that, Calculate the statistical features of the high-quality, effective sample set, input the statistical features into a pre-trained deep belief network, output the optimal correction strategy for the current data distribution, and transform and calculate the sample data according to the optimal correction strategy to output representative weight values, including: Calculate the statistical features of the high-quality effective sample set to obtain a multi-dimensional feature vector that comprehensively describes the distribution pattern; Policy decision-making is performed using an improved deep belief network. In the training of the deep belief network at the bottom layer, intrinsic features are learned from the distribution of unlabeled weight data. A supervised learning mechanism is introduced at the top layer. The network is trained using an optimal correction policy dataset labeled with aquaculture knowledge to learn the mapping from statistical features to the optimal correction policy. Multidimensional feature vectors are imported into a pre-trained deep belief network, an attention mechanism is introduced, different weights are assigned to different features, and the deep nonlinear structure of the deep belief network is activated using weighted features. An understanding vector of the distribution pattern of high-quality effective sample set data is generated in the highest hidden layer, and the probability distribution of strategy selection is obtained using the understanding vector. The strategy with the highest screening probability is selected as the optimal correction strategy. Based on the optimal correction strategy, the sample data is transformed and calculated accordingly to finally obtain the representative weight value.

7. The poultry weight data analysis method based on multi-level cleaning and distribution correction according to claim 1, characterized in that, Using historical and current corrected representative weight data, a spatiotemporal graph of chicken growth trends in different areas of the farm is constructed. A spatiotemporal graph convolutional network model is then used to predict the trends in flock weight and evenness at a specified future time point, including: The farm is divided into several regions, and each region is defined as a graph node. Each node has a feature vector at each time step, which includes the region's representative body weight, evenness, and environmental factors. Based on the distance between graph nodes and environmental flow, an edge structure and edge weights are constructed to build a multidimensional spatiotemporal graph of the growth status of chicken flocks in various areas of the farm. A gated spatiotemporal convolution module is constructed based on graph convolutional networks and gated one-dimensional convolutions. This module learns and performs spatial correlation and temporal pattern perception in parallel on the multidimensional spatiotemporal graph data. In spatial dependency perception, graph convolutions are used to propagate information on the graph structure of time slices. Multi-hop spatial correlations are captured by stacked graph convolutions. In temporal dynamics capture, one-dimensional causal convolutions with gated mechanisms are used to extract temporal patterns from the historical feature vectors of each node based on gated one-dimensional convolutions. By stacking gated spatiotemporal convolutional modules, shallow and deep spatiotemporal features at different scales are obtained. Multi-scale spatiotemporal features are fused, and a multi-task prediction output head is used to simultaneously predict the group weight and group evenness at future time points, thereby obtaining the future trends of weight and evenness in each region.

8. The poultry weight data analysis method based on multi-level cleaning and distribution correction according to claim 1, characterized in that, Based on the predicted trend results, combined with preset breeding goals and business data, the system automatically generates breeding decision-making suggestions, including feed adjustment, health management early warning, and slaughter time planning, among others. The system obtains the future trends of body weight and uniformity in various areas of the farm, and constructs a decision context by combining the farm's production goals, real-time status and market data. The system then imports a large language model that has been fine-tuned with knowledge of the aquaculture field. In the large language model, correlation analysis is used to diagnose changes in predicted weight and predicted uniformity, distinguishing between nutritional, health, and management issues. Based on the difference between predicted and target weight, feed adjustment suggestions are generated, including the types of nutrients to be adjusted, the adjustment range, and the execution time window. Based on the trend of predicted uniformity, health management early warning instructions are generated, including group management, health inspections, and equipment checks. Combining the predicted slaughter weight, market conditions, and cost data, the optimal slaughter time window suggestion and batch slaughter strategy are output. Output a set of decision recommendations with natural language explanations, and push the set of decision recommendations directly to the administrators through the user terminal.

9. A poultry weight data analysis system based on multi-level cleaning and distribution correction, characterized in that, To implement the poultry weight data analysis method based on multi-level cleaning and distribution correction as described in any one of claims 1-8, the system includes: a data acquisition and edge preprocessing module, a cloud-based multi-level data cleaning module, an intelligent statistical correction and calculation module, a growth prediction and intelligent decision-making module, and a data service and user interaction module. The data acquisition and edge preprocessing module collects the raw weight data of poultry in real time. The edge computing gateway performs preliminary filtering and data compression using a preset dynamic threshold model, and sends the data to the Internet of Things cloud platform through an adaptive transmission protocol. The cloud-based multi-level data cleaning module is responsible for performing a three-level progressive cleaning process, including dynamic fine filtering based on large language models and business rules, time-series-aware deduplication of adjacent data, and intelligent slicing of stable intervals to extract the most densely distributed stable intervals of data and form a high-quality and effective sample set. The intelligent statistical correction and calculation module calculates the statistical features of the high-quality effective sample set, inputs the statistical features into the pre-trained deep belief network, outputs the optimal correction strategy for the current data distribution, and transforms and calculates the sample data according to the optimal correction strategy to output representative weight values. The growth prediction and intelligent decision-making module uses historical and current corrected representative weight data to construct a spatiotemporal map of the growth status of chicken flocks in various areas of the farm. Through a spatiotemporal map convolutional network model, it predicts the trend of flock weight and evenness at a specified future time point. Based on the predicted trend results, combined with preset breeding goals and business data, it automatically generates breeding decision suggestions, including feed adjustment, health management early warning, and slaughter time planning. The data service and user interaction module visualizes the data analysis results, predicted trends, and decision-making suggestions, and pushes aquaculture decision-making suggestions to the user terminal through a standardized interface.

Images