Yangtze river basin illegal fishing early warning method based on deep learning

Abstract

The invention provides a Yangtze river basin illegal fishing early warning method based on deep learning. The method comprises the following steps: acquiring illegal catching case data, extracting keywords in a case condition text by adopting a TF-IDF method, and constructing an illegal catching keyword data set; obtaining keyword frequency data based on a third-party network data platform, and forming an illegal fishing keyword frequency matrix; early warning keywords are screened in combination with information gain evaluation and a genetic algorithm; keyword frequency features are extracted, public opinion levels are marked, and a training data set is constructed; and inputting the training data set into a neural network model combining a long short-term memory (LSTM) network and a self-attention mechanism for training to obtain an illegal fishing public opinion early warning model. And finally, based on the real-time keyword frequency data, outputting the public opinion early warning level of the area to be subjected to early warning, and generating illegal fishing early warning information. The accuracy and timeliness of illegal fishing behavior early warning can be effectively improved, and the manual analysis burden is reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of illegal fishing early warning, and specifically relates to an illegal fishing early warning method in the Yangtze River Basin based on deep learning. Background Art

[0002] As one of China's major river systems, the Yangtze River Basin's ecological environment is crucial to the country's water conservation, fishery resource management, and ecological balance. In recent years, with the rapid development of internet technology, public concern about ecological and environmental issues such as illegal fishing has been expressed online, generating a wealth of online public opinion data. This data is crucial for timely identification and early warning of illegal fishing activities.

[0003] Currently, monitoring and early warning of illegal fishing activities mostly rely on traditional manual inspections, reporting, and local government law enforcement. However, these methods often suffer from slow response, insufficient coverage, and poor real-time performance. Furthermore, traditional public opinion monitoring methods can only identify hot topics within a specific range, making it difficult to effectively predict potential public opinion risks. Furthermore, most methods rely on single data sources and simple algorithms, making in-depth analysis difficult.

[0004] In the prior art, Chinese patent CN119273150A discloses a method and system for predicting illegal fishing risks in the Yangtze River Basin. This method, based on a fault tree analysis method, identifies illegal fishing activities in the Yangtze River Basin as top events, decomposes the causes of illegal fishing incidents, and uses a hierarchical analysis method based on the fault tree to determine the weights of each cause. The fault tree and hierarchical analysis results are combined to predict the causes of illegal fishing. However, this method still has the following problems: It lacks full utilization of online public opinion data and fails to integrate public feedback for dynamic perception. Secondly, the analysis method is relatively traditional and cannot effectively mine the deep-level time series features contained in public opinion data. Existing methods generally lack efficient processing mechanisms for text data on illegal fishing incidents, resulting in low keyword extraction accuracy and weak feature expression capabilities, affecting the accuracy and real-time performance of predictions. Furthermore, when faced with massive, high-noise network data, traditional feature extraction and early warning strategies based on rules or shallow models are difficult to adapt to dynamic and complex environmental changes, and suffer from feature redundancy and insufficient classification performance. At the same time, since illegal fishing incidents are obviously time-sensitive, existing methods generally do not consider the time decay factor and the dynamic changes in the importance of keywords at different time scales, resulting in insufficient ability of early warning models to identify emerging illegal fishing risks. Summary of the Invention

[0005] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide an illegal fishing early warning method in the Yangtze River Basin based on deep learning.

[0006] The purpose of the present invention can be achieved by the following technical solutions:

[0007] On the one hand, the present invention provides an early warning method for illegal fishing in the Yangtze River Basin based on deep learning, comprising the following steps:

[0008] Acquiring case data of multiple illegal fishing cases, wherein the case data includes the time of occurrence of the multiple illegal fishing cases, the location of the case, and case situation text data describing the case situation;

[0009] The text data of all illegal fishing cases were processed, and keywords were extracted using the TF-IDF method to construct an illegal fishing keyword dataset.

[0010] Based on the illegal fishing keyword dataset, obtaining frequency values of each keyword in the locations corresponding to each illegal fishing case and within a first preset time period through a third-party network data platform, and constructing an illegal fishing keyword frequency matrix;

[0011] Perform feature selection on the illegal fishing keyword frequency matrix, using a method combining information gain evaluation with genetic algorithm to screen out a first preset number of warning keywords;

[0012] Based on the screened warning keywords and corresponding frequency data, a training dataset for online public opinion warning on illegal fishing was constructed. This training dataset was then input into a neural network model based on a combination of a long short-term memory (LSTM) network and a self-attention mechanism for training, resulting in an illegal fishing online public opinion warning model.

[0013] Obtain warning keyword frequency data for the area to be warned within a preset time period, input it into the trained illegal fishing online public opinion warning model, and output the illegal fishing public opinion warning level corresponding to the area to be warned;

[0014] According to the public opinion warning level, corresponding illegal fishing warning information is generated.

[0015] Furthermore, the text data of all illegal fishing cases are processed, keywords are extracted based on the TF-IDF method, and an illegal fishing keyword dataset is constructed, which specifically includes:

[0016] Perform Chinese word segmentation on the case text data to obtain the word sequence after word segmentation;

[0017] Perform part-of-speech tagging on the words after segmentation, and select words that meet the preset part-of-speech requirements as candidate keywords;

[0018] For each candidate keyword, the word frequency is calculated. The word frequency TF calculation formula is:

[0019]

[0020] Among them, f t,d is the number of times the candidate keyword t appears in the text data of the d-th illegal fishing case, ∑ k f k,d is the sum of the occurrence times of all candidate keywords in the text data of the d-th illegal fishing case;

[0021] Calculate the inverse document frequency (IDF) of each candidate keyword. The IDF calculation formula is:

[0022]

[0023] Where N is the total number of cases in the case data, n t is the number of cases containing the candidate keyword t;

[0024] Based on the term frequency TF and the inverse document frequency IDF, the TF-IDF weight value of each candidate keyword is calculated. The calculation formula of the TF-IDF weight value is:

[0025] TF-IDF(t,d)=TF(t,d)×IDF(t)

[0026] All candidate keywords are sorted according to the TF-IDF weight values, and keywords with weight values higher than a first preset threshold are selected to construct an illegal fishing keyword dataset.

[0027] Furthermore, the illegal fishing keyword frequency matrix is a matrix with each keyword in the illegal fishing keyword data set as the column of the matrix, and the combination of the location corresponding to each illegal fishing case and the first preset time period of the time of occurrence as the row of the matrix. Each element in the matrix is a frequency curve of the corresponding keyword in the corresponding location and time period, wherein the first preset time period refers to the time interval formed by extending the second preset time period before and after the time of occurrence of the case.

[0028] Furthermore, the illegal fishing keyword frequency matrix is expressed as:

[0029]

[0030] Among them, M is the frequency matrix of illegal fishing keywords, F(t j ,l i ,T i ) indicates the keyword t j At the location l corresponding to the i-th illegal fishing case i and time period T iThe frequency curve within the first preset time period represents the trend of the frequency of keyword occurrence at different time points over time, m is the number of illegal fishing cases, and n is the number of keywords in the illegal fishing keyword data set.

[0031] Furthermore, the feature selection of the illegal fishing keyword frequency matrix is performed, and a method combining information gain evaluation with a genetic algorithm is used to screen out a first preset number of warning keywords, specifically including:

[0032] Based on the illegal fishing keyword frequency matrix, a feature data set is constructed, in which each data sample corresponds to the frequency curve F(t j ,l i ,T i ), label each data sample, the label is whether the case text data of the case contains the keyword, if it does, the label is 1; if not, the label is 0;

[0033] Based on the information gain evaluation method, the frequency curve F(t j ,l i ,T i ) and the information gain value IG(t j );

[0034] According to the information gain value IG(t j ) to sort and filter out keywords whose information gain value is greater than a second preset threshold to form a candidate keyword set;

[0035] Based on the candidate keyword set, a genetic algorithm is used to further optimize the feature subset and screen out warning keywords.

[0036] Furthermore, the information gain value IG(t j ) is calculated as:

[0037] IG(t j )=H(Y)-H(Y|F(t j ,l i ,T i ))

[0038] Among them, IG(t j ) is the keyword t j The information gain value of , Y represents the public opinion event label variable, 1 represents the existence of illegal fishing public opinion, 0 represents the absence, H(Y) is the entropy of the public opinion event label Y, H(Y∣F(t j ,l i ,T i )) represents the curve F(t j,l i ,T i ) under the condition of conditional entropy;

[0039] The H(Y) calculation formula is:

[0040]

[0041] Among them, p(y) is the probability that label y appears in all data samples;

[0042] The H(Y|F(t j ,l i ,T i ))The calculation formula is:

[0043]

[0044] Where p(f) is the frequency curve F(t j ,l i ,F i ) is the proportion of samples with the value f, H(Y|F(t j ,l i ,T i )=f) is the j ,l i ,T i )=the entropy of the public opinion event label Y under the condition f.

[0045] Furthermore, based on the candidate keyword set, a genetic algorithm is used to further optimize the feature subset, specifically including:

[0046] Initialize the population by randomly generating multiple individuals, each of which represents a feature subset. The feature subset S is a part of the candidate keyword set. Each individual is a binary vector of length g, where each binary value indicates whether the corresponding keyword is selected, specifically:

[0047] S=[s1,s2,…,s g ]

[0048] Among them, si=1 means selecting keyword t i , si=0 means not selecting keyword t i , g is the number of keywords in the candidate keyword set;

[0049] The fitness of each individual is evaluated. The fitness function is determined by calculating the comprehensive value of the combined information gain value of the selected feature subset and the penalty factor of the feature subset complexity. The fitness function is:

[0050] Fitness(S)=IGcomb(S)-λ·P(S)

[0051] in, is the combined information gain between the frequency curves of all keywords in the feature subset S and the label of the illegal fishing public opinion event, Fitness(S) is the fitness of the feature subset S, P(S) is the penalty factor, P(S) = |S| / g, |S| is the number of keywords selected in the feature subset S;

[0052] According to the fitness value of each individual, the parent individual is selected by roulette wheel selection or tournament selection method, and a crossover operation is performed to generate a new offspring individual. The crossover operation is a single-point or multi-point crossover operation, and a new feature subset is generated after the crossover;

[0053] Performing a mutation operation on the new generation of individuals, reversing certain binary bits with a first preset probability to generate mutant individuals;

[0054] By repeating the selection, crossover, and mutation operations, a new generation of feature subsets is iteratively generated until the stopping condition is met, and the feature subset with the highest fitness value is obtained. The keywords selected from the feature subset with the highest fitness value are used as warning keywords.

[0055] Furthermore, the illegal fishing online public opinion warning training dataset is constructed based on the screened warning keywords and corresponding frequency data, specifically including:

[0056] Obtain frequency curve data of warning keywords in various time periods, where the time periods are divided by month;

[0057] Perform feature extraction on the frequency curve data of each warning keyword in each time period to obtain the characteristic values of each warning keyword in each time period. The characteristic values include the frequency mean, standard deviation, maximum value and minimum value. The characteristic values are used as the feature vectors of the illegal fishing network public opinion warning training data set;

[0058] For each time period, the number of illegal fishing cases and the number of warning keywords in the case descriptions are counted. Based on the number of illegal fishing cases and the number of warning keywords in the case descriptions, the public opinion warning level for the time period is determined, specifically including:

[0059] Get the number of illegal fishing cases C in each time period t ;

[0060] Get the number of cases K involving warning keywords in the case descriptions in each time period t , where K t The calculation formula is:

[0061] (The warning keyword i appears in the case description)

[0062] According to the number of cases C t and the number of cases involving keywords K t , set the public opinion warning level L t :

[0063] like Then L t =3, indicating high-risk public opinion; if Then L t =2, indicating medium-risk public opinion; if Then L t =1, indicating low-risk public opinion;

[0064] Construct a training data set and compare the characteristic values of each warning keyword and the corresponding public opinion warning level L in each time period t as training samples.

[0065] Furthermore, the training data set is input into a neural network model based on a combination of a long short-term memory network (LSTM) and a self-attention mechanism for training to obtain an illegal fishing online public opinion early warning model, specifically including:

[0066] The feature vector of each time period in the training data set is input as a time series into the illegal fishing network public opinion warning model for prediction, and its corresponding public opinion warning level is used as the target label. The parameters of the illegal fishing network public opinion warning model are updated through the loss function to obtain the trained illegal fishing network public opinion warning model.

[0067] Furthermore, the loss function is:

[0068]

[0069] Where x is the number of samples in the training data set, L i is the true public opinion level of the i-th sample, is the predicted public opinion level of the i-th sample, α, β, γ are preset parameters, t i Indicates the time span from the i-th sample to the current time, F(t j ,l i ,T i ) indicates the keyword t j At the location l corresponding to the i-th illegal fishing case i and time period T i Frequency curve inside, δ(t j ) is the keyword weight, IG(t j ) is the keyword t j The information gain value is , and u is the number of warning keywords.

[0070] Compared with the prior art, the present invention has the following advantages:

[0071] (1) The present invention achieves the technical effects and advantages of efficient perception of online public opinion dynamics, accurate early warning and intelligent risk level assessment of illegal fishing activities in the Yangtze River Basin by introducing technical means of extracting keywords from illegal fishing case texts based on TF-IDF, combining information gain and genetic algorithm for feature selection, and using a neural network model that integrates long short-term memory network (LSTM) and self-attention mechanism for training and prediction.

[0072] (2) The present invention effectively extracts key information of illegal fishing incidents and improves the accuracy of early warning indicators through keyword extraction and high-frequency feature screening of case text data; through the feature optimization strategy combining information gain with genetic algorithm, it avoids the interference of redundant features on model performance, ensures the representativeness of early warning keywords and the efficiency of model input features; by constructing a keyword frequency change matrix and extracting its time series features, it fully explores the temporal and spatial variation patterns of illegal fishing public opinion events; at the same time, by adopting an LSTM neural network model combined with a self-attention mechanism, the model can focus on key time nodes and important feature changes when processing time series features, thereby improving the model's perception and prediction accuracy of illegal fishing public opinion risk trends.

[0073] (3) Based on the illegal fishing keyword dataset, the present invention captures the frequency of occurrence of each keyword in a specific location and time period on a third-party network data platform, constructs an illegal fishing keyword frequency matrix, and quantifies public opinion data into computable structured data, which facilitates subsequent regularity mining and modeling, and improves the systematicness and operability of data processing.

[0074] (4) The present invention defines the row and column structure of the illegal fishing keyword frequency matrix, with the combination of location and time period as rows, keywords as columns, and keyword frequency curves as elements, thereby achieving a continuous and serialized description of the dynamic characteristics of illegal fishing. This enables the data to have the ability to express continuous changing trends in time and space dimensions, which helps to capture the potential evolution patterns of illegal fishing behavior.

[0075] (5) The present invention adopts a feature selection method that combines information gain evaluation with a genetic algorithm to screen the illegal fishing keyword frequency matrix and select a first preset number of warning keywords. This effectively reduces redundant features while ensuring sufficient information, thereby improving the model training efficiency and the generalization ability of the final warning model.

[0076] (6) The present invention specifically calculates the information gain value between each keyword frequency curve and the illegal fishing public opinion label in the feature selection stage, screens keywords with higher information gain, and then optimizes the feature subset based on the genetic algorithm to further balance the information volume and complexity of the feature subset. Finally, the optimal warning keyword set is selected, which realizes the scientificity and optimization of feature selection and improves the sensitivity and accuracy of subsequent models to public opinion risks.

[0077] (7) The present invention constructs an illegal fishing online public opinion warning training data set, extracts the frequency curve characteristic values of each warning keyword, including the mean, standard deviation, maximum value and minimum value, and divides the public opinion warning level into two levels according to the number of cases and the keyword correlation, thereby realizing the digitalization and structured expression of the public opinion evolution characteristics, which helps to train an efficient and accurate warning model. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] Figure 1 is a flow chart of the method of the present invention;

[0079] Figure 2 It is the frequency curve diagram of each keyword of the present invention. DETAILED DESCRIPTION

[0080] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0081] This embodiment provides an early warning method for illegal fishing in the Yangtze River basin based on deep learning, such as Figure 1 As shown, the following steps are included:

[0082] Step S1: Acquire case data of multiple illegal fishing cases, where the case data includes the time and location of occurrence of the multiple illegal fishing cases and case situation text data describing the case situation;

[0083] Step S2: Process the text data of all illegal fishing cases, extract keywords based on the TF-IDF method, and construct an illegal fishing keyword dataset;

[0084] Step S2 specifically includes:

[0085] Perform Chinese word segmentation on the case text data to obtain the word sequence after word segmentation;

[0086] Perform part-of-speech tagging on the words after segmentation, and select words that meet the preset part-of-speech requirements as candidate keywords;

[0087] For each candidate keyword, calculate the word frequency. The word frequency TF calculation formula is:

[0088]

[0089] Among them, f t,d is the number of times the candidate keyword t appears in the text data of the d-th illegal fishing case, ∑ k f k,d is the sum of the occurrence times of all candidate keywords in the text data of the d-th illegal fishing case;

[0090] Calculate the inverse document frequency of each candidate keyword. The inverse document frequency IDF calculation formula is:

[0091]

[0092] Where N is the total number of cases in the case data, n t is the number of cases containing the candidate keyword t;

[0093] Based on the term frequency TF and inverse document frequency IDF, the TF-IDF weight value of each candidate keyword is calculated. The calculation formula of the TF-IDF weight value is:

[0094] TF-IDF(t,d)=TF(t,d)×IDF(t)

[0095] All candidate keywords are sorted according to TF-IDF weight values, and keywords with weight values higher than the first preset threshold are selected to construct an illegal fishing keyword dataset.

[0096] Step S3: Based on the illegal fishing keyword dataset, obtain the frequency value of each keyword in the location corresponding to each illegal fishing case and the first preset time period through a third-party network data platform to construct an illegal fishing keyword frequency matrix; wherein the third-party network data platform is Baidu Index, and the frequency value is as follows: Figure 2 shown.

[0097] The illegal fishing keyword frequency matrix is composed of the keywords in the illegal fishing keyword dataset as the columns of the matrix, and the locations of each illegal fishing case and the combination of the first preset time period of the occurrence time as the rows of the matrix. Each element in the matrix is the frequency curve of the corresponding keyword in the corresponding location and time period, where the first preset time period refers to the time interval formed by extending the second preset time period before and after the case occurred. The illegal fishing keyword frequency matrix is expressed as:

[0098]

[0099] Among them, M is the frequency matrix of illegal fishing keywords, F(t j ,l i ,T i ) indicates the keyword t j At the location l corresponding to the i-th illegal fishing case i and time period T i The frequency curve within represents the trend of the frequency of keywords at different time points within the first preset time period, m is the number of illegal fishing cases, and n is the number of keywords in the illegal fishing keyword dataset.

[0100] Step S4: performing feature selection on the illegal fishing keyword frequency matrix, using a method combining information gain evaluation with a genetic algorithm to screen out a first preset number of warning keywords;

[0101] Step S4 specifically includes:

[0102] Based on the illegal fishing keyword frequency matrix, a feature data set is constructed, in which each data sample corresponds to the frequency curve F(t j ,l i ,T i ), label each data sample, the label is whether the case text data of the case contains the keyword, if it does, the label is 1; if not, the label is 0;

[0103] Based on the information gain evaluation method, the frequency curve F(t j ,l i ,T i ) and the information gain value IG(t j ); Information gain value IG(t j ) is calculated as:

[0104] IG(t j )=H(Y)-H(Y|F(t j ,l i ,T i ))

[0105] Among them, IG(t j ) is the keyword t j The information gain value of , Y represents the public opinion event label variable, 1 represents the existence of illegal fishing public opinion, 0 represents the absence, H(Y) is the entropy of the public opinion event label Y, H(Y∣F(t j ,l i ,T i )) represents the curve F(t j ,l i ,T i) under the condition of conditional entropy;

[0106] The calculation formula for H(Y) is:

[0107]

[0108] Among them, p(y) is the probability that label y appears in all data samples;

[0109] H(Y|F(t j ,l i ,T i ))The calculation formula is:

[0110]

[0111] Where p(f) is the frequency curve F(t j ,l i ,T i ) is the proportion of samples with the value f, H(Y|F(t j ,l i ,T i )=f) is the j ,l i ,T i )=the entropy of the public opinion event label Y under the condition f.

[0112] According to the information gain value IG(t j ) to sort and filter out keywords whose information gain value is greater than a second preset threshold to form a candidate keyword set;

[0113] Based on the candidate keyword set, a genetic algorithm is used to further optimize the feature subset and screen out warning keywords.

[0114] Based on the candidate keyword set, a genetic algorithm is used to further optimize the feature subset, including:

[0115] Initialize the population by randomly generating multiple individuals. Each individual represents a feature subset. The feature subset S is a part of the candidate keyword set. Each individual is a binary vector of length g, where each binary value indicates whether the corresponding keyword is selected. Specifically:

[0116] S=[s1,s2,…,s g ]

[0117] Among them, si=1 means selecting keyword t i , si=0 means not selecting keyword t i , g is the number of keywords in the candidate keyword set;

[0118] The fitness of each individual is evaluated. The fitness function is determined by calculating the combined information gain value of the selected feature subset and the penalty factor of the feature subset complexity. The fitness function is:

[0119] Fitness(S)=IGcomb(S)-λ·P(S)

[0120] in, is the combined information gain between the frequency curves of all keywords in the feature subset S and the label of the illegal fishing public opinion event, Fitness(S) is the fitness of the feature subset S, P(S) is the penalty factor, P(S) = |S| / g, |S| is the number of keywords selected in the feature subset S;

[0121] According to the fitness value of each individual, the parent individuals are selected through roulette wheel selection or tournament selection method, and a crossover operation is performed to generate new offspring individuals. The crossover operation is a single-point or multi-point crossover operation, and a new feature subset is generated after crossover;

[0122] Performing a mutation operation on the new generation of individuals, reversing certain binary bits with a first preset probability to generate mutant individuals;

[0123] By repeating the selection, crossover, and mutation operations, a new generation of feature subsets is iteratively generated until the stopping condition is met, and the feature subset with the highest fitness value is obtained. The keywords selected from the feature subset with the highest fitness value are used as warning keywords.

[0124] Step S5: Based on the screened warning keywords and corresponding frequency data, a training dataset for online public opinion warning on illegal fishing is constructed. The training dataset is then input into a neural network model based on a combination of a long short-term memory (LSTM) network and a self-attention mechanism for training, thereby obtaining an online public opinion warning model for illegal fishing.

[0125] Among them, based on the screened warning keywords and corresponding frequency data, an illegal fishing network public opinion warning training dataset is constructed, specifically including:

[0126] Obtain frequency curve data of warning keywords in each time period, with the time period divided by month;

[0127] Feature extraction was performed on the frequency curve data of each warning keyword in each time period to obtain the characteristic values of each warning keyword in each time period. The characteristic values include the frequency mean, standard deviation, maximum value and minimum value. The characteristic values were used as the feature vectors of the illegal fishing network public opinion warning training dataset;

[0128] For each time period, the number of illegal fishing cases and the number of warning keywords in the case descriptions are counted. Based on the number of illegal fishing cases and the number of warning keywords in the case descriptions, the public opinion warning level for the time period is determined, specifically including:

[0129] Get the number of illegal fishing cases C in each time period t ;

[0130] Get the number of cases K involving warning keywords in the case descriptions in each time period t , where K t The calculation formula is:

[0131] (The warning keyword i appears in the case description)

[0132] According to the number of cases C t and the number of cases involving keywords K t , set the public opinion warning level L t :

[0133] like

[0134] Then L t =3, indicating high-risk public opinion; if Then L t =2, indicating medium-risk public opinion; if Then L t =1, indicating low-risk public opinion;

[0135] Construct a training data set and compare the characteristic values of each warning keyword and the corresponding public opinion warning level L in each time period t as training samples.

[0136] The training dataset was input into a neural network model based on a combination of a long short-term memory network (LSTM) and a self-attention mechanism for training, resulting in an illegal fishing online public opinion early warning model. Specifically, the model includes:

[0137] The feature vector of each time period in the training data set is input as a time series into the illegal fishing network public opinion warning model for prediction, and its corresponding public opinion warning level is used as the target label. The parameters of the illegal fishing network public opinion warning model are updated through the loss function to obtain the trained illegal fishing network public opinion warning model.

[0138] The loss function is:

[0139]

[0140] Where x is the number of samples in the training data set, L i is the true public opinion level of the i-th sample, is the predicted public opinion level of the i-th sample, α, β, γ are preset parameters, t i Indicates the time span from the i-th sample to the current time, F(t j ,l i ,T i ) indicates the keyword t j At the location l corresponding to the i-th illegal fishing case i and time period T i Frequency curve inside, δ(t j ) is the keyword weight, IG(t j ) is the keyword t j The information gain value is , and u is the number of warning keywords.

[0141] Step S6: Obtaining warning keyword frequency data for the area to be warned within a preset time period, inputting the data into the trained illegal fishing online public opinion warning model, and outputting the illegal fishing public opinion warning level corresponding to the area to be warned;

[0142] Step S7: Generate corresponding illegal fishing warning information according to the public opinion warning level.

[0143] This embodiment effectively transforms public opinion information on illegal fishing from unstructured text into structured features, and through deep learning technology, accurately categorizes public opinion risks and provides intelligent early warnings. Compared with existing technologies, this embodiment offers the following technical benefits and advantages: it effectively improves the accuracy and timeliness of public opinion warnings on illegal fishing, reduces the workload of manual analysis, enhances sensitivity to changing trends in illegal fishing behavior, and exhibits good scalability and practical application value.

[0144] If the above functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0145] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and such modifications or substitutions are intended to be within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be subject to the scope of protection of the claims.

Claims

1. A deep learning-based early warning method for illegal fishing in the Yangtze River Basin, characterized by: The following steps are involved: Acquiring case data of multiple illegal fishing cases, wherein the case data includes the time of occurrence of the multiple illegal fishing cases, the location of the case, and case situation text data describing the case situation; The text data of all illegal fishing cases were processed, and keywords were extracted using the TF-IDF method to construct an illegal fishing keyword dataset. Based on the illegal fishing keyword dataset, obtaining frequency values of each keyword in the locations corresponding to each illegal fishing case and within a first preset time period through a third-party network data platform, and constructing an illegal fishing keyword frequency matrix; Perform feature selection on the illegal fishing keyword frequency matrix, using a method combining information gain evaluation with genetic algorithm to screen out a first preset number of warning keywords; Based on the screened warning keywords and corresponding frequency data, a training dataset for online public opinion warning on illegal fishing was constructed. This training dataset was then input into a neural network model based on a combination of a long short-term memory (LSTM) network and a self-attention mechanism for training, resulting in an illegal fishing online public opinion warning model. Obtain warning keyword frequency data for the area to be warned within a preset time period, input it into the trained illegal fishing online public opinion warning model, and output the illegal fishing public opinion warning level corresponding to the area to be warned; According to the public opinion warning level, corresponding illegal fishing warning information is generated.

2. The method for early warning of illegal fishing in the Yangtze River basin based on deep learning according to claim 1 is characterized in that: The text data of all illegal fishing cases are processed, and keywords are extracted based on the TF-IDF method to construct an illegal fishing keyword dataset, which specifically includes: Perform Chinese word segmentation on the case text data to obtain the word sequence after word segmentation; Perform part-of-speech tagging on the words after segmentation, and select words that meet the preset part-of-speech requirements as candidate keywords; For each candidate keyword, the word frequency is calculated. The word frequency TF calculation formula is: Among them, f t，d is the number of times the candidate keyword t appears in the text data of the d-th illegal fishing case, ∑ k f k，d is the sum of the occurrence times of all candidate keywords in the text data of the d-th illegal fishing case; Calculate the inverse document frequency (IDF) of each candidate keyword. The IDF calculation formula is: Where N is the total number of cases in the case data, n t is the number of cases containing the candidate keyword t; Based on the term frequency TF and the inverse document frequency IDF, the TF-IDF weight value of each candidate keyword is calculated. The calculation formula of the TF-IDF weight value is: TF-IDF(t,d)=TF(t,d)×IDF(t) All candidate keywords are sorted according to the TF-IDF weight values, and keywords with weight values higher than a first preset threshold are selected to construct an illegal fishing keyword dataset.

3. The method for early warning of illegal fishing in the Yangtze River basin based on deep learning according to claim 1 is characterized in that: The illegal fishing keyword frequency matrix is a matrix with each keyword in the illegal fishing keyword dataset as the matrix column, and the combination of the location corresponding to each illegal fishing case and the first preset time period of the time of occurrence as the matrix row. Each element in the matrix is the frequency curve of the corresponding keyword in the corresponding location and time period, wherein the first preset time period refers to the time interval formed by extending the second preset time period before and after the time of the case.

4. The method for early warning of illegal fishing in the Yangtze River basin based on deep learning according to claim 3 is characterized in that: The illegal fishing keyword frequency matrix is expressed as: Among them, M is the frequency matrix of illegal fishing keywords, F(t j , l i , T i ) indicates the keyword t j At the location l corresponding to the i-th illegal fishing case i and time period T i The frequency curve within the first preset time period represents the trend of the frequency of keyword occurrence at different time points over time, m is the number of illegal fishing cases, and n is the number of keywords in the illegal fishing keyword data set.

5. The method for early warning of illegal fishing in the Yangtze River basin based on deep learning according to claim 1 is characterized in that: The feature selection of the illegal fishing keyword frequency matrix is performed by combining information gain evaluation with a genetic algorithm to screen out a first preset number of warning keywords, specifically including: Based on the illegal fishing keyword frequency matrix, a feature data set is constructed, in which each data sample corresponds to the frequency curve F(t j , l i , T i ), label each data sample, the label is whether the case text data of the case contains the keyword, if it does, the label is 1; if not, the label is 0; Based on the information gain evaluation method, the frequency curve F(t j , l i , T i ) and the information gain value IG(t j ); According to the information gain value IG(t j ) to sort and filter out keywords whose information gain value is greater than a second preset threshold to form a candidate keyword set; Based on the candidate keyword set, a genetic algorithm is used to further optimize the feature subset and screen out warning keywords.

6. The method for early warning of illegal fishing in the Yangtze River basin based on deep learning according to claim 5 is characterized in that: The information gain value IG(t j ) is calculated as: IG(t j )=H(Y)-H(Y∣F(t j ,l i ,T i )) Among them, IG(t j ) is the keyword t j The information gain value of , Y represents the public opinion event label variable, 1 represents the existence of illegal fishing public opinion, 0 represents the absence, H(Y) is the entropy of the public opinion event label Y, H(Y∣F(t j , l i , T i )) represents the curve F(t j , l i , T i ) under the condition of conditional entropy; The H(Y) calculation formula is: Among them, p(y) is the probability that label y appears in all data samples; The H(Y|F(t j , l i , T i ))The calculation formula is: Where p(f) is the frequency curve F(t j , l i , T i ) is the proportion of samples with the value f, H(Y|F(t j , l i , T i )=f) is the j , l i , T i )=the entropy of the public opinion event label Y under the condition f.

7. The method for early warning of illegal fishing in the Yangtze River basin based on deep learning according to claim 5 is characterized in that: Based on the candidate keyword set, the genetic algorithm is used to further optimize the feature subset, specifically including: Initialize the population by randomly generating multiple individuals, each of which represents a feature subset. The feature subset S is a part of the candidate keyword set. Each individual is a binary vector of length g, where each binary value indicates whether the corresponding keyword is selected, specifically: S=[s1,s2,…,s g ] Among them, si=1 means selecting keyword t i , si=0 means not selecting keyword t i , g is the number of keywords in the candidate keyword set; The fitness of each individual is evaluated. The fitness function is determined by calculating the comprehensive value of the combined information gain value of the selected feature subset and the penalty factor of the feature subset complexity. The fitness function is: Fitness(S)=IGcomb(S)-λ·P(S) in, is the combined information gain between the frequency curves of all keywords in the feature subset S and the label of the illegal fishing public opinion event, Fitness(S) is the fitness of the feature subset S, P(S) is the penalty factor, P(S) = |S| / g, |S| is the number of keywords selected in the feature subset S; According to the fitness value of each individual, the parent individual is selected by roulette wheel selection or tournament selection method, and a crossover operation is performed to generate a new offspring individual. The crossover operation is a single-point or multi-point crossover operation, and a new feature subset is generated after the crossover; Performing a mutation operation on the new generation of individuals, reversing certain binary bits with a first preset probability to generate mutant individuals; By repeating the selection, crossover, and mutation operations, a new generation of feature subsets is iteratively generated until the stopping condition is met, and the feature subset with the highest fitness value is obtained. The keywords selected from the feature subset with the highest fitness value are used as warning keywords.

8. The method for early warning of illegal fishing in the Yangtze River basin based on deep learning according to claim 1 is characterized in that: The illegal fishing online public opinion warning training dataset is constructed based on the screened warning keywords and corresponding frequency data, specifically including: Obtain frequency curve data of warning keywords in various time periods, where the time periods are divided by month; Perform feature extraction on the frequency curve data of each warning keyword in each time period to obtain the characteristic values of each warning keyword in each time period. The characteristic values include the frequency mean, standard deviation, maximum value and minimum value. The characteristic values are used as the feature vectors of the illegal fishing network public opinion warning training data set; For each time period, the number of illegal fishing cases and the number of warning keywords in the case descriptions are counted. Based on the number of illegal fishing cases and the number of warning keywords in the case descriptions, the public opinion warning level for the time period is determined, specifically including: Get the number of illegal fishing cases C in each time period t ; Get the number of cases K involving warning keywords in the case descriptions in each time period t , where K t The calculation formula is: According to the number of cases C t and the number of cases involving keywords K t , set the public opinion warning level L t : like Then L t =3, indicating high-risk public opinion; if Then L t =2, indicating medium-risk public opinion; if Then L t =1, indicating low-risk public opinion; Construct a training data set and compare the characteristic values of each warning keyword and the corresponding public opinion warning level L in each time period t as training samples.

9. The method for early warning of illegal fishing in the Yangtze River basin based on deep learning according to claim 1 or 8, characterized in that: The training data set is input into a neural network model based on a combination of a long short-term memory network (LSTM) and a self-attention mechanism for training to obtain an illegal fishing network public opinion early warning model, specifically including: The feature vector of each time period in the training data set is input as a time series into the illegal fishing network public opinion warning model for prediction, and the corresponding public opinion warning level is used as the target label. The parameters of the illegal fishing network public opinion warning model are updated through the loss function to obtain the trained illegal fishing network public opinion warning model.

10. The method for early warning of illegal fishing in the Yangtze River basin based on deep learning according to claim 9 is characterized in that: The loss function is: Where x is the number of samples in the training data set, L i is the true public opinion level of the i-th sample, is the predicted public opinion level of the i-th sample, α, β, γ are preset parameters, t i Indicates the time span from the i-th sample to the current time, F(t j , l i , T i ) indicates the keyword t j At the location l corresponding to the i-th illegal fishing case i and time period T i Frequency curve inside, δ(t j ) is the keyword weight, IG(t j ) is the keyword t j The information gain value is , and u is the number of warning keywords.

Images