A method and system for predicting multi-size dust concentration in factory workshops based on adaptive spatiotemporal fusion network

Abstract

The present invention relates to a method and system for predicting multi-particle size dust concentration in a factory workshop based on an adaptive spatiotemporal fusion network, including: preprocessing the collected workshop multi-particle size dust concentration data, obtaining a spatial feature matrix through a graph neural network, and extracting multi-scale features from time series data using a multi-scale convolutional neural network to obtain a temporal feature matrix; integrating the extracted spatial feature matrix and temporal feature matrix through a self-attention mechanism, outputting a feature matrix that integrates temporal and spatial information, jointly optimizing dust concentration prediction tasks of different particle sizes under a multi-task learning framework, and outputting a predicted value of the time step of each particle size dust based on the integrated spatiotemporal features and the multi-task learning framework, and evaluating the prediction results through evaluation indicators. The present invention is suitable for predicting multi-particle size dust concentration in a complex environment of a factory workshop, can effectively improve prediction accuracy, and has certain industrial application value.

Description

Technical Field

[0001] The present invention relates to the technical field of dust concentration prediction, and in particular to a method and system for predicting multi-particle size dust concentration in a factory workshop based on an adaptive spatiotemporal fusion network. Background Art

[0002] With the rapid development of industrialization, the problem of dust pollution in factory workshops is becoming increasingly serious. Dust particles suspended in the air not only pose a serious threat to workers' health, but also have a significant impact on the equipment in the workshop. Dust particles of different particle sizes have significantly different effects on equipment and production processes due to their differences in physical properties. Large-sized particles are easily deposited on the surface of equipment, causing equipment wear, blockage, and even failure, while small-sized particles are more likely to enter the interior of the equipment, affecting its precision and service life. Therefore, accurately predicting the concentration of dust of each particle size and taking corresponding protective measures are of great significance to ensuring the normal operation of equipment, reducing maintenance costs, and extending the life of equipment.

[0003] Most of the current dust concentration prediction methods focus on the prediction of dust of a single particle size, ignoring the potential correlation between dusts of different particle sizes. These methods usually rely on traditional linear models, such as the autoregressive moving average model (ARIMA) and linear regression models, which are effective in dealing with simple time series problems, but are obviously insufficient when dealing with complex nonlinear dust concentration data in factory workshop environments. At the same time, when analyzing dust concentration data, existing methods often ignore the deep-level features in the time series and the spatial correlation between particle sizes, and are unable to fully capture the comprehensive impact of dust on equipment and workshop environment. Summary of the invention

[0004] In view of the shortcomings of the prior art, the present invention provides a method for predicting multi-size dust concentration in factory workshops based on an adaptive spatiotemporal fusion network to solve the problem that the prior art fails to fully utilize the spatiotemporal correlation in multi-size dust concentration data, thereby significantly improving the prediction accuracy of dust concentrations of various particle sizes. The present invention adopts a variety of advanced deep learning technologies and systematically integrates spatiotemporal features to achieve accurate prediction of dust concentration in factory workshop environments, providing a scientific basis for environmental monitoring and equipment maintenance.

[0005] To solve the above technical problems, the present invention provides the following technical solutions: a method for predicting multi-particle dust concentration in a factory workshop based on an adaptive spatiotemporal fusion network, comprising the following steps:

[0006] S1. Preprocess the collected workshop multi-particle dust concentration data, and output it as the processed time series data X to ensure data quality;

[0007] S2. Model the spatial correlation between dust particle size data through graph neural network, capture the spatial relationship between different particle sizes, and obtain the spatial feature matrix H;

[0008] S3, using a multi-scale convolutional neural network to extract multi-scale features of the time series data X, capturing trends and changes at different time scales, and obtaining a time feature matrix F;

[0009] S4, integrate the extracted spatial feature matrix H and temporal feature matrix F through the self-attention mechanism, and finally output the feature matrix Z that integrates temporal and spatial information, thereby improving the global dependency capture capability of the model;

[0010] S5. Under the multi-task learning framework, the dust concentration prediction tasks of different particle sizes are jointly optimized to improve the overall prediction accuracy;

[0011] S6. Based on the integrated spatiotemporal features Z and the multi-task learning framework, the model outputs the predicted value of each dust particle size in the future time step The prediction results are evaluated through evaluation indicators to ensure the accuracy and stability of the model.

[0012] Furthermore, in step S1, the specific process includes the following steps:

[0013] S11. Collect multi-size dust concentration data in indoor factory workshop environment and construct input data matrix x∈R N ×T×m , where N is the number of samples, T is the time step, and m represents the data of various dust particles with different particle sizes;

[0014] S12. Perform data cleaning, standardization, and missing value processing operations on the original input data matrix x. The standardization calculation formula is as follows:

[0015]

[0016] Among them, X norm is the standardized data, μ and σ are the mean and standard deviation respectively;

[0017] The output of data preprocessing is the processed time series data X, which is in the form of X = (X1, X2, ..., X N ).

[0018] Furthermore, in step S2, the specific process includes the following steps:

[0019] S21. Based on the processed time series data X, an adjacency matrix A representing the spatial correlation between a plurality of dust data with different particle sizes is constructed. The elements of the adjacency matrix A are defined as:

[0020]

[0021] Among them, x i and x j They represent the data characteristics of particle sizes i and j, respectively, and σ is the standard deviation;

[0022] S22. Build a spatial relationship model through graph neural network GNN and use the adjacency matrix A∈R 4×4 Update node features, and the node feature matrix is ​​represented by H (l) ∈R 4×d , the update formula is:

[0023] H (l+1) =σ(AH (l) W (l) + b);

[0024] Among them, W (l) ∈R d×d is the weight matrix of the lth layer, σ(·) is the activation function, and b is the bias term;

[0025] S23. Through multiple iterations of graph convolution operations, the node feature matrix H (l) It is updated step by step, and finally the spatial characteristic matrix H representing each dust particle size at each time step is obtained.

[0026] Furthermore, in step S3, the specific process includes the following steps:

[0027] S31. Use multi-scale convolutional neural network MSCNN to extract time series features and define convolution kernels of different scales. Convolution operations are performed on different time windows k = 3, 5, 7 to extract short-term, medium-term and long-term time series features. The convolution operation is defined as:

[0028] F i =ReLU(X*K i +b i );

[0029] Among them, F i is a multi-scale feature map, k i is the size of the i-th convolution kernel, b i is the bias, ReLU(·) is the activation function;

[0030] S32, multi-scale feature map F i Perform splicing and fusion to form the final time series feature matrix F. The operation is as follows:

[0031] F=Concat(F1,F2,…,F i );

[0032] Among them, Concat represents the feature concatenation operation.

[0033] Furthermore, in step S2, the time feature matrix F is reduced in dimension, and the feature matrix after dimension reduction is d reduced is the feature length after dimensioning, N is the number of samples, and T is the time step.

[0034] Furthermore, in step S4, the specific process includes the following steps:

[0035] S41. Introduce the self-attention mechanism to integrate the spatial feature matrix H and the temporal feature matrix F by calculating the global dependency. The calculation formula of the self-attention mechanism is:

[0036]

[0037] Among them, Q, K, and V represent the query matrix, key matrix, and value matrix respectively. k is the dimension of the key;

[0038] S42. Use the multi-head attention mechanism in Transformer to fuse spatiotemporal features and project the spatial feature matrix H and the temporal feature matrix V into Q, K, and V uniformly. For each head i , calculate self-attention:

[0039]

[0040] in, is the trainable parameter matrix;

[0041] S43, all the heads i The output of is concatenated to capture the global dependency, and finally output the feature matrix Z that integrates time and space information. The multi-head attention calculation formula is:

[0042] Z=MultiHead(Q,K,V)=Concat(head1,…,head h )W O ;

[0043] Where h is the number of heads, W O is the output weight matrix.

[0044] Furthermore, in step S5, the specific process includes the following steps:

[0045] S51. Construct a separate task loss function for the prediction task of various particle sizes It is defined as:

[0046]

[0047] in, is the predicted value of the i-th particle size in the n-th sample, y i (n) is the corresponding true value;

[0048] S52. By sharing the underlying feature representation and information transfer between tasks, the total loss function is defined as the weighted sum of the losses of each task. The mathematical expression of the total loss function is:

[0049]

[0050] in, is the loss function of the i-th particle size, α i is the corresponding weight coefficient;

[0051] S53. By optimizing the total loss function The model can simultaneously optimize the prediction accuracy of multiple particle sizes during the training process. During the training process, the total loss function is minimized by the gradient descent Adam optimizer. The process of gradient calculation is as follows:

[0052]

[0053] S54. By updating the model parameters, the loss function of each subtask is minimized, thereby improving the overall prediction performance.

[0054] Furthermore, in step S6, the prediction results are evaluated by evaluation indicators, specifically including: evaluating the prediction results by mean square error MSE and mean absolute error MAE.

[0055] Furthermore, the present invention also provides a system for predicting the concentration of multi-particle dust in a factory workshop based on an adaptive spatiotemporal fusion network, comprising:

[0056] Data processing module, the data acquisition module is used for the data space association module to pre-process the collected workshop multi-particle dust concentration data, and its output is the processed time series data X to ensure data quality;

[0057] The data spatial association module is used to model the spatial correlation between dust particle size data through a graph neural network, capture the spatial relationship between different particle sizes, and obtain the spatial feature matrix H;

[0058] A data multi-scale extraction module, wherein the data multi-scale extraction module uses a multi-scale convolutional neural network to perform multi-scale feature extraction on the time series data X, capture trends and changes at different time scales, and obtain a time feature matrix F;

[0059] A data integration module is used to integrate the extracted spatial feature matrix H and temporal feature matrix F through a self-attention mechanism, and finally output a feature matrix Z that integrates temporal and spatial information, thereby improving the global dependency capture capability of the model;

[0060] A prediction optimization module, which is used to jointly optimize dust concentration prediction tasks of different particle sizes under a multi-task learning framework to improve overall prediction accuracy;

[0061] The prediction and evaluation module is used to output the predicted value of each dust particle size in the future time step based on the integrated spatiotemporal feature Z and the multi-task learning framework. The prediction results are evaluated through evaluation indicators to ensure the accuracy and stability of the model.

[0062] By means of the above technical solution, the present invention provides a method for predicting multi-particle size dust concentration in a factory workshop based on an adaptive spatiotemporal fusion network, which has at least the following beneficial effects:

[0063] (1) The present invention can effectively capture the complex spatiotemporal relationship in dust concentration data by constructing an adaptive spatiotemporal fusion network and combining spatial correlation modeling with time series feature extraction, thus overcoming the limitation that traditional linear models are difficult to deal with nonlinear characteristics;

[0064] (2) The present invention introduces the graph neural network (GNN) to construct the spatial correlation between dust particles of various particle sizes, which can accurately reflect the interaction between different particle sizes, thereby improving the prediction accuracy. Traditional methods often ignore the correlation between multiple particle sizes, and the present invention makes up for this deficiency;

[0065] (3) The present invention jointly optimizes the prediction tasks of dust with different particle sizes through multi-task learning, which not only improves the prediction accuracy of each subtask, but also improves the overall performance by utilizing the correlation between different tasks. This method is more flexible and can handle the complexity of multi-particle size data;

[0066] (4) The present invention has certain application value in the field of dust concentration monitoring and prediction in industrial workshops. It can help enterprises monitor and predict dust concentration in real time, protect workers' health, extend equipment service life, reduce maintenance costs, and promote more environmentally friendly production practices. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative 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:

[0068] Figure 1This is a flow chart of a method for predicting multi-particle size dust concentration in a factory workshop based on an adaptive spatiotemporal fusion network according to the present invention;

[0069] Figure 2 This is a system framework diagram of a method for predicting multi-particle size dust concentration in a factory workshop based on an adaptive spatiotemporal fusion network according to the present invention.

[0070] In the figure: 100, data processing module; 200, data spatial association module; 300, data multi-scale extraction module; 400, data integration module; 500, prediction optimization module; 600, prediction evaluation module. DETAILED DESCRIPTION

[0071] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific implementation methods, so that the implementation process of how the present application uses technical means to solve technical problems and achieve technical effects can be fully understood and implemented accordingly.

[0072] Those skilled in the art can understand that all or part of the steps in the above-mentioned embodiment method can be completed by instructing the relevant hardware through a program, so the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0073] Please refer to Figure 1-Figure 2 , shows a specific implementation of this embodiment. This embodiment integrates the spatial feature matrix H and the temporal feature matrix F through a multi-task learning framework sharing layer, and finally outputs a feature matrix Z that integrates temporal and spatial information, thereby improving the model's global dependency capture capability; based on the integrated spatiotemporal feature Z and the multi-task learning framework, the model outputs the predicted value of each particle size of dust in the future time step The prediction results are evaluated through evaluation indicators to ensure the accuracy and stability of the model; through the systematic integration of spatiotemporal characteristics and a variety of advanced deep learning technologies, accurate prediction of dust concentration in factory workshop environments can be achieved, providing a scientific basis for environmental monitoring and equipment maintenance.

[0074] Please refer to Figure 1 This embodiment proposes a method for predicting multi-size dust concentration in a factory workshop based on an adaptive spatiotemporal fusion network, and the method includes the following steps:

[0075] S1. Preprocess the collected workshop multi-particle dust concentration data, and output it as the processed time series data X to ensure data quality;

[0076] As a preferred implementation of step S1, the specific process includes the following steps:

[0077] S11. Install sensors capable of collecting dust particles of four particle sizes in a factory workshop, collect dust concentration data of four particle sizes in the indoor factory workshop environment, and construct the input data matrix x∈R N×T×4 , where N is the number of samples, T is the time step, and 4 represents the data of 4 different dust particle sizes;

[0078] S12. Perform data cleaning, standardization, and missing value processing operations on the original input data matrix x. The standardization calculation formula is as follows:

[0079]

[0080] Among them, X norm is the standardized data, μ and σ are the mean and standard deviation respectively;

[0081] The output of data preprocessing is the processed time series data X, which is in the form of X = (X1, X2, ..., X N ).

[0082] S2. Model the spatial correlation between dust particle size data through graph neural network, capture the spatial relationship between different particle sizes, and obtain the spatial feature matrix H;

[0083] As a preferred implementation of step S2, the specific process includes the following steps:

[0084] S21. Based on the processed time series data X, an adjacency matrix A representing the spatial correlation between the four different dust particle size data is constructed. The elements of the adjacency matrix A are defined as:

[0085]

[0086] Among them, x i and x j They represent the data characteristics of particle sizes i and j, respectively, and σ is the standard deviation;

[0087] The adjacency matrix form is as follows:

[0088]

[0089] In which, each element a in the adjacency matrix ij It represents the degree of correlation between the i-th particle size and the j-th particle size. These values ​​are determined by the experience of domain experts or data-driven methods.

[0090] S22. Build a spatial relationship model through graph neural network GNN and use the adjacency matrix A∈R 4×4 Update node features, and the node feature matrix is ​​represented by H (l) ∈R 4×d , the update formula is:

[0091] H (l+1) =σ(AH (l) W (l) + b);

[0092] Among them, W (l) ∈R d×d is the weight matrix of the lth layer, σ(·) is the activation function, and b is the bias term;

[0093] S23. Through multiple iterations of graph convolution operations, the node feature matrix H (l) The spatial feature matrix H representing each dust particle size at each time step is finally obtained by gradual updating. The final spatial feature matrix H of each node can be used to characterize the spatial relationship between particle sizes and for subsequent spatiotemporal feature integration.

[0094] In this embodiment, the present invention introduces a graph neural network (GNN) to construct the spatial correlation between the four particle sizes of dust, which can accurately reflect the interaction between different particle sizes, thereby improving the prediction accuracy. Traditional methods often ignore the correlation between multiple particle sizes, and the present invention makes up for this deficiency.

[0095] S3, using a multi-scale convolutional neural network to extract multi-scale features of the time series data X, capturing trends and changes at different time scales, and obtaining a time feature matrix F;

[0096] As a preferred implementation of step S3, the specific process includes the following steps:

[0097] S31. Use multi-scale convolutional neural network MSCNN to extract time series features and define convolution kernels of different scales. Convolution operations are performed on different time windows k = 3, 5, 7 to extract short-term (3 hours), medium-term (5 hours) and long-term (7 hours) time series features. The convolution operation is defined as:

[0098] F i =ReLU(X*K i +b i );

[0099] Among them, F i is a multi-scale feature map, k i is the size of the i-th convolution kernel, b iis the bias, ReLU(·) is the activation function;

[0100] S32, multi-scale feature map F i Perform splicing and fusion to form the final time series feature matrix F. The operation is as follows:

[0101] F=Concat(F1,F2,…,F i );

[0102] Among them, Concat represents the feature concatenation operation.

[0103] More specifically, the time feature matrix F is reduced in dimension, and the feature matrix after dimension reduction is d reduced is the feature length after dimensioning, N is the number of samples, and T is the time step.

[0104] S4, integrate the extracted spatial feature matrix H and temporal feature matrix F through the self-attention mechanism, and finally output the feature matrix Z that integrates temporal and spatial information, thereby improving the global dependency capture capability of the model;

[0105] As a preferred implementation of step S4, the specific process includes the following steps:

[0106] S41. Introduce the self-attention mechanism to integrate the spatial feature matrix H and the temporal feature matrix F by calculating the global dependency. The calculation formula of the self-attention mechanism is:

[0107]

[0108] Among them, Q, K, and V represent the query matrix, key matrix, and value matrix respectively. k is the dimension of the key;

[0109] S42, using the multi-head attention mechanism in Transformer to fuse spatiotemporal features, thereby building a multi-task learning framework - shared layer, such as Figure 1 As shown, the spatial feature matrix H and the temporal feature matrix F are uniformly projected into Q, K, V. For each head i , calculate self-attention:

[0110]

[0111] in, is the trainable parameter matrix;

[0112] S43, all the heads i The output of is concatenated to capture the global dependency, and finally output the feature matrix Z that integrates time and space information. The multi-head attention calculation formula is:

[0113] Z=MultiHead(Q,K,V)=Concat(head1,…,head h )W O ;

[0114] Where h is the number of heads, W O is the output weight matrix.

[0115] In this embodiment, the present invention constructs an adaptive spatiotemporal fusion network and combines spatial correlation modeling with time series feature extraction to effectively capture the complex spatiotemporal relationship in dust concentration data, thereby overcoming the limitation that traditional linear models are difficult to cope with nonlinear characteristics.

[0116] S5. Under the multi-task learning framework, the dust concentration prediction tasks of four different particle sizes are jointly optimized to improve the overall prediction accuracy;

[0117] As a preferred implementation of step S5, the specific process includes the following steps:

[0118] S51. Construct separate task loss functions for the four particle size prediction tasks For example, the mean square error loss function MSE is defined as:

[0119]

[0120] in, is the predicted value of the i-th particle size in the n-th sample, y i (n) is the corresponding true value;

[0121] S52. By sharing the underlying feature representation and information transfer between tasks, the total loss function is defined as the weighted sum of the losses of each task. The mathematical expression of the total loss function is:

[0122]

[0123] in, is the loss function of the i-th particle size, α i is the corresponding weight coefficient;

[0124] S53. By optimizing the total loss function The model can simultaneously optimize the prediction accuracy of multiple particle sizes during the training process. During the training process, the total loss function is minimized by the gradient descent Adam optimizer. The process of gradient calculation is as follows:

[0125]

[0126] S54. By updating the model parameters, the loss function of each subtask is minimized, thereby improving the overall prediction performance.

[0127] In this embodiment, the present invention jointly optimizes the prediction tasks of dust of different particle sizes through multi-task learning, which not only improves the prediction accuracy of each subtask, but also improves the overall performance by utilizing the correlation between different tasks. This method is more flexible and can handle the complexity of multi-particle size data.

[0128] S6. Based on the integrated spatiotemporal features Z and the multi-task learning framework, a prediction layer is designed for each particle size, and a dedicated layer of the multi-task learning framework is built, such as Figure 1 As shown, the model outputs the predicted values ​​of four dust particle sizes (size 1, size 2, size 3, size 4) in the future time step. The prediction results are evaluated through evaluation indicators to ensure the accuracy and stability of the model.

[0129] More specifically, in step S6, the prediction results are evaluated by evaluation indicators, including: the prediction results are evaluated by mean square error MSE and mean absolute error MAE. The calculation formula of the evaluation indicators is as follows:

[0130]

[0131] in, Represents the predicted concentration value of the i-th particle size.

[0132] In this embodiment, the present invention has certain application value in the field of dust concentration monitoring and prediction in industrial workshops. It can help enterprises monitor and predict dust concentration in real time, protect workers' health, extend equipment life, reduce maintenance costs, and promote more environmentally friendly production practices.

[0133] The present invention also provides a system for predicting the concentration of multi-particle dust in a factory workshop based on an adaptive spatiotemporal fusion network. Figure 2 As shown, including:

[0134] The data processing module 100 is used by the data space association module to pre-process the collected workshop multi-particle dust concentration data, and its output is the processed time series data X to ensure data quality;

[0135] The data spatial association module 200 is used to model the spatial correlation between dust particle size data through a graph neural network, capture the spatial relationship between different particle sizes, and obtain a spatial feature matrix H;

[0136] A data multi-scale extraction module 300, wherein the data multi-scale extraction module 300 uses a multi-scale convolutional neural network to perform multi-scale feature extraction on the time series data X, capture trends and changes at different time scales, and obtain a time feature matrix F;

[0137] A data integration module 400, which is used to integrate the extracted spatial feature matrix H and temporal feature matrix F through a self-attention mechanism, and finally output a feature matrix Z that integrates temporal and spatial information, thereby improving the global dependency capture capability of the model;

[0138] A prediction optimization module 500, which is used to jointly optimize dust concentration prediction tasks of different particle sizes under a multi-task learning framework to improve overall prediction accuracy;

[0139] The prediction and evaluation module 600 is used to output the predicted value of each dust particle size in the future time step based on the integrated spatiotemporal feature Z and the multi-task learning framework. The prediction results are evaluated through evaluation indicators to ensure the accuracy and stability of the model.

[0140] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine different embodiments or examples described in this specification and the features of different embodiments or examples, unless they are contradictory.

[0141] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, which can be embodied in any computer-readable medium for use by an instruction execution system, apparatus or device (such as a computer-based system, a system including a processor or other system that can fetch instructions from an instruction execution system, apparatus or device and execute instructions), or used in combination with these instruction execution systems, apparatuses or devices.

[0142] The above implementation methods have been described in detail. Specific examples are used herein to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

Claims

1. A method for predicting multi-size dust concentration in factory workshops based on an adaptive spatiotemporal fusion network, characterized in that: The following steps are involved: S1. Preprocess the collected workshop multi-particle dust concentration data, and output it as the processed time series data X to ensure data quality; S2. Model the spatial correlation between dust particle size data through graph neural network, capture the spatial relationship between different particle sizes, and obtain the spatial feature matrix H; S21. Based on the processed time series data X, an adjacency matrix A representing the spatial correlation between a plurality of dust data with different particle sizes is constructed. The elements of the adjacency matrix A are defined as: Among them, x i and x j They represent the data characteristics of particle sizes i and j, respectively, and σ is the standard deviation; S22. Build a spatial relationship model through graph neural network GNN and use the adjacency matrix A∈R 4×4 Update node features, and the node feature matrix is ​​represented by H (l) ∈R 4×d , the update formula is: H (l+1) =σ(AH (l) W (l) +b); Among them, W (l) ∈R d×d is the weight matrix of the lth layer, σ(·) is the activation function, and b is the bias term; S23. Through multiple iterations of graph convolution operations, the node feature matrix H (l) Update gradually, and finally get the spatial characteristic matrix H representing each dust particle size at each time step; S3, using a multi-scale convolutional neural network to extract multi-scale features of the time series data X, capturing trends and changes at different time scales, and obtaining a time feature matrix F; S31. Use multi-scale convolutional neural network MSCNN to extract time series features and define convolution kernels of different scales. Convolution operations are performed on different time windows k = 3, 5, 7 to extract short-term, medium-term and long-term time series features. The convolution operation is defined as: F i =ReLU(X*K i +b i ); Among them, F i is a multi-scale feature map, k i is the size of the i-th convolution kernel, b i is the bias, ReLU(·) is the activation function; S32, multi-scale feature map F i Perform splicing and fusion to form the final time series feature matrix F. The operation is as follows: F=Concat(F1,F2,…,F i ); Among them, Concat represents the feature concatenation operation; S4, integrate the extracted spatial feature matrix H and temporal feature matrix F through the self-attention mechanism, and finally output the feature matrix Z that integrates temporal and spatial information, thereby improving the global dependency capture capability of the model; S41. Introduce the self-attention mechanism to integrate the spatial feature matrix H and the temporal feature matrix F by calculating the global dependency. The calculation formula of the self-attention mechanism is: Among them, Q, K, and V represent the query matrix, key matrix, and value matrix respectively. k is the dimension of the key; S42, using the multi-head attention mechanism in Transformer to fuse spatiotemporal features, the spatial feature matrix H and the temporal feature matrix F are uniformly projected into Q, K, V. For each head i , calculate self-attention: in, is the trainable parameter matrix; S43, all the heads i The output of is concatenated to capture the global dependency, and finally output the feature matrix Z that integrates time and space information. The multi-head attention calculation formula is: Z=MultiHead(Q,K,V)=Concat(head1,…,head h )W O ; Where h is the number of heads, W O is the output weight matrix; S5. Under the multi-task learning framework, the dust concentration prediction tasks of different particle sizes are jointly optimized to improve the overall prediction accuracy; S6. Based on the integrated spatiotemporal features Z and the multi-task learning framework, the model outputs the predicted value of each dust particle size in the future time step The prediction results are evaluated through evaluation indicators to ensure the accuracy and stability of the model.

2. According to claim 1, a method for predicting multi-particle size dust concentration in a factory workshop based on an adaptive spatiotemporal fusion network is characterized by: In step S1, the specific process includes the following steps: S11. Collect multi-size dust concentration data in indoor factory workshop environment and construct input data matrix x∈R N×T×m , where N is the number of samples, T is the time step, and m represents the data of various dust particles with different particle sizes; S12. Perform data cleaning, standardization, and missing value processing operations on the original input data matrix x. The standardization calculation formula is as follows: Among them, X norm is the standardized data, μ and σ are the mean and standard deviation respectively; The output of data preprocessing is the processed time series data X, which is in the form of X = (X1, X2, ..., X N ).

3. According to claim 1, a method for predicting multi-particle size dust concentration in a factory workshop based on an adaptive spatiotemporal fusion network is characterized by: In step S2, the time feature matrix F is dimensionally reduced, and the feature matrix after dimension reduction is d reduced is the feature length after dimensioning, N is the number of samples, and T is the time step.

4. According to claim 1, a method for predicting multi-size dust concentration in a factory workshop based on an adaptive spatiotemporal fusion network is characterized by: In step S5, the specific process includes the following steps: S51. Construct a separate task loss function for the prediction task of various particle sizes It is defined as: in, is the predicted value of the i-th particle size in the n-th sample, y i (n) is the corresponding true value; S52. By sharing the underlying feature representation and information transfer between tasks, the total loss function is defined as the weighted sum of the losses of each task. The mathematical expression of the total loss function is: in, is the loss function of the i-th particle size, α i is the corresponding weight coefficient; S53. By optimizing the total loss function The model can simultaneously optimize the prediction accuracy of multiple particle sizes during the training process. During the training process, the total loss function is minimized by the gradient descent Adam optimizer. The process of gradient calculation is as follows: S54. By updating the model parameters, the loss function of each subtask is minimized, thereby improving the overall prediction performance.

5. According to claim 1, a method for predicting multi-particle size dust concentration in a factory workshop based on an adaptive spatiotemporal fusion network is characterized by: In step S6, the prediction results are evaluated by evaluation indicators, specifically including: evaluating the prediction results by mean square error MSE and mean absolute error MAE.

6. A system for implementing the method for predicting multi-size dust concentration in a factory workshop based on an adaptive spatiotemporal fusion network as described in any one of claims 1 to 5, characterized in that: include: A data processing module (100), the data processing module (100) is used to pre-process the collected workshop multi-particle size dust concentration data, and its output is the processed time series data X to ensure data quality; A data space association module (200) is used to model the spatial correlation between dust particle size data through a graph neural network, capture the spatial relationship between different particle sizes, and obtain a spatial feature matrix H; S21, based on the processed time series data X, construct an adjacency matrix A representing the spatial correlation between a plurality of dust data with different particle sizes, wherein the elements of the adjacency matrix A are defined as: Among them, x i and x j They represent the data characteristics of particle sizes i and j, respectively, and σ is the standard deviation; S22. Build a spatial relationship model through graph neural network GNN and use the adjacency matrix A∈R 4×4 Update node features, and the node feature matrix is ​​represented by H (l) ∈R 4×d , the update formula is: H (l+1) =σ(AH (l) W (l) +b); Among them, W (l) ∈R d×d is the weight matrix of the lth layer, σ(·) is the activation function, and b is the bias term; S23. Through multiple iterations of graph convolution operations, the node feature matrix H (l) Update gradually, and finally obtain the spatial characteristic matrix H representing each dust particle size at each time step; A data multi-scale extraction module (300), wherein the data multi-scale extraction module (300) uses a multi-scale convolutional neural network to perform multi-scale feature extraction on the time series data X, captures trends and changes at different time scales, and obtains a time feature matrix F; S31, uses a multi-scale convolutional neural network MSCNN to extract time series features, and defines convolution kernels of different scales Convolution operations are performed on different time windows k = 3, 5, 7 to extract short-term, medium-term and long-term time series features. The convolution operation is defined as: F i =ReLU(X*K i +b i ); Among them, F i is a multi-scale feature map, k i is the size of the i-th convolution kernel, b i is the bias, RELU(·) is the activation function; S32, multi-scale feature map F i Perform splicing and fusion to form the final time series feature matrix F. The operation is as follows: F=Concat(F1,F2,…,F i ); Among them, Concat represents the feature concatenation operation; A data integration module (400) is used to integrate the extracted spatial feature matrix H and temporal feature matrix F through a self-attention mechanism, and finally output a feature matrix Z that integrates temporal and spatial information, thereby improving the global dependency capture capability of the model; S41, introducing a self-attention mechanism, integrating the spatial feature matrix H and the temporal feature matrix F by calculating the global dependency relationship, and the calculation formula of the self-attention mechanism is: Among them, Q, K, and V represent the query matrix, key matrix, and value matrix respectively. k is the dimension of the key; S42, using the multi-head attention mechanism in Transformer to fuse spatiotemporal features, the spatial feature matrix H and the temporal feature matrix F are uniformly projected into Q, K, V. For each head i , calculate self-attention: in, is the trainable parameter matrix; S43, all the heads i The output of is concatenated to capture the global dependency, and finally output the feature matrix Z that integrates time and space information. The multi-head attention calculation formula is: Z=MultiHead(Q,K,V)=Concat(head1,…,head h )W O ; Where h is the number of heads, W O is the output weight matrix; A prediction optimization module (500), wherein the prediction optimization module (500) is used to jointly optimize dust concentration prediction tasks of different particle sizes under a multi-task learning framework to improve overall prediction accuracy; A prediction and evaluation module (600), wherein the prediction and evaluation module (600) is used to output the predicted value of each dust particle size in the future time step based on the integrated spatiotemporal feature Z and the multi-task learning framework. The prediction results are evaluated through evaluation indicators to ensure the accuracy and stability of the model.

Images