Construction method of occupational health analysis model based on big data

Abstract

The invention discloses an occupational health analysis model construction method based on big data, and the method comprises the steps: obtaining occupational health data, carrying out the preprocessing of the occupational health data, carrying out the feature extraction, obtaining health features, labeling the health features with feature tags, and carrying out the feature extraction. Inputting the health features into the multi-modal machine learning model to obtain a health risk probability, obtaining interaction data, constructing a feature transformation matrix and a health coefficient according to the interaction data and the health features, fusing the feature transformation matrix and the health coefficient to obtain fused features, and obtaining a health risk probability; and constructing an occupational health analysis model based on big data according to the fusion features and the feature tags, and inputting occupational health data to be analyzed into the occupational health analysis model to obtain occupational health scores and health evaluations. The method not only can improve the precision and speed of occupational health analysis, but also has good interpretability, and can be directly applied to an occupational health analysis system.

Description

Technical Field

[0001] The present invention relates to the field of big data processing and artificial intelligence technology, and in particular to a method for constructing an occupational health analysis model based on big data. Background Art

[0002] Occupational health refers to the protection of workers' physical and mental health and the prevention of occupational diseases through preventive measures and health management in the workplace. Analysis models based on big data mainly provide scientific basis for decision-making, risk assessment, trend forecasting, etc. by collecting, integrating, processing and analyzing large-scale data sets.

[0003] In the context of modern industrialization and informatization, occupational health issues have received widespread attention. Traditional occupational health management methods often rely on regular physical examinations and questionnaires, which are limited in efficiency and real-time, and ignore the role of mental health and environmental factors. The rise of big data technology has provided new ideas and methods for occupational health management. Although some occupational health analysis models based on big data have been invented, there are still the following problems: insufficient data processing capabilities, single analysis models, lack of dynamic adjustment and self-learning capabilities. Therefore, it is of great significance to design an occupational health analysis model based on big data to overcome the shortcomings of existing analysis models, to achieve comprehensive and dynamic analysis of occupational health, and to provide accurate and personalized health management services for enterprises and employees. Summary of the invention

[0004] The purpose of the present invention is to provide a method for constructing an occupational health analysis model based on big data.

[0005] To achieve the above object, the present invention is implemented according to the following technical solutions:

[0006] The present invention comprises the following steps:

[0007] Acquire occupational health data and pre-process the occupational health data; the occupational health data includes historical data and environmental data;

[0008] Classifying the historical data, performing feature extraction on the historical data to obtain health features, labeling the health features with feature tags according to the values ​​of the health features, and building a feature tag library according to the feature tags;

[0009] Constructing a multimodal machine learning model, and inputting the health characteristics into the multimodal machine learning model to obtain health risk probability;

[0010] According to the health risk probability, the health feature interacts with the environmental database composed of the environmental data to obtain interactive data, constructs a feature transformation matrix according to the interactive data and the health feature, and inputs the interactive data and the mental health feature into a health function to obtain a health coefficient;

[0011] The feature transformation matrix and the health coefficient are fused to obtain fusion features, an occupational health analysis model based on big data is constructed according to the fusion features and feature labels, and the occupational health data to be analyzed is input into the occupational health analysis model to obtain an occupational health score and health evaluation.

[0012] Furthermore, the method of extracting features from the historical data to obtain health features includes:

[0013] Hot-encode non-numeric data and use the random forest algorithm to divide the data. The data types include physical health data and mental health data.

[0014] Using principal component analysis to extract features from the physical health data to obtain physical health features; the physical health features are comprehensive features related to physical health;

[0015] Factor analysis is used to extract features from mental health data to obtain mental health characteristics; the mental health characteristics are potential factors that reflect the meaning of mental health scores;

[0016] Divide a plurality of feature labels according to the values ​​of the health characteristics and label the corresponding health characteristics with feature labels; the feature labels are health evaluations of the health characteristics;

[0017] The feature labels are combined to form a feature label library, which consists of a physical health label group, a mental health label group and an environmental label group. Each label group contains multiple feature labels with different feature indicators, and the same feature indicator contains multiple feature labels with different values.

[0018] The physical health features and mental health features extracted from all historical data form a health feature library, which is divided into a probability model set and a health feature set. The probability model set is divided into a probability training set and a probability test set.

[0019] Further, the method for obtaining the health risk probability includes:

[0020] Define that the first probability training set and the first probability test set only contain physical health features, the second probability training set and the second probability test set only contain mental health features, and the third probability training set and the third probability test set contain both physical health features and mental health features;

[0021] Gradient boosting tree, multi-layer perceptron and convolutional neural network combined with long short-term memory network are used as base models to construct multimodal machine learning models;

[0022] The first probability training set is input into the gradient boosting tree, the second probability training set is input into the multilayer perceptron, and the third probability training set is input into the convolutional neural network combined with the long short-term memory network for independent training. The parameters of the base model are optimized using k-fold cross validation. After the training is completed, the first probability test set, the second probability test set, and the third probability test set are used to evaluate the performance of the base model respectively. The weighted average method is used as an integration strategy to combine the prediction results of the three base models to calculate the final prediction probability, and output the multimodal machine learning model.

[0023] The health feature set is input into the multimodal machine learning model to predict the health risk probability.

[0024] Furthermore, the method of constructing a feature transformation matrix according to the interaction data and the health feature includes:

[0025] When the health risk probability is less than the preset probability threshold, it is judged as occupational health;

[0026] When the health risk probability is greater than a preset probability threshold, it is determined that there is an occupational health risk, the health characteristics are interacted with the environmental database composed of the environmental data according to the personal information to obtain interactive data, the interactive data is preprocessed, and random forest is used to extract features from the interactive data to obtain environmental features, and multiple feature labels are divided according to the value of each environmental feature, and the feature labels are marked according to the value of the environmental feature; the interactive data includes working environment temperature, environmental quality, working hours, salary level, occupational hazard probability and occupational injury rate;

[0027] Physical health features, mental health features, and environmental features are combined into feature samples to define the original training matrix It means that the sample data matrix X contains n samples and d features. By adding the regularized sparse regression framework, the non-negative matrix decomposition framework, retaining the local structure and constraining the similarity matrix, the objective function is constructed. The feature transformation matrix W is obtained by optimizing the objective function. The objective function expression is:

[0028] stΣ j S ij =1,0≤S ij ≤1, for all i, P, Q, T ≥ 0, A T A=I,B T B=I

[0029] in is the feature transformation matrix, m=c is the number of clusters, is the cluster center matrix, is the coefficient matrix, η≥0 is the regularization parameter, ‖W‖ 2,1 To utilize l 2,1 The norm constraint matrix W, the first and second terms are the regularized sparse regression framework; XP is the basis matrix, k is the cardinality, is the coefficient matrix, the third term is the non-negative matrix decomposition framework; α≥0 is the regular coefficient, is the auxiliary matrix, and the role of B is to ensure that T is orthogonal; L is the graph Laplacian matrix of the original data matrix, R = l k×k -l k , γ≥0, λ≥0 and ε≥0 are regularization parameters, the fifth and sixth terms are local preservation structures, and the seventh term is a constraint term to reduce the correlation of the coefficient matrix B; S is the similarity matrix, O is the initial similarity matrix, and β is the learning weight parameter of the model;

[0030] The objective function involves the optimization of six variables (W, A, B, P, Q, S) in total, which makes the objective function non-convex. Fixed variables (A, B, P, Q, S) make the objective function convex, set the derivative of variable W to 0, and find the optimal solution to the problem to obtain the update formula of the variable feature transformation matrix W:

[0031] W←(XX T +γXLX T +ηD)XBA T

[0032] in is a diagonal matrix, and the matrix elements d ij =(1 / 2‖W i ‖ 2 ), ‖W‖ 2,1 The derivative of variable W is 2DW.

[0033] Furthermore, the method of inputting the interaction data and the mental health characteristics into a health function to obtain a health coefficient includes:

[0034] Align the environmental characteristics and mental health characteristics in the characteristic dimension, input the characteristic data into the health function to obtain the health coefficient, and the health function expression is:

[0035]

[0036] where δ i is the health coefficient of the i-th group of data, w 1 and w 2 is the weight coefficient, S i is the similarity between the i-th group of environmental feature vectors and the mental health feature vectors, is the mean of the similarity of n groups of feature vectors, n is the number of feature vector groups, σ s is the standard deviation of the similarity of n groups of feature vectors.

[0037] Furthermore, the method of fusing the feature transformation matrix and the health coefficient to obtain a fusion feature includes:

[0038] The health factor δ i According to the corresponding data group, add it to the i-th row of the feature transformation matrix W to form a fusion matrix m represents the number of data sets, (d+1) represents the row vector dimension, and the input vector is set to the row vector y i , the kernel function is K(·), the feature vector set is Y, and the output vector is u i , the optimal hyperplane expression for obtaining data is:

[0039] u=τ i u i K(y i , Y)+θ

[0040] Where u is the optimal hyperplane obtained, τ i is the weight coefficient of the i-th row vector, and θ is the bias term;

[0041] Based on the obtained optimal hyperplane, the fusion matrix data is divided into h groups. Different groups correspond to a new fusion feature. The actual data value is set to u, the training output is U, and the generalization parameter is χ. The health data fusion model of SVM is established. The expression is:

[0042]

[0043] Where J is the health data fusion model, l i and l j is the optimal solution marking form for data i and j, and is the optimal solution mean marking form;

[0044] By determining the optimal solution of the characteristic data, the decision function of the characteristic data is established, the model is solved, and data fusion is realized. The expression is:

[0045]

[0046] Where f(x) is the constructed decision function, is the mean value set of h groups of data, l h Mark the data in the optimal solution format;

[0047] Fusion matrix Y by row vector y iThe input is decomposed into h groups of data, and m groups of fusion features with dimension h are obtained through the health data fusion model.

[0048] Furthermore, a method for constructing an occupational health analysis model based on big data according to the fusion features and historical labels includes:

[0049] The fused features and feature labels are combined into a feature set, and the feature set is divided into an attention training set and an attention test set to build an occupational health analysis model. The specific structure includes an input layer, a feature preprocessing layer, a multi-head attention fusion layer, an adaptive feature fusion layer, and an output layer.

[0050] The attention training set is input into the occupational health analysis model. The attention training set data is linearly transformed in the feature preprocessing layer to obtain a feature representation with unified dimension and form. The linearly transformed data is subjected to feature extraction and then input into the multi-head attention fusion layer. The APReLU activation function is used to adjust the weight of each head to generate weighted features. The weighted features and original features are input into the adaptive feature fusion layer. The features of different layers are integrated into the output layer using ASFF. Two fully connected layers are set in the output layer. One fully connected layer generates the occupational health score, and the other fully connected layer is followed by a softmax activation function to predict the feature label, and the health score and health evaluation are output.

[0051] The mean square error loss and SGD optimizer are used to evaluate the model and optimize the weights to obtain the attention mechanism function. The attention test set is used to evaluate the attention mechanism function and output the attention mechanism model.

[0052] The occupational health data to be analyzed is input into the occupational health analysis model to obtain the occupational health score and health evaluation.

[0053] The beneficial effects of the present invention are:

[0054] The present invention is a method for constructing an occupational health analysis model based on big data. Compared with the prior art, the present invention has the following technical effects:

[0055] The present invention can improve the ability of data preprocessing and enhance the adaptability of the model in occupational health data analysis through the steps of feature extraction, feature labeling, feature transformation matrix construction, health coefficient acquisition, feature fusion, and model construction, thereby improving the efficiency and accuracy of occupational health data analysis, optimizing the occupational health data analysis technology, greatly saving resources, and improving work efficiency, and can realize the analysis of occupational health data, providing strong support for occupational health management, which is of great significance to occupational health data analysis, and can adapt to different occupational health analysis systems based on big data and the terminal occupational health data analysis needs of occupational health analysis systems based on big data of different users, and has a certain universality. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 This is a flowchart of the steps of a method for constructing an occupational health analysis model based on big data in the present invention. DETAILED DESCRIPTION

[0057] The present invention is further described below by means of specific embodiments. The illustrative embodiments and descriptions of the present invention are used to explain the present invention but are not intended to limit the present invention.

[0058] The present invention provides a method for constructing an occupational health analysis model based on big data, comprising the following steps:

[0059] like Figure 1 As shown, in this embodiment, the following steps are included:

[0060] Acquire occupational health data and pre-process the occupational health data; the occupational health data includes historical data and environmental data;

[0061] Classifying the historical data, performing feature extraction on the historical data to obtain health features, labeling the health features with feature tags according to the values ​​of the health features, and building a feature tag library according to the feature tags;

[0062] Constructing a multimodal machine learning model, and inputting the health characteristics into the multimodal machine learning model to obtain health risk probability;

[0063] According to the health risk probability, the health feature interacts with the environmental database composed of the environmental data to obtain interactive data, constructs a feature transformation matrix according to the interactive data and the health feature, and inputs the interactive data and the mental health feature into a health function to obtain a health coefficient;

[0064] The feature transformation matrix and the health coefficient are fused to obtain fusion features, an occupational health analysis model based on big data is constructed according to the fusion features and feature labels, and the occupational health data to be analyzed is input into the occupational health analysis model to obtain an occupational health score and health evaluation.

[0065] In this embodiment, the method of extracting features from the historical data to obtain health features includes:

[0066] Hot-encode non-numeric data and use the random forest algorithm to divide the data. The data types include physical health data and mental health data.

[0067] Using principal component analysis to extract features from the physical health data to obtain physical health features; the physical health features are comprehensive features related to physical health;

[0068] Factor analysis is used to extract features from mental health data to obtain mental health characteristics; the mental health characteristics are potential factors that reflect the meaning of mental health scores;

[0069] Divide a plurality of feature labels according to the values ​​of the health characteristics and label the corresponding health characteristics with feature labels; the feature labels are health evaluations of the health characteristics;

[0070] The feature labels are combined to form a feature label library, which consists of a physical health label group, a mental health label group and an environmental label group. Each label group contains multiple feature labels with different feature indicators, and the same feature indicator contains multiple feature labels with different values.

[0071] The physical health features and mental health features extracted from all historical data are combined into a health feature library, which is divided into a probability model set and a health feature set, and the probability model set is divided into a probability training set and a probability test set;

[0072] In the actual evaluation, taking the occupational health data of employees of Company X as an example, a group of historical data is randomly given for classification to obtain health data. After extracting the features of the health data, a group of health features are obtained: gender / male, weight / 75kg, age / 25 years old, (physical health characteristics: pulse rate 72 times / minute, blood pressure 120 / 80mmHg, vision-normal, hearing-normal, facial features-normal, electrocardiogram-sinus arrhythmia, slight abnormality, need regular review, B-ultrasound-normal, blood-normal blood routine, uric acid 450μmol / L is high, internal medicine-normal, surgery-normal, nervous system-normal), (mental health characteristics: work pressure-moderate pressure, within the controllable range, job satisfaction-overall satisfaction with the work, emotional stability-slight fluctuations in emotions due to work pressure or personal affairs, coping ability-strong), and labeling operations are performed on the health data.

[0073] In this embodiment, the method for obtaining the health risk probability includes:

[0074] Define that the first probability training set and the first probability test set only contain physical health features, the second probability training set and the second probability test set only contain mental health features, and the third probability training set and the third probability test set contain both physical health features and mental health features;

[0075] Gradient boosting tree, multi-layer perceptron and convolutional neural network combined with long short-term memory network are used as base models to construct multimodal machine learning models;

[0076] The first probability training set is input into the gradient boosting tree, the second probability training set is input into the multilayer perceptron, and the third probability training set is input into the convolutional neural network combined with the long short-term memory network for independent training. The parameters of the base model are optimized using k-fold cross validation. After the training is completed, the first probability test set, the second probability test set, and the third probability test set are used to evaluate the performance of the base model respectively. The weighted average method is used as an integration strategy to combine the prediction results of the three base models to calculate the final prediction probability, and output the multimodal machine learning model.

[0077] Input the health feature set into the multimodal machine learning model to predict the health risk probability;

[0078] In the actual assessment, the random health characteristics of this group were scored on a 10-point scale: gender / male, weight / 75kg, age / 25 years old, (physical health characteristics: pulse rate 10 points, blood pressure 10 points, vision 10 points, hearing 10 points, facial features 10 points, electrocardiogram 8 points, B-ultrasound 10 points, blood 8 points, internal medicine 10 points, surgery 10 points, nervous system 10 points), (psychological health characteristics: work pressure 7 points, job satisfaction 8 points, emotional stability 8 points, coping ability 9 points);

[0079] This set of data was input into the trained multimodal machine learning model. The health risk probabilities given by the three base models were 5%, 10%, and 8%, respectively. The weighted average method (the weight coefficients of the output results of the three base models were 0.3, 0.2, and 0.5, respectively) was used as the integration strategy to combine the prediction results of the three base models to calculate the final predicted health risk probability of 7.5%.

[0080] In this embodiment, the method of constructing a feature transformation matrix according to the interaction data and the health feature includes:

[0081] When the health risk probability is less than the preset probability threshold, it is judged as occupational health;

[0082] When the health risk probability is greater than a preset probability threshold, it is determined that there is an occupational health risk, the health characteristics are interacted with the environmental database composed of the environmental data according to the personal information to obtain interactive data, the interactive data is preprocessed, and random forest is used to extract features from the interactive data to obtain environmental features, and multiple feature labels are divided according to the value of each environmental feature, and the feature labels are marked according to the value of the environmental feature; the interactive data includes working environment temperature, environmental quality, working hours, salary level, occupational hazard probability and occupational injury rate;

[0083] Physical health features, mental health features, and environmental features are combined into feature samples to define the original training matrix It means that the sample data matrix X contains n samples and d features. By adding the regularized sparse regression framework, the non-negative matrix decomposition framework, retaining the local structure and constraining the similarity matrix, the objective function is constructed. The feature transformation matrix W is obtained by optimizing the objective function. The objective function expression is:

[0084] st∑ j S ij =1,0≤S ij ≤1, for all i, P, Q, T ≥ 0, A T A=I,B T B=I

[0085] in is the feature transformation matrix, m=c is the number of clusters, is the cluster center matrix, is the coefficient matrix, η≥0 is the regularization parameter, ‖W‖ 2,1 To utilize l 2,1 The norm constraint matrix W, the first and second terms are the regularized sparse regression framework; XP is the basis matrix, k is the cardinality, is the coefficient matrix, the third term is the non-negative matrix decomposition framework; α≥0 is the regular coefficient, is the auxiliary matrix, and the role of B is to ensure that T is orthogonal; L is the graph Laplacian matrix of the original data matrix, R = l k×k -l k , γ≥0, λ≥0 and ε≥0 are regularization parameters, the fifth and sixth terms are local preservation structures, and the seventh term is a constraint term to reduce the correlation of the coefficient matrix B; S is the similarity matrix, O is the initial similarity matrix, and β is the learning weight parameter of the model;

[0086] The objective function involves the optimization of six variables (W, A, B, P, Q, S) in total, which makes the objective function non-convex. Fixed variables (A, B, P, Q, S) make the objective function convex, set the derivative of variable W to 0, and find the optimal solution to the problem to obtain the update formula of the variable feature transformation matrix W:

[0087] W←(XX T +γXLX T +ηD)XBA T

[0088] in is a diagonal matrix, and the matrix elements d ij =(1 / 2‖W i ‖ 2 ), ‖W‖ 2,1 The derivative of variable W is 2DW;

[0089] In the actual assessment, the health risk probability of this group of data is 7.5%, which is greater than the preset probability threshold of 5%. It is determined that there is an occupational health risk. According to the personal information, the interactive data is matched in the big database: (working environment temperature - 22°C, environmental quality - good, working hours - 60h / week, slightly higher than the industry average working hours, salary level - 10% lower than the industry average, occupational hazard probability - medium, occupational injury rate - slightly higher than the industry average, and increased cardiovascular burden caused by long-term high-intensity work);

[0090] The same 10-point system is used to score: (10 points for working environment temperature, 10 points for environmental quality, 4 points for working hours, 5 points for salary level, 7 points for occupational hazard probability, and 6 points for occupational injury rate). This results in a complete set of feature samples including physical health indicators, mental health indicators, and environmental indicators. Multiple groups of data samples are processed to form a feature sample set. The corresponding row vector of this group of data in the training matrix is ​​[10,10,10,10,10,8,10,8,10,10,10,7,8,8,9,10,10,4,5,7,6], and the corresponding row vector in the feature transformation matrix W after feature transformation is [100,100,100,100,100,64,100,64,100,100,100,49,6464,81,100,10,0,16,25,49,36].

[0091] In this embodiment, the method of inputting the interaction data and the mental health characteristics into a health function to obtain a health coefficient includes:

[0092] Align the environmental characteristics and mental health characteristics in the characteristic dimension, input the characteristic data into the health function to obtain the health coefficient, and the health function expression is:

[0093]

[0094] where δ i is the health coefficient of the i-th group of data, w 1 and w 2 is the weight coefficient, S i is the similarity between the i-th group of environmental feature vectors and the mental health feature vectors, is the mean of the similarity of n groups of feature vectors, n is the number of feature vector groups, σ s is the standard deviation of the similarity of n groups of feature vectors;

[0095] In the actual assessment, the environmental characteristic dimension is 6 and the mental health characteristic dimension is 4. The mental health characteristic dimension is expanded to 6, and the characteristic data is input into the health function to calculate the health coefficient of this group of data to be 1.0469.

[0096] In this embodiment, the method of fusing the feature transformation matrix and the health coefficient to obtain a fusion feature includes:

[0097] The health coefficient η i According to the corresponding data group, add it to the i-th row of the feature transformation matrix W to form a fusion matrix m represents the number of data sets, (d+1) represents the row vector dimension, and the input vector is set to the row vector y i , the kernel function is K(·), the feature vector set is Y, and the output vector is u i , the optimal hyperplane expression for obtaining data is:

[0098] u=τ i u i K(y i , Y)+θ

[0099] Where u is the optimal hyperplane obtained, τ i is the weight coefficient of the i-th row vector, and θ is the bias term;

[0100] Based on the obtained optimal hyperplane, the fusion matrix data is divided into h groups. Different groups correspond to a new fusion feature. The actual data value is set to u, the training output is U, and the generalization parameter is χ. The health data fusion model of SVM is established. The expression is:

[0101]

[0102] Where J is the health data fusion model, l i and l j is the optimal solution marking form for data i and j, and is the optimal solution mean marking form;

[0103] By determining the optimal solution of the characteristic data, the decision function of the characteristic data is established, the model is solved, and data fusion is realized. The expression is:

[0104]

[0105] Where f(x) is the constructed decision function, is the mean value set of h groups of data, l h Mark the data in the optimal solution format;

[0106] Fusion matrix Y by row vector y i The input is decomposed into h groups of data, and m groups of fusion features with dimension h are obtained through the health data fusion model;

[0107] In the actual evaluation, the fusion matrix is ​​divided into 12 groups of data for data fusion, and the row vectors of the fusion matrix corresponding to the above data are fused to obtain the fusion features [100, 100, 64, 100, 100, 100, 100, 49, 64, 100, 16, 12].

[0108] In this embodiment, an occupational health analysis model based on big data is constructed according to the fusion features and historical labels, including:

[0109] The fused features and feature labels are combined into a feature set, and the feature set is divided into an attention training set and an attention test set to build an occupational health analysis model. The specific structure includes an input layer, a feature preprocessing layer, a multi-head attention fusion layer, an adaptive feature fusion layer, and an output layer.

[0110] The attention training set is input into the occupational health analysis model. The attention training set data is linearly transformed in the feature preprocessing layer to obtain a feature representation with unified dimension and form. The linearly transformed data is subjected to feature extraction and then input into the multi-head attention fusion layer. The APReLU activation function is used to adjust the weight of each head to generate weighted features. The weighted features and original features are input into the adaptive feature fusion layer. The features of different layers are integrated into the output layer using ASFF. Two fully connected layers are set in the output layer. One fully connected layer generates the occupational health score, and the other fully connected layer is followed by a softmax activation function to predict the feature label, and the health score and health evaluation are output.

[0111] The mean square error loss and SGD optimizer are used to evaluate the model and optimize the weights to obtain the attention mechanism function. The attention test set is used to evaluate the attention mechanism function and output the attention mechanism model.

[0112] Inputting the occupational health data to be analyzed into the occupational health analysis model to obtain occupational health scores and health evaluations;

[0113] In the actual evaluation, a group of health data to be analyzed was randomly generated: gender / male, weight / 80kg, age / 45 years old, (physical health characteristics: pulse rate 85 times / minute, slightly fast, blood pressure 145 / 95mmHg, stage I hypertension, vision-normal, hearing-normal, facial features-normal, electrocardiogram-sinus arrhythmia, left ventricular high voltage, B-ultrasound-normal, blood-normal blood routine, uric acid 500μmol / L is high, internal medicine-mild fatty liver, surgery-normal, nervous system-neurasthenia symptoms, easy fatigue, inattention, sleep disorders), (mental health characteristics: work pressure-high pressure, difficult to control, job satisfaction-overall dissatisfaction with work, emotional stability - large mood swings, coping ability - average), (working environment temperature - 20℃, environmental quality - average, working hours - 50h / week, slightly higher than the industry average working hours, salary level - 20% higher than the industry average, occupational hazard probability - high, occupational injury rate - slightly higher than the industry average, cervical spine problems and muscle strain caused by long-term sitting), the health score obtained by inputting into the occupational health analysis model is: 57 points, the overall health status is at a lower-middle level, and the health evaluation obtained according to the predicted feature labels is: physical health has problems such as high blood pressure, high uric acid and fatty liver, there are further mental health problems, there are additional work pressure and expectations, and attention should be paid to occupational health and safety.

[0114] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for constructing an occupational health analysis model based on big data, characterized in that: The following steps are involved: S1. Acquire occupational health data and pre-process the occupational health data; the occupational health data includes historical data and environmental data; S2, classifying the historical data, performing feature extraction on the historical data to obtain health features, labeling the health features with feature tags according to the values ​​of the health features, and building a feature tag library according to the feature tags; S3. Building a multimodal machine learning model, and inputting the health features into the multimodal machine learning model to obtain health risk probability; S4, according to the health risk probability, the health feature interacts with the environmental database composed of the environmental data to obtain interactive data, constructs a feature transformation matrix according to the interactive data and the health feature, and inputs the interactive data and the mental health feature into a health function to obtain a health coefficient; S5. Fusing the feature transformation matrix and the health coefficient to obtain a fusion feature, constructing an occupational health analysis model based on big data according to the fusion feature and feature labels, and inputting the occupational health data to be analyzed into the occupational health analysis model to obtain an occupational health score and health evaluation.

2. According to the method for constructing an occupational health analysis model based on big data as described in claim 1, it is characterized in that: The method of extracting features from the historical data to obtain health features includes: Hot-encode non-numeric data and use the random forest algorithm to divide the data. The data types include physical health data and mental health data. Using principal component analysis to extract features from the physical health data to obtain physical health features; the physical health features are comprehensive features related to physical health; Factor analysis is used to extract features from mental health data to obtain mental health characteristics; the mental health characteristics are potential factors that reflect the meaning of mental health scores; Divide a plurality of feature labels according to the values ​​of the health characteristics and label the corresponding health characteristics with feature labels; the feature labels are health evaluations of the health characteristics; The feature labels are combined to form a feature label library, which consists of a physical health label group, a mental health label group and an environmental label group. Each label group contains multiple feature labels with different feature indicators, and the same feature indicator contains multiple feature labels with different values. The physical health features and mental health features extracted from all historical data form a health feature library, which is divided into a probability model set and a health feature set. The probability model set is divided into a probability training set and a probability test set.

3. The method for constructing an occupational health analysis model based on big data according to claim 1 is characterized in that: Methods for obtaining health risk probabilities include: Define that the first probability training set and the first probability test set only contain physical health features, the second probability training set and the second probability test set only contain mental health features, and the third probability training set and the third probability test set contain both physical health features and mental health features; Gradient boosting tree, multi-layer perceptron and convolutional neural network combined with long short-term memory network are used as base models to construct multimodal machine learning models; The first probability training set is input into the gradient boosting tree, the second probability training set is input into the multilayer perceptron, and the third probability training set is input into the convolutional neural network combined with the long short-term memory network for independent training. The parameters of the base model are optimized using k-fold cross validation. After the training is completed, the first probability test set, the second probability test set, and the third probability test set are used to evaluate the performance of the base model respectively. The weighted average method is used as an integration strategy to combine the prediction results of the three base models to calculate the final prediction probability, and output the multimodal machine learning model. The health feature set is input into the multimodal machine learning model to predict the health risk probability.

4. The method for constructing an occupational health analysis model based on big data according to claim 1, characterized in that: The method for constructing a feature transformation matrix according to the interaction data and the health feature comprises: When the health risk probability is less than the preset probability threshold, it is judged as occupational health; When the health risk probability is greater than a preset probability threshold, it is determined that there is an occupational health risk, the health characteristics are interacted with the environmental database composed of the environmental data according to the personal information to obtain interactive data, the interactive data is preprocessed, and random forest is used to extract features from the interactive data to obtain environmental features, and multiple feature labels are divided according to the value of each environmental feature, and the feature labels are marked according to the value of the environmental feature; the interactive data includes working environment temperature, environmental quality, working hours, salary level, occupational hazard probability and occupational injury rate; Physical health features, mental health features, and environmental features are combined into feature samples to define the original training matrix It means that the sample data matrix X contains n samples and d features. By adding the regularized sparse regression framework, the non-negative matrix decomposition framework, retaining the local structure and constraining the similarity matrix, the objective function is constructed. The feature transformation matrix W is obtained by optimizing the objective function. The objective function expression is: in is the feature transformation matrix, m=c is the number of clusters, is the cluster center matrix, is the coefficient matrix, η≥0 is the regularization parameter, ||W|| 2,1 To utilize l 2,1 The norm constraint matrix W, the first and second terms are the regularized sparse regression framework; XP is the basis matrix, k is the cardinality, is the coefficient matrix, the third term is the non-negative matrix decomposition framework; α≥0 is the regular coefficient, is the auxiliary matrix, and the role of B is to ensure that T is orthogonal; L is the graph Laplacian matrix of the original data matrix, R = l k×k -l k , γ≥0, λ≥0 and ε≥0 are regularization parameters, the fifth and sixth terms are local preservation structures, and the seventh term is a constraint term to reduce the correlation of the coefficient matrix B; S is the similarity matrix, O is the initial similarity matrix, and β is the learning weight parameter of the model; The objective function involves the optimization of six variables (W, A, B, P, Q, S) in total, which makes the objective function non-convex. Fixed variables (A, B, P, Q, S) make the objective function convex, set the derivative of variable W to 0, and find the optimal solution to the problem to obtain the update formula of the variable feature transformation matrix W: W←(XX T +γXLX T +ηD)XBA T in is a diagonal matrix, and the matrix elements d ij =(1 / 2||W i ||2), ||W|| 2,1 The derivative of variable W is 2DW.

5. The method for constructing an occupational health analysis model based on big data according to claim 1, characterized in that: The method of inputting the interaction data and the mental health characteristics into a health function to obtain a health coefficient comprises: Align the environmental characteristics and mental health characteristics in the characteristic dimension, input the characteristic data into the health function to obtain the health coefficient, and the health function expression is: where δ i is the health coefficient of the i-th group of data, w1 and w2 are weight coefficients, S i is the similarity between the i-th group of environmental feature vectors and the mental health feature vectors, is the mean of the similarity of n groups of feature vectors, n is the number of feature vector groups, σ s is the standard deviation of the similarity of n groups of feature vectors.

6. The method for constructing an occupational health analysis model based on big data according to claim 1, characterized in that: The method of fusing the feature transformation matrix and the health coefficient to obtain a fusion feature includes: The health factor δ i According to the corresponding data group, add it to the i-th row of the feature transformation matrix W to form a fusion matrix m represents the number of data sets, (d+1) represents the row vector dimension, and the input vector is set to the row vector y i , the kernel function is K(·), the feature vector set is Y, and the output vector is u i , the optimal hyperplane expression for obtaining data is: Where u is the optimal hyperplane obtained, τ i is the weight coefficient of the i-th row vector, is the bias term; Based on the obtained optimal hyperplane, the fusion matrix data is divided into h groups. Different groups correspond to a new fusion feature. The actual data value is set to u, the training output is U, and the generalization parameter is χ. The health data fusion model of SVM is established. The expression is: Where J is the health data fusion model, l i and l j is the optimal solution marking form for data i and j, and is the optimal solution mean marking form; By determining the optimal solution of the characteristic data, the decision function of the characteristic data is established, the model is solved, and data fusion is realized. The expression is: Where f(x) is the constructed decision function, is the mean value set of h groups of data, l h Mark the data in the optimal solution format; Fusion matrix Y by row vector y i The input is decomposed into h groups of data, and m groups of fusion features with dimension h are obtained through the health data fusion model.

7. The method for constructing an occupational health analysis model based on big data according to claim 1, characterized in that: The method for constructing an occupational health analysis model based on big data according to the fusion features and historical labels includes: The fused features and feature labels are combined into a feature set, and the feature set is divided into an attention training set and an attention test set to build an occupational health analysis model. The specific structure includes an input layer, a feature preprocessing layer, a multi-head attention fusion layer, an adaptive feature fusion layer, and an output layer. The attention training set is input into the occupational health analysis model. The attention training set data is linearly transformed in the feature preprocessing layer to obtain a feature representation with unified dimension and form. The linearly transformed data is subjected to feature extraction and then input into the multi-head attention fusion layer. The APReLU activation function is used to adjust the weight of each head to generate weighted features. The weighted features and original features are input into the adaptive feature fusion layer. The features of different layers are integrated into the output layer using ASFF. Two fully connected layers are set in the output layer. One fully connected layer generates the occupational health score, and the other fully connected layer is followed by a softmax activation function to predict the feature label, and the health score and health evaluation are output. The mean square error loss and SGD optimizer are used to evaluate the model and optimize the weights to obtain the attention mechanism function. The attention test set is used to evaluate the attention mechanism function and output the attention mechanism model. The occupational health data to be analyzed is input into the occupational health analysis model to obtain the occupational health score and health evaluation.

Images