Model training method and apparatus, construction safety evaluation method and apparatus, and device

Abstract

The present application discloses a model training method and apparatus, a construction safety evaluation method and apparatus, and a device. The model training method comprises: acquiring sample data; on the basis of association information between training samples and true value labels, determining label missing ratios of the training samples in different dimensions and first weight values of the training samples in different dimensions; determining second weight values of the training samples on the basis of the label missing ratios; and inputting the sample data, the first weight values, and the second weight values into a preset neural network model for training until a loss value of a target loss function of the preset neural network model meets a model convergence condition, so as to obtain a target prediction model. In this way, by improving a weighted loss function, when sample data having a partially missing label is kept, a loss value of a missing label of a sample is calculated with reference to a weight, thereby solving the problem of reduced prediction accuracy caused by missing samples, improving the accuracy of a model, and simultaneously and accurately predicting prediction values of multiple dimensions.

Description

Model training method, construction safety evaluation method, device and equipment TECHNICAL FIELD

[0001] The present application relates to the technical field of computers, and in particular to a model training method, a construction safety evaluation method, a device and equipment. BACKGROUND

[0002] In actual applications, due to equipment failure, data acquisition failure and other reasons, the data set often has different degrees of missing. Usually, by deleting samples containing missing values or using mean, median and other methods for interpolation set training for deep neural network model training, but the complexity and diversity of data missing are not fully considered, which introduces errors due to the loss of authenticity and part of the information of the data set, thereby affecting the prediction accuracy of the model and causing the problem of unsatisfactory prediction effect. SUMMARY

[0003] Therefore, the present application provides a model training method, a construction safety evaluation method, a device and equipment, which retain sample data with partial missing labels by improving the weighted loss function, thereby solving the problem of affecting the prediction accuracy of the model due to the existence of partial missing data in the data set.

[0004] According to a first aspect of the present application, a model training method is provided, comprising:

[0005] Obtaining sample data, wherein the sample data includes training samples and true value labels of the training samples in different dimensions;

[0006] According to the association information of the training samples and the true value labels, determining a label missing ratio of the training samples in different dimensions and a first weight value of the training samples in different dimensions;

[0007] According to the label missing ratio, determining a second weight value of the training samples;

[0008] Inputting the sample data, the first weight value and the second weight value into a preset neural network model for training until the loss value of the target loss function of the preset neural network model meets the model convergence condition, and obtaining a target prediction model.

[0009] Optionally, according to the association information of the training samples and the true value labels, determining the first weight value of the training samples in different dimensions comprises:

[0010] If the association information is that the training samples are associated with the true value labels in any dimension, the first weight value in the any dimension is assigned as 1;

[0011] If the association information indicates that the training sample is not associated with the true value label in any dimension, the first weight value of the training sample in the any dimension is set to 0.

[0012] Optionally, the determining of the label missing ratio of the training sample in different dimensions according to the association information of the training sample and the true value label comprises:

[0013] counting a label missing number of the training sample whose association information indicates that the training sample is not associated with the true value label in any dimension;

[0014] calculating a quotient of the dimension label missing number and a total number of training samples in the sample data as the label missing ratio of the any dimension.

[0015] Optionally, the determining of the second weight value of the training sample according to the label missing ratio comprises:

[0016] calculating an information entropy and a mutual information amount of the true value label;

[0017] determining the second weight value according to the information entropy, the mutual information amount and the label missing ratio.

[0018] Optionally, a calculation formula of the information entropy of the true value label is represented as:

[0019]

[0020] wherein, E j represents the information entropy of the jth true value label, p ij represents a probability of the value of the ith training sample on the jth true value label, and n represents a total number of training samples.

[0021] a calculation formula of the mutual information amount of the true value label is represented as:

[0022]

[0023] wherein, I(Y j ;Y k ) represents the mutual information amount between labels Y j and Y k , p(y j ,y k ) represents a joint probability mass function of labels Y j and Y k , and p(y j ) and p(y k ) respectively represent marginal probability mass functions of labels Y j and Y k .

[0024] The second weight value is determined according to the information entropy, the mutual information and the label missing ratio, and the following formula is used:

[0025]

[0026] In the formula, w j represents the second weight value of the jth real value label, m represents the total number of real value labels in any dimension, E j represents the information entropy of the jth real value label, I(Y j ; Y k ) represents the mutual information between labels Y j and Y k , I(Y k ; Y l ) represents the mutual information between labels Y k and Y l , and a represents the label missing ratio in any dimension.

[0027] Optionally, the target loss function is represented as:

[0028]

[0029] In the formula, loss represents the loss function, y represents the real value label, represents the predicted value output by the model, n represents the total number of training samples, l represents the number of predicted labels, i represents the number of training samples, a ij represents the first weight value of the jth real value label of the ith training sample, w j represents the second weight value of the jth real value label.

[0030] Optionally, the model training method further comprises:

[0031] normalizing the second weight value.

[0032] According to a second aspect of the present application, a construction safety evaluation method for a wind power project is provided, comprising:

[0033] obtaining a target prediction model and safety factor information of a target wind power project, wherein the target prediction model uses safety factor information and safety scores of historical wind power projects as sample data, and is trained by using the model training method provided in the first aspect;

[0034] dimensionless processing the safety factor information according to a preset conversion rule to generate a safety index value;

[0035] inputting the safety index value into the target prediction model to obtain a safety score of the target wind power project;

[0036] if the safety score is less than or equal to a score threshold, displaying safety warning information.

[0037] According to a third aspect of the present application, a model training apparatus is provided, comprising:

[0038] an acquisition module configured to acquire sample data, wherein the sample data comprises training samples and true value labels of the training samples in different dimensions;

[0039] a determination module configured to determine, according to association information of the training samples and the true value labels, a label missing rate of the training samples in different dimensions and a first weight value of the training samples in different dimensions, and determine a second weight value of the training samples according to the label missing rate;

[0040] a training module configured to input the sample data, the first weight value and the second weight value into a preset neural network model for training until a loss value of a target loss function of the preset neural network model meets a model convergence condition, to obtain a target prediction model.

[0041] Optionally, the determination module is specifically configured to, if the association information is that the training sample is associated with the true value label in any dimension, assign the first weight value in the any dimension as 1; and if the association information is that the training sample is not associated with the true value label in any dimension, assign the first weight value in the any dimension as 0.

[0042] Optionally, the determination module is specifically configured to count a label missing number of the training sample that is not associated with the true value label in any dimension according to the association information, and calculate a quotient of the dimension label missing number and a total number of training samples in the sample data as the label missing rate of the any dimension.

[0043] Optionally, the determination module is specifically configured to calculate an information entropy and a mutual information amount of the true value label, and determine the second weight value according to the information entropy, the mutual information amount and the label missing rate.

[0044] Optionally, the model training apparatus further comprises:

[0045] a preprocessing module configured to perform normalization processing on the second weight value.

[0046] According to a fourth aspect of the present application, a construction safety evaluation apparatus of a wind power project is provided, comprising:

[0047] obtain a target prediction model and safety factor information of a target wind power project, wherein the target prediction model takes safety factor information and a safety score of a historical wind power project as sample data, and is trained by using the model training method provided in the first aspect;

[0048] The data processing module is configured to perform non-dimensional processing on the safety factor information according to a preset conversion rule to generate a safety index value.

[0049] The evaluation module is configured to input the safety index value into the target prediction model to obtain a safety score of the target wind power project.

[0050] The early warning module is configured to display safety early warning information if the safety score is less than or equal to a score threshold.

[0051] According to a fifth aspect of the present application, a readable storage medium is provided, which stores a program or instructions, and the program or instructions are executed by a processor to implement the steps of the model training method and the construction safety evaluation method of the wind power project.

[0052] According to a sixth aspect of the present application, a computer device is provided, which includes a storage medium, a processor, and a computer program stored in the storage medium and executable on the processor, and the processor implements the steps of the model training method and the construction safety evaluation method of the wind power project when executing the program.

[0053] By the above technical solution, the weight of the training sample in different dimensions is calculated based on the association information of whether the training sample is associated with a true value label, and the importance of the sample data with a missing label part is represented by the weight. In this way, when calculating the model loss value, the loss value of the missing label of the sample with a missing label part can be calculated by referring to the weight while retaining the sample with a missing label part. Thus, the problem of prediction accuracy reduction caused by deleting the sample with a missing part in the data set is effectively solved, the accuracy of the model is improved, and the prediction values of multiple dimensions can be accurately predicted at the same time.

[0054] The above description is only a summary of the technical solutions of the present application, in order to more clearly understand the technical means of the present application, the specific embodiments of the present application can be implemented according to the content of the description, and in order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the following specific embodiments of the present application are described. BRIEF DESCRIPTION OF DRAWINGS

[0055] The drawings described herein are used to provide further understanding of the present application, and form a part of the present application. The schematic embodiments of the present application and their descriptions are used to explain the present application, and do not constitute an improper limitation on the present application. In the drawings:

[0056] FIG. 1 shows a flowchart of a model training method according to an embodiment of the present application;

[0057] FIG. 2 shows a flowchart of a construction safety assessment method for a wind power project according to an embodiment of the present application;

[0058] FIG. 3 shows a structural block diagram of a model training device according to an embodiment of the present application;

[0059] FIG. 4 shows a structural block diagram of a construction safety assessment device for a wind power project according to an embodiment of the present application. DETAILED DESCRIPTION

[0060] Hereinafter, the present application will be described in detail with reference to the accompanying drawings and embodiments. It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0061] The embodiments of the present application will be described in detail below, examples of which are shown in the accompanying drawings, in which the same or similar reference numerals represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by reference to the accompanying drawings are exemplary and are only used to explain the present application, and cannot be interpreted as limiting the present application.

[0062] Those skilled in the art can understand that, unless specifically stated, the singular forms "a", "an" and "the" used herein also include the plural forms. It should be further understood that the phrase "comprising" used in the specification of the present application means that the features, integers, steps, operations, elements and / or components exist, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof. It should be understood that when we say an element is "connected" or "joined" to another element, it can be directly connected or joined to the other element, or there can be intermediate elements. In addition, "connected" or "joined" used herein can include wireless connection or wireless connection. The phrase "and / or" used herein includes all or any single unit and all combinations of the associated listed items.

[0063] Now, exemplary embodiments according to the present application will be described in more detail with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in various different forms, and should not be interpreted as being limited to the embodiments set forth herein. It should be understood that these embodiments are provided so that the disclosure of the present application is complete and complete, and the ideas of these exemplary embodiments are sufficiently conveyed to those skilled in the art.

[0064] A model training method is provided in the present embodiment, as shown in FIG. 1, which comprises:

[0065] In step 110, sample data is obtained.

[0066] wherein the sample data comprises training samples and real value labels of the training samples in different dimensions, that is, the same training sample can be associated with different real value labels in different dimensions. Specifically, the sample data can be represented in matrix format, for example, the training samples real value labels Y = [y1, y2, … y n ] T ∈R n×l wherein xi is the i-th sample, d is the feature dimension, n is the total number of samples, and l is the number of labels to be predicted. i ∈R d wherein xi is the i-th sample, d is the feature dimension, n is the total number of samples, and l is the number of labels to be predicted.

[0067] Step 120, according to the association information of the training samples and the real value labels, determining the label missing rate of the training samples in different dimensions and the first weight value of the training samples in different dimensions.

[0068] In this embodiment, by the association information of whether the training samples are associated with the real value labels, the label missing rate representing the label missing situation of each training sample in each feature dimension is determined, and the first weight value representing whether there is a real value label missing for each training sample in a certain feature dimension. In order to accurately evaluate the reference importance of each training sample by the label missing rate and the first weight value, in order to perform weighted loss calculation, so that the trained target prediction model can meet the accuracy requirements of the actual scene more.

[0069] In a specific embodiment, according to the association information of the training samples and the real value labels, the first weight value of the training samples in different dimensions is determined in step 120, which specifically includes the following steps:

[0070] Step 121-1, if the association information is that the training sample is associated with the real value label in any dimension, the first weight value in any dimension is assigned as 1.

[0071] Step 121-2, if the association information is that the training sample is not associated with the real value label in any dimension, the first weight value in any dimension is assigned as 0.

[0072] In this embodiment, the first weight value is assigned as 1 or 0 according to whether the training sample is associated with the true value label. In order to better distinguish the training samples with the true value label and the training samples without the true value label. When the training sample has the true value label in a certain dimension, the first weight value in this dimension is assigned as 1, indicating that the training sample in this dimension plays a positive role in the training of the model. On the contrary, when the training sample does not have the true value label in a certain dimension, the first weight value in this dimension is assigned as 0, indicating that the training sample in this dimension does not play a role in the training of the model. Thus, when the model is trained, even if the training sample is missing the label in a certain dimension, the training sample will be retained in order to enable the model to learn the features of the training sample in other dimensions. Thus, the learning ability and generalization ability of the model to the data during training are improved, which helps to enhance the processing ability of the model to the data with label missing.

[0073] For example, if the element in the i-th row and the j-th column of the true value label matrix Y is missing, the first weight value a ij is recorded as 0, and vice versa.

[0074] In a specific embodiment, the label missing rate of the training sample in different dimensions is determined according to the association information of the training sample and the true value label in step 120, which specifically includes the following steps:

[0075] In step 122-1, the label missing number of the training sample which is not associated with the true value label in any dimension is counted according to the association information.

[0076] In step 122-2, the quotient of the dimension label missing number and the total number of training samples in the sample data is calculated as the label missing rate of any dimension.

[0077] In this embodiment, the quotient of the label missing number of each dimension and the total number of training samples in the sample data in this dimension can obtain the label missing rate of each dimension. The label missing rate reflects the degree of label missing in each dimension, so as to adjust the weight during subsequent processing and training of the neural network model, thereby improving the accuracy and generalization ability of the model.

[0078] For example, taking 5 training samples and 3 feature dimensions as an example, the sample data matrix is represented as:

[0079]

[0080] If only the true value label y 12 is empty, it means that the training sample x1 is missing the true value label in the second feature dimension, and the label missing rate of the second feature dimension is 1 / 5. If the true value label y 12 and y 32If the second feature dimension of the training samples x1 and x3 is missing the true value label, the label missing rate of the second feature dimension is 2 / 5. The greater the label missing rate, the more serious the label missing in the corresponding dimension.

[0081] In step 130, the second weight value of the training sample is determined according to the label missing rate.

[0082] In this embodiment, by dynamically adjusting the second weight value when training the neural network model according to the degree of label missing in each dimension, the processing capability of the model for data with label missing can be further enhanced, and the accuracy and robustness of the model can be improved.

[0083] In a specific embodiment, step 130, i.e., determining the second weight value of the training sample according to the label missing rate, specifically includes the following steps:

[0084] In step 131, the information entropy and mutual information of the true value label are calculated.

[0085] Wherein, the entropy is an index for representing the amount of information, and the higher the entropy, the greater the information content and the higher the uncertainty, and it is more difficult to predict. The mutual information is a statistical quantity for measuring the closeness of the relationship between two true value labels, and it shows the correlation between the true value labels. The higher the information entropy indicates the lower reliability of the label data, and the greater the mutual information indicates the higher correlation between the labels.

[0086] Specifically, the calculation formula of the information entropy of the true value label is:

[0087]

[0088] In the formula, E j represents the information entropy of the jth true value label, p ij represents the probability of the value of the ith training sample on the jth true value label, and n represents the total number of training samples.

[0089] The calculation formula of the mutual information of the true value label is:

[0090]

[0091] In the formula, I(Y j ;Y k ) represents the mutual information between the labels Y j and Y k , p(y j ,y k ) represents the joint probability mass function of the labels Y j and Y k , p(y j ) and p(y kY j and Y k , respectively, represent the edge probability mass functions of labels Y

[0092] In step 132, the second weight value is determined according to the information entropy, the mutual information and the label missing ratio.

[0093] Specifically, the second weight value is determined according to the information entropy, the mutual information and the label missing ratio, and the following formula is used:

[0094]

[0095] In the formula, w j represents the second weight value of the jth real value label, m represents the total number of real value labels in any dimension, E j represents the information entropy of the jth real value label, E k represents the information entropy of the kth real value label, I(Y j ; Y k ) represents the mutual information between labels Y j and Y k , I(Y k ; Y l ) represents the mutual information between labels Y k and Y l , and a represents the label missing ratio in any dimension.

[0096] In this embodiment, the information entropy and the mutual information of each real value label are calculated by the entropy weight method, the uncertainty of the label and the correlation between the labels are measured by calculating the information entropy and the mutual information of the real value label, and then the second weight value is obtained by comprehensively weighting different labels in combination with the label missing ratio. Thus, the weight distribution is based on the information entropy, and the mutual information and the label missing ratio are used to further adjust the weight, so that the label with higher correlation with other real value labels or lower missing ratio obtains higher weight. Further, by reasonably determining the second weight value, the model can learn the potential distribution of the missing data and the correlation between the data, reduce the risk of overfitting and the influence of label missing on the model performance, at the same time, the subjectivity dependence of the sample data is reduced, the weight distribution is more objective and reasonable, the balance of the importance of each label in different dimensions is ensured, the generalization ability of the model is further improved, the model can achieve good prediction effect on different data sets, and the performance and stability of the target prediction model are improved.

[0097] It is worth mentioning that the generative adversarial network can also be constructed by the real value label and its information entropy and mutual information, so that the generative adversarial network learns the potential distribution of the missing label. The reconstructed label conforming to the distribution rule of the real value label is generated by using the trained generative adversarial network, and the reconstructed label is filled into the missing label position of the corresponding training sample, so that the data characteristics are learned from the irregular discrete measurement data by the generative adversarial network in an unsupervised manner, the sample data required for training is more complete, and high data accuracy can be maintained under different missing rates.

[0098] Specifically, the generative adversarial network can be a CNN (Convolutional Neural Networks) or a ViT (Vision Transformer) or the like.

[0099] In an embodiment, after step 130, the model training method further includes: performing normalization processing on the second weight value.

[0100] In this embodiment, the second weight value is linearly transformed by the normalization processing to be mapped to the range of [0, 1], and the sum of the second weight values of all real value labels in the same dimension is 1. Thus, the influence of the correlation between different real value labels on the model training quality can be better balanced, and the model has better interpretability and stability in the training process.

[0101] In step 140, the sample data, the first weight value, and the second weight value are input into a preset neural network model for training until the loss value of the target loss function of the preset neural network model meets the model convergence condition, and a target prediction model is obtained.

[0102] Specifically, the preset neural network model can adopt a single neural network model, for example, a deep neural network (DNN) model, a convolutional network (CNN) model, a recurrent neural network (RNN) model, a residual network (ResNet) model, or a BERT (Bidirectional Encoder Representation from Transformers) model.

[0103] The model training method provided in the embodiments of the present application can calculate the weight of a training sample under different dimensions based on the association information of whether the training sample is associated with a true value label, and express the importance of sample data with a part of labels missing through the weight. In this way, when calculating the model loss value, the loss value of the missing label of a sample with a part of labels missing can be calculated by referring to the weight under the condition that the sample exists. Thus, the problem of prediction accuracy decline caused by deleting the sample with a part of labels missing in the data set is effectively solved, and the accuracy of the model is improved. In addition, since the sample data is divided according to the dimensions, after the target prediction model is trained by using the sample data, the target prediction model can accurately predict the prediction values of multiple dimensions at the same time, which helps to improve the prediction efficiency.

[0104] In an actual application scenario, the target loss function is represented as:

[0105]

[0106] In the formula, loss represents the loss function, y represents the true value label, represents the prediction value output by the model, n represents the total number of training samples, l represents the number of predicted labels, i represents the number of training samples, j represents the number of true value labels, a ij represents the first weight value of the jth true value label of the ith training sample, w j represents the second weight value of the jth true value label.

[0107] By introducing the first weight value of whether the label is missing and the second weight value of the degree of relevance into the loss function, when calculating the loss, the loss of the missing label of the sample is not calculated, and the loss is weighted according to the relevance, so that the model pays more attention to the training sample with missing labels and the training sample with higher label relevance in the training process. Even if there is a part of labels missing, the model can also understand and adapt to the data characteristics and distribution under different dimensions, so that the training effect of the target prediction model is similar to that of the model without missing sample, the generalization ability of the target language model is ensured, and the accuracy and reliability are improved.

[0108] For example, the sample data in the data set is analyzed to calculate the label missing ratio. Then, the sample data is divided into a training set and a test set according to a preset ratio. The data of the test set is imported into a deep neural network model of a custom model loss function for training. Finally, the mean square error is used as the error value between the prediction output and the expected output of the loss function, and the model accuracy is evaluated, adjusted and optimized on the validation set. The parameters of the deep neural network model and the weight of the custom loss function are adjusted and optimized until the loss function reaches the minimum.

[0109] In the embodiment, a construction safety evaluation method for a wind power project is provided, as shown in FIG. 2. The method comprises:

[0110] Step 210, obtaining sample data.

[0111] The sample data includes safety factor information and safety scores of historical wind power projects. The safety factor information includes index values corresponding to different safety factors. The safety factors can be reasonably set according to the actual wind power project engineering requirements. For example, the safety factors can include wind turbine power, operating time, remaining effective period of electrical equipment such as power transmission lines, transformers, and switch devices, maintenance frequency, environmental parameters, fire prevention facility deployment, and the like. The embodiments of the present application will not be listed one by one.

[0112] Step 220, determining the label missing ratio of the training sample under different dimensions and the first weight value of the training sample under different dimensions according to the association information of the training sample and the true value label.

[0113] In a specific embodiment, in step 220, the first weight value of the training sample under different dimensions is determined according to the association information of the training sample and the true value label, which specifically includes the following steps: if the association information is that the training sample is associated with the true value label under any dimension, the first weight value under any dimension is assigned as 1; if the association information is that the training sample is not associated with the true value label under any dimension, the first weight value under any dimension is assigned as 0.

[0114] In this embodiment, the first weight value is assigned as 1 or 0 according to whether the training sample is associated with the true value label. In order to better distinguish the training samples with and without the true value label. When the training sample has the true value label under a certain dimension, the first weight value under the dimension is assigned as 1, indicating that the training sample under the dimension plays a positive role in the training of the model. On the contrary, when the training sample does not have the true value label under a certain dimension, the first weight value under the dimension is assigned as 0, indicating that the training sample under the dimension does not play a role in the training of the model. Thus, the learning ability and generalization ability of the model during training are improved, which helps to enhance the processing ability of the model for data with label missing.

[0115] In a specific embodiment, in step 220, the label missing ratio of the training sample under different dimensions is determined according to the association information of the training sample and the true value label, which specifically includes the following steps: counting the number of label missing of the training sample whose association information is that the training sample is not associated with the true value label under any dimension; calculating the quotient of the dimension label missing number and the total number of training samples in the sample data as the label missing ratio of any dimension.

[0116] In this embodiment, the label missing ratio of each dimension can be obtained by calculating the quotient of the number of label missing in each dimension and the total number of training samples in the sample data in this dimension. The label missing ratio reflects the degree of label missing in each dimension, so as to improve the accuracy and generalization ability of the model in subsequent processing and weight adjustment of the neural network model.

[0117] In step 230, the second weight value of the training sample is determined according to the label missing ratio.

[0118] In a specific embodiment, step 230, i.e., determining the second weight value of the training sample according to the label missing ratio, specifically includes the following steps: calculating the information entropy and mutual information of the true value label; and determining the second weight value according to the information entropy, mutual information and label missing ratio.

[0119] Specifically, the calculation formula of the information entropy of the true value label is:

[0120]

[0121] In the formula, E j represents the information entropy of the jth true value label, p ij represents the probability of the value of the ith training sample on the jth true value label, and n represents the total number of training samples.

[0122] The calculation formula of the mutual information of the true value label is:

[0123]

[0124] In the formula, I(Y j ;Y k ) represents the mutual information between the labels Y j and Y k , p(y j ,y k ) represents the joint probability mass function of the labels Y j and Y k , and p(y j ) and p(y k ) represent the marginal probability mass functions of the labels Y j and Y k , respectively.

[0125] According to the information entropy, mutual information and label missing ratio, the second weight value is determined by using the following formula:

[0126]

[0127] In the formula, w jthe second weight value of the jth real value label, m represents the total number of real value labels in any dimension, E j the information entropy of the jth real value label, E k the information entropy of the kth real value label, I(Y j ; Y k ) represents the mutual information between labels Y j and Y k , I(Y k ; Y l ) represents the mutual information between labels Y k and Y l , and a represents the label missing rate in any dimension.

[0128] In this embodiment, the information entropy and mutual information of each real value label are calculated by the entropy weight method, the uncertainty of the label and the correlation between the labels are measured by calculating the information entropy and mutual information of the real value label, and then the different labels are comprehensively weighted by combining the label missing rate to obtain the second weight value. Thus, the weight distribution is based on the information entropy, and the mutual information and the label missing rate are used to further adjust the weight, so that the label with higher correlation or lower missing rate with other real value labels obtains higher weight. Further, by reasonably determining the second weight value, the model can learn the potential distribution of the missing data and the correlation between the data, reduce the risk of overfitting and the influence of label missing on the model performance, at the same time, reduce the subjective dependence of the sample data, so that the weight distribution is more objective and reasonable, and the balance of the importance of each label in different dimensions is ensured. Further improve the generalization ability of the model, so that the model can achieve good prediction effect on different data sets, improve the performance and stability of the target prediction model.

[0129] In step 240, the sample data, the first weight value and the second weight value are input into a preset neural network model for training until the loss value of the target loss function of the preset neural network model meets the model convergence condition, and a target prediction model is obtained.

[0130] wherein the target loss function is represented as:

[0131]

[0132] In the formula, loss represents the loss function, y represents the real value label, represents the predicted value output by the model, n represents the total number of training samples, l represents the number of predicted labels, i represents the number of training samples, j represents the number of real value labels, a ij represents the first weight value of the jth real value label of the ith training sample, w j represents the second weight value of the jth real value label.

[0133] Step 250, obtaining safety factor information of the target wind power project.

[0134] Step 260, dimensionless processing the safety factor information according to a preset conversion rule to generate a safety index value.

[0135] In this embodiment, the complex safety factor information can be converted into a unified and concise index, facilitating the model to compare and analyze, and improving the data processing efficiency.

[0136] Step 270, inputting the safety index value into a target prediction model to obtain a safety score of the target wind power project.

[0137] Step 280, if the safety score is less than or equal to a score threshold, displaying safety warning information.

[0138] The score threshold can be reasonably set according to the safety requirements of the project, and the embodiments of the present application are not limited specifically.

[0139] The construction safety evaluation method of the wind power project provided in the embodiments of the present application takes the safety factor information and the safety score of the historical wind power project as samples, and trains the target prediction model in combination with the first weight value and the second weight value related to whether the safety score is missing. The safety condition of the target wind power project is scored by the trained target prediction model. When the safety score is low, the system automatically prompts the user of the complete problem. Even if the sample data is partially missing, the accurate prediction of the operation safety of the wind power project can be realized, effectively solving the problems of high complexity of data in the safety field and low prediction accuracy of the network model, improving the effectiveness of the safety prediction, and thus ensuring the safe construction and maintenance of the wind power project.

[0140] The model training method and the construction safety evaluation method of the wind power project provided in the embodiments of the present application can be applied in a terminal, can also be applied in a server, and can also be software running in the terminal or the server. In some embodiments, the terminal can be a smart phone, a tablet computer, a notebook computer, a desktop computer, etc.; the server can be configured as an independent physical server, can also be configured as a server cluster or a distributed system composed of multiple physical servers, can also be configured as a cloud server providing basic cloud computing services such as cloud service, cloud database, cloud computing, cloud function, cloud storage, network service, cloud communication, middleware service, domain name service, security service, CDN, and big data and artificial intelligence platform; and the software can be an application for implementing the intent recognition method, but is not limited to the above forms.

[0141] It should be noted that the size of the serial number of each step in the above embodiment does not mean the order of execution, the execution order of each process should be determined according to its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0142] Further, as shown in FIG. 3, as a specific implementation of the above model training method, the embodiments of the present application provide a model training device 300, which comprises an acquisition module 301, a determination module 302 and a training module 303.

[0143] The acquisition module 301 is configured to acquire sample data, wherein the sample data comprises training samples and real value labels of the training samples in different dimensions.

[0144] The determination module 302 is configured to determine, according to the association information of the training samples and the real value labels, a label missing rate of the training samples in different dimensions and a first weight value of the training samples in different dimensions, and determine a second weight value of the training samples according to the label missing rate.

[0145] The training module 303 is configured to input the sample data, the first weight value and the second weight value into a preset neural network model for training until a loss value of a target loss function of the preset neural network model meets a model convergence condition, so as to obtain a target prediction model.

[0146] In this embodiment, the weight of the training samples in different dimensions is calculated based on the association information of whether the training samples are associated with the real value labels, and the importance of the sample data with partially missing labels is represented by the weight. In this way, when calculating the model loss value, the loss value of the missing labels of the sample data with partially missing labels can be calculated by referring to the weight under the condition that the sample data with partially missing labels is retained. Thus, the problem of prediction accuracy decline caused by deleting the sample data with partially missing labels in the data set is effectively solved, the accuracy of the model is improved, and the prediction values of multiple dimensions can be accurately predicted at the same time.

[0147] Further, the determination module 302 is specifically configured to assign the first weight value in any dimension to 1 if the association information is that the training samples are associated with the real value labels in any dimension, and assign the first weight value in any dimension to 0 if the association information is that the training samples are not associated with the real value labels in any dimension.

[0148] Further, the determination module 302 is specifically configured to count the number of labels missing of the training samples whose association information is that the training samples are not associated with the real value labels in any dimension, and calculate the quotient of the dimension label missing number and the total number of the training samples in the sample data as the label missing rate of any dimension.

[0149] Further, the determining module 302 is specifically configured to calculate the information entropy and the mutual information of the true value label; and determine the second weight value according to the information entropy, the mutual information and the label missing rate.

[0150] Further, the model training apparatus 300 further comprises a preprocessing module (not shown in the figure).

[0151] The preprocessing module is configured to perform normalization processing on the second weight value.

[0152] Further, as shown in FIG. 4, as a specific implementation of the construction safety assessment method of the wind power project, the embodiment of the present application provides a construction safety assessment apparatus 400 of a wind power project, which comprises an acquisition module 401, a data processing module 402, an assessment module 403 and a warning module 404.

[0153] The acquisition module 401 is configured to acquire safety factor information of a target wind power project and a target prediction model, wherein the target prediction model is trained by using the model training method provided in the first aspect and taking the safety factor information and the safety score of a historical wind power project as sample data.

[0154] The data processing module 402 is configured to perform dimensionless processing on the safety factor information according to a preset conversion rule to generate a safety index value.

[0155] The assessment module 403 is configured to input the safety index value into the target prediction model to obtain a safety score of the target wind power project.

[0156] The warning module 404 is configured to display safety warning information if the safety score is less than or equal to a score threshold.

[0157] In this embodiment, the safety factor information and the safety score of a historical wind power project are taken as samples, and the first weight value and the second weight value related to whether the safety score is missing are combined to train the target prediction model. The safety condition of a target wind power project is scored by using the trained target prediction model. When the safety score is low, the system automatically prompts the user to find out the problem. Even if the sample data is partially missing, the wind power project operation safety can be accurately predicted, effectively solving the problem that the safety field is difficult to process due to high data complexity and the network model prediction accuracy is not high, improving the effectiveness of the safety prediction, and thus ensuring the safety of the wind power project construction and maintenance.

[0158] The specific limitations of the model training apparatus can be referred to the limitations of the model training method in the foregoing, which will not be repeated here. Each module in the model training apparatus can be realized by software, hardware, and combinations thereof, in whole or in part. Each module can be embedded in or independent of the processor in the computer device in hardware form, or stored in the memory in the computer device in software form, so as to be called and executed by the processor to perform the operations corresponding to each module.

[0159] Based on the above method as shown in FIG. 1 and FIG. 2, accordingly, the embodiment of the present application also provides a readable storage medium, which stores a computer program, and the program is executed by a processor to realize the model training method as shown in FIG. 1 and the construction safety evaluation method of the wind power project as shown in FIG. 2.

[0160] Based on such understanding, the technical solution of the present application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a U disk, a mobile hard disk, etc.), and includes a plurality of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute the method described in each implementation scenario of the present application.

[0161] Based on the above method as shown in FIG. 1 and FIG. 2, and the virtual apparatus embodiments as shown in FIG. 3 and FIG. 4, in order to achieve the above purpose, the embodiment of the present application also provides a computer device, which can be a personal computer, a server, a network device, etc., and the computer device includes a storage medium and a processor; the storage medium is used to store a computer program; the processor is used to execute the computer program to realize the model training method as shown in FIG. 1 and the construction safety evaluation method of the wind power project as shown in FIG. 2.

[0162] Optionally, the computer device can also include a user interface, a network interface, a camera, a radio frequency (RF) circuit, a sensor, an audio circuit, a WI-FI module, etc. The user interface can include a display screen, an input unit such as a keyboard, etc. The optional user interface can also include a USB interface, a card reader interface, etc. The network interface can optionally include a standard wired interface, a wireless interface (such as a Bluetooth interface, a WI-FI interface), etc.

[0163] Those skilled in the art can understand that the structure of the computer device provided by the embodiment does not constitute a limitation on the computer device, and can include more or fewer components, or combine certain components, or different component arrangements.

[0164] The storage medium can further include an operating system and a network communication module. The operating system is a program for managing and saving computer device hardware and software resources, and supports the running of information processing programs and other software and / or programs. The network communication module is used to realize communication between components in the storage medium, and communication with other hardware and software in the entity device.

[0165] Those skilled in the art can clearly understand from the description of the above embodiments that the present application can be implemented by means of software and a necessary general hardware platform, or by hardware.

[0166] Those skilled in the art can understand that the accompanying drawings are only schematic diagrams of a preferred implementation scenario, and the modules or flows in the drawings are not necessarily essential for implementing the present application. Those skilled in the art can understand that the modules in the device in the implementation scenario can be distributed in the device in the implementation scenario according to the description of the implementation scenario, or can be changed and located in one or more devices different from the implementation scenario. The modules of the above implementation scenario can be combined into one module, or can be further split into multiple sub-modules.

[0167] The above serial numbers of the present application are only for description, and do not represent the advantages and disadvantages of the implementation scenario. The above disclosure is only several specific implementation scenarios of the present application, but the present application is not limited thereto, and any changes that those skilled in the art can think of should fall within the protection scope of the present application.

Claims

1. A model training method, characterized in that, The method comprises: acquiring sample data, wherein the sample data comprises training samples and true value labels of the training samples in different dimensions, the training samples comprise safety factor information of historical wind power projects, and the true value labels comprise safety scores of the historical wind power projects; determining a label missing ratio of the training samples in different dimensions and first weight values of the training samples in different dimensions according to association information of the training samples and the true value labels; determining second weight values of the training samples according to the label missing ratio; inputting the sample data, the first weight values and the second weight values into a preset neural network model for training until a loss value of a target loss function of the preset neural network model meets a model convergence condition, to obtain a target prediction model, wherein the target prediction model is used for predicting safety scores of target wind power projects; the determining of the label missing ratio of the training samples in different dimensions and the first weight values of the training samples in different dimensions according to the association information of the training samples and the true value labels comprises: if the association information is that the training samples are associated with the true value labels in any dimension, assigning the first weight value in the any dimension as 1; if the association information is that the training samples are not associated with the true value labels in any dimension, assigning the first weight value in the any dimension as 0; counting a label missing number of the training samples whose association information is that the training samples are not associated with the true value labels in any dimension; and calculating a quotient of the dimension label missing number and a total number of training samples in the sample data as the label missing ratio of the any dimension; the determining of the second weight values of the training samples according to the label missing ratio comprises: calculating information entropy and mutual information of the true value labels; and determining the second weight values according to the information entropy, the mutual information and the label missing ratio; The target loss function is represented as: In the formula, loss represents a loss function, y represents a real value label, represents a predicted value of a model output, n represents a total number of training samples, l represents a number of predicted labels, i represents a number of a training sample, a ij represents a first weight value of a jth real value label of the ith training sample, w j represents a second weight value of the jth real value label.

2. The model training method according to claim 1, further comprising: The formula for calculating the information entropy of the true value label is represented as: In the formula, E j represents the information entropy of the jth real value label, p ij represents the probability of the value of the ith training sample appearing on the jth real value label, and n represents the total number of training samples; The formula for calculating the mutual information amount of the true value label is represented as: where I(Y j ;Y k ) denotes the mutual information between labels Y j and Y k , p(y j ,y k ) denotes the joint probability mass function of labels Y j and Y k , and p(y j ) and p(y k ) denote the marginal probability mass functions of labels Y j and Y k , respectively. The second weight value is determined according to the information entropy, the mutual information amount and the label missing ratio, and the following formula is adopted: where w j represents the second weight value of the jth real value label, m represents the total number of real value labels in any dimension, E j represents the information entropy of the jth real value label, I(Y j ; Y k ) represents the mutual information between labels Y j and Y k , I(Y k ; Y l ) represents the mutual information between labels Y k and Y l , and a represents the label missing rate in any dimension.

3. The model training method of claim 1, wherein, normalizing the second weight values. The method further comprises:

4. A method for construction safety assessment of a wind power project, characterized in that, acquiring a target prediction model and safety factor information of a target wind power project, wherein the target prediction model takes safety factor information and safety scores of historical wind power projects as sample data, and is trained by using the model training method according to any one of claims 1 to 3; performing dimensionless processing on the safety factor information of the target wind power project according to a preset conversion rule to generate a safety index value; inputting the safety index value into the target prediction model to obtain a safety score of the target wind power project; if the safety score is less than or equal to a score threshold, displaying safety warning information. The device comprises:

5. A model training apparatus characterized by comprising: an acquisition module configured to acquire sample data, wherein the sample data comprises training samples and true value labels of the training samples in different dimensions, the training samples comprise safety factor information of historical wind power projects, and the true value labels comprise safety scores of the historical wind power projects; ​ determine, according to the association information of the training sample and the true value label, a label missing ratio of the training sample in different dimensions and a first weight value of the training sample in different dimensions; and determine a second weight value of the training sample according to the label missing ratio; a training module configured to input the sample data, the first weight value and the second weight value into a preset neural network model for training until a loss value of a target loss function of the preset neural network model meets a model convergence condition, to obtain a target prediction model, the target prediction model being configured to predict a safety score of a target wind power project; The determination module is specifically configured to: if the association information is that the training sample is associated with the true value label in any dimension, assign the first weight value in the any dimension as 1; if the association information is that the training sample is not associated with the true value label in any dimension, assign the first weight value in the any dimension as 0; count a label missing number of training samples whose association information is that the training sample is not associated with the true value label in any dimension; calculate a quotient of the dimension label missing number and a total number of training samples in the sample data as the label missing ratio of the any dimension; calculate information entropy and mutual information of the true value label; and determine the second weight value according to the information entropy, the mutual information and the label missing ratio. The target loss function is represented as: In the formula, loss represents a loss function, y represents a real value label, represents a predicted value of a model output, n represents a total number of training samples, l represents a number of predicted labels, i represents a number of a training sample, a ij represents a first weight value of a jth real value label of the ith training sample, w j represents a second weight value of the jth real value label.

6. A construction safety assessment device for a wind power project, characterized by, The device comprises: an acquisition module configured to acquire a target prediction model and safety factor information of a target wind power project, wherein the target prediction model takes safety factor information and a safety score of a historical wind power project as sample data and is trained by using the model training method according to any one of claims 1 to 3; a data processing module configured to perform dimensionless processing on the safety factor information of the target wind power project according to a preset conversion rule, to generate a safety index value; an evaluation module configured to input the safety index value into the target prediction model, to obtain the safety score of the target wind power project; an early warning module configured to display safety early warning information if the safety score is less than or equal to a score threshold.

7. A computer device comprising a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, characterized in that, The processor implements the model training method according to any one of claims 1 to 3 or the construction safety evaluation method of the wind power project according to claim 4 when executing the program. The processor implements the model training method according to any one of claims 1 to 3 or the construction safety evaluation method of the wind power project according to claim 4 when executing the program.

Images