Intelligent interpretation system and method of six-in-one mycotoxin test card based on multidimensional data analysis

Abstract

The present invention relates to the technical field of biological toxin detection, and discloses a system and method for intelligent interpretation of a six-in-one mycotoxin detection card based on multidimensional data analysis. The method comprises collecting first toxin detection data of six mycotoxins, including fluorescence response time series, colloidal gold response time series and environmental variable time series, and performing feature extraction to obtain second toxin detection data; then establishing a historical detection database, constructing and training a mycotoxin classification model, and obtaining preliminary identification results of the six mycotoxins based on the second toxin detection data and the classification model; finally, weighted fusion of the preliminary identification results is performed in combination with the environmental variable time series to obtain accurate identification results; the present invention realizes intelligent interpretation of mycotoxin detection and greatly improves detection accuracy and efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of biological toxin detection, and more specifically, to a system and method for intelligently interpreting a six-in-one mycotoxin detection card based on multidimensional data analysis. Background Art

[0002] In order to effectively monitor and control mycotoxins in agricultural products such as grain and feed, it is urgent to develop fast, convenient and accurate mycotoxin detection technology. At present, some rapid mycotoxin detection methods have been proposed. For example, the Chinese patent application with publication number CN109587337A discloses a method and a smart phone for rapid detection of mycotoxins. The invention uses a smart phone camera to scan the QR code of the test paper to identify the type of mycotoxin to be detected, and then takes the image of the test paper after the test sample is added, performs image recognition and analysis, obtains the mycotoxin concentration, and uploads the test results to a remote device. This method uses a smart phone to achieve portable and rapid detection, but this method mainly relies on the color change of the test paper image to judge the toxin concentration, which is easily affected by the quality of the test paper and the shooting environment; the single-dimensional detection data limits the further improvement of the detection accuracy.

[0003] Another Chinese patent application with publication number CN115774101A proposes a high-sensitivity method for simultaneous rapid detection of multiple mycotoxins. This method synthesizes complete fumonisin antigens, prepares high-affinity monoclonal antibodies, and establishes a rapid detection method based on enzyme-linked immunosorbent assay and nanogold probes. This method has made a breakthrough in the sensitivity of single toxin detection, but for the scenario of simultaneous detection of multiple mycotoxins, multiple separate tests are still required, which makes it difficult to meet the needs of high-throughput, multi-dimensional toxin detection.

[0004] In summary, the existing rapid mycotoxin detection technology has a single detection dimension and is difficult to fully reflect the mycotoxin contamination status in complex matrices; it is easily interfered by the external environment, and the detection stability and reliability need to be improved; it lacks intelligent interpretation of test results and cannot fully explore the internal connections of test data. Summary of the invention

[0005] In order to overcome the above-mentioned defects of the prior art, the present invention provides a system and method for intelligent interpretation of a six-in-one mycotoxin detection card based on multidimensional data analysis. The method collects multi-dimensional detection data such as fluorescence response, colloidal gold response and environmental variables of six kinds of mycotoxins, combines machine learning algorithms, and establishes a mycotoxin classification model, thereby achieving highly accurate and robust intelligent interpretation of mycotoxins.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] Intelligent interpretation method of the six-in-one mycotoxin test card based on multidimensional data analysis, including:

[0008] Collecting first toxin detection data of six fungal toxins, performing feature extraction on the first toxin detection data, and obtaining second toxin detection data; the first toxin detection data includes a fluorescence response time series, a colloidal gold response time series, and an environmental variable time series;

[0009] Establish a historical detection database, construct and train a mycotoxin classification model, and obtain preliminary identification results of six mycotoxins based on the second toxin detection data and the mycotoxin classification model;

[0010] Based on the preliminary identification results of six fungal toxins, accurate identification results were obtained by combining the time series of environmental variables.

[0011] Further, the second toxin detection data includes fluorescence signal intensity characteristics, fluorescence response rate characteristics, colloidal gold color change rate characteristics and environmental variable volatility characteristics;

[0012] The feature extraction of the first toxin detection data includes:

[0013] Extracting fluorescence signal intensity features and fluorescence response rate features from the fluorescence response time series;

[0014] Extract the color change rate characteristics of colloidal gold from the response time series of colloidal gold;

[0015] Extract environmental variable volatility characteristics from environmental variable time series;

[0016] Extracting the fluorescence signal intensity feature and the fluorescence response rate feature from the fluorescence response time series includes:

[0017] Calculate the maximum value, minimum value, average value and integrated area of ​​the fluorescence response time series to obtain the fluorescence signal intensity characteristics;

[0018] The slope, rise time, and fall time of the fluorescence response timing are calculated to obtain the fluorescence response rate characteristics.

[0019] Further, the extracting of the colloidal gold color change rate feature from the colloidal gold response time series includes:

[0020] Perform differential operation on the colloidal gold response time series to obtain the color change speed sequence;

[0021] Calculate the maximum value, minimum value, average value and standard deviation of the color change rate sequence to obtain the color change rate characteristics of colloidal gold;

[0022] Extracting environmental variable volatility features from the environmental variable time series includes:

[0023] Perform sliding window analysis on the environmental variable time series to obtain the volatility series;

[0024] The maximum, minimum, average and dynamic range of the volatility series are calculated to obtain the volatility characteristics of the environmental variables.

[0025] Furthermore, the preliminary identification results of the six mycotoxins obtained based on the second toxin detection data and the mycotoxin classification model include:

[0026] The second toxin detection data is predicted using the mycotoxin classification model to obtain the first recognition results of the six mycotoxins; the first recognition results are the prediction probabilities of the six mycotoxins. , , , , , ];in represents the predicted probability of the i-th mycotoxin among the six mycotoxins, 1≤i≤6;

[0027] Setting the confidence threshold , represents the confidence threshold of the i-th mycotoxin among the six mycotoxins; if the predicted probabilities of the six mycotoxins are all greater than the confidence thresholds of the corresponding mycotoxins, the first recognition result is taken as the preliminary recognition result; otherwise, reference data similar to the second toxin detection data is retrieved from the historical detection database;

[0028] Performing feature fusion on the second toxin detection data and the reference data to obtain a toxin feature fusion vector;

[0029] The toxin feature fusion vector was input into the mycotoxin classification model to obtain preliminary identification results.

[0030] Furthermore, the confidence threshold is set include:

[0031] Through the historical detection database, calculate the The accuracy of mycotoxins and recall ;

[0032] Obtain the detection feature of the i-th mycotoxin from the second toxin detection data and calculate the Sensitivity scores of mycotoxin detection features;

[0033] Calculate the predicted probability of the i-th mycotoxin The predicted probability of the jth mycotoxin The correlation coefficient ρ ij , 1≤j≤6, i≠j.

[0034] Further, the historical detection database includes n1 records;

[0035] The retrieving reference data similar to the second toxin detection data from the historical detection database comprises:

[0036] Calculate the similarity between the second toxin detection data and each record in the historical detection database to obtain a similarity list;

[0037] Sort the similarity list and select the record with the highest similarity as the reference data.

[0038] Furthermore, the feature fusion of the second toxin detection data and the reference data includes:

[0039] Construct a toxin feature fusion model based on a convolutional neural network, and set up the input layer, convolution layer, pooling layer, fully connected layer, and output layer;

[0040] splicing the second toxin detection data with the reference data to form a toxin data fusion matrix;

[0041] The toxin data fusion matrix is ​​input into the toxin feature fusion model to obtain the toxin feature fusion vector.

[0042] Furthermore, the step of inputting the toxin data fusion matrix into the toxin feature fusion model to obtain the toxin feature fusion vector includes:

[0043] Input the toxin data fusion matrix into the input layer of the toxin feature fusion model, and convert the toxin data fusion matrix into a one-dimensional toxin feature vector;

[0044] The convolution layer performs a convolution operation on the one-dimensional toxin feature vector through a convolution kernel to extract local toxin features;

[0045] The pooling layer compresses the extracted local toxin features to obtain toxin compression features;

[0046] The compressed features of toxins are flattened and nonlinearly transformed in the fully connected layer to generate global toxin high-level features;

[0047] The output layer outputs the global toxin high-level features as a toxin feature fusion vector.

[0048] Furthermore, obtaining an accurate recognition result includes:

[0049] Perform pattern recognition on the time series of environmental variables to obtain the environmental state sequence;

[0050] Calculate the historical accuracy of the preliminary recognition results under different environmental states in the environmental state sequence to obtain the environmental adaptive weight matrix;

[0051] The environment-adaptive weight matrix is ​​used to perform weighted fusion on the preliminary recognition results to obtain accurate recognition results;

[0052] The obtaining of the environmental state sequence comprises:

[0053] Cluster the time series of environmental variables to obtain environmental state categories;

[0054] Perform time series labeling on the environmental state categories to obtain the environmental state sequence;

[0055] The obtaining of the environment adaptive weight matrix comprises:

[0056] Count the accuracy of the preliminary recognition results under each environmental state in the historical detection database to form an environmental state-accuracy mapping table;

[0057] The environment state-accuracy mapping table is converted into an environment adaptive weight matrix, wherein the matrix elements in the environment adaptive weight matrix are weight coefficients of the preliminary recognition results under the corresponding environment state.

[0058] The intelligent interpretation system of the six-in-one mycotoxin detection card based on multidimensional data analysis is used to implement the above-mentioned intelligent interpretation method of the six-in-one mycotoxin detection card based on multidimensional data analysis, and the system comprises:

[0059] Data preprocessing module: used to collect first toxin detection data of six fungal toxins, perform feature extraction on the first toxin detection data, and obtain second toxin detection data; the first toxin detection data includes fluorescence response time series, colloidal gold response time series and environmental variable time series;

[0060] Preliminary identification module: used to establish a historical detection database, build and train a mycotoxin classification model, and obtain preliminary identification results of six mycotoxins based on the second toxin detection data and the mycotoxin classification model;

[0061] Precise identification module: Based on the preliminary identification results of six fungal toxins and combined with the time series of environmental variables, accurate identification results are obtained.

[0062] An electronic device comprises a memory, a central processing unit and a computer program stored in the memory and executable on the central processing unit. When the central processing unit executes the computer program, the above-mentioned intelligent interpretation method of the six-in-one mycotoxin detection card based on multidimensional data analysis is implemented.

[0063] A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed, the above-mentioned method for intelligent interpretation of the six-in-one mycotoxin detection card based on multidimensional data analysis is implemented.

[0064] Compared with the prior art, the present invention has the following beneficial effects:

[0065] The intelligent interpretation system and method of the six-in-one mycotoxin detection card of the present invention comprehensively utilizes the multi-dimensional data information obtained by the two detection methods of fluorescence method and colloidal gold method, and realizes the efficient and accurate identification of six key mycotoxins through intelligent algorithms. Compared with the traditional manual interpretation method, this method reduces human errors and improves the consistency and reliability of detection quality. At the same time, the method fully considers the influence of environmental factors and introduces an environmental adaptive mechanism, so that the test results can dynamically adapt to different application scenarios. In addition, the intelligent interpretation method of the present invention is versatile and can be easily transplanted to other types of multi-link detection cards, and has broad application prospects. In short, the present invention realizes the automation and intelligence of mycotoxin detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0067] Figure 1 It is a principle flow chart of the intelligent interpretation method of the six-in-one mycotoxin detection card based on multidimensional data analysis in the present invention;

[0068] Figure 2 It is a flow chart of a method for extracting features of first toxin detection data in the intelligent interpretation method of the six-in-one mycotoxin detection card based on multidimensional data analysis of the present invention;

[0069] Figure 3 It is a flow chart of a method for extracting fluorescence signal intensity characteristics and fluorescence response rate characteristics from fluorescence response time series in the intelligent interpretation method of the six-in-one mycotoxin detection card based on multidimensional data analysis of the present invention;

[0070] Figure 4 It is a flow chart of a method for extracting the color change rate characteristics of colloidal gold from the colloidal gold response time series in the intelligent interpretation method of the six-in-one mycotoxin detection card based on multidimensional data analysis of the present invention;

[0071] Figure 5 It is a flow chart of a method for extracting environmental variable volatility characteristics from environmental variable time series in the method for intelligent interpretation of the six-in-one mycotoxin detection card based on multidimensional data analysis of the present invention;

[0072] Figure 6 A flow chart of a method for obtaining preliminary identification results of six mycotoxins in the intelligent interpretation method of the six-in-one mycotoxin detection card based on multidimensional data analysis of the present invention;

[0073] Figure 7 A flow chart of a method for retrieving reference data similar to the second toxin detection data from a historical detection database in the intelligent interpretation method of the mycotoxin six-in-one detection card based on multidimensional data analysis of the present invention;

[0074] Figure 8 It is a flow chart of a method for obtaining a toxin feature fusion vector in the intelligent interpretation method of the six-in-one mycotoxin detection card based on multidimensional data analysis of the present invention;

[0075] Fig. 9 A flow chart of a method for obtaining accurate identification results in the intelligent interpretation method of the mycotoxin six-in-one test card based on multidimensional data analysis of the present invention;

[0076] Fig.10 This is a functional module diagram of the intelligent interpretation system of the six-in-one mycotoxin detection card based on multidimensional data analysis in the present invention. DETAILED DESCRIPTION

[0077] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0078] Example 1

[0079] See also Figure 1 As shown, this embodiment provides an intelligent interpretation method of a six-in-one mycotoxin test card based on multidimensional data analysis, including:

[0080] Step S1000, collecting first toxin detection data of six fungal toxins, performing feature extraction on the first toxin detection data, and obtaining second toxin detection data;

[0081] Further, step S1000 includes:

[0082] Step S1100, collecting first toxin detection data of six fungal toxins, wherein the first toxin detection data includes a fluorescence response time series, a colloidal gold response time series, and an environmental variable time series;

[0083] Specifically, the Mycotoxin 6-in-1 Test Card is an innovative multi-test tool specifically designed to detect six important mycotoxins commonly found in grains, oils, feed ingredients and compound feeds. These toxins include aflatoxin B1, vomitoxin, zearalenone, fumonisins (B1 and B2), ochratoxin A and T-2 toxin. The presence of these toxins poses a serious threat to food security and human and animal health, so the development of efficient and accurate testing tools is of great significance.

[0084] The design of the test card integrates two detection methods: fluorescence method and colloidal gold method. Among them, aflatoxin B1, vomitoxin and zearalenone are detected by fluorescence method; fumonisins (B1 and B2), ochratoxin A and T-2 toxin (T-2) are detected by colloidal gold method. This combination of two methods not only improves the sensitivity of detection, but also solves the problem that a single method is difficult to detect multiple toxins at the same time. In terms of technical design, the six-in-one mycotoxin test card uses the detection technology of red microspheres and time-resolved fluorescent microspheres, and supports dual-function detection instrument readings.

[0085] The Mycotoxin 6-in-1 Test Card is an efficient, accurate and versatile testing tool. It can detect six major mycotoxins at the same time and has a wide range of applicability in practical applications. It can not only be used for professional testing in laboratories and industrial fields, but can also be developed into a home version to meet the needs of ordinary users.

[0086] The fluorescence response in the fluorescence response time series includes the fluorescence signal intensity, the fluorescence signal appearance time and the fluorescence signal duration; the colloidal gold response in the colloidal gold response time series includes the color change degree, the color change rate and the color change time; the environmental variables in the environmental variable time series include temperature, humidity and light intensity.

[0087] The six-in-one mycotoxin test card was scanned using a dual-function detection instrument to obtain fluorescence response time series data. The fluorescence response time series data reflects the process of the fluorescence signals of the three toxins, aflatoxin B1, vomitoxin and zearalenone, changing over time, including information such as fluorescence signal intensity, fluorescence signal appearance time, and fluorescence signal duration. These data can be used to analyze toxin concentration, reaction kinetics, etc. The fluorescence signal intensity indicates the light intensity emitted by the fluorescent substance at a specific moment, which is positively correlated with the concentration of the analyte. By analyzing the changing trend of the fluorescence signal intensity, the concentration of the toxin can be inferred. The fluorescence signal appearance time indicates the time point when the fluorescence signal intensity begins to rise, which is negatively correlated with the concentration of the analyte. The higher the concentration, the earlier the fluorescence signal appears. By comparing the fluorescence signal appearance time of different toxins, the type of toxin can be preliminarily determined. The fluorescence signal duration indicates the length of time that the fluorescence signal intensity remains high, which is positively correlated with the concentration of the analyte. The higher the concentration, the longer the fluorescence signal duration. By analyzing the fluorescence signal duration, the toxin concentration can be further determined.

[0088] The dual-function detection instrument was used to image the mycotoxin six-in-one test card to obtain the colloidal gold response time series data. The colloidal gold response time series data reflects the color change process of colloidal gold caused by three toxins, fumonisins (B1 and B2), ochratoxin A and T-2 toxin, including information such as the degree of color change, the rate of color change, and the time of color change. These data can be used to analyze the presence or absence of toxins, the content of toxins, etc. The degree of color change indicates the magnitude of the change in the color of colloidal gold relative to the initial state, which is positively correlated with the concentration of the analyte. The higher the concentration, the more obvious the color change. By analyzing the degree of color change, the concentration of the toxin can be inferred. The rate of color change indicates how fast the color of colloidal gold changes, which is positively correlated with the concentration of the analyte. The higher the concentration, the faster the color change. By comparing the color change rates at different times, the toxin concentration can be dynamically evaluated. The color change time indicates the time point when the color of colloidal gold begins to change, which is negatively correlated with the concentration of the analyte. The higher the concentration, the earlier the color change occurs. By comparing the color change time of different toxins, the type of toxin can be preliminarily determined.

[0089] Environmental sensors are used to collect time series data of environmental variables during the detection process. Environmental variables include temperature, humidity, light intensity, etc. These factors may affect the accuracy and stability of the test results. By recording the changes in environmental variables, the impact of environmental factors on the test can be analyzed and corresponding compensation or correction can be made. Temperature affects the rate of enzymatic reaction. Too high or too low temperature will reduce enzyme activity and lead to decreased detection sensitivity. Temperature time series data can be used to evaluate whether the temperature is constant within the optimal range during the test. Humidity affects the water absorption performance of paper-based chips. Too high humidity will dilute the sample and reduce the detection concentration; too low humidity will slow down the migration speed of the sample and prolong the detection time. Humidity time series data can be used to evaluate whether the humidity is constant within the appropriate range during the test. Light intensity can affect the detection of fluorescence signals. Too strong background light will drown out weak fluorescence signals; too weak background light will reduce the signal-to-noise ratio and increase the difficulty of detection. Light intensity time series data can be used to evaluate whether the detection environment meets the instrument requirements.

[0090] By collecting and analyzing the above three types of time series data, we can comprehensively evaluate the multi-dimensional response characteristics of the six-in-one mycotoxin test card, explore the reaction laws and discrimination patterns under different toxins, different concentrations, and different environments, and provide data support for subsequent intelligent interpretation.

[0091] By continuously monitoring the intensity of the fluorescence signal and the color change of colloidal gold, the toxin concentration can be dynamically analyzed, which is conducive to the timely detection of low-concentration toxins. By comparing the time of appearance of the fluorescence signal of different toxins and the time of color change of colloidal gold, the type of toxin can be cross-verified to reduce false positives. By introducing environmental variable monitoring, the impact of external factors on the test results can be evaluated to provide a reference for the interpretation of the results. Through the comprehensive analysis of multi-dimensional data, the discrimination rules and quantitative models for toxin identification can be established to achieve standardization and automation of the detection process.

[0092] In summary, this method uses machine vision and sensor technology to achieve multi-parameter and full-process monitoring of the mycotoxin detection process, providing a new idea for improving detection performance and optimizing the detection process. With the development of the Internet of Things and big data technology, intelligent analysis based on time series data streams will become an important development direction in the field of mycotoxin detection.

[0093] Step S1200, preprocessing the first toxin detection data;

[0094] Specifically, preprocessing includes data cleaning, data normalization and data synchronization. The median filter algorithm is used to smooth the fluorescence response time series data to remove random noise and spike interference. The median filter algorithm can effectively suppress pulse noise and retain the edge characteristics of the signal. The threshold method is used to binarize the colloidal gold response time series data to remove background interference and artifacts. The threshold method can divide the image into foreground (colloidal gold color change area) and background according to the color change threshold, highlighting the area of ​​interest. The outlier detection algorithm is used to screen the environmental variable time series data and eliminate outliers. The outlier detection algorithm can identify outliers that deviate from the normal range based on the data distribution characteristics, thereby improving the reliability of the data.

[0095] The cleaned data was normalized to scale the data of different dimensions to the same scale. The maximum and minimum normalization method was used to normalize the fluorescence response time series data. Maximum and minimum normalization can map the fluorescence signal intensity to the [0,1] interval, eliminate the dimension effect, and facilitate the comparison between different toxins. The Z-score normalization method was used to normalize the colloidal gold color change rate data. Z-score normalization can convert the data into a standard normal distribution with a mean of 0 and a variance of 1, highlighting the relative degree of change of the data, which is convenient for outlier detection and threshold setting. The environmental variable time series data was normalized using the decimal calibration normalization method. Decimal calibration normalization can scale the data to the [-1,1] interval while retaining the original distribution characteristics of the data, which is convenient for feature extraction and pattern recognition. The normalized data was synchronized to align the data of different time series to the same time axis. Taking the fluorescence response time series data as the benchmark, the colloidal gold response time series data was time-aligned using the interpolation algorithm. The interpolation algorithm can estimate the corresponding fluorescence response intensity value at the sampling time point of the colloidal gold response time series data, and realize the time synchronization of the two data. Taking the fluorescence response time series data as the benchmark, the nearest neighbor algorithm is used to time align the environmental variable time series data. The nearest neighbor algorithm can find the intensity value of the nearest fluorescence response time series data point according to the sampling time point of the environmental variable, and use this intensity value to time align the environmental variable time series data.

[0096] Step S1300, extracting features from the first toxin detection data to obtain second toxin detection data; the second toxin detection data includes fluorescence signal intensity features, fluorescence response rate features, colloidal gold color change rate features and environmental variable volatility features.

[0097] Furthermore, if Figure 2 As shown, step S1300 includes:

[0098] Step S1310, extracting fluorescence signal intensity characteristics and fluorescence response rate characteristics from the fluorescence response time series;

[0099] Furthermore, if Figure 3 As shown, step S1310 includes:

[0100] Step S1311, calculating the maximum value, minimum value, average value and integrated area of ​​the fluorescence response time series to obtain the fluorescence signal intensity characteristics;

[0101] Step S1312, calculating the slope, rise time and fall time of the fluorescence response timing to obtain the fluorescence response rate characteristics.

[0102] Specifically, the maximum value refers to the highest point on the fluorescence response time series curve, which reflects the peak level of the fluorescence signal. The higher the peak level, the higher the concentration of the analyte and the better the detection sensitivity. By comparing the maximum values ​​of different detection objects, the relative content of mycotoxins in the sample can be preliminarily determined. The minimum value refers to the lowest point on the fluorescence response time series curve, which reflects the background level of the fluorescence signal. The background level mainly comes from the autofluorescence of the detection environment and the instrument background noise, and is independent of the concentration of the analyte. By analyzing the change trend of the minimum value, the stability of the detection environment and instrument performance can be monitored. The average value refers to the arithmetic mean of all data points on the fluorescence response time series curve, which reflects the overall level of the fluorescence signal. The overall level takes into account the peak level and the background level, and can evaluate the test results more comprehensively. By comparing the average value with the threshold, it can be determined whether the sample to be tested contains mycotoxins. The integral area refers to the area enclosed by the fluorescence response time series curve and the time axis, which reflects the cumulative amount of the fluorescence signal. The larger the cumulative amount, the more fluorescent substances are generated and the more complete the reaction process. By calculating the integral area, the integrity and effectiveness of the detection process can be quantitatively evaluated.

[0103] The above four indicators characterize the intensity characteristics of the fluorescence signal from different aspects. The maximum value and average value focus on the signal amplitude, the minimum value mainly reflects the background interference, and the integral area emphasizes the signal duration. By comprehensively analyzing these indicators, it is possible to accurately determine whether there are mycotoxins in the sample to be tested and roughly estimate its concentration range. At the same time, the change rules of these indicators can be used to guide the optimization of detection parameters, such as adjusting the exposure time, controlling the ambient temperature, etc., so as to further improve the detection performance.

[0104] The slope refers to the ratio of the change in signal intensity to the time difference between any two points on the fluorescence response time series curve, which reflects the speed of the fluorescence signal change. The faster the fluorescence signal changes, the larger the absolute value of the slope, indicating that the reaction kinetics are faster and the detection sensitivity is higher. By analyzing the positive and negative signs of the slope, it is possible to determine whether the fluorescence signal is in the rising or falling phase. The rise time refers to the time required for the fluorescence signal to rise from the minimum value to the maximum value, reflecting the speed of the fluorescence emission process. The shorter the rise time, the higher the concentration of the analyte and the faster the reaction. By comparing the rise time of different samples, high-concentration and low-concentration toxins can be accurately distinguished. The fall time refers to the time required for the fluorescence signal to fall from the maximum value to the minimum value, reflecting the speed of the fluorescence quenching process. The longer the fall time, the greater the interference with the fluorescence quenching, and there may be matrix effects or interfering substances. By analyzing the abnormality of the fall time, suspicious samples can be identified and false positives can be excluded.

[0105] The above three indicators characterize the response rate characteristics of the fluorescence signal from a dynamic perspective. The slope reflects the instantaneous speed of the signal change, while the rise time and fall time reflect the starting and ending points of the signal change. By tracking the dynamic changes of these indicators, the detection process can be monitored in real time and abnormal conditions can be discovered in time. At the same time, by comparing the differences in response rates of different samples, the complexity of the sample matrix can be inferred and the reliability of the detection results can be evaluated.

[0106] Step S1320, extracting the colloidal gold color change rate feature from the colloidal gold response time series;

[0107] Furthermore, if Figure 4 As shown, step S1320 includes:

[0108] Step S1321, performing a differential operation on the colloidal gold response time sequence to obtain a color change speed sequence;

[0109] Step S1322, calculating the maximum value, minimum value, average value and standard deviation of the color change rate sequence to obtain the colloidal gold color change rate characteristics.

[0110] Specifically, differential operation refers to subtracting the color value of the previous moment from the color value of the next moment to obtain the color change between two adjacent moments. The color change of the entire time series is arranged in chronological order to obtain the color change rate sequence. The color change rate reflects the instantaneous progress of the colloidal gold color development reaction. The faster the speed, the faster the combination of the color group and the gold particles, and the higher the concentration of the analyte. By analyzing the dynamic distribution of the color change rate, the progress of the color development reaction can be tracked in real time and the detection node can be accurately grasped.

[0111] The maximum value refers to the maximum value in the color change rate sequence, which reflects the peak speed of the color development reaction. The greater the peak speed, the stronger the binding ability of the chromogenic group to the gold particles, and the higher the detection sensitivity. By comparing the maximum values ​​of different samples, the relative concentration of the analyte can be identified. The minimum value refers to the minimum value in the color change rate sequence, which reflects the starting speed of the color development reaction. The colloidal gold color development reaction usually has a slow start-up process and a low starting speed. If the minimum value is abnormally high, it may indicate nonspecific binding or background staining. The average value refers to the arithmetic mean of all data in the color change rate sequence, which reflects the average speed of the color development reaction. The higher the average speed, the faster the color development reaction is overall and the shorter the detection time. By comparing the difference between the average value and the empirical value, the degree of optimization of the detection process can be evaluated. The standard deviation refers to the root mean square of the sum of the squares of the differences between all data in the color change rate sequence and the average value, which reflects the degree of dispersion of the color development reaction speed. The greater the dispersion, the more obvious the fluctuation of the color development reaction in the time series, which may be affected by interference factors. By analyzing the abnormality of the standard deviation, the unstable factors in the detection process can be found.

[0112] The above four indicators characterize the color change rate characteristics of colloidal gold from a statistical perspective. The maximum value reflects the upper limit of the color development ability, the minimum value reflects the background interference level, the average value represents the overall speed level, and the standard deviation reveals the risk of speed fluctuation. By comprehensively considering these indicators, the kinetic characteristics of the colloidal gold color development reaction can be quantitatively evaluated and the test results can be accurately interpreted. At the same time, establishing a mathematical model between these indicators and the concentration of the analyte can achieve quantitative detection and expand the scope of application of the method.

[0113] Step S1330, extracting environmental variable volatility features from the environmental variable time series.

[0114] Furthermore, if Figure 5 As shown, step S1330 includes:

[0115] Step S1331, performing sliding window analysis on the environmental variable time series to obtain a volatility series;

[0116] Step S1332, calculate the maximum value, minimum value, average value and dynamic range of the volatility sequence to obtain the environmental variable volatility characteristics.

[0117] Specifically, sliding window analysis refers to selecting a time window of fixed length, calculating the standard deviation of environmental variables within the window, and then sliding the window in chronological order to obtain a series of standard deviation values, which is the volatility sequence. Volatility reflects the fluctuation of environmental variables in a local time period. The larger the volatility, the more unstable the environment is, and the more likely it is to affect the test results. By analyzing the dynamic changes of volatility, environmental anomalies can be discovered and located in a timely manner, providing a basis for quality control of detection.

[0118] The maximum value refers to the maximum value in the volatility sequence, reflecting the most drastic fluctuation of the environmental variable. If the maximum value exceeds the applicable range of the instrument, it indicates that the detection process may exceed the environmental tolerance range and the reliability of the results is reduced. The minimum value refers to the minimum value in the volatility sequence, reflecting the most stable state of the environmental variable. If the minimum value is significantly higher than the normal level, it indicates that the detection environment may have been in an unstable state and the experimental conditions need to be optimized. The average value refers to the arithmetic mean of all data in the volatility sequence, reflecting the overall level of fluctuation of the environmental variable. The higher the average value, the more unstable the environment is overall and the greater the uncertainty of the test results. By setting the warning line of the average value in advance, the substandard test batches can be automatically screened out. The dynamic range refers to the difference between the maximum and minimum values ​​in the volatility sequence, reflecting the amplitude of the fluctuation of the environmental variable. The larger the dynamic range, the greater the change from extreme stability to extreme chaos in the detection process, which may endanger the normal progress of the detection. By analyzing the distribution of the dynamic range, the overall stability of the detection environment can be evaluated.

[0119] The above four indicators characterize the volatility characteristics of environmental variables from two dimensions: dynamic and static. The maximum value and dynamic range reflect the sharp fluctuations in the short term, while the minimum value and average value reflect the overall level in the long term. By tracking and analyzing these indicators, we can grasp the dynamic changes of the detection environment, control the environmental conditions in a targeted manner, and improve the reliability and repeatability of the detection results. It is worth noting that since the fluctuations of different environmental variables (such as temperature, humidity, pressure, etc.) have different effects on the detection process, they need to be treated differently and assigned different weight coefficients when extracting and analyzing volatility characteristics. Through mathematical methods such as weighted average or principal component analysis, a comprehensive environmental volatility index can be obtained to more comprehensively evaluate the applicability of the detection environment.

[0120] By extracting the characteristic indicators defined in steps S1310, S1320, and S1330, the physical and chemical properties of the sample to be tested and the dynamic information of the detection process can be quantitatively described from four aspects: fluorescence signal intensity, fluorescence response rate, colloidal gold color change rate, and environmental variable volatility. These indicators cover multiple dimensions such as the static level, dynamic changes, and time series distribution of the detection data, and can fully reflect the inherent laws and influencing factors of mycotoxin detection. In subsequent data analysis and model construction, these indicators can be fully utilized to establish a quantitative relationship between the concentration of the test object and the signal response, and to design adaptive algorithms and decision rules to dynamically optimize the detection process and improve the sensitivity, accuracy, and robustness of the detection. At the same time, these indicators can also be used to construct quality control charts, realize real-time monitoring and early warning of the detection process, promptly discover and diagnose abnormal conditions, and ensure the reliability of detection data and analysis results.

[0121] Step S1000 realizes the collection, preprocessing and feature extraction of multi-dimensional detection data of the six-in-one mycotoxin detection card, and obtains feature data of multiple dimensions such as fluorescence signal intensity, fluorescence response rate, colloidal gold color change rate and environmental variable volatility. These feature data reflect the detection of six mycotoxins from different angles, providing comprehensive and reliable data support for subsequent intelligent interpretation. At the same time, through preprocessing operations such as data cleaning, normalization, and synchronization, the data quality is improved, the dimensional differences and time offsets between data of different dimensions are eliminated, and the foundation for feature extraction and fusion is laid. In the feature extraction process, a variety of statistics and dynamic indicators are used to fully explore the time domain characteristics and change laws of the data, and enhance the representation ability and discrimination of the features. The comprehensive application of these technical means effectively solves the complexity, diversity and inconsistency of mycotoxin detection data, provides high-quality input for subsequent intelligent interpretation algorithms, and improves the accuracy and reliability of test results.

[0122] Step S2000, establishing a historical detection database, constructing and training a mycotoxin classification model, and obtaining preliminary identification results of six mycotoxins based on the second toxin detection data and the mycotoxin classification model;

[0123] Furthermore, if Figure 6 As shown, step S2000 includes:

[0124] Step S2100, establish a historical detection database, construct and train a mycotoxin classification model, use the trained mycotoxin classification model to predict the second toxin detection data, and obtain the first recognition results of the six mycotoxins, wherein the first recognition results are the predicted probabilities of the six mycotoxins. , , , , , ];in represents the predicted probability of the i-th mycotoxin among the six mycotoxins, 1≤i≤6; the historical detection database includes n1 records;

[0125] Specifically, a historical detection database is established. Each record in the historical detection database represents a complete mycotoxin detection process, including the second toxin detection data of the detection (i.e., fluorescence signal intensity characteristics, fluorescence response rate characteristics, colloidal gold color change rate characteristics, and environmental variable volatility characteristics) and the test results (i.e., the types of mycotoxins actually contained in the sample). The historical detection data comes from a large number of previous mycotoxin detection experiments. Each experiment uses samples with known mycotoxin components. After the test card completes the test, the detection process data and sample component information are recorded together and accumulated to form a database. The test results are converted into six-dimensional binary vectors by manual annotation. Each dimension represents a mycotoxin. A value of 1 indicates that the sample contains the mycotoxin, and a value of 0 indicates that the sample does not contain the mycotoxin.

[0126] Construct and train a mycotoxin classification model. This embodiment uses a support vector machine (SVM) algorithm to construct a classification model. SVM is a binary classification model, and its basic idea is to find a hyperplane in the feature space so that positive and negative samples are separated by the hyperplane, and the distance from the sample point (support vector) closest to the hyperplane to the hyperplane is as large as possible. SVM can convert nonlinear problems into linear problems through kernel functions, and has a good classification effect. Since this embodiment involves six kinds of mycotoxins, it is necessary to construct six binary classification SVM models, each of which is used to determine whether a sample contains a specific mycotoxin. Each record in the historical detection database is randomly divided into a training set and a test set, wherein the training set is used for model training, and the test set is used for performance evaluation. The input of each SVM model is the second toxin detection data of a record, and the output is the binary label of the corresponding mycotoxin type in the record. Through the grid search and cross-validation method, the hyperparameters (such as penalty factor, kernel function type, etc.) of each SVM model are optimized to obtain a model with optimal generalization performance.

[0127] The trained mycotoxin classification model is used to predict the second toxin detection data. The second toxin detection data of the current test sample is input into the six SVM classification models in sequence. Each model gives a probability estimate (predicted probability) of the sample containing the mycotoxin. The greater the probability, the more likely the sample contains the mycotoxin. The six probability estimates are combined in sequence into a six-dimensional probability vector [y 1 ,y 2 ,y 3 ,y4 ,y 5 ,y 6 ], as the first recognition result of the sample, where y i It represents the predicted probability of the i-th mycotoxin among the six mycotoxins, that is, the predicted probability that the sample belongs to the i-th mycotoxin.

[0128] The reason why the SVM classification model can predict new samples is that it learns the association pattern between sample features and mycotoxin categories from historical test data. During the training phase, the SVM model maximizes the classification interval and finds an optimal decision boundary in the feature space so that samples of different categories can be separated as much as possible. This decision boundary essentially corresponds to a discriminant function, which can calculate the probability of a sample belonging to each category based on its features. When a new sample is input, the model substitutes its features into the discriminant function and calculates the function value belonging to each category. The one with the largest function value is the most likely category. The size of the function value can be converted into a probability estimate, which indicates the confidence that the sample belongs to each category. Since SVM is built based on statistical learning theory and the complexity of the model is controlled by the principle of structural risk minimization, the trained classification model often has good generalization ability and can more accurately predict the category of unknown samples.

[0129] By training the SVM classification model, the correspondence between sample features and mycotoxin categories can be established. When a feature description of a new sample is input, the model can give the probability that the sample belongs to each mycotoxin based on the learned classification rules, thereby achieving preliminary identification of the toxin type. Compared with the fixed threshold judgment method, this method can make full use of the feature distribution information contained in the historical detection data, automatically generate a discriminant model in a data-driven manner, and has stronger adaptability and robustness. When the amount of accumulated historical data is large enough, the trained classification model often has a high generalization ability and can better predict the category of unknown samples.

[0130] Step S2200: Setting the confidence threshold , represents the confidence threshold of the i-th mycotoxin among the six mycotoxins; if the predicted probabilities of the six mycotoxins are all greater than the confidence thresholds of the corresponding mycotoxins, the first recognition result is used as the preliminary recognition result; otherwise, step S2300 is executed;

[0131] Setting the confidence threshold include:

[0132] Through the historical detection database, calculate the The accuracy of mycotoxins and recall ;

[0133] Obtain the detection feature of the i-th mycotoxin from the second toxin detection data and calculate the Sensitivity scores of mycotoxin detection features;

[0134] Calculate the predicted probability of the i-th mycotoxin The predicted probability of the jth mycotoxin The correlation coefficient ρ ij , 1≤j≤6, i≠j.

[0135]

[0136] in:

[0137] : The confidence threshold of the i-th mycotoxin, which indicates the minimum prediction probability requirement of the mycotoxin classification model when judging whether the i-th toxin exists. Range: [0,1], the higher the value, the stricter the identification requirement for the toxin.

[0138] : No. The accuracy of the detection of various mycotoxins is calculated based on statistics from the historical detection database and cross-validation or validation sets.

[0139] Calculation method: ; is the number of true positive samples of the i-th mycotoxin in the historical detection database, is the number of false positive samples of the i-th mycotoxin in the historical detection database.

[0140] : No. Recall rate of mycotoxins, calculated by: , is the number of false negative samples of the i-th mycotoxin in the historical detection database.

[0141] : The influence weight of accuracy on confidence threshold.

[0142] : The influence weight of recall rate on confidence threshold; Used to balance the trade-off between high accuracy (reducing false positive rate) and high recall (reducing missed detection rate); set by experimental optimization to meet .

[0143] : No. Sensitivity scores of the mycotoxin detection features.

[0144] Calculation formula: ;

[0145] in:

[0146] : Standard deviation of fluorescence signal intensity, reflecting the sensitivity of fluorescence method to the characteristic response of the toxin. Calculated according to the fluorescence signal intensity of the i-th fungal toxin.

[0147] : The standard deviation of the colloidal gold color change rate reflects the sensitivity of the colloidal gold method to the characteristic response of the toxin. It is calculated based on the colloidal gold color change rate of the i-th fungal toxin.

[0148] : The influence of environmental variable fluctuation on the detection of the toxin. Calculated according to the environmental variable fluctuation during the detection of the i-th mycotoxin.

[0149] yes The weight coefficient of yes The weight coefficient of yes The weight coefficient of .

[0150] : Adjustment coefficient, which controls the influence of the sensitivity score on the confidence threshold and is determined through experimental optimization.

[0151] : No. The ratio of the characteristic value of a fungal toxin to its historical average value; used to dynamically adjust the threshold.

[0152] Calculation formula:

[0153]

[0154] : The characteristic value of the i-th fungal toxin currently detected (such as fluorescence signal intensity).

[0155] : The average value of the i-th mycotoxin characteristic in the historical detection database.

[0156] : Correlation coefficient between the predicted probabilities of the ith mycotoxin and the jth mycotoxin, 1≤j≤6. Indicates the co-occurrence probability or mutual influence degree of two toxins in the test data. Statistical data from the historical test database, calculated based on mutual information or correlation analysis.

[0157] : The predicted probability of the j-th mycotoxin, output by the mycotoxin classification model.

[0158] : Prediction probability of the ith mycotoxin.

[0159] : Coupling relationship term;

[0160] δ: The influence weight of the coupling relationship term on the confidence threshold.

[0161] η: Normalization parameter that controls the upper limit of the confidence threshold, usually a positive number (such as 10 or 100).

[0162] and right Impact:

[0163] when When increases (false positives decrease), Increase, detection becomes more stringent; when When it increases (false negatives decrease), Reduced, more sensitive detection.

[0164] Coupling relationship between toxins Impact:

[0165] When ij and the predicted probability y of other toxins j When increasing, Increase to avoid misjudgment due to strong coupling relationship.

[0166] Environmental sensitivity to Impact:

[0167] When the environment fluctuates ( ) increases, Increase, reduce the impact of environmental disturbances; when environmental fluctuations decrease, Reduce.

[0168] This formula combines the classification performance ( , )、Toxin characteristic sensitivity( , ), coupling relationship between toxins (ρ ij ), which significantly improves the comprehensiveness and adaptability of confidence threshold calculation. By introducing nonlinear processing of feature sensitivity (sigmoid function), the formula's ability to respond to dynamic changes is further enhanced. Feature dynamic change rate ( ) and the coupling relationship between toxins (ρ ij ) enables the confidence threshold to be dynamically adjusted according to the characteristics of the current detection data. ij ) is strong, Increase, avoid misjudgment caused by coupling. The introduction of α and β weight parameters allows the formula to flexibly adjust the priority according to actual needs (such as focusing on reducing false positives or reducing missed detections). Introduce environmental factors and toxin characteristic sensitivity ( , ), effectively reducing the impact of environmental disturbances on detection results and improving system robustness.

[0169] Specifically, the confidence threshold setting method in step S2200 comprehensively considers historical detection data, current detection features and the correlation between different toxins, and can improve the detection accuracy while taking into account the recall rate, thereby achieving dynamic optimization and adaptive adjustment of detection performance.

[0170] Among them, the historical detection database records the test results of a large number of known samples, including the real toxin types, predicted toxin types, predicted probabilities of various toxins, etc. Through statistical analysis of historical data, the accuracy (the proportion of the number of correctly predicted positive samples to all samples predicted to be this type of toxin) and recall (the proportion of the number of correctly predicted positive samples to all samples that are actually this type of toxin) of each fungal toxin can be calculated. Accuracy and recall are two important indicators for evaluating the performance of binary classification models. For multi-classification problems, each category can be regarded as a binary classification problem, and the accuracy and recall are calculated separately. A high accuracy rate indicates that the prediction result is highly credible and has few false positives; a high recall rate indicates that the missed detection rate is low and the false negatives are few. Ideally, both the accuracy and recall rates should be close to 1. But in reality, accuracy and recall rates are often contradictory and need to be weighed.

[0171] The historical accuracy and recall rate reflect the performance of the classifier in the past, but they cannot fully represent the characteristics of the current sample. Therefore, it is also necessary to extract the detection features of each toxin from the second toxin detection data (i.e., the first toxin detection data after feature extraction) and calculate its sensitivity score. The detection features include fluorescence signal intensity, fluorescence response rate, colloidal gold color change rate, environmental variable volatility, etc., which reflect the response pattern of different toxins under different detection conditions. The sensitivity score can be measured by mathematical statistics methods (such as t-test, variance analysis, etc.) to measure the recognition contribution of a certain feature to a certain toxin. The higher the sensitivity score, the stronger the discriminative power of the feature for the toxin. Incorporating the sensitivity score into the calculation of the confidence threshold can automatically adapt to the characteristics of the sample and dynamically adjust the threshold. When a sample contains a high concentration of a certain toxin, its sensitivity score will be significantly higher, and the confidence threshold of the toxin will be increased accordingly to avoid missed detection. When a sample does not contain a certain toxin, its sensitivity score will be significantly lower, and the confidence threshold of the toxin will be reduced accordingly to avoid false detection.

[0172] In addition, since there may be cross-reactions or synergistic effects between the six mycotoxins, their predicted probabilities are not completely independent. In order to characterize this correlation, step S2200 introduces a correlation coefficient. ij Measures the linear correlation between the predicted probability of the ith mycotoxin and the predicted probability of the jth mycotoxin, with a value range of [-1,1]. ij The closer it is to 1, the more positively correlated the predicted probabilities of the two toxins are, that is, an increase (decrease) in the predicted probability of one toxin is often accompanied by an increase (decrease) in the predicted probability of the other toxin; ρ ij The closer it is to -1, the negative correlation between the predicted probabilities of the two toxins, that is, an increase (decrease) in the predicted probability of one toxin is often accompanied by a decrease (increase) in the predicted probability of the other toxin; ρ ij The closer it is to 0, the more the predicted probabilities of the two toxins are independent, that is, the change in the predicted probability of one toxin does not affect the predicted probability of the other toxin.

[0173] The advantages of the adaptive threshold determination method in step S2200 are:

[0174] The historical information and current information are used comprehensively. The historical accuracy and recall rate reflect the basic performance of the classifier, while the current sensitivity score and correlation coefficient reflect the particularity of the sample to be tested. The two types of information complement each other and can more comprehensively evaluate the credibility of the classification.

[0175] Automatically adapt to sample characteristics. When the sample to be tested contains a high concentration of toxins, the sensitivity score will be high, and the confidence threshold will also increase accordingly, which can suppress false positives; when the sample to be tested does not contain a certain toxin, the sensitivity score will be low, and the confidence threshold will also decrease accordingly, which can suppress false negatives.

[0176] The correlation between toxins is taken into account. When two toxins tend to appear at the same time (positive correlation), the detection standard is appropriately relaxed; when two toxins are mutually exclusive (negative correlation), the detection standard is appropriately raised. This helps to improve the detection rate of multiple toxin contamination and reduce the misjudgment rate caused by cross-interference.

[0177] Step S2300, retrieving reference data similar to the second toxin detection data from a historical detection database;

[0178] Furthermore, if Figure 7 As shown, step S2300 includes:

[0179] Step S2310, calculating the similarity between the second toxin detection data and each record in the historical detection database to obtain a similarity list;

[0180] Step S2320, sort the similarity list and select the record with the highest similarity as reference data.

[0181] Specifically, step S2300 uses a similarity matching algorithm to retrieve the reference record closest to the current test data features in the historical test database. The core of this step is to measure the similarity between the current test data and the historical data, and then find the most relevant reference sample. The similarity here is a quantitative indicator for comparing the similarity between two sets of data, which can be calculated based on a variety of mathematical models.

[0182] Commonly used similarity measurement methods include Euclidean distance, Mahalanobis distance, cosine similarity, etc. Among them, Euclidean distance is the simplest and most commonly used similarity measurement method, which calculates the straight-line distance between two vectors in multidimensional space. Euclidean distance is intuitive and easy to understand, and is simple to calculate. It is suitable for data with few feature dimensions and independent features. However, when the number of features is large or there is correlation between features, the effect of Euclidean distance will be affected. Mahalanobis distance is a similarity measurement method that takes into account the covariance between features. Compared with Euclidean distance, Mahalanobis distance not only measures the difference between two samples in each dimension, but also considers the correlation between different dimensions. Cosine similarity measures the cosine value of the angle between two vectors. The smaller the angle, the closer the cosine value is to 1, indicating that the two vectors are more similar. In practical applications, different similarity measurement methods can be selected or combined according to the characteristics of the data and domain knowledge.

[0183] Taking mycotoxin detection data as an example, samples usually have multi-dimensional indicators such as fluorescence signal intensity characteristics, fluorescence response rate characteristics, colloidal gold color change rate characteristics, and environmental variable volatility characteristics. Different types of toxins may have different differences in different indicators. Therefore, a suitable similarity measurement method can be designed for each type of feature:

[0184] For the fluorescence signal intensity feature, since the fluorescence values ​​of different toxins vary greatly and the intensity value has a clear physical meaning, using Euclidean distance as a similarity measurement indicator is simple, effective, and easy to interpret; for the fluorescence response rate feature, the temporal change trends of different toxins may be more distinguishable, and you can consider using a time series similarity algorithm, such as the dynamic time warping (DTW) distance, to measure the shape differences of the response curves; for the colloidal gold color change rate feature and the environmental variable volatility feature, since there may be a certain correlation between the color value and the environmental parameters, using the Mahalanobis distance can more accurately characterize the differences between samples.

[0185] After obtaining the quantified results of the similarity of various features, it is necessary to perform weighted aggregation at the feature level to obtain the overall similarity between samples. The weight of the aggregation can be set based on domain knowledge or learned in a data-driven way.

[0186] By comprehensively considering the similarity of indicators in each dimension, the final similarity score between the current detection data and each record in the historical database can be obtained through strategies such as weighted summation. Then, the candidate results are sorted from high to low according to the similarity, and the most similar historical records are selected as the final reference data. The reference data includes reference fluorescence signal intensity characteristics, reference fluorescence response rate characteristics, reference colloidal gold color change rate characteristics, and reference environmental variable volatility characteristics, which are consistent with the characteristic dimensions of the second detection data. Using similar historical data can provide valuable reference information for the current detection and help improve the accuracy of recognition.

[0187] The recognition strategy based on similar historical data has the following advantages:

[0188] Strong interpretability. Each recognition result can be supported by intuitive historical reference samples, which makes it easier for analysts to understand and verify the detection process.

[0189] Good generalization ability. It makes full use of the useful information in historical data and can deal with new samples to be tested. Even if the current test data is different from the existing pattern, as long as similar references can be found in the historical database, a reliable judgment can be made.

[0190] Easy to update and maintain. Every new test result can be added to the database as historical reference data, which can be accumulated over time to continuously improve and expand the knowledge base. The similarity measurement and sorting rules in the test process can also be flexibly adjusted based on feedback.

[0191] High computational efficiency. Similarity matching avoids the complex model training process, and the efficiency of retrieval in massive historical data is also high. Real-time or quasi-real-time detection response can be achieved.

[0192] In summary, step S2300 makes full use of historical detection results through similarity matching, providing an interpretable, generalizable, updateable and efficient reference basis for the identification of the current sample, which is a simple, practical and flexible auxiliary strategy for toxin detection. It not only improves detection accuracy and speeds up response, but also provides analysts with intuitive reference and explanation, enhancing the auditability and credibility of the detection process.

[0193] Step S2400, using a convolutional neural network to perform feature fusion on the second toxin detection data and the reference data to obtain a toxin feature fusion vector;

[0194] Furthermore, if Figure 8 As shown, step S2400 includes:

[0195] Step S2410, constructing a toxin feature fusion model based on a convolutional neural network, setting an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer;

[0196] Step S2420, splicing the second toxin detection data with the reference data to form a toxin data fusion matrix;

[0197] Step S2430, input the toxin data fusion matrix into the toxin feature fusion model to obtain the toxin feature fusion vector.

[0198] Further, step S2430 includes:

[0199] Step S2431, inputting the toxin data fusion matrix into the input layer of the toxin feature fusion model, and converting the toxin data fusion matrix into a one-dimensional toxin feature vector;

[0200] Step S2432, the convolution layer performs a convolution operation on the one-dimensional toxin feature vector through a convolution kernel to extract local toxin features;

[0201] Step S2433, the pooling layer compresses the extracted local toxin features to obtain toxin compression features;

[0202] Step S2434, the toxin compression features are flattened and nonlinearly transformed in the fully connected layer to generate global toxin high-level features;

[0203] Step S2435, the output layer outputs the global toxin high-level features as a toxin feature fusion vector.

[0204] Specifically, step S2400 uses a convolutional neural network to perform feature fusion on the second toxin detection data and the reference data to obtain a toxin feature fusion vector. Convolutional Neural Network (CNN) is a deep learning model widely used in image recognition, speech recognition and other fields. It can effectively extract local and global features of data through local connections and weight sharing, and has the advantages of translation invariance and combinatoriality. In the present invention, a convolutional neural network is used to fuse the second toxin detection data and the reference data to generate a toxin feature fusion vector that comprehensively characterizes the characteristics of the two types of data.

[0205] In the specific implementation, the second toxin detection data is first spliced ​​with the reference data to form a toxin data fusion matrix. Each row of the matrix corresponds to a sample, which contains the fluorescence signal intensity characteristics, fluorescence response rate characteristics, colloidal gold color change rate characteristics, environmental variable volatility characteristics and reference data of the sample. Then, the toxin data fusion matrix is ​​input into the pre-built toxin feature fusion model. The model adopts the classic convolutional neural network architecture, including input layer, convolution layer, pooling layer, fully connected layer and output layer.

[0206] The method for constructing a toxin feature fusion model includes: the training samples come from the historical detection database, and each sample includes two parts of data: one is the second toxin detection data of the sample, that is, the fluorescence signal intensity feature, fluorescence response rate feature, colloidal gold color change rate feature and environmental variable volatility feature extracted in step S1300; the other is the reference data corresponding to the sample; the two parts of data are spliced ​​in a fixed format to form a two-dimensional matrix as the input of the CNN model. The true label of the sample is a six-dimensional binary vector; each dimension represents a fungal toxin, and a value of 1 indicates that the sample contains the fungal toxin, and a value of 0 indicates that the sample does not contain the fungal toxin. The samples are randomly divided into a training set, a validation set and a test set. The training set is used for model parameter learning, the validation set is used for hyperparameter selection, and the test set is used for performance evaluation. Using the training set samples, the convolution kernel weights and bias terms of the CNN model are continuously adjusted through forward propagation and backpropagation algorithms to make the model prediction output as close as possible to the true label of the sample. Through the validation set samples, the generalization performance of the model under different hyperparameters (such as the number of convolution kernels, convolution kernel size, pooling method, etc.) is evaluated, and the best one is selected. Using the test set samples, the prediction effect of the trained model on new samples is evaluated to provide a reference for model application.

[0207] When fusion is performed, first, the input layer receives the toxin data fusion matrix and converts it into a one-dimensional toxin feature vector. During the conversion process, each row of the matrix is ​​flattened to convert the two-dimensional matrix into a one-dimensional vector. In this way, the spatial structure information in the original data is preserved.

[0208] Then, the one-dimensional toxin feature vector passes through multiple convolutional layers and pooling layers in sequence to extract local features and gradually compress the feature scale. Each convolutional layer contains multiple convolution kernels, which slide on the input data, perform convolution operations, and obtain a set of feature maps. The convolution kernel can be regarded as a template for extracting local patterns of input data. By changing the weights of the convolution kernel, the network can adaptively learn different feature patterns. The pooling layer is located between two adjacent convolutional layers, and downsamples the convolution results to reduce the size of the feature map. Commonly used pooling operations include maximum pooling and average pooling, which take the maximum value and average value in each area as the representative value of the area. The pooling operation can not only compress the feature dimension, but also improve the translation invariance and robustness of the feature.

[0209] After multiple rounds of convolution and pooling, local features of the data are extracted layer by layer. These local features are aggregated in the fully connected layer to generate global feature representation. The fully connected layer consists of several ordinary neurons, each of which is connected to all neurons in the previous layer to perform nonlinear transformations. By adjusting the weight matrix, the fully connected layer can automatically learn the combination relationship between local features and integrate local features into high-level semantic features.

[0210] Finally, the output layer maps the global features into the required output format. In the present invention, the output layer contains only one neuron, which is used to generate a real-valued vector, namely the toxin feature fusion vector. This vector condenses the rich semantic information of the second toxin detection data and the reference data, and can fully reflect the multi-dimensional toxin characteristics of the sample to be tested.

[0211] The use of convolutional neural networks for feature fusion fully utilizes the powerful capabilities of deep learning and can automatically extract and combine distributed features without consuming a lot of manpower for feature engineering. Through end-to-end joint training, convolutional neural networks can mine implicit associations from massive heterogeneous data and build more robust and generalized feature representations. The fused toxin feature vector integrates the discriminant information of fluorescence data and colloidal gold data, overcomes the limitations of a single detection method, and is of great significance for improving the detection rate and accuracy of multiple toxins. In addition, the introduction of environmental factors as reference data helps to eliminate external interference and improve the reliability and repeatability of the results. In general, the feature fusion strategy based on convolutional neural networks provides a new idea for the multidimensional detection of mycotoxins and has broad application prospects in the field of food safety.

[0212] Step S2500, input the toxin feature fusion vector into the mycotoxin classification model to obtain a preliminary recognition result.

[0213] Specifically, the toxin feature fusion vector output from step S2400 is input into the mycotoxin SVM classification model trained in step S2100 to obtain the existence probability of each mycotoxin in the sample as the final identification result of the six mycotoxins. Convolutional neural networks can adaptively learn multi-scale and multi-level feature patterns and have powerful feature representation and fusion capabilities. By sharing the weights of the convolution kernel, CNN can efficiently process high-dimensional data and reduce the number of model parameters. Through the downsampling of the pooling operation, CNN can achieve feature compression and dimensionality reduction, which not only retains the main features but also reduces the risk of overfitting. After multiple layers of convolution and pooling, the high-level features extracted by CNN often contain more semantic information than the original features, which is more conducive to subsequent classification and discrimination.

[0214] Based on the extraction of high-level features by CNN, this embodiment uses SVM to achieve the final classification of high-level features. SVM is a classic discriminant classifier with a good theoretical basis and practical effect. SVM directly models the high-level features extracted by CNN, avoiding the instability of deep learning parameter learning and making the classification model more robust. At the same time, SVM can handle the feature vectors after CNN fusion very well, and a higher classification accuracy can be achieved with fewer samples. The experimental results show that the "CNN feature fusion + SVM feature classification" framework constructed in this embodiment can achieve excellent performance in the task of fungal toxin identification, and its generalization performance is significantly better than traditional methods such as single models and artificial rules.

[0215] In the field of pattern recognition and machine learning, this method of using deep learning models to extract high-level features and then using traditional machine learning models (such as SVM) for classification is called a fusion framework of "deep features + shallow classification". Shallow classification models directly model deep features, avoiding parameter learning for high-dimensional input space, reducing the risk of overfitting, and improving classification generalization capabilities. At the same time, because deep features undergo multiple layers of nonlinear transformations, they contain richer and more abstract discriminant information, and are often easier to distinguish than original features by linear models such as SVM. Therefore, connecting CNN and SVM in series can combine the advantages of both models and jointly improve toxin recognition performance at both the feature representation and classification decision levels.

[0216] Step S3000, based on the preliminary identification results of the six fungal toxins and combined with the time series of environmental variables, obtain accurate identification results.

[0217] Furthermore, if Fig. 9 As shown, step S3000 includes:

[0218] Step S3100, performing pattern recognition on the environmental variable time series to obtain an environmental state sequence;

[0219] Further, step S3100 includes:

[0220] Step S3110, clustering the environmental variable time series to obtain environmental state categories;

[0221] Step S3120, time-series labeling of the environment state categories to obtain an environment state sequence.

[0222] Specifically, step S3100 uses a method combining clustering and Hidden Markov Model (HMM) to achieve pattern recognition and state labeling of environmental variable time series. This process is divided into two steps: environmental state clustering and environmental state sequence labeling.

[0223] In step S3110, the environmental variable time series is first clustered using an unsupervised clustering algorithm. Unsupervised clustering is a commonly used data mining technology that can divide data objects into different clusters based on the similarity between them without prior knowledge. The data objects in each cluster are similar to each other, while the data objects between different clusters are quite different. Commonly used unsupervised clustering algorithms include K-means, hierarchical clustering, and DBSCAN. Among them, the K-means algorithm divides the data set into a pre-specified number of K clusters by iterative optimization, and each cluster is represented by its centroid (i.e., the mean of all points in the cluster). The hierarchical clustering algorithm generates a tree-like nested clustering structure by recursively merging or splitting clusters. The DBSCAN algorithm is based on the concept of density and can find clusters of any shape and automatically determine the number of clusters. By clustering the environmental variable time series, its inherent distribution pattern can be found, and each cluster corresponds to a specific environmental state, such as high temperature and low humidity, low temperature and high humidity, etc. The clustering results provide a basis for subsequent state sequence labeling.

[0224] In step S3120, the environmental state categories are labeled in time series using a hidden Markov model. A hidden Markov model is a statistical model used to describe a Markov process with hidden unknown parameters, and is composed of three parts: an initial probability distribution, a state transition probability distribution, and an observation probability distribution. HMM is widely used in fields such as speech recognition and natural language processing, and is an important tool for time series pattern recognition. In this step, each environmental state is regarded as a hidden state in the HMM, and the actual observed value of the environmental variable is regarded as the observed state. By establishing a transition probability matrix between hidden states and a transmission probability matrix from hidden states to observed states, the law of dynamic evolution of environmental states over time can be characterized. The HMM parameters are trained using the Baum-Welch algorithm, and then the environmental state sequence is decoded using the Viterbi algorithm to obtain the most likely hidden state sequence, i.e., the environmental state labeling result. This result can reflect the dynamic change trend of environmental variables and provide a basis for subsequent trend prediction and anomaly detection.

[0225] Through step S3100, the automatic classification and labeling of the environmental variable time series can be realized, revealing the dynamic evolution law of the environmental state. Compared with the traditional judgment method based on fixed thresholds, this method can adaptively cluster according to data distribution, does not rely on prior knowledge and manual experience, and has stronger robustness and adaptability. At the same time, the introduction of the hidden Markov model to model the environmental state sequence fully considers the temporal correlation between states, and can more accurately characterize the dynamic change characteristics of environmental variables.

[0226] In summary, step S3100 provides a method for recognizing environmental variable time series patterns based on machine learning. This method realizes the automatic discovery and sequence labeling of environmental states through the organic combination of clustering and HMM, and provides important data support for subsequent trend prediction, anomaly detection and other analyses. Compared with traditional methods, this method has the advantages of strong adaptability and accurate modeling of time series correlation, and can more comprehensively and accurately characterize the dynamic changes of environmental variables, providing strong support for quality control and result interpretation of the fungal toxin detection process.

[0227] Step S3200, calculating the historical accuracy of the preliminary recognition results under different environmental states in the environmental state sequence to obtain an environmental adaptive weight matrix;

[0228] Further, step S3200 includes:

[0229] Step S3210, counting the accuracy of the preliminary recognition results under each environmental state in the historical detection database to form an environmental state-accuracy mapping table;

[0230] Step S3220, converting the environment state-accuracy mapping table into an environment adaptive weight matrix, wherein the matrix elements in the environment adaptive weight matrix are weight coefficients of the preliminary recognition results under the corresponding environment state.

[0231] Specifically, first, the data in the historical detection database are grouped according to the environmental state. The environmental state can be divided according to the characteristic indicators in the environmental variable time series, such as temperature range, humidity range, light intensity range, etc. Each environmental state corresponds to a set of detection records under specific conditions. Then, for each set of data, the number of samples whose preliminary identification results are consistent with the true toxin type is counted, and the accuracy is calculated. Let the total number of samples under the kth environmental state be N k , the number of samples initially identified correctly is n k , then the accuracy of the preliminary recognition result under this environmental state is P k P k =n k / N kFinally, various environmental states and corresponding accuracy rates are mapped to form an environmental state-accuracy mapping table. This mapping table reflects the credibility of the preliminary recognition results under different environmental conditions and provides a basis for the calculation of environmental adaptive weights.

[0232] Assume the number of environmental states is The number of mycotoxins is , then the environment adaptation weight matrix The dimension is The elements in the matrix Indicates Under the environmental conditions, the The weight coefficient of each mycotoxin is calculated as follows:

[0233]

[0234] in, For the The accuracy of the preliminary recognition result under the environmental state. Through the above formula, the accuracy under each environmental state is normalized into a weight coefficient, and the sum of the weight coefficients under all environmental states is 1. The higher the accuracy of the environmental state, the larger its weight coefficient, which means that the preliminary recognition result obtained under the environmental condition is more reliable, and a higher weight should be given in the final recognition.

[0235] The introduction of the environmental adaptive weight matrix can make full use of the influence of environmental factors on the test results, dynamically adjust the importance of the preliminary identification results under different environmental conditions, and improve the adaptability and accuracy of identification. When the detection environment changes, the identification weights of various fungal toxins under the current environmental conditions can be obtained by querying the weight matrix without retraining the model, which greatly improves the real-time and flexibility of the system.

[0236] The environmental adaptive weight matrix is ​​a data-driven weight adjustment mechanism. By mining the environment-result correlation contained in historical detection data, it achieves adaptive matching between detection algorithms and environmental factors, improves the environmental adaptability and result reliability of mycotoxin detection, and lays the foundation for intelligent and precise mycotoxin detection.

[0237] Step S3300, using the environment adaptive weight matrix to perform weighted fusion on the preliminary recognition results to obtain accurate recognition results.

[0238] Furthermore, step S3300 includes:

[0239] Step S3310, representing the preliminary identification result as a probability distribution vector, where the vector elements are the probabilities that the sample belongs to each mycotoxin category;

[0240] Step S3320, searching the weight vector corresponding to the current environment state from the environment adaptive weight matrix;

[0241] Step S3330, performing element-by-element multiplication of the weight vector and the probability distribution vector to obtain a corrected probability distribution vector;

[0242] Step S3340, normalize the corrected probability distribution vector to obtain an accurate recognition result.

[0243] Specifically, step S3300 performs weighted fusion of the preliminary recognition results through the environment adaptive weight matrix to obtain an accurate recognition result. This step comprehensively considers the original probability distribution of the preliminary recognition result and the dynamic influence of the environmental state, realizes the dynamic correction of the recognition result through adaptive weighting, and improves the accuracy and adaptability of recognition.

[0244] First, the preliminary recognition results are expressed as a probability distribution vector. The probability distribution vector is a six-dimensional vector, and each element of the vector represents the probability that the sample belongs to the corresponding mycotoxin category. Then, the weight vector corresponding to the current environmental state is found from the environmental adaptive weight matrix. Each row of the matrix corresponds to a typical environmental state, each column corresponds to a mycotoxin category, and the matrix elements are the weights of the impact of the environmental state on the recognition performance. Given the actual value of the current environmental variable, by looking up the table in the matrix, the row closest to the current environmental state can be obtained. The elements of this row constitute the weight vector, which quantitatively characterizes the difficulty of identifying each mycotoxin category under the current environmental conditions.

[0245] Next, the weight vector is multiplied element by element with the probability distribution vector to obtain the corrected probability distribution vector. The elements with larger weight values ​​in the weight vector indicate that the recognition accuracy of the corresponding fungal toxin category is higher under the current environmental conditions, and should occupy a larger proportion in the fusion result; conversely, the elements with smaller weight values ​​correspond to categories with lower recognition accuracy, and their influence in the fusion result should be weakened. Through the operation of element-by-element multiplication, the adaptive weighted adjustment of the preliminary recognition results by environmental factors is realized, and the corrected probability distribution that comprehensively considers the sample characteristics and environmental status is obtained.

[0246] Finally, the corrected probability distribution vector is normalized to obtain the accurate identification result. The normalization process makes the corrected probability distribution satisfy the constraint that the sum of probabilities is equal to 1. The obtained accurate identification result is a reasonable probability distribution, which indicates the most likely fungal toxin category of the sample after environmental adaptive correction. The category with the largest probability value in the accurate identification result is the final identification conclusion. At the same time, the probability values ​​of other categories also provide a quantitative measure of the uncertainty of the identification decision, which can be used to assist judgment and risk assessment.

[0247] The beneficial effect of this step is that, through the adaptive weighted fusion strategy, the environmental state information can be fully utilized to dynamically adjust the recognition results, reduce the impact of environmental variable fluctuations on recognition accuracy, and maintain the robustness and reliability of the system in complex and changeable practical application scenarios. When the environmental state changes, the algorithm can automatically adapt and timely correct the recognition probability, and always give the optimal recognition result that matches the current environmental conditions. This adaptive ability comes from the construction of the environmental adaptive weight matrix. Through offline historical data analysis, the intrinsic connection between environmental factors and recognition performance is excavated, forming a set of weighted fusion strategies for dynamic environments, which have strong environmental adaptability and generalization ability.

[0248] In summary, this step uses the environment-adaptive weight matrix to achieve weighted fusion of preliminary identification results. The technical routes adopted, such as matrix construction, table lookup weighting, and normalization processing, have strong environmental adaptability, can effectively cope with the identification challenges brought by complex dynamic environments, and improve the accuracy and practicality of mycotoxin detection, and have broad application prospects.

[0249] Example 2

[0250] This embodiment provides, based on the first embodiment, an intelligent interpretation system for the mycotoxin six-in-one detection card based on multidimensional data analysis, such as Fig.10 As shown, including:

[0251] Data preprocessing module: used to collect first toxin detection data of six fungal toxins, perform feature extraction on the first toxin detection data, and obtain second toxin detection data; the first toxin detection data includes fluorescence response time series, colloidal gold response time series and environmental variable time series;

[0252] Preliminary identification module: used to establish a historical detection database, build and train a mycotoxin classification model, and obtain preliminary identification results of six mycotoxins based on the second toxin detection data and the mycotoxin classification model;

[0253] Precise identification module: Based on the preliminary identification results of six fungal toxins and combined with the time series of environmental variables, accurate identification results are obtained.

[0254] In the data preprocessing module, the first toxin detection data includes fluorescence response timing, colloidal gold response timing and environmental variable timing; the second toxin detection data includes fluorescence signal intensity characteristics, fluorescence response rate characteristics, colloidal gold color change rate characteristics and environmental variable volatility characteristics.

[0255] In the data preprocessing module, the feature extraction of the first toxin detection data includes:

[0256] Step S1310, extracting fluorescence signal intensity characteristics and fluorescence response rate characteristics from the fluorescence response time series;

[0257] Step S1320, extracting the colloidal gold color change rate feature from the colloidal gold response time series;

[0258] Step S1330, extracting environmental variable volatility features from the environmental variable time series.

[0259] The step S1310 includes:

[0260] Step S1311, calculating the maximum value, minimum value, average value and integrated area of ​​the fluorescence response time series to obtain the fluorescence signal intensity characteristics;

[0261] Step S1312, calculating the slope, rise time and fall time of the fluorescence response timing to obtain the fluorescence response rate characteristics.

[0262] The step S1320 includes:

[0263] Step S1321, performing a differential operation on the colloidal gold response time sequence to obtain a color change speed sequence;

[0264] Step S1322, calculating the maximum value, minimum value, average value and standard deviation of the color change rate sequence to obtain the colloidal gold color change rate characteristics.

[0265] The step S1330 includes:

[0266] Step S1331, performing sliding window analysis on the environmental variable time series to obtain a volatility series;

[0267] Step S1332, calculate the maximum value, minimum value, average value and dynamic range of the volatility sequence to obtain the environmental variable volatility characteristics.

[0268] In the preliminary identification module, the preliminary identification results of the six mycotoxins obtained according to the second toxin detection data and the mycotoxin classification model include:

[0269] Step S2100, establish a historical detection database, construct and train a mycotoxin classification model, use the trained mycotoxin classification model to predict the second toxin detection data, and obtain the first recognition results of the six mycotoxins, wherein the first recognition results are the predicted probabilities of the six mycotoxins. , , , , , ];in represents the predicted probability of the i-th mycotoxin among the six mycotoxins, 1≤i≤6; the historical detection database includes n1 records;

[0270] Step S2200: Setting the confidence threshold , represents the confidence threshold of the i-th mycotoxin among the six mycotoxins; if the predicted probabilities of the six mycotoxins are all greater than the confidence thresholds of the corresponding mycotoxins, the first recognition result is used as the preliminary recognition result; otherwise, step S2300 is executed;

[0271] Step S2300, retrieving reference data similar to the second toxin detection data from a historical detection database;

[0272] Step S2400, using a convolutional neural network to perform feature fusion on the second toxin detection data and the reference data to obtain a toxin feature fusion vector;

[0273] Step S2500, input the toxin feature fusion vector into the mycotoxin classification model to obtain a preliminary recognition result.

[0274] The step S2300 includes:

[0275] Step S2310, calculating the similarity between the second toxin detection data and each record in the historical detection database to obtain a similarity list;

[0276] Step S2320, sort the similarity list and select the record with the highest similarity as reference data.

[0277] The step S2400 includes:

[0278] Step S2410, constructing a toxin feature fusion model based on a convolutional neural network, setting an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer;

[0279] Step S2420, splicing the second toxin detection data with the reference data to form a toxin data fusion matrix;

[0280] Step S2430, input the toxin data fusion matrix into the toxin feature fusion model to obtain the toxin feature fusion vector.

[0281] The step S2430 includes:

[0282] Step S2431, inputting the toxin data fusion matrix into the input layer of the toxin feature fusion model, and converting the toxin data fusion matrix into a one-dimensional toxin feature vector;

[0283] Step S2432, the convolution layer performs a convolution operation on the one-dimensional toxin feature vector through a convolution kernel to extract local toxin features;

[0284] Step S2433, the pooling layer compresses the extracted local toxin features to obtain toxin compression features;

[0285] Step S2434, the toxin compression features are flattened and nonlinearly transformed in the fully connected layer to generate global toxin high-level features;

[0286] Step S2435, the output layer outputs the global toxin high-level features as a toxin feature fusion vector.

[0287] In the accurate identification module, obtaining the accurate identification result includes:

[0288] Step S3100, performing pattern recognition on the environmental variable time series to obtain an environmental state sequence;

[0289] Step S3200, calculating the historical accuracy of the preliminary recognition results under different environmental states in the environmental state sequence to obtain an environmental adaptive weight matrix;

[0290] Step S3300, using the environment adaptive weight matrix to perform weighted fusion on the preliminary recognition results to obtain accurate recognition results.

[0291] The step S3100 includes:

[0292] Step S3110, clustering the environmental variable time series to obtain environmental state categories;

[0293] Step S3120, time-series labeling of the environment state categories to obtain an environment state sequence.

[0294] The step S3200 includes:

[0295] Step S3210, counting the accuracy of the preliminary recognition results under each environmental state in the historical detection database to form an environmental state-accuracy mapping table;

[0296] Step S3220, converting the environment state-accuracy mapping table into an environment adaptive weight matrix, wherein the matrix elements in the environment adaptive weight matrix are weight coefficients of the preliminary recognition results under the corresponding environment state.

[0297] The step S3300 includes:

[0298] Step S3310, representing the preliminary identification result as a probability distribution vector, where the vector elements are the probabilities that the sample belongs to each mycotoxin category;

[0299] Step S3320, searching the weight vector corresponding to the current environment state from the environment adaptive weight matrix;

[0300] Step S3330, performing element-by-element multiplication of the weight vector and the probability distribution vector to obtain a corrected probability distribution vector;

[0301] Step S3340, normalize the corrected probability distribution vector to obtain an accurate recognition result.

[0302] Example 3

[0303] This embodiment discloses an electronic device, which may include one or more processors and one or more memories. The memories store computer-readable codes, which, when executed by the one or more processors, can execute the above-mentioned method for intelligent interpretation of the mycotoxin six-in-one detection card based on multidimensional data analysis.

[0304] The method or system according to the embodiment of the present application can also be implemented with the help of the architecture of an electronic device. The electronic device may include a bus, one or more CPUs, a read-only memory (ROM), a random access memory (RAM), a communication port connected to a network, an input / output component, a hard disk, etc. The storage device in the electronic device, such as a ROM or a hard disk, can store the intelligent interpretation method of the six-in-one detection card for mycotoxins based on multidimensional data analysis provided in this application. The intelligent interpretation method of the six-in-one detection card for mycotoxins based on multidimensional data analysis may, for example, include: collecting the first toxin detection data of six mycotoxins, performing feature extraction on the first toxin detection data, and obtaining the second toxin detection data; the first toxin detection data includes a fluorescence response time series, a colloidal gold response time series, and an environmental variable time series; establishing a historical detection database, constructing and training a mycotoxins classification model, and obtaining preliminary identification results of six mycotoxins based on the second toxin detection data and the mycotoxins classification model; based on the preliminary identification results of the six mycotoxins, combined with the environmental variable time series, an accurate identification result is obtained.

[0305] Furthermore, the electronic device may also include a user interface. Of course, the architecture disclosed in the present invention is only exemplary, and when implementing different devices, one or more components in the electronic device disclosed in the present invention may be omitted according to actual needs.

[0306] Example 4

[0307] This embodiment discloses a computer-readable storage medium, on which computer-readable instructions are stored. When the computer-readable instructions are executed by a processor, the method for intelligent interpretation of the six-in-one mycotoxin detection card based on multidimensional data analysis of the embodiment of the present application can be executed. The storage medium includes, but is not limited to, for example, volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and cache memory (cache), etc. Non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc.

[0308] In addition, according to the implementation of the present application, the process described in the above reference flow chart can be implemented as a computer software program. For example, the present application provides a non-transitory machine-readable storage medium, which stores machine-readable instructions, and the machine-readable instructions can be executed by a processor to execute instructions corresponding to the method steps provided in the present application, for example: collecting the first toxin detection data of six fungal toxins, extracting features from the first toxin detection data, and obtaining the second toxin detection data; the first toxin detection data includes a fluorescence response time series, a colloidal gold response time series, and an environmental variable time series; establishing a historical detection database, constructing and training a fungal toxin classification model, and obtaining preliminary identification results of six fungal toxins based on the second toxin detection data and the fungal toxin classification model; based on the preliminary identification results of the six fungal toxins, combined with the environmental variable time series, obtaining an accurate identification result. When the computer program is executed by a central processing unit (CPU), the above functions defined in the method of the present application are executed.

[0309] The methods, systems, and devices of the present application may be implemented in many ways. For example, the methods, systems, and devices of the present application may be implemented by software, hardware, firmware, or any combination of software, hardware, and firmware. The above order of steps for the method is for illustration only, and the steps of the method of the present application are not limited to the order specifically described above unless otherwise specified. In addition, in some embodiments, the present application may also be implemented as programs recorded in a recording medium, which include machine-readable instructions for implementing the method according to the present application. Therefore, the present application also covers recording media storing programs for executing the method according to the present application.

[0310] In addition, the parts of the above-mentioned technical solutions provided in the embodiments of the present application that are consistent with the implementation principles of the corresponding technical solutions in the prior art are not described in detail to avoid excessive redundancy.

[0311] The specific implementation modes as described above further describe the purpose, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation mode of the present invention and is not intended to limit the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for intelligently interpreting a six-in-one mycotoxin detection card based on multidimensional data analysis, which is used to detect mycotoxins in grains, oils and feeds, and is characterized in that: The method comprises: The first toxin detection data of six kinds of mycotoxins are collected, and the features of the first toxin detection data are extracted to obtain the second toxin detection data; the first toxin detection data include fluorescence response time series, colloidal gold response time series and environmental variable time series; the mycotoxins are aflatoxin B1, vomitoxin, zearalenone, fumonisins B1 and B2, ochratoxin A and T-2 toxin; the detection method of the mycotoxins is: aflatoxin B1, vomitoxin and zearalenone are detected by fluorescence method; fumonisins B1 and B2, ochratoxin A and T-2 toxin are detected by colloidal gold method; The second toxin detection data includes a fluorescence signal intensity characteristic, a fluorescence response rate characteristic, a colloidal gold color change rate characteristic, and an environmental variable fluctuation rate characteristic; The feature extraction of the first toxin detection data includes: extracting the fluorescence signal intensity feature and the fluorescence response rate feature from the fluorescence response time series; extracting the colloidal gold color change rate feature from the colloidal gold response time series; extracting the environmental variable volatility feature from the environmental variable time series; the extraction of the fluorescence signal intensity feature and the fluorescence response rate feature from the fluorescence response time series includes: calculating the maximum value, minimum value, average value and integrated area of ​​the fluorescence response time series to obtain the fluorescence signal intensity feature; calculating the slope, rise time and fall time of the fluorescence response time series to obtain the fluorescence response rate feature; The method of extracting the colloidal gold color change rate feature from the colloidal gold response time series includes: performing a differential operation on the colloidal gold response time series to obtain a color change rate sequence; calculating the maximum value, minimum value, average value and standard deviation of the color change rate sequence to obtain the colloidal gold color change rate feature; The step of extracting environmental variable volatility characteristics from environmental variable time series includes: performing sliding window analysis on the environmental variable time series to obtain a volatility sequence; calculating the maximum value, minimum value, average value and dynamic range of the volatility sequence to obtain environmental variable volatility characteristics; Establish a historical detection database, build and train a mycotoxin classification model, and obtain preliminary identification results of six mycotoxins based on the second toxin detection data and the mycotoxin classification model; Based on the preliminary identification results of the six mycotoxins, combined with the time series of environmental variables, obtain accurate identification results; The method of obtaining preliminary identification results of the six mycotoxins based on the second toxin detection data and the mycotoxin classification model includes: using the mycotoxin classification model to predict the second toxin detection data to obtain first identification results of the six mycotoxins; the first identification results are the predicted probabilities of the six mycotoxins. , , , , , ];in represents the predicted probability of the i-th mycotoxin among the six mycotoxins, 1≤i≤6; Setting the confidence threshold , represents the confidence threshold of the i-th mycotoxin among the six mycotoxins; if the predicted probabilities of the six mycotoxins are all greater than the confidence thresholds of the corresponding mycotoxins, the first recognition result is taken as the preliminary recognition result; otherwise, reference data similar to the second toxin detection data is retrieved from the historical detection database; the second toxin detection data and the reference data are feature fused to obtain a toxin feature fusion vector; the toxin feature fusion vector is input into the mycotoxin classification model to obtain a preliminary recognition result.

2. The method for intelligent interpretation of the mycotoxin six-in-one detection card based on multidimensional data analysis according to claim 1, characterized in that: Setting the confidence threshold include: Through the historical detection database, calculate the The accuracy of mycotoxins and recall ; Obtain the detection feature of the i-th mycotoxin from the second toxin detection data and calculate the Sensitivity scores of mycotoxin detection features; Calculate the predicted probability of the i-th mycotoxin The predicted probability of the jth mycotoxin The correlation coefficient ρ ij , 1≤j≤6, i≠j.

3. The method for intelligent interpretation of the mycotoxin six-in-one detection card based on multidimensional data analysis according to claim 1, characterized in that: The historical detection database includes n1 records; The retrieving reference data similar to the second toxin detection data from the historical detection database comprises: Calculate the similarity between the second toxin detection data and each record in the historical detection database to obtain a similarity list; Sort the similarity list and select the record with the highest similarity as the reference data.

4. The method for intelligent interpretation of the mycotoxin six-in-one detection card based on multidimensional data analysis according to claim 1, characterized in that: The feature fusion of the second toxin detection data and the reference data includes: Construct a toxin feature fusion model based on a convolutional neural network, and set up the input layer, convolution layer, pooling layer, fully connected layer, and output layer; splicing the second toxin detection data with the reference data to form a toxin data fusion matrix; The toxin data fusion matrix is ​​input into the toxin feature fusion model to obtain the toxin feature fusion vector.

5. The method for intelligent interpretation of the mycotoxin six-in-one detection card based on multidimensional data analysis according to claim 4, characterized in that: The step of inputting the toxin data fusion matrix into the toxin feature fusion model to obtain the toxin feature fusion vector includes: Input the toxin data fusion matrix into the input layer of the toxin feature fusion model, and convert the toxin data fusion matrix into a one-dimensional toxin feature vector; The convolution layer performs a convolution operation on the one-dimensional toxin feature vector through a convolution kernel to extract local toxin features; The pooling layer compresses the extracted local toxin features to obtain toxin compression features; The compressed features of toxins are flattened and nonlinearly transformed in the fully connected layer to generate global toxin high-level features; The output layer outputs the global toxin high-level features as a toxin feature fusion vector.

6. The method for intelligent interpretation of the mycotoxin six-in-one detection card based on multidimensional data analysis according to claim 1, characterized in that: Obtaining accurate recognition results includes: Perform pattern recognition on the time series of environmental variables to obtain the environmental state sequence; Calculate the historical accuracy of the preliminary recognition results under different environmental states in the environmental state sequence to obtain the environmental adaptive weight matrix; The environment-adaptive weight matrix is ​​used to perform weighted fusion on the preliminary recognition results to obtain accurate recognition results; The obtaining of the environmental state sequence comprises: Cluster the time series of environmental variables to obtain environmental state categories; Perform time series labeling on the environmental state categories to obtain the environmental state sequence; The obtaining of the environment adaptive weight matrix comprises: Count the accuracy of the preliminary recognition results under each environmental state in the historical detection database to form an environmental state-accuracy mapping table; The environment state-accuracy mapping table is converted into an environment adaptive weight matrix, wherein the matrix elements in the environment adaptive weight matrix are weight coefficients of the preliminary recognition results under the corresponding environment state.

7. A system for intelligent interpretation of a six-in-one mycotoxin detection card based on multidimensional data analysis, which is used to implement the method for intelligent interpretation of a six-in-one mycotoxin detection card based on multidimensional data analysis as claimed in any one of claims 1 to 6, characterized in that: The system comprises: Data preprocessing module: used to collect first toxin detection data of six fungal toxins, perform feature extraction on the first toxin detection data, and obtain second toxin detection data; the first toxin detection data includes fluorescence response time series, colloidal gold response time series and environmental variable time series; Preliminary identification module: used to establish a historical detection database, build and train a mycotoxin classification model, and obtain preliminary identification results of six mycotoxins based on the second toxin detection data and the mycotoxin classification model; Precise identification module: Based on the preliminary identification results of six fungal toxins and combined with the time series of environmental variables, accurate identification results are obtained.

Images