A deep learning-based regional mouse density intelligent monitoring and dynamic early warning method

Abstract

This invention relates to the field of rodent density monitoring and early warning technology, and discloses a deep learning-based intelligent monitoring and dynamic early warning method for regional rodent density, comprising: S1, dividing the area to be monitored into multiple regional units and establishing a time series index; S2, acquiring multi-source monitoring data and corresponding sampling efforts of each regional unit at each time point, and generating observation features; S3, extracting source reliability features and inputting them into an evaluation model to obtain source reliability weights; S4, weighting and fusing the observation features based on the source reliability weights to obtain a fused representation; S5, outputting the estimated rodent density and uncertainty; S6, generating a dynamic early warning threshold based on the historical rodent density baseline and adjusting it in conjunction with exogenous environmental variables; S7, calculating the over-threshold confidence level and triggering a dynamic early warning, and outputting the early warning result. By introducing uncertainty and a dynamic threshold in a linked manner, a stable early warning criterion can still be output under conditions of missing, abnormal, or equipment fluctuations.

Description

Technical Field

[0001] This invention relates to the field of rodent density monitoring and early warning technology, specifically a method for intelligent monitoring and dynamic early warning of regional rodent density based on deep learning. Background Technology

[0002] Regional rodent density intelligent monitoring and dynamic early warning refers to the continuous observation, summary analysis, and risk alerts of rodent activity intensity or density levels within a certain spatial range. It provides a basis for vector-borne disease control, rodent prevention in agriculture and storage, and public health inspections. The regional rodent density intelligent monitoring and dynamic early warning method based on deep learning is a technical solution that estimates the rodent density of regional units at different time points using a deep learning model based on multi-source monitoring data, and combines it with threshold strategies to form early warning outputs. It usually involves regional unit division, time-series data organization, multi-source information fusion, and early warning triggering. Existing regional rodent density monitoring and early warning methods mostly adopt point-based collection and experience-based threshold alarms. For example, monitoring data is obtained through trapping surveys, equipment-triggered counting, or video recognition, and then summarized by region or time and compared with fixed thresholds to issue early warning alerts.

[0003] However, with current technology, due to the frequent occurrence of missing or abnormal multi-source monitoring data in different regions and time periods, as well as fluctuations in equipment status, it is difficult to maintain stable and consistent data quality. This makes the early warning judgment based on the aggregated results and empirical thresholds unreliable. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a deep learning-based method for intelligent monitoring and dynamic early warning of regional rat density, which solves the problem that the basis for early warning judgment is not reliable due to missing, abnormal, and fluctuating multi-source monitoring data.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a deep learning-based intelligent monitoring and dynamic early warning method for regional mouse density, comprising: S1. Divide the area to be monitored into multiple regional units and establish a time series index; S2. Obtain multi-source monitoring data of each regional unit at each time point and the sampling effort corresponding to the multi-source monitoring data, and generate observation features based on the multi-source monitoring data and the sampling effort. S3. For the multi-source monitoring data, extract the source reliability features that characterize the data quality, input the source reliability features into the source reliability evaluation model, and output the source reliability weights of each data source at the corresponding regional unit and time point. S4. Based on the source reliability weight, the observation features of each data source are weighted and fused to obtain the fused representation of the corresponding regional unit and time point; S5. Input the fused representation into the mouse density estimation model, and output the mouse density estimate of the corresponding regional unit and time point, as well as the uncertainty corresponding to the mouse density estimate. S6. Generate dynamic early warning thresholds for corresponding regional units and time points based on historical rat density baselines, and adjust the dynamic early warning thresholds according to exogenous environmental variables; S7. Based on the estimated rat density, the uncertainty, and the dynamic warning threshold, calculate the over-threshold confidence level when the rat density exceeds the dynamic warning threshold, and trigger a dynamic warning based on the over-threshold confidence level and output the warning result.

[0006] Preferably, dividing the area to be monitored into multiple regional units includes: Obtain the geographic boundary information of the area to be monitored, and determine the spatial resolution or granularity of the regional unit division; Based on the geographic boundary information and the spatial resolution or division granularity, the area to be monitored is divided into regular grid area units or administrative area units, and a unique identifier is generated for each area unit. Establish a mapping relationship between the unique identifier and the geographical location parameters of the corresponding regional unit, so as to collect regional units of subsequent multi-source monitoring data.

[0007] Preferably, the acquisition of multi-source monitoring data and the sampling effort corresponding to the multi-source monitoring data includes: Acquire multi-source monitoring data from at least two data sources, and associate the multi-source monitoring data with data source identifiers, acquisition time information, and acquisition location parameters; Based on the acquisition location parameters and the mapping relationship, the multi-source monitoring data is collected into the corresponding regional unit, and based on the acquisition time information, the multi-source monitoring data is collected into the corresponding time point. Obtain the sampling effort corresponding to the multi-source monitoring data, and associate the sampling effort with the corresponding data source identifier, regional unit identifier, and time point.

[0008] Preferably, the step of generating observation features based on the multi-source monitoring data and sampling effort includes: The multi-source monitoring data are normalized based on the sampling effort to obtain the observations corresponding to a unit sampling effort. The observations are subjected to dimensional unification and missing labeling to generate the observation features of the corresponding data source at the corresponding regional unit and time point; The observed features are associated with the corresponding data source identifier, regional unit identifier, and time point for subsequent source reliability feature extraction and weighted fusion.

[0009] Preferably, the source reliability features used to extract and characterize data quality include: Based on the integrity of the multi-source monitoring data, missing correlation features are generated, and the missing correlation features include at least the missing rate and the disconnection rate. Anomaly-related features are generated based on the distribution characteristics of the multi-source monitoring data. The anomaly-related features include at least the outlier ratio and fluctuation amplitude indicators. Device health-related features are generated based on the data source's operating status. These device health-related features include at least one or more of the following: power status, communication status, and fault status markers. At least two of the missing related features, abnormal related features, and equipment health related features are used as the source reliability features, and the source reliability features are associated with the corresponding data source identifier, regional unit identifier, and time point.

[0010] Preferably, the step of inputting the source reliability characteristics into the source reliability evaluation model and outputting the source reliability weights includes: The source reliability characteristics corresponding to the same time point and the same regional unit are respectively input into the source reliability evaluation model to obtain the initial source reliability weight of each data source. The initial source reliability weights of each data source are normalized to obtain the source reliability weights used for weighted fusion. When any data source is detected to meet the preset abnormal conditions, the corresponding source reliability weight is adjusted to a reduced value or a masked value. The preset abnormal conditions include at least one of the following: missing related features exceeding a preset threshold, abnormal related features exceeding a preset threshold, and equipment health related features indicating a fault state.

[0011] Preferably, the weighted fusion of observation features from each data source includes: For observation features from different data sources at the same time point and within the same regional unit, feature alignment is performed to ensure they have consistent feature dimensions. The observed features of each aligned data source are weighted and summed based on the corresponding source reliability weights to obtain the fused representation; The fusion representation is associated with the corresponding regional unit identifier and time point as input to the mouse density estimation model.

[0012] Preferably, in step S5, the fusion representation of the input mouse density estimation model includes: The fused representations of the same region unit at multiple consecutive time points are obtained to form a fused representation sequence; The fused representation sequence is input into the mouse density estimation model, and the mouse density estimate at the corresponding time point is output. The mouse density estimation model simultaneously outputs uncertainty parameters corresponding to the estimated mouse density value, and determines the uncertainty based on the uncertainty parameters.

[0013] Preferably, the dynamic early warning threshold for generating corresponding regional units and time points includes: Historical rat density data for the corresponding regional unit is obtained, and the historical rat density data is statistically analyzed according to a preset time window to obtain the historical rat density baseline. The dynamic early warning threshold is determined based on the historical mouse density baseline and in combination with a preset sensitivity coefficient. The dynamic warning threshold is associated with the corresponding regional unit identifier and time point for dynamic warning triggering.

[0014] Preferably, in step S7, triggering a dynamic early warning based on the over-threshold confidence level and outputting the early warning result includes: The over-threshold confidence level is determined based on the estimated mouse density, the uncertainty, and the dynamic early warning threshold. When the overthreshold confidence level is greater than the preset trigger threshold, the dynamic warning is triggered, and the warning level is determined based on the deviation of the estimated mouse density value from the dynamic warning threshold and the overthreshold confidence level. Output the warning result corresponding to the warning level. The warning result includes at least the warning level and at least two of the following: the estimated rat density, uncertainty, dynamic warning threshold, and source reliability weight.

[0015] This invention provides a deep learning-based method for intelligent monitoring and dynamic early warning of regional mouse density. It offers the following advantages: 1. This invention generates reliability weights for each data source through source reliability assessment, and introduces uncertainty and dynamic thresholds in conjunction with fusion, density estimation and early warning triggering, thereby achieving the effect of outputting stable early warning criteria even under conditions of missing, abnormal or equipment fluctuations.

[0016] 2. This invention utilizes sampling effort to normalize multi-source monitoring data and combines dimensional uniformity and missing markers to generate observation features, achieving the effect of uniformity of observation results under different sampling intensities, comparable features, and convenient regional and temporal dimension modeling.

[0017] 3. This invention constructs a dynamic early warning threshold based on historical rat density baselines and adjusts the threshold in conjunction with exogenous environmental variables. At the same time, it outputs early warning results containing information such as early warning level, density estimate, uncertainty and dynamic threshold. It achieves the effect of threshold adaptively updating with region and time period, early warning classification rules that can be implemented, and results that can be directly used for judgment and disposal. Attached Figure Description

[0018] Figure 1 This is a flowchart of a deep learning-based method for intelligent monitoring and dynamic early warning of regional mouse density according to the present invention. Detailed Implementation

[0019] The technical solution of the present invention will now be clearly and completely described 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.

[0020] Please see the appendix Figure 1 This invention provides a method for intelligent monitoring and dynamic early warning of regional mouse density based on deep learning, comprising: S1. Divide the area to be monitored into multiple regional units and establish a time series index; Furthermore, the area to be monitored is divided into multiple regional units, including: Obtain the geographic boundary information of the area to be monitored, and determine the spatial resolution or granularity of the regional unit division; Based on geographic boundary information and spatial resolution or granularity, the area to be monitored is divided into regular grid area units or administrative area units, and a unique identifier is generated for each area unit. Establish a mapping relationship between unique identifiers and the geographical location parameters of corresponding regional units to facilitate the aggregation of regional units from subsequent multi-source monitoring data.

[0021] Specifically, the geographic boundary information of the area to be monitored is obtained, and the spatial resolution or granularity of the area unit division is determined. Specifically, boundary polygons can be obtained by importing the monitoring area boundary file (such as Shapefile / GeoJSON) into a GIS platform, and a unified coordinate system is established. The spatial resolution or granularity is set according to monitoring management needs and data density, for example, using a fixed grid side length or administrative level as the granularity, to subsequently map monitoring data from different locations into a unified spatial unit. Based on geographic boundary information and spatial resolution or granularity, the area to be monitored is divided into regular grid area units or administrative region area units, and a unique identifier is generated for each area unit. When using a regular grid, a fixed-size grid is generated within the boundary range and units outside the boundary are clipped. The unique identifier can be generated by row and column numbers or by auto-incrementing numbers. When using administrative regions, the administrative region boundary polygon is directly used as the area unit. The unique identifier can be the administrative region code or a custom code, which is used to form an indexable set of area units, which is convenient for subsequent storage, statistics and retrieval by area unit. Establishing a mapping relationship between a unique identifier and the geographical location parameters of the corresponding regional unit can be used to collect regional units of subsequent multi-source monitoring data. Specifically, for each regional unit, its polygonal geometric information, center point coordinates, and circumscribed rectangle range can be saved, and a mapping table of "regional unit identifier - geometric parameters" can be established. Subsequently, when monitoring data carries the collection location parameters, the regional unit identifier to which it belongs can be determined by point-to-grid calculation or point-to-polygon judgment, and the collection can be completed accordingly. For example, if a county is used as the monitoring area and a regular grid of 500m×500m is used, a grid is generated within the county boundary and identified by "row number-column number". At the same time, the geometric boundary of each grid is saved. When a detector uploads a monitoring record with coordinates, the grid ID it falls into can be calculated based on the coordinates, and the record can be collected into the corresponding grid area unit.

[0022] S2. Obtain multi-source monitoring data of each regional unit at each time point and the sampling effort corresponding to the multi-source monitoring data, and generate observation features based on the multi-source monitoring data and the sampling effort. Furthermore, acquiring multi-source monitoring data and the corresponding sampling effort includes: Acquire multi-source monitoring data from at least two data sources, and associate the multi-source monitoring data with data source identifiers, acquisition time information, and acquisition location parameters; Based on the location parameters and mapping relationship, the multi-source monitoring data is collected to the corresponding regional unit, and based on the collection time information, the multi-source monitoring data is collected to the corresponding time point. Obtain the sampling effort corresponding to the multi-source monitoring data, and associate the sampling effort with the corresponding data source identifier, regional unit identifier, and time point.

[0023] Specifically, in regional rat density intelligent monitoring, data from different front-ends need to be uniformly accessed and identified first. During implementation, a uniform data access format or interface can be set for each data source. When the data enters the platform, the data source identifier is written and the collection time information and collection location parameters are retained. Among them, the collection time information is used for subsequent alignment according to a uniform statistical granularity, and the collection location parameters are used for subsequent spatial attribution judgment, thereby providing a data source and spatiotemporal positioning basis for regional rat density analysis. To enable monitoring data to be organized by regional units and time points, during implementation, the collection location parameters of each monitoring data are converted into the corresponding regional unit identifier based on the mapping relationship between regional unit identifier and geographical location parameters. At the same time, the collection time information is mapped to a preset time point index, such as by day or by hour, so that multi-source monitoring data within the same regional unit and the same time point are merged into the same index, which facilitates the subsequent generation of observation features and entry into the rat density estimation and early warning process. Since the working duration and effective sampling coverage of different data sources vary at different time periods, the sampling effort corresponding to the monitoring data is acquired synchronously during implementation and stored in association at the data source identifier, regional unit identifier, and time point granularity. The sampling effort can be determined according to the data source type. For example, for online devices, the online duration or effective working duration is taken; for images or videos, the effective sampling duration or effective frame count is taken; for manual surveys, the number of deployments is taken by multiplying the duration or the survey route length, etc. By binding the sampling effort with the collected monitoring data, input is provided for subsequent generation of observation features based on the monitoring data and sampling effort, so that data from different sources have a consistent statistical caliber when modeling regional rat density. For example, when infrared cameras and smart detectors are deployed within a county, the data reported by both types of devices includes coordinates and timestamps. After the platform accesses the data, it writes the data source identifiers into the corresponding grid cells through the regional unit mapping relationship. Then, it aligns the timestamps to the corresponding date and time point by day. At the same time, it calculates the effective working time of the infrared camera and the online time of the detector on the same day as the sampling effort, and stores them in association with the corresponding data source identifier, grid cell identifier, and date and time point. This provides basic data for generating observation features and making dynamic early warnings within the grid and on the same date.

[0024] Furthermore, observational features generated based on multi-source monitoring data and sampling effort include: The multi-source monitoring data are normalized based on the sampling effort to obtain the observations corresponding to a unit sampling effort. The observations are processed for dimensional unification and missing labeling to generate the observation characteristics of the corresponding data source in the corresponding regional unit and time point; The observed features are associated with the corresponding data source identifier, regional unit identifier, and time point for subsequent source reliability feature extraction and weighted fusion.

[0025] Specifically, to ensure the comparability of monitoring data obtained from different data sources under varying online durations or effective sampling coverage conditions, the multi-source monitoring data is first normalized using sampling effort to obtain the observation corresponding to a unit sampling effort. In practice, under the same data source, the same regional unit, and the same time point granularity, the ratio of the original measurement to the corresponding sampling effort is calculated to obtain the observation per unit effort. ; in, For raw measurements, the corresponding data fields can be the number of events or the number of triggers, etc. For sampling effort, the corresponding data fields can be online duration, effective sampling duration, or effective frame count, etc. For unit sampling effort observation; Subsequently, the unit effort force observations are processed to unify the dimensions and add missing markers to generate observation features. In practice, unit effort force observations from different sources can be scaled or standardized according to unified rules, and missing markers are set for cases where the sampling effort is zero, the data is not reported, or the data is invalid, so as to retain data availability information and facilitate subsequent data quality-related processing. Finally, the observation features are associated with and stored with the data source identifier, regional unit identifier, and time point. The observation features can be saved using a record or tensor structure with the data source identifier, regional unit identifier, and time point as index keys. This allows the corresponding observation features to be read directly according to the above index for source reliability feature extraction and weighted fusion. For example, if the infrared camera triggers 20 times and has an effective working time of 10 hours in the same grid cell on a certain day, the unit effort force observation is 2 times per hour. If the smart detector has 15 events and is online for 5 hours, the unit effort force observation is 3 times per hour. After unifying the scale of both and recording the missing markers, corresponding observation features are formed and associated with their respective data source identifiers, grid identifiers, and date indexes.

[0026] S3. For multi-source monitoring data, extract source reliability features that characterize data quality, input the source reliability features into the source reliability assessment model, and output the source reliability weights of each data source at the corresponding regional unit and time point. Furthermore, the source reliability features that characterize data quality include: Based on the integrity of multi-source monitoring data, missing correlation features are generated. Missing correlation features include at least the missing rate and the disconnection rate. Anomaly-related features are generated based on the distribution characteristics of multi-source monitoring data. These anomaly-related features include at least the proportion of outliers and the fluctuation amplitude index. Based on the data source's operating status, device health-related features are generated. These features include at least one or more of the following: power status, communication status, and fault status markers. At least two of the following categories are selected as source reliability features: missing relevant features, abnormal relevant features, and equipment health relevant features. The source reliability features are then associated with the corresponding data source identifier, regional unit identifier, and time point.

[0027] Specifically, to describe the data availability and stability of different data sources in regional rat density monitoring, this implementation generates missing-related features based on the integrity of multi-source monitoring data. The missing rate is used to characterize the proportion of missing entries within a preset statistical window, which can be directly obtained from the observation feature sequence organized by data source identifier, regional unit identifier, and time point. The disconnection rate is used to characterize the situation of continuous missing or continuous no reporting, which can be obtained by counting the number of occurrences of continuous missing segments or their proportion. These features are used to characterize the data supply stability of the data source in the time dimension and provide input for subsequent reliability assessment. Anomaly-related features are generated based on the distribution characteristics of multi-source monitoring data. The outlier ratio is used to characterize the proportion of times the observed features exceed the preset reasonable range or deviate significantly from the historical normal. It can be obtained by threshold discrimination or deviation from the historical baseline. The fluctuation amplitude index is used to characterize the degree of change of the observed features within the statistical window. It can be calculated by the statistical measure of the difference between the maximum and minimum values ​​or the difference between adjacent time points. This is used to reflect whether there are sudden changes, drifts or instabilities in the data, and to provide a basis for subsequent reliability assessment. Based on the data source's operational status, device health-related features are generated. Status information reported by the device or maintained by the platform can be used to generate power status, communication status, and fault status markers, and stored in alignment with regional units and time points. Among them, power status can be power level or low power marker, communication status can be signal strength or disconnection marker, and fault status marker can be given by device self-test code or platform rules. These features are used to characterize the availability status at the hardware and communication levels.

[0028] When forming source reliability features, at least two types of features, including missing features, abnormal features, and equipment health features, are combined into a feature vector and stored in association with the data source identifier, regional unit identifier, and time point, so that the subsequent source reliability assessment model can read the corresponding features and output the weights at the same index granularity.

[0029] For example, if an infrared camera in a grid cell fails to report for two days within a week, the missing rate can be obtained from this, and the existence of consecutive non-reporting can be recorded to form the disconnection rate. If the trigger value on a certain day exceeds the preset reasonable range, it is counted as an outlier and the outlier ratio is updated. At the same time, the fluctuation range index is obtained by combining the maximum and minimum difference of the trigger values ​​within that week. If the device reports low battery or disconnection, the corresponding battery status or communication status field is generated and included in the source reliability characteristics.

[0030] Furthermore, the source reliability characteristics are input into the source reliability assessment model, and the output source reliability weights include: The source reliability characteristics corresponding to the same time point and the same regional unit are input into the source reliability assessment model to obtain the initial source reliability weights of each data source. The initial source reliability weights of each data source are normalized to obtain the source reliability weights used for weighted fusion. When any data source is detected to meet the preset abnormal conditions, the corresponding source reliability weight is adjusted to a reduced value or a masked value. The preset abnormal conditions include at least one of the following: missing related features exceeding a preset threshold, abnormal related features exceeding a preset threshold, and equipment health related features indicating a fault state.

[0031] Specifically, the source reliability assessment model is used to map source reliability features into weights that can be used for fusion. For different data sources at the same time point and in the same regional unit, the corresponding source reliability feature vectors are read as model inputs, and the model outputs the initial source reliability weights of each data source. The source reliability assessment model can be implemented using a deep learning model, such as a regression network composed of several fully connected layers. The input is a vector composed of missing relevant features, abnormal relevant features and equipment health relevant features, and the output is a single scalar weight to characterize the availability of the data source in the region and at the time point. To facilitate subsequent weighted fusion, after obtaining the initial source reliability weights of each data source, they are normalized to obtain the source reliability weights used for weighted fusion. The normalization process can be carried out by summing the weights so that the sum of the weights of each data source within the same regional unit and the same time point is 1, which facilitates the proportional weighting of the observed features. Accordingly, the normalized weights are stored in association with the data source identifier, regional unit identifier, and time point, and are used as parameters directly called in the fusion stage. In actual operation, when the data quality of a certain data source is obviously abnormal, it is necessary to suppress or remove its weight to prevent abnormal data from entering the fusion process. Therefore, abnormal condition detection can be performed before or after normalization: when the detected missing related features exceed the preset threshold, or the abnormal related features exceed the preset threshold, or the equipment health related features indicate a fault state, the source reliability weight corresponding to the data source is adjusted to a reduced value or a masked value. The reduced value can be the weight multiplied by a preset attenuation coefficient, and the masked value can be the weight reset to 0 and the weights of the remaining data sources are redistributed during subsequent normalization. For example, on a certain day in a certain grid cell, both infrared cameras and smart detectors provide observation features. At the same time, the platform statistics show that the infrared camera has a high missing rate and abnormal communication status. In this case, the source reliability features corresponding to the infrared camera are input into the source reliability assessment model to obtain a low initial weight. If the missing relevant features or communication abnormality meets the preset abnormality conditions, the weight is further reset to the mask value or reduced by the attenuation coefficient. Then the remaining weights are normalized to obtain the source reliability weights used for subsequent weighted fusion.

[0032] S4. Based on the source reliability weight, the observation features of each data source are weighted and fused to obtain the fused representation of the corresponding regional unit and time point; Furthermore, the weighted fusion of observation features from various data sources includes: For observation features from different data sources at the same time point and within the same regional unit, feature alignment is performed to ensure they have consistent feature dimensions. The observed features of each aligned data source are weighted and summed based on the corresponding source reliability weights to obtain the fused representation; The fusion representation is associated with the corresponding regional unit identifier and time point as input to the mouse density estimation model.

[0033] Specifically, in order to form a uniformly processed input from different data sources in the same regional unit and at the same time point, the observation features of each data source are first aligned to ensure that they have consistent feature dimensions. The alignment method can adopt a unified feature template: the observation features output by each data source at this time point are mapped to a feature vector of the same dimension. If a certain dimension does not exist in the data source, it is filled with a preset default value and the missing value is retained. For features with different dimensions, the aforementioned dimension unification rule can be used to ensure that the meaning represented by the same dimension is consistent, thereby eliminating the input incompatibility problem caused by the inconsistency of feature dimensions of different data sources. After feature alignment, the observed features from each data source are weighted and summed according to their corresponding source reliability weights to obtain the fused representation. The fused representation can be calculated using the following formula: ; in, For data source At the point of time Regional Units The aligned observed feature vectors correspond to data fields that are the aligned feature vector column groups. This corresponds to the source reliability weight, and the corresponding data field contains the weight value. The set of data sources currently participating in the integration. For fusion representation, the corresponding data fields are the fused feature vector column groups, which can integrate multi-source information at the same spatial and temporal granularity and make the data sources with higher reliability contribute more to the fusion result; Subsequently, the fused representation is stored in association with the corresponding regional unit identifier and time point, serving as input to the mouse density estimation model. In implementation, the fused representation can be written into a feature table or feature tensor using the regional unit identifier and time point as index keys. The subsequent mouse density estimation model can then read the fused representation sequence using the same index to complete the input preparation required for regional mouse density estimation and early warning calculation. For example, on a certain day in a certain grid cell, the aligned observation features of the infrared camera are in vector form, and the intelligent detector also generates a vector of the same dimension. If the infrared camera has a higher weight due to its lower missing rate, then the infrared camera will contribute more to the fusion representation when weighted summation. After the fusion representation is generated, it is associated with the grid identifier and date index and saved for direct use by the subsequent mouse density estimation model.

[0034] S5. Input the fusion representation into the mouse density estimation model, and output the mouse density estimate of the corresponding regional unit and time point, as well as the uncertainty corresponding to the mouse density estimate. Furthermore, in S5, the fusion representation of the input mouse density estimation model includes: The fused representations of the same region unit at multiple consecutive time points are obtained to form a fused representation sequence; The fused representation sequence is input into the mouse density estimation model, and the mouse density estimate at the corresponding time point is output. The rat density estimation model simultaneously outputs uncertainty parameters corresponding to the estimated rat density, and determines the uncertainty based on the uncertainty parameters.

[0035] Specifically, to enable rat density estimation to utilize the temporal continuity of rat activity in a region, fused representations from multiple consecutive time points are read for the same regional unit and a fused representation sequence is formed. The sequence length can be determined by a preset sliding window, for example, by taking fused representations from the most recent several days or hours and arranging them in chronological order. When fused representations for certain time points are missing, missing markers or preset padding values ​​can be used to place the missing data, ensuring that the sequence length remains consistent and that the input format for subsequent models is stable.

[0036] The fused representation sequence is input into the mouse density estimation model, which outputs the estimated mouse density value at the corresponding time point. The mouse density estimation model can be implemented using a deep learning temporal model, such as a temporal network based on a recurrent neural network, temporal convolutional network, or attention mechanism. The model input is the fused representation sequence, and the model output is the estimated mouse density value of the unit in that region at the current time point. This step maps the multi-source fused information to a unified numerical scale for mouse density, facilitating subsequent threshold determination and early warning classification.

[0037] The mouse density estimation model simultaneously outputs uncertainty parameters corresponding to the estimated mouse density, and determines the uncertainty based on these parameters. To ensure feasibility, this implementation uses heteroscedastic regression: in addition to outputting the estimated mouse density, the model also outputs a variance parameter or scale parameter corresponding to that estimate. The platform uses this to obtain the uncertainty, which is then used for subsequent overthreshold confidence calculations. This output relationship can be expressed as: ; in, For regional units At the point of time The fusion representation corresponds to the fusion feature vector in the data field. For sequence length, For the mouse density estimation model, This represents the estimated mouse density; the corresponding data field is the density estimate. For uncertainty parameters or uncertainties determined by uncertainty parameters, the corresponding data field is the uncertainty value. The above output provides input for subsequent dynamic early warning threshold comparison and early warning triggering. For example, when estimating a grid cell on a daily basis, the fused representation of the most recent 7 days is used to form the sequence input model. The model outputs the estimated mouse density of the grid cell on that day, and also outputs the corresponding uncertainty parameters. If the fused representation is missing on a certain day due to the device being offline, a missing marker can be used in the sequence to place the missing part, so that the model can still complete the inference according to the fixed length sequence and obtain the corresponding output.

[0038] S6. Generate dynamic early warning thresholds for corresponding regional units and time points based on historical rat density baselines, and adjust the dynamic early warning thresholds according to exogenous environmental variables; Furthermore, the generation of dynamic early warning thresholds for corresponding regional units and time points includes: Historical rat density data for the corresponding regional unit is obtained, and the historical rat density data is statistically analyzed according to a preset time window to obtain the historical rat density baseline. Dynamic warning thresholds are determined based on historical rat density baselines and in conjunction with preset sensitivity coefficients. The dynamic warning threshold is associated with the corresponding regional unit identifier and time point for dynamic warning triggering.

[0039] Specifically, in order to ensure that different regional units have comparable early warning standards under different seasons and historical fluctuation levels, historical rat density data of the corresponding regional units are first obtained, and the historical rat density baseline is obtained by statistical analysis according to a preset time window. The historical rat density data can come from the output of the aforementioned rat density estimation model in historical periods or from the results of manual surveys and verifications. After being summarized by regional unit identifiers, samples are taken according to a preset time window, such as statistical analysis according to the most recent several weeks, several months, or the same season, to obtain the baseline mean and baseline fluctuation of the regional unit under that window. A dynamic warning threshold is determined based on a historical mouse density baseline and a preset sensitivity coefficient. The dynamic warning threshold can be constructed using the baseline mean and baseline fluctuation, and the sensitivity coefficient can be used to adjust the threshold's tightness. Its calculation can be expressed as: ; in, For regional units At the point of time The dynamic warning threshold, corresponding to the data field "dynamic threshold". This is a statistical measure of the historical baseline rat density, with the corresponding data field being the baseline mean. This is a statistical measure of the fluctuation of the historical mouse density baseline, and the corresponding data field is the baseline standard deviation or equivalent volatility index. The preset sensitivity coefficient corresponds to the configuration parameter field. If it is necessary to introduce exogenous environmental variables to adjust the threshold, the adjustment amount can be added to the above threshold. The adjustment amount is determined by exogenous environmental variables, for example: ; in, This is a vector of exogenous environmental variables, and the corresponding data fields can be temperature, rainfall, humidity, or vegetation index, etc. This is a threshold adjustment function or adjustment model, and the corresponding implementation can be a rule table or a regression model output. The dynamic warning threshold is associated with the corresponding regional unit identifier and time point for dynamic warning triggering. In implementation, the dynamic warning threshold can be written into a threshold table indexed by regional unit identifier and time point, so that the threshold can be directly read by the same index when performing over-threshold confidence calculation and warning judgment. It can participate in warning triggering together with mouse density estimate and uncertainty. For example, by statistically analyzing the historical rat density of a grid cell on a weekly basis and taking the most recent 8 weeks as a time window, the baseline mean and fluctuation of the grid are calculated. After selecting a sensitivity coefficient, a dynamic early warning threshold for that week is generated accordingly. If exogenous environmental changes such as continuous rainfall occur during that week, the threshold can be shifted accordingly based on preset rules or model adjustments. The updated threshold is then associated with and saved along with the grid identifier and the index of the time point for that week, for use in subsequent early warning triggering phases.

[0040] S7. Based on the estimated rat density, uncertainty, and dynamic warning threshold, calculate the over-threshold confidence level when the rat density exceeds the dynamic warning threshold, and trigger dynamic warning based on the over-threshold confidence level and output the warning result.

[0041] Furthermore, S7 triggers dynamic early warnings based on over-threshold confidence levels and outputs the warning results, including: The confidence level for exceeding the threshold is determined based on the estimated rat density, uncertainty, and dynamic early warning threshold. When the confidence level exceeds the threshold, a dynamic warning is triggered, and the warning level is determined based on the degree of deviation of the estimated mouse density from the dynamic warning threshold and the confidence level. Output the warning result corresponding to the warning level. The warning result shall include at least the warning level and at least two of the following: mouse density estimate, uncertainty, dynamic warning threshold, and source reliability weight.

[0042] Specifically, the overthreshold confidence score is used to characterize the degree of confidence that the rat density exceeds the dynamic warning threshold. It is determined jointly by the estimated rat density, uncertainty, and the dynamic warning threshold. To ensure computability, this implementation uses the overthreshold probability as the overthreshold confidence score. Based on the normal distribution assumption, the estimated value and uncertainty are mapped to the probability of exceeding the threshold. The calculation method is as follows: ; in, For regional units At the point of time The overthreshold confidence level corresponds to the overthreshold probability in the data field. This represents the estimated mouse density; the corresponding data field is the density estimate. This indicates uncertainty; the corresponding data field represents an uncertain value. This is a dynamic early warning threshold; the corresponding data field is the dynamic threshold. The cumulative distribution function of the standard normal distribution. The mouse density is a real but unknown random variable. A dynamic alert is triggered when the confidence level exceeds a preset trigger threshold. The alert level is determined based on the degree of deviation. The trigger threshold can be pre-configured by management strategies, such as setting a probability threshold. ,when The time is determined as a trigger. The degree of deviation can be represented by the difference or ratio of the estimated mouse density value to the dynamic warning threshold, which is used to distinguish different degrees of severity. The warning level can be determined by segmented rules. For example, after the trigger condition is met, if the degree of deviation is in different intervals or the confidence level of exceeding the threshold is in different intervals, then different levels are output accordingly, thereby realizing the rule-based output of graded warnings. Output the warning results corresponding to the warning level. The warning results should include at least the warning level and can simultaneously output at least two of the following: mouse density estimate, uncertainty, dynamic warning threshold, and source reliability weight, for subsequent analysis and handling. The source reliability weight can be used to reflect the contribution of each data source to the fused representation in this judgment, which is convenient for tracing and verifying the basis of the warning. For example, on a certain day in a certain grid cell, the mouse density estimation model outputs the density estimate and uncertainty, while the dynamic early warning threshold module provides the threshold for that day. The platform calculates the over-threshold probability for that grid cell based on this. If the over-threshold probability exceeds the preset trigger threshold, a dynamic early warning is triggered. After triggering, the early warning level is determined based on the degree of deviation of the density estimate from the threshold, and an early warning result containing the early warning level, density estimate, uncertainty, and dynamic early warning threshold is output. The source reliability weights of each data source on that day can also be attached for analysis.

[0043] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent monitoring and dynamic early warning of regional mouse density based on deep learning, characterized in that, include: S1. Divide the area to be monitored into multiple regional units and establish a time series index; S2. Obtain multi-source monitoring data of each regional unit at each time point and the sampling effort corresponding to the multi-source monitoring data, and generate observation features based on the multi-source monitoring data and the sampling effort. S3. For the multi-source monitoring data, extract the source reliability features that characterize the data quality, input the source reliability features into the source reliability evaluation model, and output the source reliability weights of each data source at the corresponding regional unit and time point. S4. Based on the source reliability weight, the observation features of each data source are weighted and fused to obtain the fused representation of the corresponding regional unit and time point; S5. Input the fused representation into the mouse density estimation model, and output the mouse density estimate of the corresponding regional unit and time point, as well as the uncertainty corresponding to the mouse density estimate. S6. Generate dynamic early warning thresholds for corresponding regional units and time points based on historical rat density baselines, and adjust the dynamic early warning thresholds according to exogenous environmental variables; S7. Based on the estimated rat density, the uncertainty, and the dynamic warning threshold, calculate the over-threshold confidence level when the rat density exceeds the dynamic warning threshold, and trigger a dynamic warning based on the over-threshold confidence level and output the warning result.

2. The method for intelligent monitoring and dynamic early warning of regional mouse density based on deep learning according to claim 1, characterized in that, The process of dividing the area to be monitored into multiple regional units includes: Obtain the geographic boundary information of the area to be monitored, and determine the spatial resolution or granularity of the regional unit division; Based on the geographic boundary information and the spatial resolution or division granularity, the area to be monitored is divided into regular grid area units or administrative area units, and a unique identifier is generated for each area unit. Establish a mapping relationship between the unique identifier and the geographical location parameters of the corresponding regional unit, so as to collect regional units of subsequent multi-source monitoring data.

3. The method for intelligent monitoring and dynamic early warning of regional mouse density based on deep learning according to claim 2, characterized in that, The acquisition of multi-source monitoring data and the corresponding sampling effort includes: Acquire multi-source monitoring data from at least two data sources, and associate the multi-source monitoring data with data source identifiers, acquisition time information, and acquisition location parameters; Based on the acquisition location parameters and the mapping relationship, the multi-source monitoring data is collected into the corresponding regional unit, and based on the acquisition time information, the multi-source monitoring data is collected into the corresponding time point. Obtain the sampling effort corresponding to the multi-source monitoring data, and associate the sampling effort with the corresponding data source identifier, regional unit identifier, and time point.

4. The method for intelligent monitoring and dynamic early warning of regional mouse density based on deep learning according to claim 1, characterized in that, The generation of observation features based on the multi-source monitoring data and sampling effort includes: The multi-source monitoring data are normalized based on the sampling effort to obtain the observations corresponding to a unit sampling effort. The observations are subjected to dimensional unification and missing labeling to generate the observation features of the corresponding data source at the corresponding regional unit and time point; The observed features are associated with the corresponding data source identifier, regional unit identifier, and time point for subsequent source reliability feature extraction and weighted fusion.

5. The method for intelligent monitoring and dynamic early warning of regional mouse density based on deep learning according to claim 1, characterized in that, The source reliability features extracted to characterize data quality include: Based on the integrity of the multi-source monitoring data, missing correlation features are generated, and the missing correlation features include at least the missing rate and the disconnection rate. Anomaly-related features are generated based on the distribution characteristics of the multi-source monitoring data. The anomaly-related features include at least the outlier ratio and fluctuation amplitude indicators. Device health-related features are generated based on the data source's operating status. These device health-related features include at least one or more of the following: power status, communication status, and fault status markers. At least two of the missing related features, abnormal related features, and equipment health related features are used as the source reliability features, and the source reliability features are associated with the corresponding data source identifier, regional unit identifier, and time point.

6. The method for intelligent monitoring and dynamic early warning of regional mouse density based on deep learning according to claim 1, characterized in that, The step of inputting the source reliability characteristics into the source reliability evaluation model and outputting the source reliability weights includes: The source reliability characteristics corresponding to the same time point and the same regional unit are respectively input into the source reliability evaluation model to obtain the initial source reliability weight of each data source. The initial source reliability weights of each data source are normalized to obtain the source reliability weights used for weighted fusion. When any data source is detected to meet the preset abnormal conditions, the corresponding source reliability weight is adjusted to a reduced value or a masked value. The preset abnormal conditions include at least one of the following: missing related features exceeding a preset threshold, abnormal related features exceeding a preset threshold, and equipment health related features indicating a fault state.

7. The method for intelligent monitoring and dynamic early warning of regional mouse density based on deep learning according to claim 1, characterized in that, The weighted fusion of observation features from each data source includes: For observation features from different data sources at the same time point and within the same regional unit, feature alignment is performed to ensure they have consistent feature dimensions. The observed features of each aligned data source are weighted and summed based on the corresponding source reliability weights to obtain the fused representation; The fusion representation is associated with the corresponding regional unit identifier and time point as input to the mouse density estimation model.

8. The method for intelligent monitoring and dynamic early warning of regional mouse density based on deep learning according to claim 1, characterized in that, The S5 section includes the fusion representation of the input mouse density estimation model, which comprises: The fused representations of the same region unit at multiple consecutive time points are obtained to form a fused representation sequence; The fused representation sequence is input into the mouse density estimation model, and the mouse density estimate at the corresponding time point is output. The mouse density estimation model simultaneously outputs uncertainty parameters corresponding to the estimated mouse density value, and determines the uncertainty based on the uncertainty parameters.

9. The method for intelligent monitoring and dynamic early warning of regional mouse density based on deep learning according to claim 1, characterized in that, The dynamic early warning thresholds for generating corresponding regional units and time points include: Historical rat density data for the corresponding regional unit is obtained, and the historical rat density data is statistically analyzed according to a preset time window to obtain the historical rat density baseline. The dynamic early warning threshold is determined based on the historical mouse density baseline and in combination with a preset sensitivity coefficient. The dynamic warning threshold is associated with the corresponding regional unit identifier and time point for dynamic warning triggering.

10. The method for intelligent monitoring and dynamic early warning of regional mouse density based on deep learning according to claim 1, characterized in that, The step S7, which triggers a dynamic early warning based on the over-threshold confidence level and outputs the early warning result, includes: The over-threshold confidence level is determined based on the estimated mouse density, the uncertainty, and the dynamic early warning threshold. When the overthreshold confidence level is greater than the preset trigger threshold, the dynamic warning is triggered, and the warning level is determined based on the deviation of the estimated mouse density value from the dynamic warning threshold and the overthreshold confidence level. Output the warning result corresponding to the warning level. The warning result includes at least the warning level and at least two of the following: the estimated rat density, uncertainty, dynamic warning threshold, and source reliability weight.

Images