A parasitic rate intelligent calculation method and system based on multi-source data fusion

Abstract

The application discloses a kind of parasitic rate intelligent calculation method and system based on multi-source data fusion, it is related to the technical field of monitoring of agricultural and forestry pest control, the specific steps of this method are as follows: collecting monitoring data and preprocessing to obtain original data set;Fusion cleaning data, label parasitic rate true value to construct feature data set;Train deep learning regression model;After pre-processing, input model prediction for the sample to be measured, correct numerical value, calculate confidence interval and mark abnormality;Finally, statistical chart report is generated, double medium is locked and archived;The application trains offline model using deep regression network with early stopping constraint, the parasitic rate prediction value is obtained by standardization link reasoning for the sample to be measured, the abnormal sample is automatically identified by numerical correction and multi-dimensional rule, the whole process automation of measurement and calculation is realized, visual chart and multi-format report are automatically generated by the system, all data files are archived relying on local and cloud storage, and the measurement efficiency, result reliability and whole-link data traceability management are considered.

Description

Technical Field

[0001] This invention relates to the field of biological control and monitoring technology for agricultural and forestry pests, specifically to an intelligent method and system for calculating parasitism rate based on multi-source data fusion. Background Technology

[0002] Parasitic wasps are a mainstream technology in the field of green pest control in agriculture and forestry. Parasitism rate is a core quantitative indicator for evaluating the breeding quality and field control efficacy of parasitic wasps. It has long supported various scientific research and production experiments, such as wasp species selection, crossbreeding experiments, and environmental adaptability studies. At present, the industry monitoring system can simultaneously acquire multiple types of monitoring information, such as field environment, microscopic images of insect eggs, experimental batch management, long-term historical observation, and the entire operation sequence. These types of information correspond to key influencing factors such as the developmental status of insect eggs, the interaction process between wasps and insects, and experimental control conditions. With the popularization of digital monitoring equipment, massive amounts of heterogeneous monitoring data will be continuously generated in various scenarios. Different types of data storage carriers, numerical dimensions, text expression forms, and image dimensions have significant differences. Deep learning regression algorithms, with their nonlinear fitting capabilities, can explore the complex correlation between multiple variables and parasitism rate. They are gradually being applied to the field of intelligent measurement of agricultural biological indicators, providing a feasible technical path for obtaining parasitism rate automatically without manual egg counting. The industry urgently needs an integrated computing solution that adapts to multi-source heterogeneous data, integrates model training, intelligent reasoning, and standardized archiving.

[0003] Traditional parasitism rate calculations heavily rely on manual counting of eggs for numerical calculations. Analysis is based on single images or localized environmental data, failing to integrate multi-dimensional information across the entire process. This lack of data dimensions leads to significant discrepancies between the calculated results and the actual parasitism level. Traditional data processing lacks a unified standardized process, hindering vector fusion of heterogeneous data. Outliers and missing fields are manually filled, consuming significant manpower and introducing human error. Conventional modeling methods do not stratify datasets according to experimental logic, lack stable early stopping constraints during model training, and do not include independent offline inference processes after training. Data processing logic must be repeatedly built for test samples, making batch automated calculations difficult. Existing calculation processes do not eliminate numerical interference from natural mortality in the blank control group, lack confidence intervals to quantify the reliability of results, and cannot automatically identify anomalous samples caused by image acquisition or environmental exceedances. Finally, data is stored in a single table file without dual-media backup and archiving mechanisms. Statistical analysis, visualization charts, and standardized reports are all manually created, resulting in significant shortcomings in experimental data traceability and long-term secure storage. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an intelligent method and system for calculating parasitism rate based on multi-source data fusion. This system simultaneously collects multiple types of data, including environmental data, insect egg images, production management data, historical samples, and time-series monitoring data. After standardization, time-series alignment, and anomaly cleaning, multi-source feature fusion is achieved. A labeled dataset is constructed based on the actual parasitism rate obtained through manual counting. A hierarchical partitioning strategy is used to train a fixed-structure deep regression model and generate offline inference files. The test samples reuse the entire preprocessing workflow to generate feature vectors, which are then input into the model to obtain predicted values. A correction algorithm is used to eliminate interference from natural mortality. Confidence intervals are calculated simultaneously, and abnormal samples in images and the environment are automatically identified. The system automatically completes statistical analysis, generates visual charts and multi-format reports, and archives the entire process data through both local and cloud-based archiving. This complete solution automates the entire parasitism rate calculation process, solving the problems of single-dimensionality, high labor costs, large result errors, and fragmented data management inherent in traditional methods. It is applicable to various experimental monitoring scenarios for agricultural and forestry biological control.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: On one hand, a method for intelligent calculation of parasitic rate based on multi-source data fusion, the specific steps of which are as follows:

[0006] S1. Simultaneously collect multiple types of raw monitoring data, bind timestamps to the collected data and filter out invalid data to obtain a standardized raw dataset;

[0007] S2. Perform multi-source feature fusion and data cleaning operations on the standardized original dataset, calculate the parasitism rate true label based on the egg count, and construct a standardized fused feature dataset.

[0008] S3. Train a deep learning regression model using the standardized fusion feature dataset to obtain an offline inference model;

[0009] S4. Reuse steps S1 and S2 to generate the fusion feature vector of the sample to be tested, input it into the offline inference model to obtain the predicted parasitic rate value, use the correction algorithm and confidence interval algorithm to calculate the corrected parasitic rate and confidence interval, and mark the abnormal samples.

[0010] S5. Perform statistical analysis on the data obtained in step S4, generate statistical charts and multi-format reports, and perform dual-media storage and locked archiving on the reports.

[0011] Furthermore, S1, simultaneously collects multiple types of raw monitoring data, binds timestamps to the collected data and filters out invalid data to obtain a standardized raw dataset;

[0012] S2. Perform multi-source feature fusion and data cleaning operations on the standardized original dataset, calculate the parasitism rate true label based on the egg count, and construct a standardized fused feature dataset.

[0013] S3. Train a deep learning regression model using the standardized fusion feature dataset to obtain an offline inference model;

[0014] S4. Reuse steps S1 and S2 to generate the fusion feature vector of the sample to be tested, input it into the offline inference model to obtain the predicted parasitic rate value, use the correction algorithm and confidence interval algorithm to calculate the corrected parasitic rate and confidence interval, and mark the abnormal samples.

[0015] S5. Perform statistical analysis on the data obtained in step S4, generate statistical charts and multi-format reports, and perform dual-media storage and locked archiving on the reports.

[0016] Furthermore, the data synchronously collected by S1 includes aquaculture environment data, insect egg image data, production management data, historical sample data, and time-series monitoring data. The aquaculture environment data includes hourly collected temperature values, hourly collected humidity values, preset bee egg release ratio, bee arrival start time, bee arrival end time, and yellow field borer egg development duration. The insect egg image data includes image pixel size parameters, unique image code, horizontal coordinate of a single egg, vertical coordinate of a single egg, width of a single egg, height of a single egg, and the total number of manually counted eggs. The production management data includes project number, test batch number, number of repeated test groups, control group type code, sampling latitude, and sampling longitude. The historical sample data includes the total number of eggs per sample over the years, the calculated value of original parasitism over the years, and the percentage of natural mortality in the control group over the years. The time-series monitoring data includes operator identification code, operation time accurate to the second, and unique sample binding code. The collected image files are stored in an independent file partition, and the numerical data collected are stored in a relational data table, with each data table recording a global timestamp accurate to the second.

[0017] Furthermore, in S2, the complete execution process of cleaning outlier samples using the 3σ rule includes three independent operations: the first operation calculates the arithmetic mean and standard deviation of the three independent sequences: the total egg count sequence, the ambient temperature sequence, and the original parasitism rate sequence; the second operation marks samples whose single field value is greater than three times the standard deviation as outlier samples; and the third operation fills the missing fields of the marked outlier samples and the missing fields of the unmarked ordinary samples by matching the arithmetic mean of the same field of other samples in the same experimental batch and the same temperature range.

[0018] Furthermore, in S3, the hierarchical allocation rule for dividing the standardized fusion feature dataset into training, validation, and test sets according to a 7:2:1 ratio is as follows: all samples are hierarchically divided according to two classification dimensions: temperature range and control group. The number of samples within each layer is split according to a fixed numerical ratio of 7:2:1. After the hierarchical split is completed, all corresponding subsets within each layer are summarized to obtain the complete training set, complete validation set, and complete test set. The number of neurons in each layer of the constructed network is fixed. The total number of neurons in the input layer is equal to the dimension of a single fusion feature vector. The first hidden layer is configured with 256 neurons, the second hidden layer with 128 neurons, and the third hidden layer with 64 neurons. All three hidden layers are equipped with ReLU activation operation logic and 0.25 ratio Dropout operation logic. The output layer is configured with a single linear neuron.

[0019] Furthermore, in S3, the early termination judgment rule includes fixed judgment conditions; after each round of weight iteration calculation, the training set and validation set are read and substituted into the mean squared error loss algorithm to obtain the corresponding loss value. When the validation set loss value does not decrease during ten consecutive rounds of iteration calculation, the entire weight iteration process is terminated; after the iteration process is terminated, three types of static parameter files are stored synchronously. The first type is the weight array of all neurons in the network, the second type is the Min-Max scaling coefficient corresponding to each dimension feature, and the third type is the temperature and egg development duration label mapping table. The three types of static parameter files together form the offline inference model.

[0020] Furthermore, in S4, the two types of judgment thresholds for labeled images and environmental anomaly samples are fixed; the image anomaly judgment threshold is a single egg confidence value of less than 0.7, and the environmental anomaly judgment threshold includes two types of boundary conditions: temperature value less than 18 degrees Celsius and temperature value greater than 32 degrees Celsius; after any judgment condition is triggered, a fixed anomaly identifier code is written into the corresponding sample record; each record of the generated batch structured calculation dataset includes six sets of fixed fields, which are, in order, the unique code of the sample, the total number of eggs in the sample, the predicted parasitic value output by the network, the value output by the corrected parasitic rate algorithm, two sets of values ​​for the upper and lower limits of the confidence interval, and the anomaly identifier code.

[0021] Furthermore, in S4, the mathematical expression for the parasitic rate correction algorithm is:

[0022]

[0023] Within the expression, This represents the corrected percentage of parasites after removing interference from natural mortality. This is the percentage calculation result output by the original parasitism rate algorithm; The percentage of naturally occurring oocytes obtained from the blank control group sample, and .

[0024] Furthermore, in S4, the mathematical expression for the confidence interval statistical algorithm is:

[0025]

[0026] Within the expression, The half-width of the confidence interval representing the corrected parasitism rate at a 95% confidence level; To correct the percentage output of the parasitic rate algorithm; Output the percentage result for the original parasitism rate algorithm; 1.96 is the fixed quantile constant of the standard normal distribution that matches at the 95% confidence level; This is the total count of all eggs in a single sample, and the data comes from the egg image data collected in step S1.

[0027] Furthermore, in S5, the variance analysis performed adopts the one-way variance operation rule, and the multiple comparison operation adopts the TukeyHSD operation rule; the output statistical charts include two types of fixed visualization graphics, the first type is a parasitism rate time series line chart, and the second type is a grouped confidence interval bar chart; the generated multi-format report includes two types of independent files, the first type is a PDF file, which embeds all statistical charts and variance operation details tables, and the second type is an Excel file, which stores the original collection fields, original parasitism rate calculation values, corrected parasitism rate calculation values, and upper and lower limits of confidence intervals for each sample in rows.

[0028] On the other hand, a parasitic rate intelligent computing system based on multi-source data fusion, the system comprising:

[0029] The multi-source data synchronous acquisition module is used to synchronously acquire multiple types of raw monitoring data, bind timestamps to the acquired data and filter out invalid data to obtain a standardized raw dataset;

[0030] The feature fusion and label construction module is used to perform multi-source feature fusion and data cleaning operations on the standardized original dataset, calculate the parasitism rate true label based on the egg count, and construct a standardized fused feature dataset.

[0031] The model training and generation module is used to train a deep learning regression model using the standardized fusion feature dataset, obtain an offline inference model, and store the network weight array, Min-Max scaling factor, and temperature-egg development duration label mapping table.

[0032] The inference correction and anomaly labeling module is used to reuse the logic of the multi-source data synchronous acquisition module and the feature fusion and label construction module to generate the fusion feature vector of the sample to be tested, input the offline inference model to obtain the predicted parasitic rate value, use the correction algorithm and confidence interval algorithm to calculate the corrected parasitic rate and confidence interval, and label the image and environmental anomaly samples.

[0033] The statistical analysis and report archiving module is used to perform variance analysis and multiple comparison operations on the data output by the inference correction and anomaly marking module, generate statistical charts and multi-format reports, and perform dual-media storage of the reports on local solid-state storage and cloud object storage, as well as lock archiving operations.

[0034] Compared with existing technologies, this intelligent parasitism rate calculation method and system based on multi-source data fusion has the following advantages:

[0035] I. This invention overcomes the limitations of computational bias caused by single data dimensions by synchronously collecting and standardizing multi-type monitoring data. It integrates multi-dimensional information from environment, images, production, and historical time series to achieve feature fusion. Relying on unified normalization, image dimension conversion, and text feature encoding processes, it achieves the same vector space mapping of heterogeneous data. Combined with time series alignment and outlier sample cleaning mechanisms to remove invalid interference information, it constructs a complete and unified feature input system. The complementary multi-source information can fully cover the influencing factors of the entire process of insect egg development, bee species release, and field environment. It avoids the one-sided problem of parasitism rate calculation caused by relying solely on manual counting or single image data. At the same time, the unified data preprocessing process can automatically complete missing value filling and outlier sample identification, greatly reducing the workload of manual data processing. This ensures that the feature information input to the model has comprehensiveness and consistency, laying a stable data foundation for the subsequent model to accurately fit the real parasitism pattern, and improving the overall reliability of parasitism rate measurement from the data source level.

[0036] II. This invention constructs an offline model by combining hierarchical dataset partitioning with a fixed-structure deep regression network and an early stopping constraint mechanism. It uses standardized ground truth labels to complete model fitting and training, reuses the entire data processing chain to generate feature vectors for the test samples, and only performs forward inference to output prediction results. It corrects the calculated values ​​by combining the natural mortality correction formula and confidence interval statistical methods, and simultaneously sets multi-dimensional anomaly judgment rules to automatically identify unqualified samples. The entire process automates the parasitism rate from data collection, model prediction, numerical correction to anomaly identification, without the need for manual secondary correction of the calculation results. The backend integrates standardized statistical analysis logic to automatically generate visual charts and dual-format file reports. With a dual storage archiving mechanism of local and cloud storage, it locks the original data, model files and calculation records of the entire process. It can not only intuitively present the differences in parasitism rate under different experimental conditions, but also realize full data traceability and retention, taking into account the calculation output efficiency, result reliability and long-term safe management needs of experimental data.

[0037] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0039] Figure 1 This is an overall flowchart of the intelligent parasitism rate calculation method based on multi-source data fusion of the present invention;

[0040] Figure 2 This is a module architecture diagram of the parasitic rate intelligent computing system based on multi-source data fusion of the present invention;

[0041] Figure 3 This is a screenshot of the login interface of the parasitism rate test management system of the present invention;

[0042] Figure 4 This invention provides an interface diagram for creating and configuring parameters.

[0043] Figure 5 This is a diagram of the batch sample binding management interface of the present invention;

[0044] Figure 6 This is a diagram of the logic verification interface for the test parameters of this invention;

[0045] Figure 7 This is a diagram of the sample QR code label printing management interface of the present invention;

[0046] Figure 8 This is a diagram of the project archive storage management interface for this invention. Detailed Implementation

[0047] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structure, features and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0048] Reference Figure 1 , Figure 3 , Figure 4 , Figure 5 , Figure 6 , Figure 7 , Figure 8This invention proposes an intelligent parasitism rate calculation method based on multi-source data fusion. It adopts an integrated data processing mechanism that includes simultaneous acquisition of multi-type monitoring data, standardized fusion and cleaning of multi-dimensional features, hierarchical deep learning regression modeling, offline inference correction and anomaly labeling, and multi-dimensional statistical archiving and storage. Based on manual sample counting, it can automatically learn the correlation rules of multi-source environment, image, and management time-series features. It can achieve automated and high-precision calculation of parasitism rate of yellow field moth eggs without the need for manual counting of eggs throughout the process, and simultaneously complete abnormal sample identification, data statistics, and dual-media locking archiving.

[0049] The method described in this embodiment specifically includes: S1. Synchronously collecting multiple types of raw monitoring data, binding global timestamps accurate to the second to the collected data and filtering out invalid data, storing image files in independent partitions and storing numerical data in a relational data table to obtain a standardized raw dataset; S2. Performing three-stage format unification, timestamp alignment, egg development duration label mapping, 3σ outlier sample cleaning, and multi-class vector feature concatenation operations on the standardized raw dataset, calculating the parasitic rate true value label based on the artificial egg count results using the original parasitic rate calculation formula, and constructing a standardized fusion feature dataset; S3. Splitting the standardized fusion feature dataset into training set, validation set, and test set in a 7:2:1 dual-dimensional layer according to temperature range and control group, building a fixed-structure three-layer hidden layer deep learning regression network, iteratively training with the mean squared error loss function, terminating the iteration with an early stopping rule that the loss does not decrease after ten consecutive rounds of validation, synchronously storing three types of static parameters: network weights, Min-Max scaling coefficients, and egg development duration reference table, and generating an offline inference model; S4. Complete. The entire process, from S1 to S2, generates a fused feature vector for the test sample. Only the forward pass operation of the offline inference model is called to output the predicted original parasitic value. This value is then substituted into the corrected parasitic rate formula to remove interference from natural mortality, yielding the corrected parasitic rate. The upper and lower limits of the confidence interval are calculated based on a 95% confidence level formula. Image anomalies and environmental anomalies are marked according to image recognition confidence thresholds and environmental temperature boundary thresholds. A structured measurement dataset containing sample code, total number of eggs, predicted value, corrected value, confidence interval, and anomaly markers is output. In S5, one-way ANOVA and TukeyHSD multiple comparison operations are performed on the structured measurement dataset, generating two types of visualization charts: a parasitic rate time-series line chart and a grouped confidence interval bar chart. Simultaneously, PDF and Excel reports are output. The original data table, feature vector, offline model, measurement dataset, and report file are synchronously stored in a local solid-state partition and a cloud object storage partition, with read-only locking and archiving. Three sets of index fields—project code, test batch, and sample unique code—are bound to achieve full-process data traceability and retrieval.

[0050] Specifically, the operational logic of the technical process of this invention lies in building an automated parasitic rate calculation pipeline that covers the entire chain of data collection, cleaning and modeling, inference and calculation, and statistical archiving. This automated production line first collects five types of heterogeneous data in parallel: breeding environment, insect egg images, production management, historical samples, and time-series operations. The raw data is standardized through separate media storage and a unified timestamp. Then, heterogeneous data fusion and cleaning are completed through numerical normalization, image vectorization, one-hot encoding of text, time-series environmental mean matching, outlier imputation, and ordered concatenation of multiple vectors. Parasitism rate ground truth labels are generated based on manual egg counting. Subsequently, a hierarchical dataset partitioning strategy is used to train a fixed-structure deep regression network, relying on an early stopping mechanism to prevent overfitting. A complete set of normalization, mapping, and weight parameters is packaged to form an offline deployable inference model. In the online calculation stage, the entire collection and fusion process is reused, only performing the model's forward computation to output predicted values. The calculation results are then corrected sequentially using the natural mortality correction formula and the 95% confidence interval formula, with dual evaluation of abnormal images and environmental samples. Finally, batch data is automatically statistically tested, visualized, and output in two formats. Local and cloud dual copies are locked and archived, and a multi-index traceability mechanism is established. This significantly reduces the workload of manual statistics throughout the process, while eliminating interference in parasitism rate calculation caused by manual counting errors, environmental temperature, natural mortality, and image recognition biases.

[0051] Optionally, the S1 step of simultaneously collecting multiple types of raw monitoring data, binding timestamps, and filtering out invalid data includes: parallel collection of five types of monitoring data: aquaculture environment data, insect egg image data, production management data, historical sample data, and time-series monitoring data; collecting temperature and humidity values ​​hourly for aquaculture environment data, and simultaneously recording preset bee egg placement ratios, bee arrival start and end times, and yellow field borer egg development duration; recording image pixel size, unique image code, horizontal and vertical coordinates of a single egg, width and height of a single egg, and manually counting the total number of eggs for insect egg image data; recording project number, test batch, number of repeated test groups, control group code, and sampling latitude and longitude for production management data; retrieving the total number of eggs per sample over the years, the calculated value of original parasitism over the years, and the percentage of natural mortality in the control group over the years for historical sample data; and recording operator identification code, operation time point accurate to the second, and unique sample binding code for time-series monitoring data. The data acquisition hardware is synchronously triggered by a unified master clock. Each numerical data table record and each image file metadata is bound to a global timestamp accurate to the second. After acquisition, invalid data filtering rules are executed: entries with an empty or zero total number of eggs, missing operation time points, duplicate image codes, or temperature and humidity exceeding the reasonable range of 0-50℃ are directly removed. Image files are uniformly stored in an independent disk file partition, retaining only the unique image code associated with the data table. All numerical indicators are written into a relational data table, with fields corresponding one-to-one with the acquisition category, forming a standardized raw dataset.

[0052] For example, when simultaneously collecting multi-source raw monitoring data, 12 replicate experiments were carried out simultaneously for a certain experimental batch. Each group was divided into two categories: a blank control group and a Trichogramma wasp release experimental group. The system collected the breeding temperature and humidity once per hour, and simultaneously recorded the start time of bee arrival at 8:00 and the end time at 18:00 on the same day. The development time of the yellow field borer egg was matched to 72 hours based on the temperature. The microscope camera captured images of the insect eggs, and automatically generated a unique image code IMG20260719001. The coordinates, width and height parameters of a single egg were identified and stored. The total number of eggs in the sample was counted manually on site, which was 1246. The production management fields were entered as follows: project code JY202606, experimental batch P007, number of replicate groups 12, control group code CK01, and sampling latitude 34. 76°, longitude 113.65°; retrieved historical sample data from the same plot over many years, with an average of 1182 eggs per sample over the years, an average original parasitism rate of 42.6%, and an average natural mortality rate of 0.062% in the control group; operator code OP027, operation time corresponding to a complete second-level timestamp, and unique sample code SMP007-08; all data were bound to the same global timestamp, image files were stored in an independent partition path / imgs / P007 / corresponding coded image file, and numerical data were written to the data table t_monitor_P007; during verification, the total number of eggs in this sample was 1246, which was not zero, the timestamp was complete, and the temperature of 26.3℃ was within a reasonable range, so it was determined to be valid data and retained in the standardized original dataset.

[0053] Optionally, the three-stage format unification, timestamp alignment, and tag mapping operation process performed by S2 includes:

[0054] The first section applies a Min-Max linear scaling operation to all numerical indicators, including temperature, humidity, and total number of eggs, constraining all values ​​to the 0-1 range after scaling. The Min-Max scaling formula is: ,in This represents the normalized feature value after linear scaling. Represents the original collected values ​​of the sample. This represents the maximum value of this feature field across all samples within the same test batch. The first segment represents the minimum value of all samples of this feature field within the same experimental batch, using the maximum and minimum values ​​of the corresponding field for all samples within a single experimental batch as the scaling benchmark; the second segment performs uniform size scaling on the original two-dimensional pixel matrix of the microscopic image, with all images fixed at a width of 640 pixels and a height of 480 pixels. After scaling, the two-dimensional pixel matrix is ​​expanded row by row and converted into a one-dimensional pixel vector with a fixed vector length of 640×480=307200; the third segment performs one-hot encoding conversion on the text encoding of the control group and the text identifier of the Trichogramma bee species, generating an independent 0-1 vector dimension for each independent text identifier, eliminating the interference of disordered text features on model training.

[0055] When performing timestamp alignment, the system uses the single sample collection time point as the baseline node, extracts all environmental temperature and humidity values ​​within the 12-hour interval before and after the collection time point, and calculates the arithmetic mean as the environmental parameter matching for that sample. Environmental data entries with a difference of more than 24 hours between the sample collection time point and the environmental record time point are directly discarded and not included in the mean calculation. When performing egg development duration label mapping, the system pre-stores a numerical lookup table that corresponds one-to-one between temperature values ​​and the egg development duration of *Meretrix meretrix*. The average temperature matching for the sample is input into the lookup table to retrieve the corresponding egg development duration scalar, which is then concatenated as an independent one-dimensional feature to the end of the feature vector.

[0056] For example, when performing format standardization, timestamp alignment, and label mapping operations, a sample's original temperature is 26.3℃, humidity is 72.5%, and total number of eggs is 1246; the minimum value of the temperature field in this batch is 17.2, and the maximum value is 33.8; the minimum value of humidity is 41.2, and the maximum value is 89.6; the minimum value of the total number of eggs is 112, and the maximum value is 1864; substituting these values ​​into the Min-Max scaling calculation:

[0057] Temperature scaling value Humidity scaling value Total number of eggs scaling value ; Original size of the corresponding insect egg image Scaling to uniform size Expanded to length A one-dimensional pixel vector; the control group encoding of this sample. Red-eyed bee strain Perform one-hot encoding to generate two segments with lengths of respectively , of - Feature vector; the sample collection time point corresponds to a complete second-level timestamp, filtering before and after the collection time point. All temperature and humidity records within the hourly interval, total Calculate the average temperature from a set of valid data. Average humidity Search the temperature-egg development comparison table. Corresponding egg development duration , scalar As an independent one-dimensional feature, it is concatenated to the end of the vector.

[0058] Optionally, the use The complete execution flow of rule-based outlier cleaning, missing value imputation, and multi-class feature concatenation includes three independent operations: The first operation calculates the arithmetic mean and standard deviation for each of the three independent field sequences: total egg count, ambient temperature, and primitive parasitism rate. The expression for the arithmetic mean is as follows: The standard deviation expression is: ,in Represents the arithmetic mean of a single-field sample sequence. Represents the standard deviation of a single-field sample sequence. This represents the total number of valid samples under this field. Representing the The first part iterates through the original values ​​of the field corresponding to each sample; the second part iterates through each field of a single sample, and if the value of that field differs from the sequence mean by more than three standard deviations, the entire sample is marked as an outlier. The judgment expression is: ;

[0059] The third paragraph distinguishes between two types of missing field filling rules: missing fields marked as outliers and missing fields of ordinary samples without anomalies are filled by matching the arithmetic mean of the same field of other samples in the same test batch and the same temperature range.

[0060] When performing feature concatenation, four types of vectors are concatenated in a strict fixed order: the first type is the normalized environment numerical vector, the second type is the one-dimensional pixel vector of the image, the third type is the one-hot encoded vector of the text, and the fourth type is the historical parasitic statistical vector. After concatenation, a single-sample fusion feature vector with fixed dimensions is generated. All entries with a total number of eggs within the sample equal to 0 are directly removed and do not enter the subsequent dataset.

[0061] Meanwhile, S2 has a built-in formula for calculating the true value of the original parasitism rate label:

[0062]

[0063] in, The original parasitic percentage was converted from manual counting for a single sample. Count the total number of parasitized eggs that have been labeled within the sample; The total number of eggs in the sample is counted. This data comes from the egg image data collected in step S1. The calculation result is used as the ground value label for training the deep learning regression model.

[0064] For example, when performing 3σ outlier cleaning and missing value imputation, a certain experimental batch has a total of 126 samples. The mean of the total number of eggs is 1162, the standard deviation is 214, and three times the standard deviation is 642. A certain sample has a total number of eggs of 1920, and the difference between the total number of eggs and the mean is 1920-1162=758>642. This sample is marked as an outlier. The historical natural mortality rate field of this sample is missing. It is matched with the mean natural mortality rate of the other 28 samples in the same batch in the 25-27℃ temperature range, which is 0.061, and filled into the missing field of this sample. The humidity field of another normal sample without abnormalities is empty. It is matched with the mean humidity of the samples in the same temperature range, which is 72.3, and filled.

[0065] The sample vectors were spliced ​​in the following order: normalized temperature and humidity, egg development duration environmental vector (3D) + 640×480 image 1D vector (307200D) + control group, bee species unique heat encoding vector (7D) + historical statistical vectors of egg quantity, original parasitism rate, and natural mortality rate over the years (3D). The total dimension of each spliced ​​single fused feature vector was 307213. There were 7 blank sampling entries with a total of 0 eggs in the sample, all of which were removed and not included in the standardized fused feature dataset. The number of parasitized eggs in this sample was manually counted. =528, total number of eggs =1246, substitute into the formula to calculate the original parasitism rate. =528÷1246×100%≈42.38%, this value is used as the ground truth label for model training.

[0066] Optionally, the S3 process of hierarchically dividing the standardized fused feature dataset into training, validation, and test sets, building a deep learning regression network, and performing early stopping iterative training includes: Hierarchical allocation rules: all samples are hierarchically divided according to two classification dimensions: temperature range and control group. The number of samples within each layer is strictly split according to a 7:2:1 ratio. After hierarchical splitting, all training subsets, validation subsets, and test subsets within each layer are summarized to obtain complete training, validation, and test sets. Network structure construction rules: the number of neurons in the input layer is exactly equal to the total dimension of a single fused feature vector; the first hidden layer is configured with 256 neurons, the second hidden layer with 128 neurons, and the third hidden layer with 64 neurons; all three hidden layers are equipped with ReLU activation and 0.25 ratio Dropout random deactivation; the output layer is configured with a single linear neuron, outputting a parasitic rate prediction value in the 0-100 range; the loss function adopts the mean squared error loss algorithm, with the mathematical expression:

[0067]

[0068] in, This represents the overall mean square error loss for a single iteration. This represents the total number of samples in the subset. For the network forward operation Predicted parasitic values ​​for each sample; The first number obtained by manual counting True values ​​of parasitism rates for each sample;

[0069] Early stopping rule: After each round of weight iteration, the training set and validation set are substituted into the loss function to calculate the corresponding loss value; if the loss value of the validation set does not decrease during ten consecutive rounds of iteration, the entire weight iteration process is terminated directly.

[0070] After the iteration terminates, three types of static parameter files are stored synchronously: the first type is a file containing the weight array of all neurons in the network; the second type is a file containing the Min-Max scaling coefficient table corresponding to each dimension of features; and the third type is a file containing a mapping table between temperature and egg development duration labels. The three types of static parameter files are merged and packaged to form an offline inference model that can be deployed and called offline.

[0071] For example, when performing dataset stratification and model training iteration, the standardized fusion feature dataset contains 1260 samples, divided into three layers: temperature range 18-22℃, 22-28℃, and 28-32℃. Each layer is further divided into two groups: blank control group (CK) and experimental group (TY), for a total of 6 layers. Taking the 420 samples in the 22-28℃ experimental layer as an example, they are split in a 7:2:1 ratio to obtain 294 training samples, 84 validation samples, and 42 test samples. After summing all the stratified subsets, the complete training set contains 882 samples, the validation set contains 252 samples, and the test set contains 126 samples.

[0072] The single fused feature vector has a dimension of 307213, the network input layer has 307213 neurons; the three hidden layers have 256, 128, and 64 neurons respectively, each with ReLU activation + 0.25 Dropout, and the output layer has 1 linear neuron;

[0073] The first iteration had a training set loss of 18.62 and a validation set loss of 19.15. In the 32nd iteration, the validation set loss dropped to a minimum of 10.27. From the 33rd to the 42nd iteration, the validation set loss was 10.29, 10.31, 10.35, 10.38, 10.42, 10.45, 10.49, 10.53, 10.57, and 10.61 respectively, showing no decrease throughout the iteration, thus triggering the early stopping rule and terminating the iteration.

[0074] The system synchronously saves the weight array file model_weight.pth, the Min-Max scaling coefficient table scale_coeff.csv, and the temperature-egg development mapping table temp_egg_map.xlsx. These three files are packaged and compressed into an offline inference model file infer_model.zip, retaining only the forward inference operation logic and removing the code related to backward gradient updates.

[0075] Optionally, the complete execution process of S4, which reuses the S1 and S2 processes to generate the feature vector of the sample to be tested, performs offline inference prediction, calculates the corrected parasitic rate, solves the confidence interval, and marks abnormal samples, includes:

[0076] The first step is to fully reproduce the S1 acquisition, timestamp binding, invalid data filtering, and media-specific storage operations, and to acquire environmental, image, management, historical, and time-series data of the sample under test to generate the original dataset of the sample under test.

[0077] The second step is to fully reproduce the entire operation process of S2 three-segment format unification, timestamp alignment, label mapping, 3σ outlier cleaning, and multi-vector feature concatenation, and generate a single independent fused feature vector for the test sample.

[0078] The fused feature vector is input into the offline inference model. The model only performs forward numerical propagation and disables the reverse weight update logic. After the forward propagation is completed, the predicted parasitic raw value (in percentage form) in the range of 0-100 is output.

[0079] Substituting the parasitic rate into the correction formula to eliminate interference from natural mortality in the control group, the correction formula is:

[0080]

[0081] in, The corrected percentage of parasites after removing interference from natural mortality; Output the predicted parasitic raw values ​​to the model; The percentage of natural deaths obtained from the blank control group, with a range of values. The confidence interval for the corrected parasitic rate is calculated based on a 95% confidence level. The formula for calculating the confidence interval is as follows:

[0082]

[0083] Where CI is the half-width of the corrected parasitic rate confidence interval at a 95% confidence level; 1.96 is the fixed quantile constant of the standard normal distribution at a 95% confidence level. The total number of eggs in the sample to be tested was counted, and this data came from the egg image data collected in step S1. Two fixed anomaly detection thresholds were set to mark abnormal samples: the image anomaly detection threshold was a single egg confidence value of less than 0.7; the environmental anomaly detection threshold was temperature. or temperature ; When any condition is triggered, a fixed anomaly identifier code is written into the sample record: image anomaly code IMG_ERR, environment anomaly code TEMP_ERR, and dual anomaly code ALL_ERR.

[0084] After batch calculations are completed, a structured calculation dataset is output. Each record has six fixed fields: unique sample code, total number of eggs in the sample, predicted parasitic value of the network, corrected parasitic rate value, upper limit of confidence interval, lower limit of confidence interval, and anomaly identification code.

[0085] For example, during the offline inference and correction calculation of the test sample, the code SMP009-12 of the test sample was collected on-site, and the S1 and S2 processes were reproduced to generate a fusion feature vector with a dimension of 307213; the input was used for forward operation of the offline inference model, and the output was the predicted original value of parasitism of 43.12%; the percentage of natural mortality in the same batch of blank control group was retrieved. Substitute into the correction formula: ;

[0086] The total number of eggs in this sample Substitute into the confidence interval formula to calculate the half-width:

[0087] That is, the half-width of the confidence interval is 2.69%;

[0088] The final confidence interval has a lower limit of 39.42 - 2.69 = 36.73% and an upper limit of 39.42 + 2.69 = 42.11%.

[0089] The average confidence level for identifying a single egg in sample A was 0.64 < 0.7, and the average ambient temperature was 33.5°C > 32°C. Both abnormal conditions were triggered simultaneously, and the abnormality was recorded in the code ALL_ERR.

[0090] The complete fields of this structured measurement record are: SMP009-12, 1312, 43.12, 39.42, 42.11, 36.73, ALL_ERR.

[0091] Optionally, the complete process of S5 performing statistical analysis, generating visualization charts, outputting multi-format reports, and locking and archiving on dual media includes:

[0092] Statistical test operation rules: One-way ANOVA is used for between-group difference analysis, and Tukey HSD multiple comparisons are used for pairwise differences between multiple groups. The variance decomposition table and p-values ​​for multiple comparisons are fully saved. Visualization chart generation rules: Two types of fixed visualization graphs are output in batches. The first type is a parasitism rate time-series line chart, with the horizontal axis representing sampling time / sample number and the vertical axis representing the corrected parasitism rate, distinguishing between different experimental groups. The second type is a group confidence interval bar chart, with the horizontal axis representing the experimental group and the bar height representing the mean corrected parasitism rate for each group. Confidence interval error bars are marked above and below the bars.

[0093] Multi-format report generation rules: Two types of independent files are output simultaneously. The first type is a PDF report, which fully embeds time series line charts, confidence interval bar charts, variance analysis detail tables, and multiple comparison results tables. The second type is an Excel file, which stores the original collected fields, model predicted parasitic values, corrected parasitic rates, upper and lower limits of confidence intervals, anomaly identification codes, and variance statistics results in separate worksheets.

[0094] Dual-media storage and locked archiving rules: Set up two independent storage partitions: the first is a local solid-state storage partition, and the second is a cloud object storage partition; the original monitoring data table, standardized fusion feature vector file, offline inference model file, batch structured measurement dataset, PDF report, and Excel report are all synchronously written to the two storage partitions;

[0095] After the archive locking operation is completed, all archived files are switched to read-only permissions, prohibiting modification, deletion, and overwriting. Each archived record is bound to three sets of independent index fields: project code, test batch code, and sample unique code. It supports one-click retrieval of all related data records of the entire process of sample collection, cleaning, modeling, inference, and reporting through any set of index fields.

[0096] For example, during the statistical analysis and dual-media archiving process, this batch consisted of 12 repeated trials, including a CK blank control group and three placement groups (TY1-TY3). A one-way ANOVA was performed on the structured dataset, yielding a p-value of 0.0023 < 0.05, indicating that different bee egg placement ratios significantly affected parasitism rates. TukeyHSD multiple comparisons showed that the mean corrected parasitism rate of group TY2 was significantly higher than that of groups CK, TY1, and TY3. The system automatically plotted a time-series line graph of parasitism rates, with the horizontal axis representing sample numbers 1-12, and the four lines corresponding to the corrected parasitism rates of groups CK, TY1, TY2, and TY3, respectively. A group confidence interval histogram was also plotted, with the four bars labeled with the upper and lower error bars of the corresponding 95% confidence interval. A PDF report was generated. The file _P007.pdf contains two embedded visualization charts, a variance decomposition detail, and a table of significance for multiple comparisons using letter markers. An Excel report, data_P007.xlsx, is generated, consisting of three worksheets: the original collected data table, the batch calculation results table, and the statistical test details table. All files are synchronously written to the local solid-state archive partition and the corresponding project path in cloud object storage. After the archiving process is complete, the system automatically changes the permissions of all files to read-only. The archive index record is bound to project code JY202606, test batch P007, and sample code SMP009-12. Entering the project code retrieves all batch data under that project; entering the test batch retrieves all samples, models, and reports for that batch; entering the sample code retrieves the entire data chain from collection to report for that single sample.

[0097] Optionally, the method further includes: extracting the weights of key neurons related to egg images and ambient temperature in the offline inference model; analyzing the temporal iterative change trend of the weights of such key neurons; if the weights continue to decay / increase unidirectionally during the iteration process, the overall confidence of the parasitism rate calculation results of this batch is improved.

[0098] Specifically, firstly, from the weight array of neurons in the three hidden layers, key neurons corresponding to the normalized feature dimensions of image pixels and temperature in the input layer are selected. These neurons account for 1%-3% of the total number of neurons in the hidden layers. After each iteration, the average weight of this batch of key neurons is automatically recorded. The formula for calculating the average weight is as follows: ,in This represents the average weight of all key neurons in a single iteration. This represents the total number of key neurons selected. Representing the The weight values ​​corresponding to the current iteration of each key neuron are recorded; the average weight value record covering all iterations is maintained, and the weight time series is fitted using linear regression to obtain the fitting slope. The expression for solving the linear regression slope is:

[0099]

[0100] in The slope of the linear fit of the weighted time series is represented by the following: This represents the total number of complete iterations of the model. Represents the iteration round index. Representing the The average weight of the key neurons corresponding to each iteration; if the absolute value of the fitting slope is greater than the preset micro-engineering threshold of 0.001, it is determined that the weight has a continuous unidirectional changing trend; this trend indicates that the model has fully and stably learned the core features of images and temperature, and the confidence level of all samples in this batch is uniformly increased to 99%, and a high confidence level label is added to the PDF report to provide highly reliable data support for the selection of bee species and optimization of bee release ratio in field trials.

[0101] For example, in the process of improving the confidence level of the measurement, the offline inference model has a total of 448 neurons in three hidden layers. Nine key neurons associated with image pixels and temperature features were selected, accounting for 2.01% of the total. During the 42 rounds of iteration, the average weight of the nine neurons was recorded in each round to construct a 42-point time series. Linear regression was performed on the series, and the fitting slope was +0.0032. The absolute value was greater than the engineering threshold of 0.001, indicating that the weight of the key neurons showed a continuous and stable growth trend, proving that the model had learned the image and temperature features sufficiently. The system improved the confidence level of all 1260 samples in this batch to 99%, and added a label on the first page of the PDF report indicating that the model features were well fitted and the measurement results were highly confident, providing reliable data for the optimization experiment of the Trichogramma wasp release ratio.

[0102] Optionally, after generating the standardized fused feature dataset and before hierarchically dividing the training set, the method further includes: reading the natural mortality percentage sequence of the control group in all historical sample datasets over the years and calculating the arithmetic mean of the sequence; storing this global average natural mortality percentage as a backup correction parameter in the static parameter file of the offline inference model; when the blank control group data of the test sample is missing and cannot be obtained. When the global mean is automatically retrieved, it is used as the input parameter for the correction formula, ensuring that the correction parasitism rate can be calculated even without a control group sample.

[0103] Specifically, the system compiles data on the natural mortality rate of all plots and all batches of blank control groups annually, generates a global natural mortality rate sequence, updates the global mean once a year and writes it into the temperature-egg development mapping table; during the online inference stage, it automatically determines whether the sample to be tested matches the control group data of the same batch. If the control group sample is damaged or sampling is missing, it directly reads the global mean and substitutes it into the correction formula to avoid interruption of the correction parasitism rate calculation due to the lack of control group data, thus expanding the applicable scenarios of the method.

[0104] For example, in the process of configuring backup natural mortality correction parameters, the system summarizes the natural mortality rate of 127 experimental batches in blank control groups over many years, and the sequence mean is calculated to be 0.0608; the value 0.0608 is written into the backup parameter field of the offline inference model temp_egg_map.xlsx; a certain experimental batch of the sample to be tested did not have a blank control group set, so the batch-specific parameters could not be obtained. The system automatically retrieves the global mean of 0.0608 and substitutes it into the correction formula to complete the calculation, ensuring uninterrupted operation of the process.

[0105] Optionally, the hardware synchronization acquisition rules for S1 to simultaneously acquire multiple types of raw monitoring data include: using industrial temperature and humidity sensors, high-definition microscope cameras, handheld field sampling terminals, database historical data interfaces, and operation record terminals to acquire five types of data in parallel; using a unified server master clock signal to synchronously trigger the analog-to-digital conversion and data writing actions of all acquisition devices; the image camera sampling frame rate is not less than 5 frames / minute, the temperature and humidity sensor sampling frequency is 1 time / hour, and the terminal operation record is synchronized with millisecond-level timestamps; all devices are configured with a unified time synchronization verification mechanism for the data writing link, and after acquisition, the timestamp deviation of each device is verified, and data with a deviation greater than 1 second is directly marked as invalid and discarded.

[0106] Specifically, the entire data acquisition hardware is synchronized by a central server. Each data write from the temperature and humidity sensor, microscope camera, sampling terminal, and operation terminal synchronously reads the server's global timestamp. The microscope camera continuously captures images of insect eggs to preserve complete spatial features, the temperature and humidity sensor stably collects temporal environmental fluctuation data, the field terminal enters static management information such as batch and latitude / longitude, the historical interface retrieves sample data from previous years to supplement temporal features, and the operation terminal records the operator and operation time. The acquisition link automatically checks the deviation between the local clock of each device and the server's master clock every hour. If the deviation between the device's local time and the server's timestamp exceeds 1 second during a single acquisition, the acquired data is directly determined to be invalid and will not be included in the standardized raw dataset, thus eliminating feature fusion errors caused by temporal misalignment.

[0107] For example, during the multi-source hardware synchronous acquisition process, the central server's master clock corresponds to a complete second-level timestamp. The local clock deviations of the temperature and humidity sensor are 0.3s, the microscope camera deviation is 0.5s, and the sampling terminal deviation is 0.2s, all less than 1 second, and the acquired data is effectively retained. The local clock deviation of another backup microscope camera is 1.8s, exceeding the 1-second threshold. All insect egg image data captured by this camera are directly marked as invalid and not entered into the standardized original dataset to avoid time sequence misalignment interfering with subsequent timestamp alignment and feature fusion calculations.

[0108] Based on the same inventive concept, such as Figure 2 As shown, the present invention also provides a parasitic rate intelligent calculation system based on multi-source data fusion, the system being used to execute the parasitic rate intelligent calculation method based on multi-source data fusion described in any of the above claims, the system comprising:

[0109] Multi-source data synchronous acquisition module: used to synchronously acquire multiple types of raw monitoring data, including breeding environment, insect egg images, production management, historical samples, and time-series operations. It uniformly binds global timestamps accurate to the second, filters out invalid data according to rules, stores images in partitions and numerical data in separate data tables, and outputs standardized raw datasets.

[0110] Feature fusion and label construction module: It is used to perform three-stage numerical normalization, image vectorization, and one-hot text encoding on the standardized original dataset, complete timestamp alignment, egg development duration label mapping, 3σ outlier sample cleaning, missing value imputation, and multi-vector ordered concatenation. It calculates the true label based on the original parasitism rate formula using artificial egg count, and outputs the standardized fused feature dataset.

[0111] Model training and generation module: This module is used to split the fused feature dataset into training, validation, and test subsets in a 7:2:1 ratio according to temperature range and control group, build a fixed three-layer hidden deep learning regression network, iterate training with mean squared error loss function, terminate the iteration according to the rule that the loss does not decrease after ten consecutive rounds of validation, synchronously store three types of static parameters: network weights, Min-Max scaling factor, and egg development duration comparison table, and encapsulate and generate an offline inference model.

[0112] The inference correction and anomaly labeling module is used to fully reuse the logic of the acquisition module and the feature fusion module to generate the fusion feature vector of the sample to be tested. It only calls the forward pass operation of the offline model to output the predicted parasitic value, and substitutes it into the correction parasitic rate formula and the 95% confidence interval formula to complete the calculation. Based on the image recognition confidence threshold and the environmental temperature boundary threshold, it labels the image and the environmental anomaly sample, and outputs a standardized batch structured calculation dataset.

[0113] The statistical analysis and report archiving module is used to perform one-way ANOVA and TukeyHSD multiple comparison statistical operations on structured measurement datasets, automatically generate parasitic rate time series line charts and grouped confidence interval bar charts, and synchronously output PDF reports with embedded charts and tables, as well as full-field Excel detailed reports. It synchronously stores all raw data, feature files, offline models, measurement datasets, and reports into local solid-state and cloud object dual storage partitions, performs read-only locking and archiving of execution files, and binds three sets of index fields—project code, test batch, and sample unique code—to achieve full-process data traceability and retrieval.

[0114] It should be noted that the electrical connections and program calling logic between the various units described above do not necessarily represent direct calls to a single line or function, indirect data transmission, or cross-module interface calls. Any method that achieves the purpose of this invention can be applied to the embodiments of this invention. The above descriptions are merely exemplary embodiments of this invention and should not be construed as limiting the scope of this invention.

[0115] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for intelligent calculation of parasitic rate based on multi-source data fusion, characterized in that, The specific steps of this method are as follows: S1. Simultaneously collect multiple types of raw monitoring data, bind timestamps to the collected data and filter out invalid data to obtain a standardized raw dataset; S2. Perform multi-source feature fusion and data cleaning operations on the standardized original dataset, calculate the parasitism rate true label based on the egg count, and construct a standardized fused feature dataset. S3. Train a deep learning regression model using the standardized fusion feature dataset to obtain an offline inference model; S4. Reuse steps S1 and S2 to generate the fusion feature vector of the sample to be tested, input it into the offline inference model to obtain the predicted parasitic rate value, use the correction algorithm and confidence interval algorithm to calculate the corrected parasitic rate and confidence interval, and mark the abnormal samples. S5. Perform statistical analysis on the data obtained in step S4, generate statistical charts and multi-format reports, and perform dual-media storage and locked archiving on the reports.

2. The intelligent parasitic rate calculation method based on multi-source data fusion according to claim 1, characterized in that, In step S1, the synchronously collected data includes aquaculture environment data, insect egg image data, production management data, historical sample data, and time-series monitoring data. The aquaculture environment data includes hourly collected temperature values, hourly collected humidity values, preset bee egg release ratio, bee arrival start time, bee arrival end time, and yellow field borer egg development duration. The insect egg image data includes image pixel size parameters, unique image code, horizontal coordinate of a single egg, vertical coordinate of a single egg, width of a single egg, height of a single egg, and the total number of manually counted eggs. The production management data includes project number, test batch number, number of repeated test groups, control group type code, sampling latitude, and sampling longitude. The historical sample data includes the total number of eggs per sample over the years, the calculated value of primary parasitism over the years, and the percentage of natural mortality in the control group over the years. The time-series monitoring data includes operator identification codes, operation timestamps accurate to the second, and unique codes for sample binding; the acquired image files are stored in independent file partitions, and numerical data are stored in relational data tables, with each data table recording a global timestamp accurate to the second.

3. The intelligent parasitic rate calculation method based on multi-source data fusion according to claim 1, characterized in that, In S2, the complete execution process of cleaning outlier samples using the 3σ rule includes three independent operations: the first operation calculates the arithmetic mean and standard deviation of the three independent sequences: total egg count, ambient temperature, and original parasitism rate; the second operation marks samples whose single field value differs from the mean of that field sequence by more than three times the standard deviation as outlier samples; and the third operation fills the missing fields marked as outlier samples and the missing fields of unmarked ordinary samples by matching the arithmetic mean of the same field from other samples in the same experimental batch and the same temperature range.

4. The intelligent parasitic rate calculation method based on multi-source data fusion according to claim 1, characterized in that, In S3, the standardized fusion feature dataset is divided into training, validation, and test sets according to a 7:2:1 hierarchical allocation rule. All samples are hierarchically divided according to two classification dimensions: temperature range and control group. The number of samples within each layer is split according to a fixed numerical ratio of 7:2:

1. After the hierarchical split is completed, all corresponding subsets within each layer are summarized to obtain the complete training set, complete validation set, and complete test set. The number of neurons in each layer of the constructed network is fixed. The total number of neurons in the input layer is equal to the dimension of a single fusion feature vector. The first hidden layer is configured with 256 neurons, the second hidden layer with 128 neurons, and the third hidden layer with 64 neurons. All three hidden layers are equipped with ReLU activation operation logic and 0.25 ratio Dropout operation logic. The output layer is configured with a single linear neuron.

5. The intelligent parasitic rate calculation method based on multi-source data fusion according to claim 1, characterized in that, In S3, the early stop determination rule includes fixed determination conditions; After each round of weight iteration calculation, the training set and validation set are read and substituted into the mean squared error loss algorithm to obtain the corresponding loss value. When the loss value of the validation set does not decrease during ten consecutive rounds of iteration calculation, the entire weight iteration process is terminated. After the iteration process is terminated, three types of static parameter files are stored synchronously. The first type is the weight array of all neurons in the network, the second type is the Min-Max scaling coefficient corresponding to each dimension feature, and the third type is the temperature and egg development duration label mapping table. The three types of static parameter files together form the offline inference model.

6. The intelligent parasitic rate calculation method based on multi-source data fusion according to claim 1, characterized in that, In S4, the two types of judgment thresholds for labeled images and environmental anomaly samples are fixed; the image anomaly judgment threshold is that the confidence value of the intelligent identification single egg is less than 0.7, and the environmental anomaly judgment threshold includes two types of boundary conditions: temperature value less than 18 degrees Celsius and temperature value greater than 32 degrees Celsius; after any judgment condition is triggered, a fixed anomaly identifier code is written into the corresponding sample record. Each record in the generated batch structured measurement dataset includes six fixed fields: unique sample code, total number of eggs in the sample, predicted parasitic value from network output, value from the corrected parasitic rate algorithm, two sets of values ​​for the upper and lower limits of the confidence interval, and anomaly identification code.

7. The intelligent parasitic rate calculation method based on multi-source data fusion according to claim 1, characterized in that, In S4, the mathematical expression for the parasitic rate correction algorithm is: Within the expression, This represents the corrected percentage of parasites after removing interference from natural mortality. This is the percentage calculation result output by the original parasitism rate algorithm; The percentage of naturally occurring oocytes obtained from the blank control group sample, and .

8. The intelligent parasitic rate calculation method based on multi-source data fusion according to claim 1, characterized in that, In S4, the mathematical expression for the confidence interval statistical algorithm is: Within the expression, The half-width of the confidence interval representing the corrected parasitism rate at a 95% confidence level; To correct the percentage output of the parasitic rate algorithm; Output percentage results for the original parasitism rate algorithm; 1.96 is the fixed quantile constant of the standard normal distribution with a 95% confidence level match; This represents the total number of all eggs in a single sample.

9. The intelligent parasitic rate calculation method based on multi-source data fusion according to claim 1, characterized in that, In step S5, the analysis of variance performed adopts the one-way variance operation rule, and the multiple comparison operation adopts the TukeyHSD operation rule. The output statistical charts include two types of fixed visualization graphics: the first type is a parasitism rate time series line chart, and the second type is a grouped confidence interval bar chart. The generated multi-format report includes two types of independent files: the first type is a PDF file, which embeds all statistical charts and variance operation details tables, and the second type is an Excel file, which stores the original collection fields, original parasitism rate calculation values, corrected parasitism rate calculation values, and upper and lower limits of the confidence interval for each sample in rows.

10. A parasitic rate intelligent calculation system based on multi-source data fusion, the system being applicable to the parasitic rate intelligent calculation method based on multi-source data fusion as described in any one of claims 1-9, characterized in that, The system includes: The multi-source data synchronous acquisition module is used to synchronously acquire multiple types of raw monitoring data, bind timestamps to the acquired data and filter out invalid data to obtain a standardized raw dataset; The feature fusion and label construction module is used to perform multi-source feature fusion and data cleaning operations on the standardized original dataset, calculate the parasitism rate true label based on the egg count, and construct a standardized fused feature dataset. The model training and generation module is used to train a deep learning regression model using the standardized fusion feature dataset, obtain an offline inference model, and store the network weight array, Min-Max scaling factor, and temperature-egg development duration label mapping table. The inference correction and anomaly labeling module is used to reuse the logic of the multi-source data synchronous acquisition module and the feature fusion and label construction module to generate the fusion feature vector of the sample to be tested, input the offline inference model to obtain the predicted parasitic rate value, use the correction algorithm and confidence interval algorithm to calculate the corrected parasitic rate and confidence interval, and label the image and environmental anomaly samples. The statistical analysis and report archiving module is used to perform variance analysis and multiple comparison operations on the data output by the inference correction and anomaly marking module, generate statistical charts and multi-format reports, and perform dual-media storage of the reports on local solid-state storage and cloud object storage, as well as lock archiving operations.

Images