Livestock and poultry health state intelligent evaluation method based on multi-sensor fusion and AI prediction model

Abstract

The invention discloses a livestock and poultry health state intelligent evaluation method based on multi-sensor fusion and an AI prediction model, and belongs to the technical field of livestock breeding, and the method comprises the following steps: setting multiple types of sensors in a livestock and poultry breeding area, collecting multi-source data such as temperature, humidity, carbon dioxide concentration, ammonia gas concentration, illumination intensity, weight and current, and carrying out the analysis of the data; wherein the weight data is obtained through the weighing platform, and the current data is used for sensing the operation state of drinking, ingestion, ventilation or lighting equipment; various types of sensors such as temperature and humidity sensors, carbon dioxide sensors, ammonia gas sensors, illumination sensors, weight sensors and current sensors are deployed in livestock and poultry breeding areas, all-directional data acquisition of environments and behavior states is realized, and dynamic modeling and feature extraction of key behavior nodes are realized by introducing a deep neural network structure of a gating circulation unit and an attention mechanism.

Description

Technical Field

[0001] This invention relates to the field of livestock breeding technology, and more specifically, to an intelligent assessment method for the health status of livestock and poultry based on multi-sensor fusion and AI prediction models. Background Technology

[0002] With the continuous acceleration of the intelligentization process in animal husbandry, livestock and poultry health management is gradually shifting from the traditional extensive model that relies on manual experience to a refined management approach based on information technology and intelligent means.

[0003] Especially driven by the integrated development of sensor technology, wireless communication, artificial intelligence and other technologies, intelligent aquaculture systems have developed rapidly, and the capabilities of multi-source environmental data collection, real-time monitoring of growth indicators and automatic recognition of behavior patterns have been gradually realized.

[0004] However, existing technologies still have technical bottlenecks in areas such as sensor data synchronization, feature representation integrity, accuracy of abnormal behavior identification, and the timeliness of health trend modeling, which restrict the real-time performance and reliability of health status assessment in aquaculture scenarios.

[0005] While some mainstream intelligent aquaculture systems currently possess basic environmental perception and behavior classification capabilities, they are often limited to static processing of single data sources and lack the ability to deeply model the temporal correlations between multi-source data. Specifically:

[0006] Existing systems mostly rely on preset thresholds or simplified logical models for health assessment, which makes it difficult to adapt to the complexity of behavioral differences and state evolution among individual livestock and poultry, and fails to establish an effective feedback mechanism for model adaptive optimization. At the same time, most methods ignore factors such as temporal anomalies, inconsistent sampling intervals, and unstable label reliability in the collected data, resulting in large fluctuations in health assessment results and a high false alarm rate.

[0007] Furthermore, in the areas of multimodal feature extraction and behavioral anomaly recognition, there is still a lack of fusion methods that can combine complex model structures such as wavelet transform, attention mechanism, and gated neural network.

[0008] In particular, there is currently no mature system for regulating the credibility of health labels and correcting prediction models across time dimensions;

[0009] Therefore, there is an urgent need for an intelligent assessment method for the health status of livestock and poultry that can integrate multi-source heterogeneous data and has self-learning and feedback optimization capabilities. Summary of the Invention

[0010] To achieve the above objectives, the present invention provides the following technical solution:

[0011] A method for intelligent assessment of livestock and poultry health status based on multi-sensor fusion and AI prediction models includes the following steps:

[0012] Multiple types of sensors are installed in the livestock and poultry breeding area to collect multi-source data such as temperature, humidity, carbon dioxide concentration, ammonia concentration, light intensity, weight and current. Among them, weight data is obtained through a weighing platform, and current data is used to sense the operating status of drinking water, feeding, ventilation or lighting equipment.

[0013] The multi-source data is subjected to time-stamp-based synchronous fusion processing, and a sliding window resampling and interpolation completion algorithm is used to construct a unified multi-dimensional time series dataset.

[0014] A behavior recognition model is constructed based on the weight and environmental parameters to identify abnormal behavior states and establish their response relationship with environmental factors.

[0015] The dataset and recognition results are input into the health modeling module to construct a time-evolving digital twin model of health and output a multi-dimensional health status assessment index.

[0016] A causal reasoning model was constructed by combining gas concentration trends with equipment operating current to assess the risk of respiratory diseases;

[0017] The health index and causal model are input into the AI ​​model for comprehensive analysis, outputting a health status prediction result, and triggering an alarm command when the warning conditions are met.

[0018] Furthermore, the sensors include: an integrated temperature and humidity sensor, an infrared photoelectric carbon dioxide sensor, an electrochemical ammonia sensor, an illuminance sensor, a ground-embedded weighing platform sensor, and a current detection sensor installed in the drinking water device, feeding device, and ventilation motor circuit.

[0019] Furthermore, the synchronous fusion processing of the multi-source data includes:

[0020] A time sliding window with multiple parameters is constructed based on the sampling period of various sensors, data fluctuation amplitude, and historical stability scores.

[0021] For data with different sampling frequencies, a multi-level weighted interpolation algorithm that includes a trend matching factor is used to complete missing data.

[0022] An anomaly identification mechanism combining fluctuation consistency coefficient and time domain offset tolerance threshold is introduced to filter out sampled outliers and form a multi-channel fusion feature matrix under a unified time axis.

[0023] The fusion feature matrix serves as the feature input structure for subsequent health assessment models.

[0024] Furthermore, the behavior recognition model further includes a combined network structure consisting of gated recurrent units and attention mechanisms in a deep neural network, which is used to dynamically model the input time-series sensor fusion feature matrix, and extract important variable sequences related to behavioral anomalies by allocating attention weights at key time points, so as to enhance the model's ability to identify the suddenness of abnormal behavior and behavioral state transition points.

[0025] The behavior recognition model also includes a multi-objective learning module, which is used to simultaneously output the health status classification label, confidence score and behavioral intention feature vector of individual livestock and poultry;

[0026] The health status classification includes three labels: normal, sub-healthy, and abnormal; the behavioral intentions include semantic vectors of behaviors such as drinking water, eating, moving, and remaining still.

[0027] The time series feature matrix also constructs a multimodal time-frequency feature spectrum of livestock and poultry through wavelet transform, and performs adaptive anomaly identification and dynamic confidence screening on the abnormal interval of wavelet spectrum peak based on variational Bayesian inference mechanism, and constructs an abnormal behavior discrimination matrix based on attention focus;

[0028] The model output results establish a feedback loop relationship with the control module on the device side;

[0029] When the judgment result continues to meet the abnormal threshold range or the health score continues to deteriorate, the lower-level control module is output with behavior restriction or warning instructions to control the equipment operation status, including ventilation rate adjustment, water volume restriction or feeding suspension.

[0030] The feedback mechanism also includes a dynamic update mechanism for the training set data, which inputs the abnormal behavior identification results and health status labels back into the GRU network for online incremental learning and parameter optimization, thereby enabling the model to self-correct and adaptively adjust.

[0031] Furthermore, the health status identification results are incorporated into an asynchronous annotation and calibration mechanism during the model training phase. This calibration mechanism includes:

[0032] A cluster-based behavior distribution deviation estimation model is constructed to estimate the label offset between multiple batches of sensor feature sequences and manual annotation results. In combination with the sample behavior evolution trend under different time windows, a soft label correction matrix is ​​constructed by using fuzzy C-means clustering and cross-label entropy evaluation.

[0033] Based on sample conditioning, a soft label loss function is further constructed by temperature scaling and KL divergence control mechanism, and combined with a multi-layer residual correction module to perform error compression and label correction on sample features.

[0034] The above correction results are used to update the main parameters of the recognition model. The parameter optimization strategy includes a parameter freezing mechanism based on error threshold and a learning rate dynamic adjustment mechanism based on gradient fluctuation detection.

[0035] The soft labels and parameter adjustment results are fed back to the sample management module to update the sample labeling credibility index and participation weight, and to form a linkage update mechanism for samples, labels and parameters in subsequent training.

[0036] Furthermore, it includes a confidence-adaptive adjustment mechanism based on the model output results, the adjustment mechanism comprising:

[0037] An estimation model based on the confidence offset between historical sample labels is constructed to conduct trial evaluation of the offset between multiple batches of sensor feature sequences and manual labeling results, and then the sample confidence weight is adjusted accordingly.

[0038] The offset is constructed by setting a confidence threshold and combining it with a Bayesian update mechanism, a sliding mean window, or a temperature scaling method to build a dynamic confidence interval.

[0039] Based on the sample confidence matrix and combined with the distribution trend of behavioral characteristics under different time windows, a label dynamic adjustment mapping table is constructed.

[0040] The control results are fed back to the health status level adjustment control factor, which is used to link the early warning level switching and tag update rhythm matching strategy of the control system.

[0041] The label adjustment mechanism supports dynamic adjustment of sample feature input weights based on label offset trends during subsequent training.

[0042] Furthermore, the method further includes a label credibility control mechanism, which includes:

[0043] A dynamic adjustment model based on the label offset between historically identified labels and manually labeled labels is constructed to estimate the label confidence range;

[0044] A sliding time window residual statistics mechanism is introduced to dynamically calculate the deviation trend between health status identification labels and manual annotations, which is used to correct the label credibility score, and a residual trend intervention factor is constructed in the credibility score results.

[0045] The setting of the confidence interval also includes: determining the confidence shift strength and adjusting the sample confidence weight based on the KL divergence calculation results between the historical label prediction probability distribution and the current sample label distribution;

[0046] The residual between the identified label and the original label is used as a dynamic feedback parameter during the training phase to construct a reverse repair path for the credibility of the health label, and generate a residual trend matrix for labeling the signal intervention module.

[0047] The credibility weights are adjusted based on the residual trend feedback results, and the label loss function weight coefficients of the model during the training phase are dynamically corrected to complete the coupling adjustment between the credibility enhancement parameters and the core label parameters.

[0048] Furthermore, when the label confidence score exceeds the set upper limit, a label enhancement mechanism is triggered to increase the sample weight and label response coefficient of this type of label during the training process; when the label confidence score is lower than the set lower limit, an abnormal label removal mechanism is activated to prevent this type of label from disturbing the healthy model training process.

[0049] The credibility control mechanism further constructs a label time distribution offset map based on the sample residual change trend under different time periods, and jointly models it with the behavior recognition probability matrix to achieve dynamic temporal calibration of label credibility across time domains.

[0050] Furthermore, the method further includes a model credibility enhancement mechanism for supporting the predictive model in performing adaptive repair, the mechanism comprising:

[0051] Based on historical evaluation accuracy and sample residual distribution within the sliding window, a dynamically adjusted weighting factor is constructed to identify samples with high offset risk.

[0052] When the confidence level of continuous predictions is lower than a preset threshold, the model self-learning fine-tuning mechanism is triggered, and a training fine-tuning subset is constructed based on the sample distribution characteristics and residual trends. Small-step gradient correction is performed in combination with weighted transfer regularization terms.

[0053] A joint penalty factor for label loss and confidence deviation rate is constructed to enhance the feedback pathway between prediction error and label offset;

[0054] A temperature softening control module is introduced to dynamically adjust the temperature coefficient when abnormal prediction disturbances are detected, so as to improve the model's adaptability to abnormal distributions.

[0055] It supports multi-scale label confidence self-calibration strategies to achieve multi-objective confidence balance among different label categories.

[0056] In summary, the present invention has the following beneficial effects:

[0057] By deploying various types of sensors, such as temperature, humidity, carbon dioxide, ammonia, illuminance, weight, and current, in livestock and poultry breeding areas, comprehensive data collection of environmental and behavioral states is achieved. Furthermore, by incorporating a deep neural network structure with gated recurrent units and attention mechanisms, dynamic modeling and feature extraction of key behavioral nodes are realized. Combined with wavelet transform and Bayesian inference methods, potential weak abnormal signals in behavioral states can be focused and identified, improving the sensitivity of abnormal behavior identification and the ability to control confidence intervals.

[0058] By constructing a causal reasoning model that integrates environmental factors and behavioral indicators, and establishing a health digital twin and label feedback mechanism, it is possible not only to achieve continuous assessment and risk warning of livestock and poultry health status, but also to have the ability to regulate label credibility and fine-tune the model through self-learning, effectively solving the problems of data drift and label uncertainty in existing systems. Attached Figure Description

[0059] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art 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 based on these drawings without creative effort.

[0060] Figure 1 This is a diagram of the credibility control system of the present invention. Detailed Implementation

[0061] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] Example:

[0063] The following is in conjunction with the appendix Figure 1 The present invention will be described in further detail below.

[0064] Please see Figure 1 This invention provides a technical solution: an intelligent assessment method for livestock and poultry health status based on multi-sensor fusion and AI prediction models, such as... Figure 1 As shown, it includes the following steps:

[0065] Multiple types of sensors are installed in the livestock and poultry breeding area to collect multi-source data such as temperature, humidity, carbon dioxide concentration, ammonia concentration, light intensity, weight and current. Among them, weight data is obtained through a weighing platform, and current data is used to sense the operating status of drinking water, feeding, ventilation or lighting equipment.

[0066] Multi-source data are synchronously fused based on timestamps, and a sliding window resampling and interpolation completion algorithm is used to construct a unified multi-dimensional time series dataset.

[0067] A behavior recognition model is built based on weight and environmental parameters to identify abnormal behavior states and establish their response relationship with environmental factors.

[0068] Input the dataset and recognition results into the health modeling module to construct a time-evolving digital twin model of health and output a multi-dimensional health status assessment index;

[0069] A causal reasoning model was constructed by combining gas concentration trends with equipment operating current to assess the risk of respiratory diseases;

[0070] The health index and causal model are input into the AI ​​model for comprehensive analysis, outputting health status prediction results, and triggering alarm commands when the warning conditions are met.

[0071] In this embodiment: In order to model the causal relationship between the equipment operating status and the current change trend, the causal inference model adopts the Granger causal analysis method based on time series. By constructing a linear regression structure between the lagged terms and the predictive terms of multidimensional variables, the influence of each perceived variable, such as gas concentration, temperature, humidity, etc., on the predicted current of the equipment is judged, thereby forming a causal spectrum.

[0072] Furthermore, for nonlinear scenarios, a gated graph neural network can be introduced to enhance the representation of topological dependencies between multiple sensors and improve the causal prediction capability under complex conditions.

[0073] The health digital twin model constructs a three-dimensional state map of multi-sensor states, including state nodes, time edges, and behavioral label edges. It takes historical health assessment values ​​as input and dynamically updates the model by combining the output of the prediction model. In the model structure, the twin state at each time step receives new sensing data through a sliding window mechanism, uses a time-recursive state transition function for state evolution, and triggers label correction when a significant shift is detected. The system as a whole forms a closed-loop update structure, achieving simultaneous enhancement of state fitting and prediction capabilities.

[0074] like Figure 1As shown, the sensors include: an integrated temperature and humidity sensor, which includes an infrared photoelectric carbon dioxide sensor, an electrochemical ammonia sensor, an illuminance sensor, a ground-embedded weighing platform sensor, and a current detection sensor installed in the drinking water device, feeding device, and ventilation motor circuit.

[0075] In this embodiment, the weighing platform uses a ground-embedded load sensing module with high sensitivity, a measurement accuracy of ±0.1kg, and a high response frequency of ≥1Hz. This module is deployed at key passage locations to acquire real-time data on the time livestock spend in the water source area and changes in their unit weight. The drinking device is equipped with a flow meter and a switch sensor, with a sampling cycle set to 10 seconds per cycle. This ensures accurate capture of the livestock's drinking frequency and water volume fluctuations, thus correlating with health assessment indicators such as dehydration risk and water intake trends. The sampling cycle settings for all sensors have been verified through data analysis to match the target indicator requirements, ensuring the real-time nature and effectiveness of the monitoring.

[0076] like Figure 1 As shown, the synchronous fusion processing of multi-source data includes:

[0077] A time sliding window with multiple parameters is constructed based on the sampling period of various sensors, data fluctuation amplitude, and historical stability scores.

[0078] For data with different sampling frequencies, a multi-level weighted interpolation algorithm that includes a trend matching factor is used to complete missing data.

[0079] An anomaly identification mechanism combining fluctuation consistency coefficient and time domain offset tolerance threshold is introduced to filter out sampled outliers and form a multi-channel fusion feature matrix under a unified time axis.

[0080] The fused feature matrix serves as the feature input structure for subsequent health assessment models.

[0081] The sliding window features from different channels or sensor sources are merged in chronological order to construct a two-dimensional feature matrix, with rows corresponding to each time window and columns corresponding to the fused feature quantities under each channel.

[0082] In this embodiment, the sliding window is a dynamically adjustable window, and its length is adaptively adjusted according to the data fluctuation amplitude and stability evaluation index. Specifically, the window length is initially set to N, such as 10 seconds. The window size is dynamically adjusted according to the rate of change of the ratio of the standard deviation σ(t) to the historical mean μ in each time period, Δσ / μ, in order to capture sudden changes in non-stationary regions. The sliding method adopts a forward-moving scrolling strategy with a step size of 1 second. The interpolation algorithm prioritizes linear interpolation, and spline interpolation, such as cubic B splines, is used for missing values ​​with large spans to preserve data continuity and trend characteristics. The interpolation strategy is embedded in the sliding window update mechanism and executed synchronously.

[0083] The multi-channel fusion feature matrix is ​​a two-dimensional tensor structure. Rows represent different time segments, divided by a sliding window, and columns represent the fusion feature set extracted from different data sources, such as temperature, humidity, gravity sensing, and behavior recognition. Each column can be a one-dimensional vector, such as the average temperature or maximum fluctuation, or a multi-dimensional representation after feature engineering, such as FFT spectral features or coefficient of variation. This feature matrix is ​​arranged in time sequence to form a sample set, which is used for subsequent health risk assessment model training or identification model input.

[0084] like Figure 1 As shown, the behavior recognition model further includes a combined network structure consisting of gated recurrent units and attention mechanisms in a deep neural network, which is used to dynamically model the input time-series sensor fusion feature matrix, and extract important variable sequences related to behavioral anomalies through attention weight allocation at key time points, so as to enhance the model's ability to identify the suddenness of abnormal behavior and behavioral state transition points.

[0085] The behavior recognition model also includes a multi-objective learning module, which is used to simultaneously output the health status classification label, confidence score, and behavioral intention feature vector of individual livestock and poultry.

[0086] The health status classification includes three labels: normal, sub-healthy, and abnormal; the behavioral intentions include semantic vectors of behaviors such as drinking water, eating, moving, and remaining still.

[0087] The time series feature matrix also constructs a multimodal time-frequency feature spectrum of livestock and poultry through wavelet transform, and performs adaptive anomaly identification and dynamic confidence screening on the peak anomaly interval of the wavelet spectrum based on the variational Bayesian inference mechanism, and constructs an abnormal behavior discrimination matrix based on attention focus;

[0088] The model output results establish a feedback loop with the control module on the device side;

[0089] When the judgment result continues to meet the abnormal threshold range or the health score continues to deteriorate, the lower-level control module is output with behavior restriction or warning instructions to control the equipment operation status, including ventilation rate adjustment, water volume restriction or feeding suspension.

[0090] The feedback process also includes a dynamic update mechanism for the training set data, which inputs the abnormal behavior identification results and health status labels back into the GRU network for online incremental learning and parameter optimization, thereby enabling the model to self-correct and adaptively adjust.

[0091] In this embodiment, the gated deep neural network preferably adopts a gated recurrent unit (GRU) structure, which is suitable for processing time-series multi-source sensor data. The GRU network includes: an input layer that receives sensor fusion feature vectors with an input dimension of d (e.g., d=32); one to two GRU hidden layers, each containing 64 hidden units for learning temporal dependencies; and an output layer connected to a health status label classifier or regression unit. Each GRU unit includes update and reset gates, and its network structure is expressed as follows:

[0092] ;

[0093] ;

[0094] ;

[0095] ;

[0096] In the training phase, the GRU model receives labeled feature vectors as supervision signals and outputs feature vectors used to predict health risk levels or behavioral patterns.

[0097] The label feature vector is a three-dimensional vector matrix formed by structured encoding of behavior type or health status category. Construction methods include:

[0098] One-dimensional label: Behavioral category or health status type, such as drinking water, standing, walking, abnormal heart rate, with a total of L categories;

[0099] Two-dimensional tags: Tag encoding methods can include one-hot encoding, integer index encoding, or word embedding, such as Word2Vec or TransformerEmbedding;

[0100] 3D Labels: Time-series frame-level annotations, i.e., the continuous change matrix of labels at different times, used for dynamic model training; for example, when using one-hot vectors, each label... After embedding, the tag vector can be obtained through linear mapping. ,like:

[0101]

[0102] If a three-dimensional feature matrix is ​​used, it can be expressed as:

[0103]

[0104] in For the length of time, For tag categories, The embedding dimension is used for alignment supervision in temporal GRU models.

[0105] like Figure 1 As shown, the health status identification results introduce an asynchronous annotation and calibration mechanism during the model training phase. The calibration mechanism includes:

[0106] A cluster-based behavior distribution deviation estimation model is constructed to estimate the label offset between multiple batches of sensor feature sequences and manual annotation results. In combination with the sample behavior evolution trend under different time windows, a soft label correction matrix is ​​constructed by using fuzzy C-means clustering and cross-label entropy evaluation.

[0107] Based on sample conditioning, a soft label loss function is further constructed by temperature scaling and KL divergence control mechanism, and combined with a multi-layer residual correction module to perform error compression and label correction on sample features.

[0108] The above correction results are used to update the main parameters of the recognition model. The parameter optimization strategies include a parameter freezing mechanism based on error threshold and a learning rate dynamic adjustment mechanism based on gradient fluctuation detection.

[0109] The soft labels and parameter adjustment results are fed back to the sample management module to update the sample labeling confidence index and participation weights, and form a linkage update mechanism for samples, labels and parameters in subsequent training;

[0110] In this embodiment, the offset factor is used to measure the statistical deviation between the current sample characteristics and the historical health status distribution, and is defined as follows:

[0111]

[0112] in, Represents the mean of the characteristics of historical samples. This represents the mean of the current sample features. Indicates the first Mean shift of the class samples. Shift factor. Used to dynamically adjust sample confidence and weights, prioritizing the training of representative and stable samples;

[0113] Introducing sample weighting factors According to the degree of offset The contribution of samples to model training is dynamically adjusted using the following function:

[0114]

[0115] in, This is an adjustment coefficient used to control the degree of offset suppression; As a sample offset factor, the above weighting factors are used to weight the sample loss function during the training phase, correcting the impact of abnormal samples in the training set on model fitting.

[0116] The sample adjustment mechanism includes:

[0117] Sample bias evaluation module: Calculates the bias of the feature mean and the historical statistical distribution for all training samples;

[0118] Weight calculation module: Calculates weights using offset factors to form a sample weighting matrix;

[0119] Weight feedback module: Periodically updates the weights based on the model training results to achieve dynamic learning adjustment;

[0120] Abnormal sample suppression mechanism: When the offset factor exceeds the set threshold, the sample is marked as abnormal and its participation in model update is removed;

[0121] The offset factor is the Euclidean distance between the statistical characteristics of the current sample and the distribution of historical samples;

[0122] The sample weights are dynamically generated based on the offset factor using an exponential function and are used for weighting the loss function during training.

[0123] The control path includes: sample offset evaluation, weight generation, training feedback update, and abnormal sample removal mechanism.

[0124] like Figure 1 As shown, it includes a confidence-adaptive adjustment mechanism based on the model output results. The adjustment mechanism includes:

[0125] An estimation model based on the confidence offset between historical sample labels is constructed to conduct trial evaluation of the offset between multiple batches of sensor feature sequences and manual labeling results, and then the sample confidence weight is adjusted accordingly.

[0126] The offset is constructed by setting a confidence threshold and combining it with a Bayesian update mechanism, a sliding mean window, or a temperature scaling method to create a dynamic confidence interval.

[0127] Based on the sample confidence matrix and combined with the distribution trend of behavioral characteristics under different time windows, a label dynamic adjustment mapping table is constructed.

[0128] The control results are fed back to the health status level adjustment control factor, which is used to match the early warning level switching and label update rhythm of the linkage control system.

[0129] The label adjustment mechanism supports dynamic adjustment of sample feature input weights based on label offset trends during subsequent training;

[0130] In this embodiment, behavioral feature trends are extracted using a sliding window convolution method, employing the following calculation approach:

[0131] ;

[0132] in:

[0133] For behavioral channels In time eigenvalues,

[0134] For trend filtering kernels, such as Gaussian kernels or first-order difference kernels;

[0135] This outputs the trend corresponding to the behavioral channel;

[0136] Trend features are used to detect the direction of behavioral pattern evolution and are combined with the confidence matrix to construct a dynamic prediction matrix;

[0137] The model input data structure includes: behavioral temporal feature tensor ,in For the sample size, For time step, For sensor feature dimensions; trend tensor Record the trend components of each behavior; confidence matrix This reflects the credibility of the health model's output;

[0138] The input weighting path includes a sample confidence feedback module and a parameter correction mechanism. When the prediction confidence of a given input sample falls below a threshold... At that time, the input module triggers the weight suppression mechanism: where ;

[0139] ;

[0140] This weighting factor is used for channel weighting of the model input, forming an adaptive input path adjustment mechanism for the health status recognition model;

[0141] In the above text, the confidence matrix is ​​used to record the confidence probability of the model output for various behavioral features; behavioral trend features are generated by sliding window convolution and used to dynamically predict the trend direction; the weights of the input path are dynamically adjusted by the feedback of the confidence matrix, forming an input adjustment and control mechanism in the model training process.

[0142] like Figure 1 As shown, the method further includes a label credibility control mechanism, which includes:

[0143] A dynamic adjustment model based on the label offset between historically identified labels and manually labeled labels is constructed to estimate the label confidence range;

[0144] A sliding time window residual statistics mechanism is introduced to dynamically calculate the deviation trend between health status identification labels and manual annotations, which is used to correct the label credibility score, and a residual trend intervention factor is constructed in the credibility score results.

[0145] The setting of the confidence interval also includes: determining the confidence shift strength and adjusting the sample confidence weight based on the KL divergence calculation results between the historical label prediction probability distribution and the current sample label distribution;

[0146] The residual between the identified label and the original label is used as a dynamic feedback parameter during the training phase to construct a reverse repair path for the credibility of the health label, and generate a residual trend matrix for labeling the signal intervention module.

[0147] The credibility weights are adjusted based on the residual trend feedback results, and the label loss function weight coefficients of the model during the training phase are dynamically corrected to complete the coupling adjustment between the credibility enhancement parameters and the core label parameters.

[0148] Furthermore, when the label confidence score exceeds the set upper limit, a label enhancement mechanism is triggered to increase the sample weight and label response coefficient of this type of label during the training process; when the label confidence score is lower than the set lower limit, an abnormal label removal mechanism is activated to prevent this type of label from disturbing the healthy model training process.

[0149] The credibility control mechanism further constructs a label time distribution offset map based on the sample residual change trend under different time periods, and jointly models it with the behavior recognition probability matrix to achieve dynamic temporal calibration of label credibility across time domains;

[0150] The confidence weighting mechanism includes a confidence factor matrix constructed based on historical prediction errors and label distribution offsets, and a confidence adjustment threshold implemented through KL divergence calculation.

[0151] In this embodiment, the credibility weighting mechanism constructs a tag credibility matrix. The prediction reliability of each training sample under different time windows is mapped to a confidence weight factor. The calculation is as follows:

[0152] ;

[0153] in These are the model's predicted values. The actual label value. Indicates a point in time For the sample The degree of authorization;

[0154] The KL divergence calculation is used to measure the probability distribution of the current predicted label. Historical stable label distribution The degree of difference between them is calculated using the following formula:

[0155] ;

[0156] in, The predicted label distribution for the current sample. The KL value represents the historical mean distribution within the sliding time window. This value is used to construct a label credibility adjustment index. If the KL value is greater than the threshold, the current prediction is considered unreliable, and model penalties or weight adjustments are required.

[0157] like Figure 1 As shown, the method further includes a model credibility enhancement mechanism to support the predictive model in performing adaptive repairs, the mechanism including:

[0158] Based on historical evaluation accuracy and sample residual distribution within the sliding window, a dynamically adjusted weighting factor is constructed to identify samples with high offset risk.

[0159] When the confidence level of continuous predictions is lower than a preset threshold, the model self-learning fine-tuning mechanism is triggered, and a training fine-tuning subset is constructed based on the sample distribution characteristics and residual trends. Small-step gradient correction is performed in combination with weighted transfer regularization terms.

[0160] A joint penalty factor for label loss and confidence deviation rate is constructed to enhance the feedback pathway between prediction error and label offset;

[0161] A temperature softening control module is introduced to dynamically adjust the temperature coefficient when abnormal prediction disturbances are detected, so as to improve the model's adaptability to abnormal distributions.

[0162] It supports multi-scale label confidence self-calibration strategies to achieve multi-objective confidence balance among different label categories;

[0163] The dynamic weighting factors include a gradient adjustment factor constructed based on KL divergence and confidence fluctuation, and an information density penalty rate based on label frequency and sample distribution statistical estimation, which is used to control the influence of sample bias.

[0164] In this embodiment, the dynamic weight factor set refers to the set of perturbation responses of each sample when the label confidence deviates from the historical expected value, and is defined as:

[0165] ;

[0166] in, For adjustment coefficients, This is the current sample confidence value. This is the historical mean; this factor is used to adjust the gradient influence of fine-tuned samples on model weight updates. The weighted transfer regularization term is a penalty function term in transfer learning, used to guide model parameter transfer preferences, and its expression is as follows:

[0167] ;

[0168] in These are feature extraction mappings for the source model and the target model, respectively. Weighting factors are assigned to the samples.

[0169] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0170] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A method for intelligent assessment of livestock and poultry health status based on multi-sensor fusion and AI prediction models, characterized in that, Includes the following steps: Multiple types of sensors are installed in the livestock and poultry breeding area to collect multi-source data such as temperature, humidity, carbon dioxide concentration, ammonia concentration, light intensity, weight and current. Among them, weight data is obtained through a weighing platform, and current data is used to sense the operating status of drinking water, feeding, ventilation or lighting equipment. The multi-source data is subjected to time-stamp-based synchronous fusion processing, and a sliding window resampling and interpolation completion algorithm is used to construct a unified multi-dimensional time series dataset. A behavior recognition model is constructed based on the weight and environmental parameters to identify abnormal behavior states and establish their response relationship with environmental factors. The dataset and recognition results are input into the health modeling module to construct a time-evolving digital twin model of health and output a multi-dimensional health status assessment index. A causal reasoning model was constructed by combining gas concentration trends with equipment operating current to assess the risk of respiratory diseases; The health index and causal model are input into the AI ​​model for comprehensive analysis, outputting a health status prediction result, and triggering an alarm command when the warning conditions are met.

2. The intelligent assessment method for livestock and poultry health status based on multi-sensor fusion and AI prediction model according to claim 1, characterized in that: The sensors include: an integrated temperature and humidity sensor, an infrared photoelectric carbon dioxide sensor, an electrochemical ammonia sensor, an illuminance sensor, a ground-embedded weighing platform sensor, and a current detection sensor installed in the drinking water device, feeding device, and ventilation motor circuit.

3. The intelligent assessment method for livestock and poultry health status based on multi-sensor fusion and AI prediction model according to claim 2, characterized in that, The synchronous fusion processing of the multi-source data includes: A time sliding window with multiple parameters is constructed based on the sampling period of various sensors, data fluctuation amplitude, and historical stability scores. For data with different sampling frequencies, a multi-level weighted interpolation algorithm that includes a trend matching factor is used to complete missing data. An anomaly identification mechanism combining fluctuation consistency coefficient and time domain offset tolerance threshold is introduced to filter out sampled outliers and form a multi-channel fusion feature matrix under a unified time axis. The fusion feature matrix serves as the feature input structure for subsequent health assessment models.

4. The intelligent assessment method for livestock and poultry health status based on multi-sensor fusion and AI prediction model according to claim 3, characterized in that: The behavior recognition model further includes a combined network structure consisting of gated recurrent units and attention mechanisms in a deep neural network, used to dynamically model the input time-series sensor fusion feature matrix and extract important variable sequences related to behavioral anomalies. The behavior recognition model also includes a multi-objective learning module, which is used to simultaneously output the health status classification label, confidence score and behavioral intention feature vector of individual livestock and poultry; The health status classification includes three labels: normal, sub-healthy, and abnormal; the behavioral intentions include semantic vectors of behaviors such as drinking water, eating, moving, and remaining still. The time series feature matrix also constructs a multimodal time-frequency feature spectrum of livestock and poultry through wavelet transform, and performs adaptive anomaly identification and dynamic confidence screening on the abnormal interval of wavelet spectrum peak based on variational Bayesian inference mechanism, and constructs an abnormal behavior discrimination matrix based on attention focus; The model output results establish a feedback loop relationship with the control module on the device side; The feedback mechanism also includes a dynamic update mechanism for the training set data, which inputs the abnormal behavior identification results and health status labels back into the GRU network for online incremental learning and parameter optimization.

5. The intelligent assessment method for livestock and poultry health status based on multi-sensor fusion and AI prediction model according to claim 4, characterized in that: The health status identification results are used to introduce an asynchronous annotation and calibration mechanism during the model training phase. This calibration mechanism includes: A cluster-based behavior distribution deviation estimation model is constructed to estimate the label offset between multiple batches of sensor feature sequences and manual annotation results. In combination with the sample behavior evolution trend under different time windows, a soft label correction matrix is ​​constructed by using fuzzy C-means clustering and cross-label entropy evaluation. Based on sample conditioning, a soft label loss function is constructed by introducing a temperature scaling and KL divergence control mechanism, and combined with a multi-layer residual correction module to perform error compression and label correction on sample features. The correction results are used to update the main parameters of the recognition model. The parameter optimization strategy includes a parameter freezing mechanism based on an error threshold and a learning rate dynamic adjustment mechanism based on gradient fluctuation detection. The soft labels and parameter adjustment results are fed back to the sample management module to update the sample labeling credibility index and participation weight, and to form a linkage update mechanism for samples, labels and parameters in subsequent training.

6. The intelligent assessment method for livestock and poultry health status based on multi-sensor fusion and AI prediction model according to claim 5, characterized in that: This includes an adaptive confidence adjustment mechanism based on model output results, the adjustment mechanism comprising: An estimation model based on the confidence offset between historical sample labels is constructed to conduct trial evaluation of the offset between multiple batches of sensor feature sequences and manual labeling results, and then the sample confidence weight is adjusted accordingly. The offset is constructed by setting a confidence threshold and combining it with a Bayesian update mechanism, a sliding mean window, or a temperature scaling method to build a dynamic confidence interval. Based on the sample confidence matrix and combined with the distribution trend of behavioral characteristics under different time windows, a label dynamic adjustment mapping table is constructed. The control results are fed back to the health status level adjustment control factor, which is used to link the early warning level switching and tag update rhythm matching strategy of the control system. The label adjustment mechanism supports dynamic adjustment of sample feature input weights based on label offset trends during subsequent training.

7. The intelligent assessment method for livestock and poultry health status based on multi-sensor fusion and AI prediction model according to claim 6, characterized in that: The method further includes a label credibility control mechanism, which includes: A dynamic adjustment model based on the label offset between historically identified labels and manually labeled labels is constructed to estimate the label confidence range; A sliding time window residual statistics mechanism is introduced to dynamically calculate the deviation trend between health status identification labels and manual annotations, which is used to correct the label credibility score, and a residual trend intervention factor is constructed in the credibility score results. The setting of the confidence interval also includes: determining the confidence shift strength and adjusting the sample confidence weight based on the KL divergence calculation results between the historical label prediction probability distribution and the current sample label distribution; The residual between the identified label and the original label is used as a dynamic feedback parameter during the training phase to construct a reverse repair path for the credibility of the health label, and generate a residual trend matrix for labeling the signal intervention module. The credibility weights are adjusted based on the residual trend feedback results, and the label loss function weight coefficients of the model during the training phase are dynamically corrected to complete the coupling adjustment between the credibility enhancement parameters and the core label parameters.

8. The intelligent assessment method for livestock and poultry health status based on multi-sensor fusion and AI prediction model according to claim 7, characterized in that: The method further includes a model credibility enhancement mechanism to support the predictive model in performing adaptive repair, the mechanism comprising: Based on historical evaluation accuracy and sample residual distribution within the sliding window, a dynamically adjusted weighting factor is constructed to identify samples with high offset risk. When the confidence level of continuous predictions is lower than a preset threshold, the model self-learning fine-tuning mechanism is triggered, and a training fine-tuning subset is constructed based on the sample distribution characteristics and residual trends. Small-step gradient correction is performed in combination with weighted transfer regularization terms. A joint penalty factor for label loss and confidence deviation rate is constructed to enhance the feedback pathway between prediction error and label offset; A temperature softening control module is introduced to dynamically adjust the temperature coefficient when abnormal prediction disturbances are detected, so as to improve the model's adaptability to abnormal distributions. It supports multi-scale label confidence self-calibration strategies to achieve multi-objective confidence balance among different label categories.

Images