A method and device for analyzing the remediation effect of microplastic-contaminated soil

Abstract

The present application relates to the field of environmental monitoring technology, and in particular to a method and device for analyzing the restoration effect of microplastic-contaminated soil; the method comprises: obtaining current soil index data of microplastic pollution; determining, according to the current soil index data, an analysis result corresponding to the current soil index data through a first analysis model; wherein the analysis result indicates the restoration degree of the microplastic-contaminated soil; the network structure of the first analysis model is a first extreme learning machine; the first weight matrix of the hidden layer in the first extreme learning machine is determined according to the weighted sum of the first weight matrices of the hidden layers in multiple second extreme learning machines; the second extreme learning machine is obtained by training according to a training sample set including historical soil index data; the present application realizes batch analysis of current soil index data through the first analysis model, so the present application can meet the growing analysis needs of soil restoration effect.

Description

Technical Field

[0001] The present application relates to the field of environmental monitoring technology, and in particular to a method and device for analyzing the remediation effect of microplastic-contaminated soil. Background Art

[0002] Microplastics (less than 5 Once microplastics (such as plastic fragments or fibers) enter the soil, they will destroy the soil structure, affect soil microorganisms, and even affect the growth of crops. Therefore, soil remediation work targeting microplastic pollution is very important.

[0003] In actual implementation, since soil remediation work for microplastic pollution is very complex and difficult to evaluate, it is necessary to regularly analyze the soil remediation effect of microplastic pollution in order to formulate the next soil remediation plan in a timely and accurate manner.

[0004] In the prior art, the analysis method of soil remediation effect for microplastic pollution mainly relies on physical, chemical and biological analysis methods in the laboratory. In actual application, the above-mentioned physical, chemical and biological analysis methods are more dependent on high-tech personnel and high-precision analysis equipment, resulting in defects such as long time consumption and high cost. Therefore, the above-mentioned physical, chemical and biological analysis methods are difficult to meet the growing analysis needs of soil remediation effect, which ultimately affects the remediation of microplastic-contaminated soil. Summary of the invention

[0005] In view of this, the purpose of this application is to provide a method and device for analyzing the remediation effect of microplastic contaminated soil, so as to solve the technical problem that the physical, chemical, and biological analysis methods in the prior art are difficult to meet the growing analysis needs of soil remediation effects, which ultimately affects the remediation of microplastic contaminated soil.

[0006] In the first aspect, the present application provides a method for analyzing the remediation effect of microplastic contaminated soil, comprising:

[0007] Obtain current soil indicator data for microplastic pollution;

[0008] According to the current soil index data, determining an analysis result corresponding to the current soil index data through a first analysis model;

[0009] Among them, the analysis results indicate the degree of restoration of the microplastic contaminated soil; the network structure of the first analysis model is a first extreme learning machine; the first weight matrix of the hidden layer in the first extreme learning machine is determined according to the weighted sum of the first weight matrices of the hidden layers in multiple second extreme learning machines; the second extreme learning machine is trained according to a training sample set including historical soil indicator data.

[0010] In a second aspect, the present application provides a device for analyzing the remediation effect of microplastic-contaminated soil, including: a protective gear acquisition module and an analysis module;

[0011] The data acquisition module is used to obtain current soil indicator data of microplastic pollution;

[0012] The analysis module is used to determine the analysis result corresponding to the current soil index data through a first analysis model according to the current soil index data;

[0013] Among them, the analysis results indicate the degree of restoration of the microplastic contaminated soil; the network structure of the first analysis model is a first extreme learning machine; the first weight matrix of the hidden layer in the first extreme learning machine is determined according to the weighted sum of the first weight matrices of the hidden layers in multiple second extreme learning machines; the second extreme learning machine is trained according to a training sample set including historical soil indicator data.

[0014] Beneficial effects:

[0015] The present application provides a method for analyzing the restoration effect of microplastic-contaminated soil, comprising: obtaining current soil index data of microplastic pollution; determining an analysis result corresponding to the current soil index data through a first analysis model according to the current soil index data; wherein the analysis result indicates the restoration degree of the microplastic-contaminated soil; the network structure of the first analysis model is a first extreme learning machine; the first weight matrix of the hidden layer in the first extreme learning machine is determined according to the weighted sum of the first weight matrices of the hidden layers in multiple second extreme learning machines; the second extreme learning machine is obtained by training according to a training sample set including historical soil index data;

[0016] In summary, the present application realizes batch analysis of current soil index data through the first analysis model, so the present application can meet the growing analysis needs of soil remediation effects. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings used in the embodiments of the present application will be briefly introduced below. The following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0018] Figure 1 A schematic diagram of a process for analyzing the effect of microplastic contaminated soil remediation provided in an embodiment of the present application;

[0019] Figure 2 Provided in the embodiments of this application Schematic diagram of network structure;

[0020] Figure 3 A schematic diagram of the application process of the first analysis model and the second analysis model provided in the embodiment of the present application;

[0021] Figure 4 This is a schematic diagram of the structure of the device for analyzing the remediation effect of microplastic-contaminated soil provided in an embodiment of the present application. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution of the present application will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present application.

[0023] First, this application provides a method for analyzing the remediation effect of microplastic contaminated soil, such as Figure 1 As shown, Figure 1 A schematic diagram of a process for analyzing the effect of microplastic contaminated soil remediation provided in an embodiment of the present application, the method comprising: S110 to S120, the details of which are as follows:

[0024] S110: Obtain current soil indicator data for microplastic pollution.

[0025] Specifically, the current soil indicator data are physical indicator data, chemical indicator data and / or biological indicator data collected from soil contaminated by microplastics through high-precision soil analysis instruments, environmental monitoring equipment and other equipment.

[0026] In actual implementation, the physical and chemical index data included in the soil index data include at least three of the following index data: soil moisture content , organic matter content , pH value , number of microplastic particles , Heavy metal content , nitrogen (and / or phosphorus, potassium) content , soil temperature , soil conductivity , soil density or microbial activity ; Among them, soil moisture content Indicates the moisture content in the soil; organic matter content Indicates the percentage of organic matter in the soil; pH value Indicates soil pH; number of microplastic particles Indicates the amount of microplastics in the soil; heavy metal content Indicates the total content of heavy metals such as lead and mercury; nitrogen (and / or phosphorus, potassium) content Indicates soil fertility; soil temperature Affects soil chemical reaction rate and biological activity; soil conductivity Indicates soil salt content and ion activity; soil density Affects root growth and water circulation; microbial activity Indicates the activity of microbial communities in the soil;

[0027] S120: Determine, according to the current soil index data, an analysis result corresponding to the current soil index data through a first analysis model;

[0028] Among them, the analysis results indicate the degree of restoration of microplastic contaminated soil; the network structure of the first analysis model is a first extreme learning machine; the first weight matrix of the hidden layer in the first extreme learning machine is determined according to the weighted sum of the first weight matrices of the hidden layers in multiple second extreme learning machines; the second extreme learning machine is trained according to a training sample set including historical soil indicator data.

[0029] Specifically, after obtaining the current soil index data, the current soil index data is input into the first analysis model, and the first analysis model is used to identify and classify the current soil index data to obtain an analysis result; the analysis result is one of multiple labels used by the first analysis model during the training process; the labels include: poor repair effect, general repair effect, good repair effect, excellent repair effect and other categories.

[0030] In one implementation, before S110, the method further includes: Step (1) to Step (3), the details of which are as follows:

[0031] Step (1): Obtain a first training sample set including historical soil indicator data of microplastic pollution.

[0032] Specifically, the samples in the first training sample set are historical physical indicator data, historical chemical indicator data and / or historical biological indicator data of soil collected from soil contaminated by microplastics through high-precision soil analysis instruments, environmental monitoring equipment and other equipment.

[0033] In actual implementation, the physical and chemical index data included in the samples in the first training sample set include at least three of the following index data: soil moisture content , organic matter content , pH value , number of microplastic particles , heavy metal content , nitrogen (and / or phosphorus, potassium) content , soil temperature , soil conductivity , soil density or microbial activity ;

[0034] Soil moisture content Indicates the moisture content in the soil; organic matter content Indicates the percentage of organic matter in the soil; pH value Indicates soil pH; number of microplastic particles Indicates the amount of microplastics in the soil; heavy metal content Indicates the total content of heavy metals such as lead and mercury; nitrogen (and / or phosphorus, potassium) content Indicates soil fertility; soil temperature Affects soil chemical reaction rate and biological activity; soil conductivity Indicates soil salt content and ion activity; soil density Affects root growth and water circulation; microbial activity Indicates the activity of microbial communities in the soil;

[0035] In practical applications, the physical and chemical index data included in the samples in the first training sample set may also be deleted or expanded according to actual needs. This application does not limit the number and content of the physical and chemical index data included in the samples in the first training sample set.

[0036] After the first training sample set is determined, samples in the first training sample set may be manually labeled, and the labels used for the labeling may include: poor restoration effect, average restoration effect, good restoration effect, and excellent restoration effect, a total of 4 categories.

[0037] Step (2): performing sample expansion on the first training sample set using a sample expansion model to obtain a second training sample set;

[0038] Among them, the model parameters of the sample expansion model are updated according to the gradient, loss function value and learning rate during the training process.

[0039] Specifically, there are relatively few studies on microplastic pollution in the prior art, and there is also relatively little data on microplastic pollution, so the number of samples in the first training sample set used for training is usually insufficient. If the number of samples is insufficient, the generalization of the analysis model obtained through training will be poor. For this reason, the embodiment of the present application expands the samples of the first training sample set, that is, increases the number of samples to obtain the second training sample set.

[0040] In the embodiment of the present application, the network structure of the sample expansion model is a generative adversarial network, which includes a discriminator and a generator; the task of the generator is to generate fake samples similar to real samples, and the task of the discriminator is to distinguish fake samples from real samples. In order to improve the accuracy and relevance of fake samples, the embodiment of the present application optimizes the training process of the generative adversarial network through an adaptive error compensation mechanism, so that its model parameters are updated according to the gradient, loss function value and learning rate during the training process, so as to generate sufficiently realistic fake samples.

[0041] In one implementation, before step (2), the method further includes: step (4) to step (5), where steps (4) to (5) are the training process of the sample expansion model, and the details are as follows:

[0042] Step (4): Obtain a third training sample set including historical soil indicator data of microplastic pollution.

[0043] Step (5): iteratively training the generative adversarial network using the third training sample set until the number of iterative training times meets a first threshold to obtain a sample expansion model;

[0044] Specifically, the value of the first threshold can be set to 1000, that is, when When , the training of the generative adversarial network is stopped, and the generative adversarial network obtained by training is determined as the sample expansion model; the value of the first threshold can also be determined according to actual needs, and this application does not make specific limitations on this.

[0045] Among them, for the generation of adversarial networks The training process of round training includes: step (5.1) to step (5.2), the details are as follows:

[0046] Step (5.1): According to The historical model parameters, historical gradients of historical model parameters, historical loss function values, learning rates and parameter update formulas corresponding to the generator and discriminator in the round of training are determined to be used for the first The target model parameters corresponding to the generator and discriminator of the round of training.

[0047] Step (5.2): According to the target model parameters corresponding to the generator and the discriminator respectively, the generative adversarial network is trained using the third training sample set.

[0048] Specifically, is a positive integer; in the process of training the generative adversarial network, it is first necessary to randomly initialize the model parameters of the generator and discriminator in the generative adversarial network, that is, randomly generate model parameters; among them, the network functions corresponding to the generator and the discriminator are as follows:

[0049] ; ;

[0050] In the formula, represents a generator, represents a generator function; represents the discriminator, represents the discriminator function; and Represent the model parameters of the generator and discriminator respectively; and Represent the weight matrices of the generator and discriminator respectively; and Represent the bias vectors of the generator and the discriminator respectively; Represents the noise data that needs to be input to the generator during training.

[0051] In the prior art, the conditional vector is set according to the soil indicator data of microplastic pollution , the conditional vector With noisy data They are input into the generator together to ensure that the fake samples generated by the generator have characteristics corresponding to the physical and chemical indicator data. Represents the samples in the first training sample set, that is, the true samples.

[0052] In the embodiment of the present application, The target model parameters corresponding to the generator and discriminator in the round training are determined according to the historical model parameters, the historical gradients of the historical model parameters, the historical loss function values, the learning rate and the parameter update formula; wherein, the target model parameters indicate the model parameters used in the current round, and the historical model parameters are the model parameters used in the training rounds before the current round.

[0053] ① In one implementation, the first parameter update formula of the generator is as follows:

[0054] ;

[0055] In the formula, Indicated in Generator function used in round training Target model parameters; Indicated in Generator function used in round training Historical model parameters; Represents the preset learning rate of the generator and discriminator; Indicates that based on The generator determined by the round of training The influence coefficient of the weight matrix of each network layer; Indicates that based on The historical gradients of the historical model parameters of the generator determined by the rounds of training; Indicated in The historical loss function value of the generator determined in the round of training; Indicates that based on The increment of the model parameters of the generator determined by the round of training.

[0056] ② In one implementation, the second parameter update formula of the discriminator is as follows:

[0057] ;

[0058] In the formula, Indicated in The discriminator function used in the round of training Target model parameters; Indicated in The discriminator function used in the round of training Historical model parameters; Indicates that based on The discriminator determined by the round of training The influence coefficient of the weight matrix of each network layer; Indicates that based on The historical gradients of the historical model parameters of the discriminator determined by the round of training; Indicates that based on The historical loss function value of the discriminator determined by the round of training.

[0059] ③In one implementation, It is determined according to the first coefficient update formula, and the first coefficient update formula is as follows:

[0060] ; ;

[0061] In the formula, Indicates the preset adjustment coefficient, which is used to control the adjustment range; in actual implementation, The value of can be set to 0.5; Indicates that based on The generator determined by the round of training The sensitivity coefficient of the weight matrix of each network layer; in actual implementation, The initial value of can be set to 1; Indicates that based on The generator determined by the round of training The weight matrix of each network layer.

[0062] ④In one implementation, It is determined according to the parameter incremental update formula, which is as follows:

[0063] ;

[0064] ; ; ;

[0065] In the formula, represents new noise data; represents noisy data; Indicated in The generator function value determined in the training round; Indicates that based on The discriminator function value determined by the round of training The output probability of Represents the preset fusion coefficient, which is used to adjust the conditional vector The influence of the generation process can be The value of is set to 0.3. Represents the concatenation operation of vectors; represents the conditional vector.

[0066] ⑤In one implementation, is determined according to the second coefficient update formula, and the second coefficient update formula is as follows:

[0067] ; ;

[0068] In the formula, Indicates that based on The discriminator determined by the round of training The sensitivity coefficient of the weight matrix of each network layer; Indicates that based on The discriminator determined by the round of training The weight matrix of each network layer.

[0069] ⑥ In one implementation, the loss functions used by the generator and the discriminator in the iterative training are different; wherein the loss functions of the generator and the discriminator are respectively as follows:

[0070] ;

[0071] ;

[0072] In the formula, and Respectively represent the generator and discriminator in the The loss function value determined in the round of training; represents expected value calculation; Represents the divergence regularization parameter. In actual implementation, The value of is set to 0.03;

[0073] represents the L2 norm, indicating the Euclidean distance between true samples and false samples; represents the KL divergence, Indicated in The distribution of true samples in the first training sample set input to the generator in the round of training, Indicated in Distribution of fake samples generated by the generator in one round of training.

[0074] ⑦In one implementation, Used to indicate the Distribution of true samples in round training And the fake sample distribution The degree of difference between them ensures that the key features in the real data still play a dominant role in the generation process. The calculation formula is as follows:

[0075] ;

[0076] In the formula, and Represents the true sample and the false sample in Probability distribution of dimensions; Represents the feature dimension of fake samples.

[0077] In practical applications, if the number of samples included in the first training sample set is 800, 200 samples can be generated through the sample expansion model in the embodiment of the present application, that is, the first training sample set including 800 samples is expanded to the second training sample set including 1000 samples.

[0078] Step (3): Train the first extreme learning machine using the second training sample set to obtain a first analysis model.

[0079] Specifically, the first weight matrix of the hidden layer in the first extreme learning machine is determined according to the weighted sum of the first weight matrices of the hidden layers in multiple second extreme learning machines; the second extreme learning machine is trained according to a training sample set of soil indicator data on microplastic pollution.

[0080] Specifically, in the prior art, if an extreme learning machine is to be trained, it is first necessary to randomly initialize the model parameters of the extreme learning machine, that is, to randomly generate model parameters; the model parameter functions used by the hidden layer and the output layer in the extreme learning machine are respectively as follows:

[0081] ; ;

[0082] ;

[0083] ； ;

[0084] In the formula, Represents the weight matrix of the hidden layer in the extreme learning machine; Represents the weight matrix of the output layer in the extreme learning machine Represents the bias vector of the hidden layer in the extreme learning machine;

[0085] Represents the number of neurons included in the hidden layer of the extreme learning machine; Represents the dimension of the input features of the extreme learning machine; represents the standard normal distribution function, It means to randomly generate a Matrix of

[0086] represents the uniform distribution function, It means to randomly generate a A vector of represents the transposed matrix of the activation output of the hidden layer; A vector representing labels; Represents the regularization parameter to prevent overfitting; represents the identity matrix;

[0087] Represents the input features of the extreme learning machine;

[0088] Represents the input of the hidden layer after linear transformation;

[0089] is the transposed matrix of the activation output of the hidden layer;

[0090] Represents the ReLU activation function.

[0091] In the prior art, the weight matrix of the hidden layer in the extreme learning machine is randomly determined during the initialization process, but during the training process of the extreme learning machine, the weight matrix of the hidden layer will not be adjusted along with the training process, which results in that the extreme learning machine currently trained may not be the best model for currently performing the repair effect analysis function. In order to solve the limitations of the extreme learning machine in the classification field, the embodiment of the present application optimizes the training process of the extreme learning machine so that the weight matrix of the hidden layer can be updated with iterative training.

[0092] The optimization operation is specifically to determine the first weight matrix of the hidden layer in the first extreme learning machine according to the weighted sum of the first weight matrices of the hidden layers in multiple second extreme learning machines; the second extreme learning machine is trained according to the training sample set of soil indicator data on microplastic pollution; the practical significance of the above optimization operation is to migrate the weight matrix of the hidden layer in the second extreme learning with good performance to the first extreme learning machine currently being trained.

[0093] In actual operation, the second extreme learning machine can also be the first extreme learning machine, that is, the first weight matrix of the hidden layer in the first extreme learning machine trained in the embodiment of the present application can be used as the first weight matrix of the hidden layer in a second extreme learning machine, and then combined with the first weight matrices of the hidden layers in other second extreme learning machines to determine the weighted sum of multiple second extreme learning machines, and the weighted sum is determined as the first weight matrix of the hidden layer in the first extreme learning machine trained in the embodiment of the present application.

[0094] It should be emphasized that after the first weight matrix of the hidden layer in the second extreme learning machine is "migrated" to the first extreme learning machine in the embodiment of the present application, the weight "migration" process is completed. In the subsequent iterative training of the first extreme learning machine, what is continuously updated with the training rounds is the weight corresponding to each first weight matrix in the above-mentioned "weighted sum of multiple first weight matrices", rather than the first weight matrices of each second extreme learning machine in the first extreme learning machine; as the weights in the "weighted sum" are continuously updated, the first weight matrix of the hidden layer in the first extreme learning machine as a model parameter is also continuously updated accordingly.

[0095] In one implementation, step (3) includes: step (3.1), step (3.1) is a training process of a first analysis model, the details of which are as follows:

[0096] Step (3.1): Perform iterative training on the first extreme learning machine using the second training sample set until the number of iterative trainings meets a second threshold to obtain a first analysis model.

[0097] Specifically, in actual implementation, the value of the second threshold can be set to 1000, that is, when When , the training of the first extreme learning machine stops, and the trained first extreme learning machine is determined as the first analysis model; the value of the second threshold can also be determined according to actual needs, and this application does not specifically limit this. The training process of round training includes:

[0098] Step (3.1.1): According to The historical model parameters, historical loss function values, learning rate and parameter update formula of the first extreme learning machine in the round of training are determined to be used for the second extreme learning machine. The target model parameters of the first extreme learning machine in round 1 training;

[0099] The model parameters of the first extreme learning machine include: a first weight matrix of the hidden layer.

[0100] Step (3.1.2): According to the target model parameters of the first extreme learning machine, the first extreme learning machine is trained using the second training sample set.

[0101] Specifically, in the embodiments of the present application, in The target model parameters of the first extreme learning machine in the round of training are determined based on the historical model parameters, the historical loss function value, the learning rate and the parameter update formula; wherein the target model parameters indicate the model parameters used in the current round, and the historical model parameters are the model parameters used in the training rounds before the current round.

[0102] ⑧ In one implementation, the third parameter update formula of the first weight matrix is ​​as follows:

[0103] ;

[0104] ;

[0105] In the formula, Indicated in The target first weight matrix of the first extreme learning machine used in the round of training; Indicated in The target migration coefficient of the first weight matrix used in round training is used to adjust the control , The proportion in the first extreme learning machine, in actual implementation, can be The initial value of is set to 0.1; Indicated in The historical migration coefficient of the first weight matrix used in the round of training; The value of or the degree of dependence;

[0106] and Both represent the first weight matrix of the hidden layer in the second extreme learning machine;

[0107] represents the learning rate of the first extreme learning machine In actual implementation, the adjustment coefficient can be The value of is set to 0.05; Indicates that based on The historical loss function value of the first extreme learning machine determined by the training round.

[0108] ⑨ In one implementation, the model parameters of the first extreme learning machine also include: a second weight matrix of the output layer; the fourth parameter update formula of the second weight matrix is ​​as follows:

[0109] ;

[0110] ;

[0111] In the formula, Indicated in The target second weight matrix of the first extreme learning machine used in the round of training; represents the transposed matrix of the activation output of the hidden layer; A vector representing labels;

[0112] Indicated in The target regularization parameter of the second weight matrix used in the round of training. In actual implementation, The initial value of is set to 0.0001; Indicated in The history regularization parameter of the second weight matrix used in the round of training; Represents the identity matrix.

[0113] ⑩In one implementation, the loss function of the first extreme learning machine is as follows:

[0114] ;

[0115] In the formula, It means that the first extreme learning machine is trained on the second training sample set. The prediction results of samples; Indicates the first The labels corresponding to the samples;

[0116] represents the number of samples included in the second training sample set;

[0117] represents the sparse regularization term of the weight matrix of the output layer in the first extreme learning machine, norm, represents the first extreme learning machine in the output layer The weights of the neurons;

[0118] represents the number of neurons included in the output layer of the first extreme learning machine;

[0119] represents the sparse regularization term influence coefficient of the weight matrix in the output layer; it should be noted that, Indicated in The sparse regularization term used in the round training affects the coefficient, in order to be consistent with the above mentioned To distinguish, therefore Not specified ”.

[0120] ⑪In one implementation, is updated according to the third coefficient updating formula;

[0121] Among them, the third coefficient update formula is as follows:

[0122] ;

[0123] ;

[0124] In the formula, express The initial regularization parameter of The value of is set to 0.3; express The preset influence coefficient is used to adjust In actual implementation, the range of change can be The value of is set to 0.2;

[0125] Indicated in The sparsity measure of the hidden layer used in the round training is used to measure the sparsity of the hidden layer activation output, so that the first extreme learning machine can automatically adjust the sparsity in different training rounds to avoid over-regularization or under-regularization;

[0126] represents the target value of the sparsity measure; in actual implementation, The value of can be set to a value between 0 and 1 to guide the first extreme learning machine to achieve the desired sparsity level;

[0127] represents the number of neurons included in the hidden layer of the first extreme learning machine;

[0128] It means that when the hidden layer The activation value of a neuron Less than the sparsity threshold , the value is 1, otherwise it is 0. In actual execution, The value of is set to 0.5.

[0129] In one implementation, after S130, the method further includes: step (6) to step (7), the details of which are as follows:

[0130] Step (6): according to the second training sample set, determining the analysis result corresponding to each training sample in the second training sample set by using the first analysis model;

[0131] Among other things, the analysis results indicate the extent of remediation of microplastic pollution in the soil;

[0132] Step (7): Use the analysis results, related information, and analysis text corresponding to the analysis results as question-answering samples for The network is iteratively trained until the number of iterative trainings meets a third threshold to obtain a second analysis model;

[0133] Among them, the associated information at least includes: explanatory information of microplastic pollution and explanatory information of the analysis results; the second analysis model is a question-answer model; and the analysis text is a text obtained by expanding the analysis results based on the associated information.

[0134] Specifically, Figure 2 As shown, Figure 2 Provided for the embodiments of this application Schematic diagram of network structure. In the prior art, The network consists of an encoder and a decoder. The encoder maps the input data into a series of context-aware vector representations, and the decoder gradually decodes these vector representations into target features.

[0135] The encoder includes: The same encoding layer ( Figure 2 Only one encoding layer is shown in the figure), the encoding layer includes: a multi-head attention mechanism layer and a feedforward network layer. The encoder adds residual connections and normalization operations on the output side of the multi-head attention mechanism layer and the feedforward network layer in each encoding layer.

[0136] The multi-head attention mechanism layer is implemented by the multi-head self-attention mechanism. The multi-head self-attention mechanism is the core module of the encoder, which is responsible for learning the importance of each position from the data features of the input data and combining the information of all positions to obtain a global context representation, thereby improving The performance and generalization ability of the network; the specific steps are as follows:

[0137] 1) The implementation method of the multi-head self-attention mechanism is to pass the data features of the input data through three different linear transformation layers to generate "query ( )" vector, "key ( )" vector and "value ( )" vector; the linear transformation is achieved by multiplying the weight matrix; the generated query vector, key vector and value vector are divided into multiple smaller parts, namely the "head ( )”, which enables the model to process multiple different representation subspaces in parallel, thereby capturing the diversity of information in the input data;

[0138] 2) For each head, the dot product of the query vector and the key vector is calculated to obtain an attention score, which reflects the relative importance of different positions in the sequence. In order to make the value of the attention score stable, the score is divided by a scaling factor, which can be the square root of the key vector dimension.

[0139] 3) Adoption The function processes the attention scores to obtain the attention weights for each position, which is used to ensure that all weights add up to 1 so that they can be treated as probability distributions. The formula used by the function to process the attention score is as follows:

[0140] ;

[0141] In the formula, Indicates The attention weight of the head;

[0142] Indicates The query vector of the head; Indicates The key vector of the heads;

[0143] Indicates The transpose of the key vector of the head; The dimension representing the data characteristics of the input data;

[0144] represents the number of attention heads, For the A vector of values ​​for the heads.

[0145] 4) The obtained attention weights are weighted and summed to obtain the output vector;

[0146] 5) Concatenate the output vectors of all heads and pass the concatenated vectors through a linear transformation layer to generate the output of the multi-head self-attention layer.

[0147] The feedforward network layer includes two fully connected layers and an activation function, which is used to perform nonlinear transformation on the vector at each position. In the feedforward network layer, the input vector is linearly transformed by the first fully connected layer, then nonlinearly transformed by a ReLU activation function, and then linearly transformed by the second fully connected layer to improve The expressiveness and generalization ability of the model make The model can better handle complex input data.

[0148] The decoder includes: The same decoding layer ( Figure 2 Only one decoding layer is shown in the figure); Different from the encoding layer, each decoding layer has an additional masked multi-head attention mechanism layer, which is equivalent to the decoding layer including two multi-head attention mechanism layers and one feedforward network layer mechanism sublayer, in which the second multi-head attention mechanism sublayer is implemented by a multi-head cross attention mechanism.

[0149] The multi-head cross attention mechanism is a mechanism for information exchange between the encoder and the decoder. It takes the output data of the encoder and the output data of the masked multi-head attention layer in the decoder as input data, performs cross attention calculation on the input data based on the attention mechanism, and finally obtains a set of weighted decoder output data.

[0150] It should be noted that in In the encoder and decoder of the model, the output of each network layer is the result of residual connection and normalization operations. The functions used for residual connection and normalization operations are as follows:

[0151] ;

[0152] In the formula, represents the normalization function; Represents the output of the residual connection;

[0153] Represents input.

[0154] In actual implementation, the value of the third threshold can be set to 1000, that is, when When, for The training of the model is stopped and the trained The model is determined to be the second analysis model; the value of the third threshold can also be determined according to actual needs, and this application does not make any specific limitations on this.

[0155] Among them, for The network The training process of round training includes:

[0156] Step (7.1): According to In round training The historical model parameters, historical loss function values ​​and parameter update formulas of each network layer in the feedforward network layer in the network are determined to be used for the first The target model parameters of each feedforward network layer in the feedforward network layer of the round training;

[0157] Step (7.2): According to The target model parameters of each feedforward network layer in the network are Each feed-forward network layer in the network is trained;

[0158] Step (7.3): Use the gradient descent method to The remaining network layers in the network except the feed-forward network layer are trained.

[0159] Specifically, unlike the above-mentioned prior art, in order to avoid In order to avoid the gradient vanishing and gradient descent phenomena caused by the gradient descent method when the network is trained, and to avoid the phenomenon that the model parameters fall into the local optimal solution after the training is completed, the embodiment of the present application adopts the exploratory heterogeneous bee colony optimization algorithm to update the model parameters of the feedforward network layer during the training process. Different from the traditional learning method based on gradient descent, the exploratory heterogeneous bee colony optimization mechanism allows the feedforward network layer to automatically adjust the balance between exploration and utilization during the training process, thereby more effectively avoiding the local optimal solution and effectively improving the performance. The network's ability to extract complex features.

[0160] The exploratory heterogeneous swarm optimization mechanism is A bee colony is deployed for each feedforward network layer in the network. Each bee colony is responsible for exploring the optimal weight configuration of the feedforward network layer. The position of the bees represents the possible value of the weights, and the weights are updated by simulating the search behavior of the bees.

[0161] During the training process, the weights and biases of the feedforward network are initialized as follows:

[0162] ; ;

[0163] In the formula, Indicates The initial weights of the feed-forward network layers;

[0164] Indicates The initial bias of the feed-forward network layer; represents the number of nodes in the previous layer of the feedforward network layer, Represents a random number drawn from a standard normal distribution.

[0165] ⑫ Network The parameter update formula corresponding to the feedforward network layer is as follows: ;

[0166] ;

[0167] ;

[0168] ; ;

[0169] In the formula, Indicated in In the training round The target weight matrix used by the feed-forward network layer; Indicated in In the training round The history weight matrix used by the feed-forward network layer; Represents the learning rate of the preset feedforward network layer; in actual implementation, The value of is set to 0.01; Indicates that based on The first round of training determined The weight matrix increment of the feed-forward network layer; Represents the regularization parameter of the preset feedforward network layer. In actual implementation, The value of is set to 0.3; Indicates that based on The first round of training determined The historical gradients of the historical weight matrices of the feed-forward network layers;

[0170] Indicates that based on Determined by rounds of training The historical loss function value of the network; represents a preset first configuration parameter; represents a preset second configuration parameter;

[0171] Indicates the preset standard deviation; in actual implementation, and The values ​​of are all set to 0.5, The value of is set to 0.01; represents a random number drawn from a standard normal distribution; and Represents the parameters constructed based on the artificial bee colony algorithm, and each feedforward network layer corresponds to one bee colony; Indicated in The first round of training used Local optimal solution for each feed-forward network layer; Indicated in The first round of training used The global optimal solution for the feed-forward network layer;

[0172] Represents the set of weight matrices of all feed-forward network layers; represents the neighborhood set of all bee colonies; Representation weight The loss value of Indicates the currently updated weight parameter value.

[0173] ⑬ In one implementation, the formula used to draw random numbers from a standard normal distribution is as follows: ;

[0174] In the formula, The mean is 0 and the standard deviation is In actual implementation, The value of is set to 0.01.

[0175] ⑭In one implementation, The calculation formula of the network's loss function is as follows:

[0176] ;

[0177] In the formula, Indicated in In round training The historical loss function value of the network,

[0178] express The network output Output data of the class; express Network output The predicted probability of In the network The function is calculated, Represents the total number of categories.

[0179] In practical applications, when After the network is trained to obtain the second analysis model, the first analysis model and the second analysis model can be used to analyze the soil index data of the soil to be analyzed, such as Figure 3 As shown, Figure 3 A schematic diagram of the application process of the first analysis model and the second analysis model provided in the embodiment of the present application;

[0180] In actual applications, the current soil index data is input into the first analysis model, and the first analysis model is used to perform qualitative analysis on the soil index data to obtain one of the analysis results mentioned in the above example, namely, poor restoration effect, general restoration effect, good restoration effect, and excellent restoration effect; subsequently, the analysis results and related information are input into the second analysis model to obtain an analysis text corresponding to the analysis results.

[0181] In summary, the training method of the soil remediation effect analysis model provided in the embodiment of the present application has the following advantages over the prior art:

[0182] A. The first analysis model uses the first extreme learning machine based on the “migration” of the first weight matrix to enhance the model’s generalization ability for microplastic-contaminated soil data. In view of the complexity and variability of microplastic-contaminated soil data, it solves the problem of insufficient performance of traditional models on such data.

[0183] B. The second analysis model adopts The structure of the network is optimized by training methods for multiple feedback network layers to improve The network's ability to understand the context of input data enables the second analysis model to combine the output of the first analysis model with relevant domain knowledge for in-depth analysis;

[0184] C. The first analysis model works in collaboration with the second analysis model. The first analysis model is responsible for the preliminary data processing and classification of microplastic contaminated soil, while the second analysis model conducts further in-depth analysis based on the output of the first analysis model. This combination provides an analysis process from micro to macro, which enhances the accuracy and depth of the restoration effect assessment;

[0185] D. Generative adversarial networks are used for data expansion. To address the problem of insufficient training data, a generative adversarial network derived from microplastic pollution indicators is used to improve the accuracy and relevance of generated data through an adaptive error compensation mechanism.

[0186] Second, the present application embodiment provides a device for analyzing the effect of microplastic contaminated soil remediation, such as Figure 4 As shown, Figure 4This is a schematic diagram of the structure of a microplastic contaminated soil remediation effect analysis device provided in an embodiment of the present application, the device comprising: a data acquisition module 310 and an analysis module 320;

[0187] A data acquisition module 310 is used to obtain current soil indicator data of microplastic pollution;

[0188] An analysis module 320, configured to determine an analysis result corresponding to the current soil index data through a first analysis model according to the current soil index data;

[0189] Among them, the analysis results indicate the degree of restoration of microplastic contaminated soil; the network structure of the first analysis model is a first extreme learning machine; the first weight matrix of the hidden layer in the first extreme learning machine is determined according to the weighted sum of the first weight matrices of the hidden layers in multiple second extreme learning machines; the second extreme learning machine is trained according to a training sample set including historical soil indicator data.

[0190] In one implementation, the analysis module 320 is further used to obtain a first training sample set including historical soil indicator data of microplastic soil pollution;

[0191] The analysis module 320 is further used to perform sample expansion on the first training sample set by using a sample expansion model to obtain a second training sample set; wherein the model parameters of the sample expansion model are updated according to the gradient, the loss function value and the learning rate during the training process;

[0192] The analysis module 320 is further configured to train the first extreme learning machine using the second training sample set to obtain a first analysis model.

[0193] In one implementation, the network structure of the sample expansion model is a generative adversarial network, which includes: a generator and a discriminator; an analysis module 320, further used to obtain a third training sample set including historical soil indicator data of microplastic pollution;

[0194] The analysis module 320 is further configured to perform iterative training on the generative adversarial network using the third training sample set until the number of iterative trainings meets a first threshold to obtain a sample expansion model;

[0195] Among them, for the generation of adversarial networks The training process of round training includes:

[0196] The analysis module 320 is further configured to: The historical model parameters, historical gradients of historical model parameters, historical loss function values, learning rates and parameter update formulas corresponding to the generator and discriminator in the round of training are determined to be used for the first The target model parameters corresponding to the generator and discriminator of the round training respectively;

[0197] The analysis module 320 is further used to train the generative adversarial network through a third training sample set according to the target model parameters corresponding to the generator and the discriminator respectively.

[0198] In one implementation, the first parameter update formula of the generator is as follows:

[0199] ;

[0200] In the formula, Indicated in Generator function used in round training Target model parameters; Indicated in Generator function used in round training Historical model parameters; Represents the preset learning rate of the generator and discriminator; Indicates that based on The generator determined by the round of training The influence coefficient of the weight matrix of each network layer; Indicates that based on The historical gradients of the historical model parameters of the generator determined by the rounds of training; Indicated in The historical loss function value of the generator determined in the round of training; Indicates that based on The model parameter increment of the generator determined by the round of training;

[0201] And, the second parameter update formula of the discriminator is as follows:

[0202] ;

[0203] In the formula, Indicated in The discriminator function used in the round of training Target model parameters; Indicated in The discriminator function used in the round of training Historical model parameters; Indicates that based on The discriminator determined by the round of training The influence coefficient of the weight matrix of each network layer; Indicates that based on The historical gradients of the historical model parameters of the discriminator determined by the round of training; Indicates that based on The historical loss function value of the discriminator determined by the round of training; Indicates that based on The discriminator determined by the round of training The influence coefficient of the weight matrix of each network layer.

[0204] In one implementation, It is determined according to the first coefficient update formula, and the first coefficient update formula is as follows:

[0205] ; ;

[0206] In the formula, Indicates the preset adjustment coefficient; Indicates that based on The generator determined by the round of training The sensitivity coefficient of the weight matrix of each network layer; Indicates that based on The generator determined by the round of training The weight matrix of each network layer;

[0207] as well as, It is determined according to the parameter incremental update formula, which is as follows:

[0208] ;

[0209] ; ; ;

[0210] In the formula, represents new noise data; represents noisy data; Indicated in The generator function value determined in the training round; Indicates that based on The discriminator function value determined by the round of training The output probability of

[0211] Indicates the preset fusion coefficient; Represents the concatenation operation of vectors; represents the conditional vector;

[0212] as well as, is determined according to the second coefficient update formula, and the second coefficient update formula is as follows:

[0213] ; ;

[0214] In the formula, Indicates that based on The discriminator determined by the round of training The sensitivity coefficient of the weight matrix of each network layer; Indicates that based on The discriminator determined by the round of training The weight matrix of each network layer.

[0215] In one implementation, training the first extreme learning machine using the second training sample set to obtain the first analysis model includes:

[0216] Iteratively training the first extreme learning machine using the second training sample set until the number of iterative trainings meets a second threshold to obtain a first analysis model;

[0217] Among them, for the first extreme learning machine The training process of round training includes:

[0218] According to The historical model parameters, historical loss function values, learning rate and parameter update formula of the first extreme learning machine in the round of training are determined to be used for the second extreme learning machine. The target model parameters of the first extreme learning machine of the round of training; wherein the model parameters of the first extreme learning machine include: a first weight matrix of the hidden layer;

[0219] The first extreme learning machine is trained using the second training sample set according to the target model parameters of the first extreme learning machine.

[0220] In one implementation, the third parameter update formula of the first weight matrix is ​​as follows:

[0221] ;

[0222] ;

[0223] In the formula, Indicated in The target first weight matrix of the first extreme learning machine used in the round of training; Indicated in The target transfer coefficient of the first weight matrix used in the round of training; Indicated in The historical migration coefficient of the first weight matrix used in the round of training; and Both represent the first weight matrix of the hidden layer in the second extreme learning machine; Represents the preset learning rate of the first extreme learning machine The adjustment coefficient of Indicates that based on The historical loss function value of the first extreme learning machine determined by the round of training;

[0224] And, the model parameters of the first extreme learning machine also include: a second weight matrix of the output layer; the fourth parameter update formula of the second weight matrix is ​​as follows:

[0225] ; ;

[0226] In the formula, Indicated in The target second weight matrix of the first extreme learning machine used in the round of training; represents the transposed matrix of the activation output of the hidden layer; A vector representing labels; Indicated in The target regularization parameter of the second weight matrix used in the round of training; Indicated in The history regularization parameter of the second weight matrix used in the round of training; Represents the identity matrix.

[0227] In one implementation, the analysis module 320 is further configured to determine, based on the second training sample set, an analysis result corresponding to each training sample in the second training sample set by using the first analysis model;

[0228] Among other things, the analysis results indicate the extent of remediation of microplastic pollution in the soil;

[0229] The analysis module 320 is further configured to use the analysis results, the associated information, and the analysis text corresponding to the analysis results as question and answer samples for The network is iteratively trained until the number of iterative trainings meets a third threshold to obtain a second analysis model;

[0230] The associated information includes: information on microplastic pollution and information on the analysis results; the second analysis model is a question-answer model; the analysis text is a text expanded based on the associated information for the analysis results; The network The training process of round training includes:

[0231] The analysis module 320 is further configured to: In round training The historical model parameters, historical loss function values, and parameter update formulas of each feedforward network layer in the network are determined to be used for the first Wheel training The target model parameters of each feed-forward network layer in the network;

[0232] The analysis module 320 is also used to The target model parameters of each feedforward network layer in the network are Each feed-forward network layer in the network is trained;

[0233] The analysis module 320 is also used to use the gradient descent method to The remaining network layers in the network except the feed-forward network layer are trained.

[0234] In one implementation, Network The fifth parameter update formula of the feedforward network layer is as follows:

[0235] ;

[0236] ;

[0237] ;

[0238] ; ;

[0239] In the formula, Indicated in In the training round The target weight matrix used by the feed-forward network layer; Indicated in In the training round The history weight matrix used by the feed-forward network layer; Represents the learning rate of the preset feedforward network layer;

[0240] Indicates that based on The first round of training determined The weight matrix increment of the feed-forward network layer; Represents the regularization parameter of the preset feedforward network layer; Indicates that based on The first round of training determined The historical gradients of the historical weight matrices of the feed-forward network layers; Indicates that based on Determined by rounds of training The historical loss function value of the network; represents a preset first configuration parameter; represents a preset second configuration parameter; Indicates the preset standard deviation; represents a random number drawn from a standard normal distribution;

[0241] and Represents the parameters constructed based on the artificial bee colony algorithm, and each feedforward network layer corresponds to one bee colony; Indicates that based on The first round of training determined Local optimal solution for each feed-forward network layer; Indicates that based on The first round of training determined The global optimal solution for the feed-forward network layer; Represents the set of weight matrices of all feed-forward network layers;

[0242] represents the neighborhood set of all bee colonies; Represents the currently updated weight parameter value, indicating the weight parameter of the feedforward network layer corresponding to the most promising swarm position in the current iteration.

[0243] Third, the present application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, and when the processor executes the computer program, steps S110 to S120 provided in the above embodiment are implemented.

[0244] Fourth, the present application also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, steps S110 to S120 of the above embodiment are executed.

[0245] Fifth, the computer program product provided in the present application includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the method in the previous method embodiment. The specific implementation can refer to steps S110~S120 of the method embodiment, which will not be repeated here.

[0246] In the embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some communication interfaces, and the indirect coupling or communication connection of the devices or units can be electrical, mechanical or other forms.

[0247] In addition, the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0248] Furthermore, the functional modules in the various embodiments of the present application may be integrated together to form an independent part, or each module may exist separately, or two or more modules may be integrated to form an independent part.

[0249] It should be noted that if the function is implemented in the form of a software function module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, 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, including a number of instructions to enable 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 application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, and other media that can store program codes.

[0250] In this document, relational terms such as first and second, etc. are used merely to distinguish one entity or operation from another entity or operation, but do not necessarily require or imply any such actual relationship or order between these entities or operations.

[0251] The above description is only an embodiment of the present application and is not intended to limit the protection scope of the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for analyzing the effect of remediation of microplastic-contaminated soil, characterized in that: include: Obtain current soil indicator data for microplastic pollution; According to the current soil index data, determining an analysis result corresponding to the current soil index data through a first analysis model; Wherein, the analysis result indicates the restoration degree of the microplastic contaminated soil; the network structure of the first analysis model is a first extreme learning machine; the first weight matrix of the hidden layer in the first extreme learning machine is determined according to the weighted sum of the first weight matrices of the hidden layers in multiple second extreme learning machines; the second extreme learning machine is obtained by training according to a training sample set including historical soil indicator data; The method further comprises: Obtaining a first training sample set including historical soil indicator data of microplastic pollution; Performing sample expansion on the first training sample set by using a sample expansion model to obtain a second training sample set; Wherein, the model parameters of the sample expansion model are updated according to the gradient, the loss function value and the learning rate during the training process; Training the first extreme learning machine through the second training sample set to obtain the first analysis model; The method further comprises: According to the second training sample set, determining an analysis result corresponding to each training sample in the second training sample set by using the first analysis model; Wherein, the analysis result indicates the degree of restoration of microplastic pollution in the soil; The analysis results, associated information and analysis text corresponding to the analysis results are used as question and answer samples for The network is iteratively trained until the number of iterative trainings meets a third threshold to obtain a second analysis model; The associated information includes: information on microplastic pollution and information on analysis results; the second analysis model is a question-answer model; the analysis text is a text expanded based on the associated information for the analysis results; The network The training process of round training includes: According to As described in the round training The historical model parameters, historical loss function values, and parameter update formulas of each feedforward network layer in the network are determined to be used for the first Round training target model parameters of each of the feedforward network layers in the network; According to the The target model parameters of each feedforward network layer in the network, for the Each of the feedforward network layers in the network is trained; By using the gradient descent method for the The remaining network layers in the network except the feed-forward network layer are trained.

2. The method according to claim 1, characterized in that: The network structure of the sample expansion model is a generative adversarial network, which includes: a generator and a discriminator; the method also includes: obtaining a third training sample set including historical soil indicator data of microplastic pollution; Iteratively training the generative adversarial network using the third training sample set until the number of iterative trainings meets a first threshold to obtain the sample expansion model; Among them, for the generation of the adversarial network The training process of round training includes: According to The historical model parameters, historical gradients, historical loss function values, learning rates and parameter update formulas corresponding to the generator and the discriminator in the round of training are determined to be used for the first The target model parameters corresponding to the generator and the discriminator respectively in the round of training; The generative adversarial network is trained using the third training sample set according to the target model parameters corresponding to the generator and the discriminator respectively.

3. The method according to claim 2, characterized in that The first parameter update formula of the generator is as follows: ； In the formula, Indicated in Generator function used in round training The target model parameters of Indicated in Generator function used in round training The historical model parameters of Representing the preset learning rates of the generator and the discriminator; Indicates that based on The generator determined by the round of training The influence coefficient of the weight matrix of each network layer; Indicates that based on The historical gradients of the historical model parameters of the generator determined by rounds of training; Indicated in The historical loss function value of the generator determined in a round of training; Indicates that based on The model parameter increment of the generator determined by the round of training; And, the second parameter update formula of the discriminator is as follows: ； In the formula, Indicated in The discriminator function used in the round of training The target model parameters of Indicated in The discriminator function used in the round of training The historical model parameters of Indicates that based on The discriminator determined by the round of training The influence coefficient of the weight matrix of each network layer; Indicates that based on The historical gradient of the historical model parameters of the discriminator determined by rounds of training; Indicates that based on The historical loss function value of the discriminator determined by a round of training.

4. The method according to claim 3, characterized in that Said is determined according to a first coefficient updating formula, wherein the first coefficient updating formula is as follows: ； ； In the formula, Indicates the preset adjustment coefficient; Indicates that based on The generator determined by the round of training The sensitivity coefficient of the weight matrix of each network layer; Indicates that based on The generator determined by the round of training The weight matrix of the network layer; And, the It is determined according to the parameter incremental update formula, and the parameter incremental update formula is as follows: ； ； ； ； In the formula, represents new noise data; represents noisy data; Indicated in The generator function value determined in the training round; Indicates that based on The discriminator function value determined by the round of training The output probability of Indicates the preset fusion coefficient; Represents the concatenation operation of vectors; represents the conditional vector; And, the is determined according to a second coefficient updating formula, wherein the second coefficient updating formula is as follows: ； ； In the formula, Indicated in The discriminator determined in the round of training The sensitivity coefficient of the weight matrix of each network layer; Indicates that based on The discriminator determined by the round of training The weight matrix of the network layer.

5. The method according to claim 1, characterized in that: The step of training the first extreme learning machine by using the second training sample set to obtain the first analysis model includes: Performing iterative training on the first extreme learning machine using the second training sample set until the number of iterative trainings meets a second threshold to obtain the first analysis model; Among them, for the first extreme learning machine The training process of round training includes: According to The historical model parameters, historical loss function values, learning rate and parameter update formula of the first extreme learning machine in the first round of training are determined to be used for the second The target model parameters of the first extreme learning machine after round training; wherein the model parameters of the first extreme learning machine include: the first weight matrix of the hidden layer; The first extreme learning machine is trained using the second training sample set according to the target model parameters of the first extreme learning machine.

6. The method according to claim 4, characterized in that The third parameter update formula of the first weight matrix is ​​as follows: ； ； In the formula, Indicated in a target first weight matrix of the first extreme learning machine used in a round of training; Indicated in The target migration coefficient of the first weight matrix used in the round training; Indicated in The historical migration coefficient of the first weight matrix used in the round training; and Both represent the first weight matrix of the hidden layer in the second extreme learning machine; represents the preset learning rate of the first extreme learning machine The adjustment coefficient of Indicates that based on The historical loss function value of the first extreme learning machine determined by a round of training; And, the model parameters of the first extreme learning machine also include: a second weight matrix of the output layer; a fourth parameter update formula of the second weight matrix is ​​as follows: ； ； In the formula, Indicated in a target second weight matrix of the first extreme learning machine used in a round of training; a transposed matrix representing the activation output of the hidden layer; A vector representing labels; Indicated in A target regularization parameter of the second weight matrix used in the round of training; Indicated in A history regularization parameter of the second weight matrix used in the round of training; Represents the identity matrix.

7. The method according to claim 1, characterized in that Said Network The fifth parameter update formula of the feedforward network layer is as follows: ； ； ； ； ； In the formula, Indicated in In the training round The target weight matrix used by the feed-forward network layer; Indicated in In the training round a history weight matrix used by the feed-forward network layer; Represents a preset learning rate of the feedforward network layer; Indicates that based on The first round of training determined An increment of the weight matrix of the feed-forward network layer; Represents a preset regularization parameter of the feedforward network layer; Indicates that based on The first round of training determined The historical gradient of the historical weight matrix of the feed-forward network layer; Indicates that based on The The historical loss function value of the network; represents a preset first configuration parameter; represents a preset second configuration parameter; Indicates the preset standard deviation; represents a random number drawn from a standard normal distribution; and represents parameters constructed based on the artificial bee colony algorithm, and each of the feedforward network layers corresponds to one bee colony; Indicates that based on The first round of training determined A local optimal solution of the feed-forward network layer; Indicates that based on The first round of training determined The global optimal solution of the feed-forward network layer; represents a set of weight matrices of all said feed-forward network layers; represents the neighborhood set of all bee colonies; Represents the currently updated weight parameter value, indicating the weight parameter of the feedforward network layer corresponding to the most promising swarm position in the current iteration.

8. A device for analyzing the effect of soil remediation contaminated by microplastics, characterized in that: For implementing the method for analyzing the effect of microplastic-contaminated soil remediation as claimed in claim 1, the device comprises: a data acquisition module and an analysis module; The data acquisition module is used to obtain current soil indicator data of microplastic pollution; The analysis module is used to determine the analysis result corresponding to the current soil index data through a first analysis model according to the current soil index data; Among them, the analysis results indicate the degree of restoration of the microplastic contaminated soil; the network structure of the first analysis model is a first extreme learning machine; the first weight matrix of the hidden layer in the first extreme learning machine is determined according to the weighted sum of the first weight matrices of the hidden layers in multiple second extreme learning machines; the second extreme learning machine is trained according to a training sample set including historical soil indicator data.

Images