Agriculture and animal husbandry internet of things state monitoring method based on multi-source perception fusion

Abstract

The application discloses a kind of based on multi-source perception fusion's agriculture and animal husbandry internet of things state monitoring method, comprising: acquisition multi-source perception data, generate standardization multi-source time series data;Identify the operation stage of agriculture and animal husbandry internet of things, build multi-source perception benchmark distribution library;Build distribution consistency evaluation module, obtain the distribution deviation vector of each perception source;Generation each perception source corresponding confidence weight, obtain fusion state feature data;Based on improved generalized additive model, output the comprehensive quantitative index of agriculture and animal husbandry internet of things;The score of environment, equipment and animal state is handled and analyzed, and the state monitoring output of agriculture and animal husbandry internet of things is generated.The application realizes the reliable fusion of agriculture and animal husbandry internet of things multi-class perception data and the accurate monitoring of operation state by combining distribution consistency evaluation module and improved generalized additive model.

Description

Technical Field

[0001] This invention relates to the field of agricultural and animal husbandry Internet of Things (IoT) technology, and in particular to an agricultural and animal husbandry IoT status monitoring method based on multi-source sensing fusion. Background Technology

[0002] With the continuous improvement of the scale, intensification, and intelligence of agricultural production, agricultural IoT technology is widely used in environmental monitoring, production management, and equipment operation and maintenance. Traditional agricultural IoT systems typically use various sensors to continuously collect information such as temperature, humidity, soil parameters, animal physiological behavior, and equipment operating status to achieve real-time perception and remote monitoring of the production process. However, due to the complex and variable agricultural production environment, different sensing sources vary significantly in sampling frequency, data distribution characteristics, and reliability. The collected multi-source sensing data often exhibits strong non-stationarity, high noise interference, and insufficient consistency, posing significant challenges to subsequent status assessment and decision analysis.

[0003] In existing technologies, the assessment of the operational status of agricultural and livestock IoT typically employs multi-source sensing data fusion and statistical modeling methods to comprehensively quantify the status of monitored objects such as the environment, animals, and equipment. Through weighted fusion, threshold discrimination, or status scoring methods based on regression models, existing technologies have achieved, to some extent, continuous monitoring and status perception of agricultural and livestock production processes, and can assist in anomaly identification and trend analysis. However, due to significant differences in sampling frequency, data distribution characteristics, and reliability among different sensing sources in agricultural and livestock IoT, existing methods often rely on fixed weights or uniform modeling assumptions, making it difficult to effectively characterize the differences in distribution consistency among multi-source data. When some sensing source data is affected by environmental interference or abnormal shifts, it can significantly impact the overall status assessment results, leading to increased volatility in status scores, high false alarm rates, or insufficient stability of monitoring results.

[0004] Therefore, how to provide a method for agricultural and livestock IoT status monitoring based on multi-source sensing fusion is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a method for monitoring the state of agricultural and livestock IoT based on multi-source sensing fusion. This invention fully utilizes multi-source sensing data fusion, distribution consistency assessment, and statistical modeling techniques. Focusing on various sensing sources such as agricultural environment and crops, livestock animals, and equipment operation, it details the overall implementation process from multi-source time-series data acquisition, operational stage identification, distribution consistency measurement, to reliability-weighted fusion and state quantification assessment. By introducing a distribution consistency assessment method based on Hellinger distance and an improved generalized additive model, the reliability and contribution of data from different sensing sources are structurally adjusted, achieving a comprehensive quantitative representation and continuous monitoring of the operational state of agricultural and livestock IoT. This method effectively improves the stability and robustness of multi-source sensing data fusion, enhances the accuracy and interpretability of state assessment results, and possesses advantages such as adaptability to complex environmental changes, reduction of abnormal interference, and improved monitoring reliability. It is suitable for long-term operational status monitoring and intelligent management scenarios in agricultural and livestock IoT.

[0006] A method for agricultural and pastoral IoT status monitoring based on multi-source sensing fusion according to an embodiment of the present invention includes:

[0007] Collect multi-source sensing data from different sensing sources in agricultural and animal husbandry Internet of Things, preprocess the multi-source sensing data, and generate standardized multi-source time series data;

[0008] Based on standardized multi-source time-series data, the operation stages of agricultural and animal husbandry Internet of Things are identified, corresponding stage labels are assigned to each state window, and a multi-source sensing benchmark distribution library with one-to-one correspondence with the stage labels is constructed.

[0009] A distribution consistency evaluation module based on Hellinger distance is constructed. The current distribution representation is constructed for the sensing data corresponding to each sensing source within the current state window. The Hellinger distance of each sensing source is calculated by combining the multi-source sensing benchmark distribution library to obtain the distribution deviation vector of each sensing source under the current state window.

[0010] The confidence weights of each sensing source are generated based on the distribution deviation vector, and the multi-source sensing data are weighted and fused according to the confidence weights to obtain the fused state feature data.

[0011] Based on the improved generalized additive model, the fused state feature data is modeled, and the Hellinger distance is introduced to adjust the contribution terms of the corresponding sensing sources in the improved generalized additive model, and the comprehensive quantitative index of agricultural and animal husbandry Internet of Things is output.

[0012] Based on comprehensive quantitative indicators, the environmental, equipment, and animal status scores are normalized, anomaly detected, and trend analyzed to generate status monitoring outputs for agricultural and livestock IoT.

[0013] Optionally, the multi-source sensing data specifically includes agricultural environment and crop data, livestock data, and equipment operation data.

[0014] 1. The method for monitoring the status of agricultural and pastoral IoT based on multi-source sensing fusion according to claim 1, wherein the preprocessing of multi-source sensing data specifically includes data cleaning, data standardization, data synchronization and data feature extraction.

[0015] Optionally, the construction of a multi-source sensing benchmark distribution library corresponding one-to-one with stage labels includes:

[0016] Standardized multi-source time-series data is segmented according to a preset time length to form several continuous state windows, and the corresponding multi-source sensing data set is collected in each state window.

[0017] For the multi-source sensing data set within each status window, stage discrimination features reflecting the overall change level, fluctuation characteristics and distribution pattern of the data are extracted. Based on the stage discrimination features, the agricultural and animal husbandry Internet of Things operation stage to which the status window belongs is identified, and a stage label uniquely corresponding to the status window is generated.

[0018] After obtaining the stage labels, the standardized multi-source time series data corresponding to each sensing source in each state window are statistically analyzed to construct the current distribution description result that characterizes the data distribution characteristics of the sensing source in the current state window.

[0019] The current distribution description results of different state windows are classified and collected according to the stage labels. The distribution description results of corresponding sensing sources in state windows with the same stage labels are summarized and processed to form the baseline distribution description results of each sensing source in the corresponding stage.

[0020] Using stage labels and sensing source types as indexes, the baseline distribution descriptions of each sensing source corresponding to each stage are uniformly stored to construct a multi-source sensing baseline distribution library, and a one-to-one correspondence is established between the status window and the corresponding stage label and baseline distribution.

[0021] Optionally, obtaining the distribution deviation vector of each sensing source in the current state window includes:

[0022] A distribution consistency assessment module based on Hellinger distance is constructed. The distribution consistency assessment module consists of a distribution representation construction unit, a stage benchmark distribution retrieval unit, a distribution consistency measurement unit, and a multi-source deviation representation unit.

[0023] The distribution representation construction unit performs statistical modeling on the standardized multi-source time-series data corresponding to each sensing source within the current state window, forming a window distribution description result that represents the data distribution characteristics of each sensing source under the state window;

[0024] The stage baseline distribution retrieval unit retrieves the stage baseline distribution description results corresponding to each sensing source type from the multi-source sensing baseline distribution library based on the stage label corresponding to the current state window, and forms a stage baseline distribution set that matches the window distribution description results.

[0025] The distribution consistency measurement unit performs a distribution consistency measurement based on Hellinger distance on the window distribution description results of each sensing source and the corresponding stage baseline distribution description results, and introduces an uncertainty estimation dimension to generate uncertainty estimation results that reflect the confidence level of the consistency measurement.

[0026] The multi-source deviation characterization unit jointly organizes the distribution deviation measurement results of each sensing source with the corresponding uncertainty estimation results, constructs a distribution consistency matrix while maintaining the one-to-one correspondence between sensing sources, and forms a structured distribution deviation vector based on the distribution consistency matrix.

[0027] Optionally, obtaining the fused state feature data specifically involves:

[0028] Based on the distribution deviation vector, the degree of distribution deviation of each sensing source is mapped to obtain a reliability score that reflects the reliability of the data from each sensing source. The reliability score is inversely related to the degree of distribution deviation of the corresponding sensing source.

[0029] The credibility scores corresponding to each sensing source are normalized to generate credibility weights for each sensing source. The credibility weights of each sensing source are within a uniform scale range and satisfy the relative proportional relationship between the credibility weights.

[0030] Based on the credibility weight, the multi-source sensing data corresponding to each sensing source within the current status window are weighted and fused to form fused status feature data that characterizes the overall operating status of the current agricultural and animal husbandry Internet of Things.

[0031] Optionally, the comprehensive quantitative indicators for outputting agricultural and livestock IoT include:

[0032] The fused state feature data is input into the improved generalized additive model. A perception source category index structure is introduced. The fused state feature data is indexed and organized based on the perception source category to form feature subsets corresponding to agricultural environment and crop data, livestock animal data and equipment operation data, respectively. A one-to-one feature input channel is established for each feature subset.

[0033] Construct a sub-additive structure that connects to the main additive predictor in parallel. The feature subsets in the feature input channel are connected to multiple sub-additive structures respectively. Each sub-additive structure independently constructs the corresponding additive smoothing prediction sub-item for the connected feature subset, and obtains the sub-prediction output of each sub-additive structure. The sub-prediction outputs are then merged into the main additive smoothing prediction sub-structure in a modular parallel manner to form the main additive prediction output.

[0034] By introducing the confidence weight of the sensing source into the improved generalized additive model, and establishing a correspondence between the confidence weight of the sensing source and the sub-prediction output and internal smoothing function term of each sub-additive structure through the weight interface structure and the adjustment coefficient interface, the contribution intensity of each sensing source category in the main additive prediction output is structurally adjusted to obtain the contribution-adjusted main additive prediction output.

[0035] Obtain the distribution deviation vector, introduce a consistency penalty interface and a smoothing adjustment channel, introduce the distribution deviation vector into the improved generalized additive model, establish a one-to-one correspondence between the distribution deviation vector and the contribution items of each perceptual source category in the main additive prediction output, and apply consistency constraint adjustment to the contribution-adjusted main additive prediction output to obtain the consistency-adjusted main additive prediction output.

[0036] A link function mapping table is established based on the link function structure. The main additive prediction output after consistency adjustment is matched with the link function mapping table. The main additive prediction output after consistency adjustment is mapped by the link function to form a comprehensive quantitative index of agricultural Internet of Things.

[0037] Optionally, the generation of agricultural IoT status monitoring output includes:

[0038] Based on comprehensive quantitative indicators, the monitoring objects in the current status window are divided into agricultural environment and crop status, livestock status and equipment operation status, and status score data sets corresponding to the above monitoring objects are formed respectively.

[0039] Based on the status score dataset, the status scores of agricultural environment and crops, livestock animals and equipment operation status are normalized respectively. All types of status scores are within a unified scale range, and normalized status score results corresponding one-to-one with the current status window are generated.

[0040] Based on the normalized state score results, state anomaly identification is performed, and state score records that exceed the anomaly threshold range are marked to form anomaly detection outputs for environmental anomalies, animal anomalies, and equipment anomalies.

[0041] The normalized state score results are aggregated according to the state window sequence to form the state score time series of the corresponding monitoring objects. The state score time series is then processed by trend analysis to generate trend analysis output reflecting the changes in agricultural environment and crop status, livestock status and equipment operation status over time.

[0042] Based on the anomaly detection output and trend analysis output, a state monitoring output for agricultural and animal husbandry Internet of Things is generated. The state monitoring output includes normalized state score results, anomaly detection results, and trend analysis results.

[0043] The beneficial effects of this invention are:

[0044] This invention proposes a state monitoring method for agricultural and pastoral IoT based on multi-source sensing fusion. It comprehensively utilizes multi-source time-series data processing, distribution consistency assessment, credibility-weighted fusion, and statistical modeling to uniformly model and assess the state of various sensing source data, including agricultural environment and crop, livestock, and equipment operation. The method standardizes and divides the multi-source sensing data into state windows, constructs a phased multi-source sensing baseline distribution library by identifying operational stages, and introduces a distribution consistency assessment method based on Hellinger distance to quantitatively characterize the degree of distribution deviation between the current monitoring data and the baseline state. Furthermore, it generates credibility weights reflecting the reliability of each sensing source data.

[0045] This invention utilizes credibility weights to weight and fuse multi-source sensing data, then inputs the fused state features into an improved generalized additive model. By structurally adjusting the contribution terms of each sensing source and consistency constraints, it achieves comprehensive quantitative modeling and continuous monitoring of the operational status of agricultural and livestock IoT. This method effectively mitigates the impact of large differences in the distribution and inconsistent reliability of multi-source sensing data on the state assessment results, improves the stability, accuracy, and interpretability of state monitoring results, and enhances the system's adaptability to environmental changes and abnormal disturbances. It is suitable for long-term operational status monitoring and intelligent management in complex agricultural and livestock IoT scenarios. Attached Figure Description

[0046] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0047] Figure 1 This is a schematic diagram of the distributed consistency evaluation module of an agricultural and animal husbandry Internet of Things state monitoring method based on multi-source sensing fusion proposed in this invention.

[0048] Figure 2 This is a schematic diagram of an agricultural and livestock Internet of Things (IoT) status monitoring method based on multi-source sensing fusion proposed in this invention.

[0049] Figure 3 This is a flowchart of the improved generalized additive model of the agricultural and animal husbandry Internet of Things state monitoring method based on multi-source sensing fusion proposed in this invention. Detailed Implementation

[0050] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0051] refer to Figure 1 , Figure 2 and Figure 3 A method for agricultural and livestock IoT status monitoring based on multi-source sensing fusion, comprising:

[0052] Collect multi-source sensing data from different sensing sources in agricultural and animal husbandry Internet of Things, preprocess the multi-source sensing data, and generate standardized multi-source time series data;

[0053] Based on standardized multi-source time-series data, the operation stages of agricultural and animal husbandry Internet of Things are identified, corresponding stage labels are assigned to each state window, and a multi-source sensing benchmark distribution library with one-to-one correspondence with the stage labels is constructed.

[0054] A distribution consistency evaluation module based on Hellinger distance is constructed. The current distribution representation is constructed for the sensing data corresponding to each sensing source within the current state window. The Hellinger distance of each sensing source is calculated by combining the multi-source sensing benchmark distribution library to obtain the distribution deviation vector of each sensing source under the current state window.

[0055] The confidence weights of each sensing source are generated based on the distribution deviation vector, and the multi-source sensing data are weighted and fused according to the confidence weights to obtain the fused state feature data.

[0056] Based on the improved generalized additive model, the fused state feature data is modeled, and the Hellinger distance is introduced to adjust the contribution terms of the corresponding sensing sources in the improved generalized additive model, and the comprehensive quantitative index of agricultural and animal husbandry Internet of Things is output.

[0057] Based on comprehensive quantitative indicators, the environmental, equipment, and animal status scores are normalized, anomaly detected, and trend analyzed to generate status monitoring outputs for agricultural and livestock IoT.

[0058] In this embodiment, the multi-source sensing data specifically includes agricultural environment and crop data, livestock data, and equipment operation data.

[0059] In this embodiment, the preprocessing of multi-source sensing data specifically includes data cleaning, data standardization, data synchronization, and data feature extraction.

[0060] In this embodiment, the construction of a multi-source sensing benchmark distribution library that corresponds one-to-one with the stage labels includes:

[0061] Standardized multi-source time-series data is segmented according to a preset time length to form several consecutive state windows. Within each state window, a corresponding set of multi-source sensing data is collected.

[0062] The preset time length refers to the window duration of the status window, which specifies the continuous time range covered by each status window;

[0063] For the multi-source sensing data set within each status window, stage discrimination features reflecting the overall level of data change, fluctuation characteristics, and distribution pattern are extracted. Based on these stage discrimination features, the agricultural IoT operation stage to which the status window belongs is identified, generating a stage label uniquely corresponding to the status window.

[0064] The stage discrimination feature is a set of features obtained by statistically analyzing the data from each sensing source within the state window. It includes statistics representing the overall level of change of the data, dispersion representing the fluctuation characteristics, and distribution description representing the distribution pattern. The statistics representing the overall level of change include the mean and median within the window. The dispersion representing the fluctuation characteristics includes the standard deviation and the range. The distribution description representing the distribution pattern includes skewness, kurtosis, or quantile description results.

[0065] After obtaining the stage labels, the standardized multi-source time series data corresponding to each sensing source in each state window are statistically analyzed to construct the current distribution description result that characterizes the data distribution characteristics of the sensing source in the current state window.

[0066] The current distribution description results of different state windows are classified and collected according to the stage labels. The distribution description results of corresponding sensing sources in state windows with the same stage labels are summarized and processed to form the baseline distribution description results of each sensing source in the corresponding stage.

[0067] Using stage labels and sensing source types as indexes, the baseline distribution descriptions of each sensing source corresponding to each stage are uniformly stored to construct a multi-source sensing baseline distribution library, and a one-to-one correspondence is established between the status window and the corresponding stage label and baseline distribution.

[0068] In this embodiment, obtaining the distribution deviation vector of each sensing source under the current state window includes:

[0069] A distribution consistency assessment module based on Hellinger distance is constructed. This module comprises a distribution representation construction unit, a stage baseline distribution retrieval unit, a distribution consistency measurement unit, and a multi-source deviation representation unit. Specifically, the construction of the distribution consistency assessment module based on Hellinger distance is as follows:

[0070] The distribution characterization construction unit, the stage benchmark distribution retrieval unit, the distribution consistency measurement unit, and the multi-source deviation characterization unit are sequentially connected to form a distribution consistency evaluation module based on Hellinger distance;

[0071] The distribution representation construction unit performs statistical modeling on the standardized multi-source time-series data corresponding to each sensing source within the current state window, forming a window distribution description result that characterizes the data distribution features of each sensing source under the state window. The statistical modeling process specifically involves:

[0072] For each sensing source, the range of values ​​for standardized multi-source time-series data within the current state window is determined, and the range of values ​​is divided into several distribution intervals. The number of data falling into each distribution interval within the state window is counted, and the proportion of data corresponding to each distribution interval is calculated to form a window distribution description result of the sensing source under the current state window.

[0073] The stage baseline distribution retrieval unit retrieves the stage baseline distribution description results corresponding to each sensing source type from the multi-source sensing baseline distribution library based on the stage label corresponding to the current state window, and forms a stage baseline distribution set that matches the window distribution description results.

[0074] The distribution consistency measurement unit performs a Hellinger distance-based distribution consistency measurement on the window distribution description results of each sensing source and the corresponding stage baseline distribution description results. It also introduces an uncertainty estimation dimension to generate uncertainty estimation results that reflect the confidence level of the consistency measurement, where:

[0075] The distribution consistency measure based on Hellinger distance is performed as follows:

[0076] Under a unified set of distribution intervals, the window distribution description results of each sensing source are aligned with the corresponding stage baseline distribution description results. The window distribution proportions and baseline distribution proportions corresponding to each distribution interval are subjected to square root transformation. The difference in proportions after square root transformation is squared interval by interval and accumulated. The accumulated result is then normalized to the square root to obtain the Hellinger distance value of the sensing source under the current state window, which serves as the distribution deviation measurement result.

[0077] The uncertainty estimation dimension refers to the reliability characterization dimension set in parallel with the distribution consistency measurement results. It represents the stability and credibility of the distribution consistency measurement results under the current state window. The uncertainty estimation dimension corresponds one-to-one with the perception source type and serves as an additional output dimension for the distribution consistency assessment.

[0078] The generation of uncertainty estimation results that reflect the confidence level of the consistency measure specifically includes:

[0079] An uncertainty description quantity is formed by weighting the effective data volume of the sensing source within the current state window, the coverage of the distribution interval, and the dispersion of the interval proportion distribution. A one-to-one correspondence is established between the uncertainty description quantity and the distribution deviation measurement result of the sensing source, which serves as the uncertainty estimation result.

[0080] The multi-source deviation characterization unit jointly organizes the distribution deviation measurement results and corresponding uncertainty estimation results of each sensing source. While maintaining the one-to-one correspondence between sensing sources, it constructs a distribution consistency matrix and forms a structured distribution deviation vector based on the distribution consistency matrix, where:

[0081] The construction of the distribution consistency matrix is ​​as follows:

[0082] Using the sensing source as the index dimension of the distribution consistency matrix, the distribution deviation measurement results and corresponding uncertainty estimation results obtained by each sensing source under the current state window are paired and organized. According to the sensing source type, the distribution deviation measurement results and uncertainty estimation results corresponding to each sensing source are combined to form the elements of the distribution consistency matrix. Each row of the distribution consistency matrix corresponds to a sensing source, and a unified distribution consistency matrix representing the distribution consistency state of multiple sources is constructed.

[0083] The structured distribution deviation vector formed based on the distribution consistency matrix is ​​specifically as follows:

[0084] Extract and recombine the matrix elements corresponding to the same sensing source in the distribution consistency matrix, and arrange the distribution deviation measurement results and uncertainty estimation results corresponding to each sensing source in order to form a distribution deviation vector that maintains the one-to-one correspondence between sensing sources.

[0085] In this embodiment, obtaining the fusion state feature data specifically refers to:

[0086] Based on the distribution deviation vector, the degree of distribution deviation of each sensing source is mapped to obtain a reliability score reflecting the reliability of the data from each sensing source. The reliability score is inversely related to the degree of distribution deviation of the corresponding sensing source. The mapping process for the degree of distribution deviation of each sensing source specifically involves:

[0087] According to the preset monotonic mapping rule, the distribution deviation measurement result is converted into a confidence score. Perception sources with small distribution deviations correspond to high confidence scores, and perception sources with large distribution deviations correspond to low confidence scores, forming a confidence score set that corresponds one-to-one with each perception source. The monotonic mapping rule is a mapping rule that maintains the inverse relationship between the distribution deviation measurement result and the confidence score, and the amplitude of the mapping result is limited so that the confidence score is within the range of values.

[0088] The credibility scores corresponding to each sensing source are normalized to generate credibility weights for each sensing source. The credibility weights of each sensing source are within a uniform scale range and satisfy the relative proportional relationship between the credibility weights, where:

[0089] The relative proportion between the credibility weights refers to the proportional relationship between the credibility weights of each perception source and the credibility score. Perception sources with high credibility scores correspond to credibility weights greater than the weight threshold, while perception sources with low credibility scores correspond to credibility weights less than the threshold.

[0090] Based on the credibility weight, the multi-source sensing data corresponding to each sensing source within the current status window are weighted and fused to form fused status feature data that characterizes the overall operating status of the current agricultural and animal husbandry Internet of Things.

[0091] In this embodiment, the comprehensive quantitative indicators for outputting agricultural and livestock IoT include:

[0092] The fused state feature data is input into an improved generalized additive model. A sensing source category index structure is introduced, and the fused state feature data is indexed and organized based on the sensing source category to form feature subsets corresponding to agricultural environment and crop data, livestock animal data, and equipment operation data, respectively. A one-to-one feature input channel is established for each feature subset, wherein:

[0093] The sensing source category index structure refers to the indexed organizational structure that identifies and distinguishes different sensing source categories. The indexed organizational structure uses the sensing source category as the index key to establish a correspondence between the fused state feature data and the sensing source category to which it belongs. The feature data corresponding to agricultural environment and crop data, livestock data and equipment operation data are independently identified and categorized according to their categories.

[0094] The establishment of a one-to-one corresponding feature input channel for each feature subset is specifically as follows:

[0095] Based on the perception source category index structure, the fusion state feature data is classified and collected to form feature subsets corresponding to each perception source category. Each feature subset is assigned an independent data input path and identification parameters. Each feature input channel only carries the data input of the corresponding feature subset, and a one-to-one correspondence is maintained between the feature input channel and the feature subset.

[0096] A sub-additive structure is constructed to access the main additive predictor in parallel. A subset of features from the feature input channel is accessed into multiple sub-additive structures. Each sub-additive structure independently constructs a corresponding additive smoothing prediction term for the accessed feature subset, resulting in the sub-prediction output of each sub-additive structure. These sub-prediction outputs are then merged into the main additive smoothing prediction sub-structure in a modular parallel manner to form the main additive prediction output.

[0097] The sub-additive structure for constructing parallel access to the main additive predictor is specifically as follows:

[0098] Based on the perception source category index structure, corresponding sub-additive structure modules are configured for agricultural environment and crop data, livestock animal data and equipment operation data respectively. An input interface for receiving feature input channel data is set in each sub-additive structure module, and an output interface for outputting sub-prediction results to the main additive smoothing prediction substructure is set. Multiple sub-additive structure modules are connected to the main additive smoothing prediction substructure in parallel.

[0099] The construction of the corresponding additive smoothing prediction sub-item is as follows:

[0100] In each sub-additive structure, for the accessed feature subset, each feature is mapped to a corresponding smoothing function term, and the smoothing function terms are aggregated in an additive combination manner to form an additive smoothing prediction sub-term of the sub-additive structure. The additive smoothing prediction sub-term outputs the sub-prediction result of the sub-additive structure on the comprehensive quantitative index.

[0101] The formation of the principal additive prediction output specifically involves:

[0102] The sub-prediction results output by each sub-additive structure are aggregated in parallel and then connected to the main additive smoothing prediction substructure. In the main additive smoothing prediction substructure, the sub-prediction results are additively combined to form the main additive prediction output that represents the contribution of multi-sensory source fusion.

[0103] By introducing the source confidence weights into the improved generalized additive model, and establishing a correspondence between the source confidence weights and the sub-prediction outputs and internal smoothing function terms of each sub-additive structure through a weight interface structure and an adjustment coefficient interface, the contribution intensity of each source category in the main additive prediction output is structurally adjusted to obtain the contribution-adjusted main additive prediction output, where:

[0104] The credibility weight of the sensing source refers to the weight parameter that characterizes the reliability of the data of each sensing source under the current state window. The weight parameter corresponds one-to-one with the sensing source and reflects the relative contribution ratio that different sensing sources should occupy in the subsequent fusion modeling process.

[0105] The internal smoothing function term refers to the set of smoothing terms corresponding to the subset of features accessed in the sub-additive structure. The set of smoothing terms represents the nonlinear contribution relationship of each input feature to the output of the improved generalized additive model and serves as a component of the additive smoothing prediction sub-terms constructed by the sub-additive structure.

[0106] The obtained contribution-adjusted principal additive prediction output is specifically as follows:

[0107] The confidence weights of the sensing sources are applied to the sub-prediction outputs and internal smoothing function terms of each sub-additive structure through the weight interface structure and the adjustment coefficient interface, respectively, to form contribution adjustment parameters that correspond one-to-one with the sensing source categories. Based on the contribution adjustment parameters, the contribution intensity of each sensing source category in the main additive prediction output is adjusted and reorganized to form the contribution-adjusted main additive prediction output.

[0108] The distribution bias vector is obtained, and a consistency penalty interface and a smoothing adjustment channel are introduced. This distribution bias vector is then incorporated into the improved generalized additive model. A one-to-one correspondence is established between the distribution bias vector and the contribution terms of each perceptual source category in the principal additive prediction output. Consistency constraints are applied to the contribution-adjusted principal additive prediction output to obtain the consistency-adjusted principal additive prediction output, where:

[0109] The consistency penalty interface and smoothing adjustment channel refer to the structured input interface and adjustment transmission path set in the improved generalized additive model. They are used to introduce the distribution deviation vector based on Hellinger distance as external consistency constraint information into the improved generalized additive model, and to pass the consistency constraint information to the contribution item and smoothing structure corresponding to the perception source category for consistency constraint access and adjustment mapping.

[0110] Each sensing source category contribution item refers to the contribution component in the main additive prediction output that corresponds one-to-one with the sensing source category index structure. The contribution component is formed by the sub-additive structure output and internal smoothing function term of the corresponding sensing source category, and represents the contribution strength and contribution form of the sensing source category to the main additive prediction output.

[0111] The consistency constraint adjustment of the contribution-adjusted principal additive prediction output is specifically as follows:

[0112] The distribution deviation vector is obtained and accessed through the consistency penalty interface. Based on the perception source category index structure, the distribution deviation information in the distribution deviation vector is established in a one-to-one correspondence with the contribution items of each perception source category in the main additive prediction output. The distribution deviation information is applied to the corresponding perception source category contribution items through the smoothing adjustment channel. The contribution intensity of each perception source category contribution item is adjusted by consistency constraint to form the main additive prediction output after consistency adjustment. In the consistency constraint adjustment, the contribution items corresponding to perception source categories with a distribution deviation degree greater than the threshold are subject to stronger constraints.

[0113] A link function mapping table is established based on the link function structure. The consistency-adjusted main additive prediction output is matched with the link function mapping table. After the consistency-adjusted main additive prediction output is mapped by the link function, a comprehensive quantitative index for agricultural and livestock IoT is formed, where:

[0114] The establishment of the link function mapping table based on the link function structure is specifically as follows:

[0115] Based on the link function structure of the improved generalized additive model, the perception source category is used as the mapping index key, and the link function configuration items corresponding to each perception source category are registered and stored to form a link function mapping table. The link function configuration item refers to the mapping relationship that determines the main additive prediction output to the comprehensive quantitative index output space.

[0116] The formation of the comprehensive quantitative indicators for the agricultural and livestock Internet of Things is specifically as follows:

[0117] The consistent-adjusted main additive prediction output is read, and the link function configuration item corresponding to the main additive prediction output is determined according to the link function mapping table. The consistent-adjusted main additive prediction output is then input into the mapping process corresponding to the determined link function configuration item to obtain the comprehensive quantitative index.

[0118] In this embodiment, generating the status monitoring output of the agricultural and livestock Internet of Things includes:

[0119] Based on comprehensive quantitative indicators, the monitoring objects in the current status window are divided into agricultural environment and crop status, livestock status, and equipment operation status, and status score data sets corresponding to each of the above monitoring objects are formed. Specifically, forming status score data sets corresponding to each of the above monitoring objects involves:

[0120] Based on comprehensive quantitative indicators, the contribution results in the comprehensive quantitative indicators are classified and collected according to the index structure of the sensing source categories. The contribution results corresponding to agricultural environment and crop data are collected to form an agricultural environment and crop status score data set, and the contribution results corresponding to livestock animal data are collected to form a livestock animal status score data set. The contribution results corresponding to equipment operation data are collected to form an equipment operation status score data set.

[0121] Based on the status score dataset, the status scores of agricultural environment and crops, livestock animals and equipment operation status are normalized respectively. All types of status scores are within a unified scale range, and normalized status score results corresponding one-to-one with the current status window are generated.

[0122] Based on the normalized state score results, state anomaly identification is performed, and state score records that exceed the anomaly threshold range are marked to form anomaly detection outputs for environmental anomalies, animal anomalies, and equipment anomalies.

[0123] The normalized state score results are aggregated according to the state window sequence to form a state score time series for the corresponding monitoring object. Trend analysis is then performed on the state score time series to generate a trend analysis output reflecting the changes in agricultural environment and crop status, livestock status, and equipment operating status over time. Specifically, the trend analysis of the state score time series involves:

[0124] The status score time series are continuously arranged according to the time order of the status window. The status score time series are smoothed to obtain a smoothed series. The change direction and change magnitude information of the time series are extracted based on the smoothed series to generate trend discrimination results that represent upward trend, downward trend or stable trend. The trend discrimination results are then linked to the corresponding monitoring objects to form trend analysis output.

[0125] Based on the anomaly detection output and trend analysis output, a state monitoring output for the agricultural and livestock IoT is generated. This state monitoring output includes a normalized state score, anomaly detection results, and trend analysis results. Specifically, generating the state monitoring output for the agricultural and livestock IoT involves:

[0126] The normalized status score results, anomaly detection outputs, and trend analysis outputs corresponding to the current status window are uniformly collected and structured. Based on the monitoring object category, the normalized status score results, anomaly detection results, and trend analysis results are respectively merged into the output structures corresponding to agricultural environment and crop status, livestock animal status, and equipment operation status. The output structures are configured with time identifiers and window identifiers that correspond one-to-one with the current status window. The status monitoring output comprehensively reflects the score status, abnormal status, and changing trend of each monitoring object under the current status window, thus forming the status monitoring output of the agricultural and livestock Internet of Things.

[0127] Example 1:

[0128] To verify the feasibility of this invention in practice, it was applied to a large-scale integrated agricultural and livestock demonstration park in northern China. The park covers approximately 1200 acres and includes greenhouse planting areas, open-air livestock breeding areas, and centralized equipment operation areas. Various types of IoT sensing devices are deployed within the park. The agricultural environment and crop monitoring section includes temperature and humidity sensors, soil moisture sensors, and light intensity sensors, with a sampling period of 5 minutes. The livestock animal monitoring section includes wearable activity sensors and body surface temperature sensors, with a sampling period of 10 minutes. The equipment operation monitoring section includes sensors for the operating status of feeding equipment, power equipment, and ventilation equipment, with a sampling period of 1 minute. Due to inconsistent sampling frequencies and significant differences in data distribution among different sensing sources, and because some sensors are susceptible to environmental interference during actual operation, traditional methods based on fixed weights or thresholds are insufficient to accurately reflect the true operating status of the park, easily leading to false alarms or missed alarms.

[0129] In this application scenario, the method of this invention is used to uniformly process continuously collected multi-source sensing data within the park. First, the collected data is standardized and divided into time windows, with each window lasting 30 minutes. Multi-source sensing data within the same window is then aggregated. A multi-source sensing baseline distribution database for the park under normal production conditions is constructed through operational phase identification. During operation, the Hellinger distance between the data distribution of each sensing source under the current state window and the corresponding stage baseline distribution is calculated in real time to obtain the distribution deviation vector, which is then used to generate the credibility weight of each sensing source. Sensing sources with higher credibility contribute a larger proportion in subsequent fusion modeling, while the influence of sensing sources with larger distribution deviations is effectively suppressed. The fused state features are input into an improved generalized additive model, outputting a comprehensive quantitative index reflecting the overall operational status of the park, and further forming scores for environmental status, animal status, and equipment status.

[0130] In a comparative experiment conducted continuously for 30 days, the method of this invention was compared and analyzed with the traditional equal-weight fusion method. The experimental results showed that the accuracy of the method of this invention in equipment anomaly identification increased from 82.4% to 93.1% compared with the original method, and the false alarm rate decreased from 17.6% to 6.8%. In the identification of abnormal behavior in pastoral animals, the average lead time for anomaly detection was increased by about 18 minutes. In agricultural environmental monitoring, the daily fluctuation range of the comprehensive status score was reduced by about 27%, thus improving the stability of status assessment.

[0131] Actual operation shows that the present invention can effectively solve the monitoring deviation problem caused by inconsistent reliability and distribution differences of multi-source sensing data, improve the accuracy, stability and practical value of agricultural and animal husbandry Internet of Things status monitoring, and has good engineering application effects.

[0132] Table 1. Comparison of experimental data on the effectiveness of agricultural IoT state monitoring methods based on multi-source sensing fusion.

[0133] Category of monitoring objects Evaluation indicators Traditional equal-weight fusion method Method of the present invention Improvement range Agricultural environment and crop status Average daily fluctuation range of comprehensive status score 0.42 0.31 ↓ Approximately 26.2% Agricultural environment and crop status Accuracy of abnormal environment identification 85.6% 94.3% ↑8.7% Agricultural environment and crop status False alarm rate 14.1% 6.5% ↓ Approximately 53.9% Livestock status Accuracy of abnormal behavior identification 80.9% 92.6% ↑11.7% Livestock status Anomaly early identification time 6.3 minutes 24.5 minutes ↑ Approximately 18.2 minutes Livestock status Stability of state rating (standard deviation) 0.38 0.27 ↓ Approximately 28.9% Equipment operating status Equipment anomaly identification accuracy 82.4% 93.1% ↑10.7% Equipment operating status Equipment malfunction false alarm rate 17.6% 6.8% ↓ Approximately 61.4% Equipment operating status Abnormal missed detection rate 9.2% 3.5% ↓ Approximately 62.0% Comprehensive monitoring results Improved stability of comprehensive quantitative indicators — Significant improvement — Comprehensive monitoring results Multi-source sensing fusion robustness generally Significantly enhanced —

[0134] As shown in Table 1, under the same agricultural IoT operating environment and monitoring cycle, the method of this invention significantly outperforms the traditional equal-weighted fusion method in several key indicators. In agricultural environment and crop status monitoring, after adopting the method of this invention, the daily average fluctuation of the comprehensive status score decreased from 0.42 to 0.31, a decrease of approximately 26%. This indicates that after introducing the distribution consistency assessment and credibility weighting mechanism, the environmental status quantification results are more stable and can effectively suppress the interference of abnormal fluctuations from a single sensing source on the overall assessment results. The accuracy rate of environmental anomaly identification increased from 85.6% to 94.3%, and the false alarm rate decreased from 14.1% to 6.5%, indicating that the present invention is more reliable in identifying abnormal states under complex environmental changes.

[0135] In the monitoring of pastoral animal conditions, the method of this invention also demonstrates significant advantages. The accuracy rate of behavioral anomaly identification increased from 80.9% to 92.6%, and the standard deviation of state score stability decreased from 0.38 to 0.27. This reflects that the fusion modeling approach based on credibility adjustment and consistency constraints can more accurately depict the true behavioral state of animals and reduce misjudgments caused by factors such as device misalignment and signal loss. The anomaly early identification time increased from 6.3 minutes in the original method to 24.5 minutes, an increase of approximately 18 minutes, providing pastoral managers with more time to intervene and enhancing its practical application value.

[0136] In terms of equipment operation status monitoring, the method of this invention significantly improves the ability to identify equipment anomalies. The accuracy rate of equipment anomaly identification increased from 82.4% to 93.1%, the false alarm rate decreased from 17.6% to 6.8%, and the missed detection rate decreased from 9.2% to 3.5%. This indicates that by adjusting the consistency constraints on the equipment sensing data with large distribution deviations, the interference of abnormal noise on the model output can be effectively avoided, thereby improving the reliability of equipment operation status assessment.

[0137] In summary, the experimental data in Table 1 fully demonstrate that the present invention has achieved significant improvements in stability, accuracy, and anomaly identification foresight through the synergistic effects of multi-source sensing distribution consistency assessment, confidence weight adjustment, and improved generalized additive model. This verifies its practical application effect and engineering feasibility in long-term operation monitoring of complex agricultural and livestock IoT systems.

[0138] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for agricultural and pastoral IoT status monitoring based on multi-source sensing fusion, characterized in that, include: Collect multi-source sensing data from different sensing sources in agricultural and animal husbandry Internet of Things, preprocess the multi-source sensing data, and generate standardized multi-source time series data; Based on standardized multi-source time-series data, the operation stages of agricultural and animal husbandry Internet of Things are identified, corresponding stage labels are assigned to each state window, and a multi-source sensing benchmark distribution library with one-to-one correspondence with the stage labels is constructed. A distribution consistency evaluation module based on Hellinger distance is constructed. The current distribution representation is constructed for the sensing data corresponding to each sensing source within the current state window. The Hellinger distance of each sensing source is calculated by combining the multi-source sensing benchmark distribution library to obtain the distribution deviation vector of each sensing source under the current state window. The confidence weights of each sensing source are generated based on the distribution deviation vector, and the multi-source sensing data are weighted and fused according to the confidence weights to obtain the fused state feature data. Based on the improved generalized additive model, the fused state feature data is modeled, and the Hellinger distance is introduced to adjust the contribution terms of the corresponding sensing sources in the improved generalized additive model, and the comprehensive quantitative index of agricultural and animal husbandry Internet of Things is output. Based on comprehensive quantitative indicators, the environmental, equipment, and animal status scores are normalized, anomaly detected, and trend analyzed to generate status monitoring outputs for agricultural and livestock IoT.

2. The method for agricultural and pastoral IoT status monitoring based on multi-source sensing fusion according to claim 1, characterized in that, The multi-source sensing data specifically includes agricultural environment and crop data, livestock data, and equipment operation data.

3. The method for agricultural and pastoral IoT status monitoring based on multi-source sensing fusion according to claim 1, characterized in that, The preprocessing of multi-source sensing data specifically includes data cleaning, data standardization, data synchronization, and data feature extraction.

4. The method for agricultural and pastoral IoT status monitoring based on multi-source sensing fusion according to claim 1, characterized in that, The construction of a multi-source sensing benchmark distribution library that corresponds one-to-one with stage labels includes: Standardized multi-source time-series data is segmented according to a preset time length to form several continuous state windows, and the corresponding multi-source sensing data set is collected in each state window. For the multi-source sensing data set within each status window, stage discrimination features reflecting the overall change level, fluctuation characteristics and distribution pattern of the data are extracted. Based on the stage discrimination features, the agricultural and animal husbandry Internet of Things operation stage to which the status window belongs is identified, and a stage label uniquely corresponding to the status window is generated. After obtaining the stage labels, the standardized multi-source time series data corresponding to each sensing source in each state window are statistically analyzed to construct the current distribution description result that characterizes the data distribution characteristics of the sensing source in the current state window. The current distribution description results of different state windows are classified and collected according to the stage labels. The distribution description results of corresponding sensing sources in state windows with the same stage labels are summarized and processed to form the baseline distribution description results of each sensing source in the corresponding stage. Using stage labels and sensing source types as indexes, the baseline distribution descriptions of each sensing source corresponding to each stage are uniformly stored to construct a multi-source sensing baseline distribution library, and a one-to-one correspondence is established between the status window and the corresponding stage label and baseline distribution.

5. The method for agricultural and pastoral IoT status monitoring based on multi-source sensing fusion according to claim 1, characterized in that, The step of obtaining the distribution deviation vector of each sensing source in the current state window includes: A distribution consistency assessment module based on Hellinger distance is constructed. The distribution consistency assessment module consists of a distribution representation construction unit, a stage benchmark distribution retrieval unit, a distribution consistency measurement unit, and a multi-source deviation representation unit. The distribution representation construction unit performs statistical modeling on the standardized multi-source time-series data corresponding to each sensing source within the current state window, forming a window distribution description result that represents the data distribution characteristics of each sensing source under the state window; The stage baseline distribution retrieval unit retrieves the stage baseline distribution description results corresponding to each sensing source type from the multi-source sensing baseline distribution library based on the stage label corresponding to the current state window, and forms a stage baseline distribution set that matches the window distribution description results. The distribution consistency measurement unit performs a distribution consistency measurement based on Hellinger distance on the window distribution description results of each sensing source and the corresponding stage baseline distribution description results, and introduces an uncertainty estimation dimension to generate uncertainty estimation results that reflect the confidence level of the consistency measurement. The multi-source deviation characterization unit jointly organizes the distribution deviation measurement results of each sensing source with the corresponding uncertainty estimation results, constructs a distribution consistency matrix while maintaining the one-to-one correspondence between sensing sources, and forms a structured distribution deviation vector based on the distribution consistency matrix.

6. The method for agricultural and pastoral IoT status monitoring based on multi-source sensing fusion according to claim 1, characterized in that, The obtained fusion state feature data specifically includes: Based on the distribution deviation vector, the degree of distribution deviation of each sensing source is mapped to obtain a reliability score that reflects the reliability of the data from each sensing source. The reliability score is inversely related to the degree of distribution deviation of the corresponding sensing source. The credibility scores corresponding to each sensing source are normalized to generate credibility weights for each sensing source. The credibility weights of each sensing source are within a uniform scale range and satisfy the relative proportional relationship between the credibility weights. Based on the credibility weight, the multi-source sensing data corresponding to each sensing source within the current status window are weighted and fused to form fused status feature data that characterizes the overall operating status of the current agricultural and animal husbandry Internet of Things.

7. The method for agricultural and pastoral IoT status monitoring based on multi-source sensing fusion according to claim 1, characterized in that, The comprehensive quantitative indicators for outputting agricultural and livestock IoT include: The fused state feature data is input into the improved generalized additive model. A perception source category index structure is introduced. The fused state feature data is indexed and organized based on the perception source category to form feature subsets corresponding to agricultural environment and crop data, livestock animal data and equipment operation data, respectively. A one-to-one feature input channel is established for each feature subset. Construct a sub-additive structure that connects to the main additive predictor in parallel. The feature subsets in the feature input channel are connected to multiple sub-additive structures respectively. Each sub-additive structure independently constructs the corresponding additive smoothing prediction sub-item for the connected feature subset, and obtains the sub-prediction output of each sub-additive structure. The sub-prediction outputs are then merged into the main additive smoothing prediction sub-structure in a modular parallel manner to form the main additive prediction output. By introducing the confidence weight of the sensing source into the improved generalized additive model, and establishing a correspondence between the confidence weight of the sensing source and the sub-prediction output and internal smoothing function term of each sub-additive structure through the weight interface structure and the adjustment coefficient interface, the contribution intensity of each sensing source category in the main additive prediction output is structurally adjusted to obtain the contribution-adjusted main additive prediction output. Obtain the distribution deviation vector, introduce a consistency penalty interface and a smoothing adjustment channel, introduce the distribution deviation vector into the improved generalized additive model, establish a one-to-one correspondence between the distribution deviation vector and the contribution items of each perceptual source category in the main additive prediction output, and apply consistency constraint adjustment to the contribution-adjusted main additive prediction output to obtain the consistency-adjusted main additive prediction output. A link function mapping table is established based on the link function structure. The main additive prediction output after consistency adjustment is matched with the link function mapping table. The main additive prediction output after consistency adjustment is mapped by the link function to form a comprehensive quantitative index of agricultural Internet of Things.

8. The method for agricultural and pastoral IoT status monitoring based on multi-source sensing fusion according to claim 1, characterized in that, The generated agricultural IoT status monitoring output includes: Based on comprehensive quantitative indicators, the monitoring objects in the current status window are divided into agricultural environment and crop status, livestock status and equipment operation status, and status score data sets corresponding to the above monitoring objects are formed respectively. Based on the status score dataset, the status scores of agricultural environment and crops, livestock animals and equipment operation status are normalized respectively. All types of status scores are within a unified scale range, and normalized status score results corresponding one-to-one with the current status window are generated. Based on the normalized state score results, state anomaly identification is performed, and state score records that exceed the anomaly threshold range are marked to form anomaly detection outputs for environmental anomalies, animal anomalies, and equipment anomalies. The normalized state score results are aggregated according to the state window sequence to form the state score time series of the corresponding monitoring objects. The state score time series is then processed by trend analysis to generate trend analysis output reflecting the changes in agricultural environment and crop status, livestock status and equipment operation status over time. Based on the anomaly detection output and trend analysis output, a state monitoring output for agricultural and animal husbandry Internet of Things is generated. The state monitoring output includes normalized state score results, anomaly detection results, and trend analysis results.

Images