Grain storage temperature prediction method based on improved width learning model
By improving the width learning model, combining feature fusion and multi-head self-attention modules, the temperature of stored grains is predicted, which solves the problems of low accuracy and long training time in the existing technology, and achieves fast and accurate prediction of stored grain temperature.
Patent Information
- Application Number
- CN202411446885.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-16
- Publication Date
- 2025-01-21
AI Technical Summary
In the prior art, traditional machine learning accuracy is not high in grain storage temperature prediction, while deep learning models are long training time and require high-performance equipment, which is not suitable for practical applications.
Adopt the improvement of width learning model, integrate the food storage temperature data and humidity data through the characteristic fusion module, use the multi -head self -attention module to extract the time correlation between the enhanced node, and solve the connection weight through the splicing matrix to achieve fast and accurate accurate Forecast of grain storage temperature.
It achieves more accurate prediction of stored grain temperature, and the model training speed is fast, suitable for practical applications, and reduces update costs.
Smart Images
Figure CN119338061A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the field of grain storage temperature prediction, and in particular to a grain storage temperature prediction method based on an improved width learning model. Background Art
[0002] Every year, a considerable amount of stored grain suffers losses due to mildew and insect pests, among which mildew is the main cause. Extreme changes in grain storage temperature are the key factor leading to grain mildew. Therefore, knowing the changes in grain storage temperature in advance to guide the cooling of the granary can reduce the possibility of mildew in stored grain and ensure the safety of stored grain.
[0003] In the closed environment of the granary, the micro-airflow inside the grain pile will cause the temperature and humidity to diffuse. The size and distribution of temperature are affected by factors such as the temperature and humidity inside and outside the granary. In recent years, most of the research on grain storage temperature prediction has adopted traditional machine learning and deep learning methods. The accuracy of grain storage temperature prediction using traditional machine learning methods is not high and cannot meet daily needs. Deep learning methods can achieve more accurate grain storage temperature prediction, but the model training time is long and requires high-performance equipment training, which is not conducive to practical application. Summary of the invention
[0004] In order to solve the above problems, the purpose of the present invention is to provide a method for predicting the temperature of stored grain based on an improved width learning model, which can predict the temperature of stored grain more accurately and has a simple model training method.
[0005] Based on this, the present invention provides a method for predicting stored grain temperature based on an improved width learning model, the method comprising:
[0006] Collecting grain storage data, inputting the grain storage data into a preset grain storage temperature prediction model, and the grain storage temperature prediction model outputs a grain storage temperature prediction result;
[0007] The grain storage temperature prediction model is an improved width learning model. The acquisition process of the grain storage temperature prediction model includes: pre-processing the grain storage data to divide it into a training set and a training set; inputting the grain storage data in the training set into a feature fusion module to obtain a feature fusion result, performing feature mapping on the feature fusion result to obtain a feature node, performing an enhancement transformation on the feature node to obtain an enhanced node, and processing the enhanced node through a multi-head self-attention module to generate a new enhanced node; performing node splicing on the new enhanced node and the feature node to obtain a splicing matrix, and solving the connection weight according to the splicing matrix and the output layer;
[0008] The test set verifies the prediction effect of the grain storage temperature prediction model using the connection weights, and uses the verification result to determine whether to retrain the grain storage temperature prediction model.
[0009] Wherein, the training set and the training set divided after preprocessing the grain storage data include:
[0010] Removing abnormal points of the grain storage data and supplementing missed points of the grain storage data;
[0011] Dividing the grain storage data into a training set and a test set according to a timeline;
[0012] The stored grain data in the training set and the test set are normalized respectively.
[0013] Wherein, inputting the grain storage data in the training set into a feature fusion module to obtain a feature fusion result comprises:
[0014] Obtaining grain storage temperature data weight information features and grain storage humidity data weight information features corresponding to grain storage temperature data and grain storage humidity data, respectively, wherein the grain storage data includes: the grain storage temperature data and the grain storage humidity data;
[0015] The weight information feature of the grain storage temperature data and the weight information feature of the grain storage humidity data are feature spliced to obtain a splicing result;
[0016] The splicing results are processed to obtain the grain storage temperature and humidity data to be fused;
[0017] The stored grain temperature and humidity data to be fused are added to the stored grain data when input into the feature fusion module to obtain the output data of the feature fusion module.
[0018] The process of acquiring the grain storage temperature weight information feature includes:
[0019] The grain storage temperature data is sliced according to the timeline, and the grain storage temperature data after slicing is globally averaged pooled to obtain a global average pooling result. After the global average pooling result passes through the first fully connected layer, a rule activation function is used to obtain an activation result. The activation result passes through the second fully connected layer and is used as an input of a sigmoid function to obtain a weight vector. The weight vector is multiplied by the grain storage temperature data before slicing to obtain weight information features.
[0020] The splicing result is processed to obtain the stored grain temperature and humidity data to be fused, including:
[0021] The splicing result is sliced according to the timeline, and the splicing result after slicing is subjected to global average pooling and global maximum pooling respectively to obtain a global average pooling result and a global maximum pooling result, and the global average pooling result is added to the global maximum pooling result to obtain an addition result, and the addition result is used as the input of the sigmoid function, and the output of the sigmoid function passes through the third fully connected layer, and then uses the rule activation function to obtain an activation result, and the activation result passes through the fourth fully connected layer and is again used as the input of the sigmoid function to obtain a weight vector, and the weight vector is multiplied by the splicing result before slicing to obtain the stored grain temperature and humidity data to be fused.
[0022] Wherein, performing feature mapping on the feature fusion result to obtain feature nodes includes:
[0023]
[0024] Among them, the θ i is a linear or nonlinear activation function, and stated are randomly generated weights and biases, n is the number of feature nodes, X is the feature fusion result, and Z i is the characteristic node.
[0025] The step of performing an enhancement transformation on the characteristic node to obtain an enhanced node includes:
[0026]
[0027] Among them, the is a nonlinear activation function, and stated are randomly generated weights and biases, m is the number of feature nodes, and Z n is the characteristic node set, the H j is the enhanced node.
[0028] The step of processing the enhanced node by a multi-head self-attention module to generate a new enhanced node includes:
[0029]
[0030] MSA(H m ,H m ,H m )
[0031] =Concat(Self_attention1,…,Self_attention h )W'
[0032]
[0033] Among them, the H m is the set of enhanced nodes, the MSA is the multi-head self-attention mechanism function, the Norm is the normalization function, the Tanh is the tanh nonlinear activation function, and H m’ is a new enhanced node, the Self_attention is a self-attention mechanism function, Concat is a concatenation function that aggregates multiple self-attentions together, and the softmax is a softmax activation function. For the H m The transposed matrix of , m is the input augmented node dimension.
[0034] The number of self_attention is h, which can be set manually and can be 2, 6, ..., etc., as long as the enhanced node latitude m is divisible by h.
[0035] The step of performing node splicing on the new enhanced node and the characteristic node to obtain a splicing matrix includes:
[0036] Y=[Z1,Z2,…Z n |H'1,H'2,…H' m ]W=[Z n |H m’ ]W=AW
[0037] The [Z1, Z2, ... Z n ] is the characteristic node set, and [H'1,H'2,…H' m ] is the new enhanced node set, and the [Z n |H m’ ] That is, A is the concatenation matrix, and W is the connection weight.
[0038] Wherein, solving the connection weight according to the splicing matrix and the output layer includes:
[0039] W=A + Y
[0040]
[0041] Wherein, λ is the regularization coefficient, I is the unit matrix, Y is the true value matrix of grain storage temperature, and A T is the transposed matrix of A, + is the pseudo-inverse matrix of A.
[0042] The test set verifies the prediction effect of the grain storage temperature prediction model using the connection weights, and uses the verification result to determine whether to retrain the grain storage temperature prediction model, including:
[0043] Obtain the predicted value of grain storage temperature and the actual value of grain storage temperature;
[0044] The error value between the predicted value of the grain storage temperature and the true value of the grain storage temperature is determined by using a mean absolute error function, a root mean square error function and a mean absolute percentage error function;
[0045] Whether to retrain the grain storage temperature prediction model is determined based on the error value.
[0046] In the acquisition process, i.e., the training process, of the grain storage temperature prediction model of the present invention, firstly, the abnormal data of the grain storage data is removed and the missing data is supplemented to enhance the trainability of the data, and the grain storage data is divided into a training set and a test set. For the training set, firstly, the grain storage temperature data and the grain storage humidity data are effectively fused by using a feature fusion module, and secondly, the time correlation between the enhanced nodes is extracted by using a multi-head self-attention module, and finally, the mapped feature nodes are spliced with the new enhanced nodes, and the connection weights of the grain storage temperature prediction model are solved according to the splicing matrix and the output layer to obtain a better grain storage temperature prediction result, and the grain storage temperature is quickly predicted using the model. The grain storage temperature prediction model of the present invention has a fast training speed and a low updating cost, and provides a very feasible method for predicting grain storage temperature. BRIEF DESCRIPTION OF THE DRAWINGS
[0047] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0048] Figure 1 is a flow chart of a process for acquiring a grain storage temperature prediction model provided by an embodiment of the present invention;
[0049] Figure 2 is a flow chart of a test process of a grain storage temperature prediction model provided by an embodiment of the present invention;
[0050] Figure 3 is a schematic diagram of a multi-head self-attention module provided by an embodiment of the present invention;
[0051] Figure 4is a flow chart of inputting the grain storage data in the training set into a feature fusion module to obtain a feature fusion result provided by an embodiment of the present invention;
[0052] Figure 5 A 30-day processing temperature prediction experiment was conducted for the trained grain storage temperature prediction model. DETAILED DESCRIPTION
[0053] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0054] An embodiment of the present invention provides a method for predicting stored grain temperature based on an improved width learning model, the method comprising:
[0055] Collecting grain storage data, inputting the grain storage data into a preset grain storage temperature prediction model, and the grain storage temperature prediction model outputs a grain storage temperature prediction result;
[0056] The grain storage temperature prediction model is an improved width learning model. Figure 1 is a schematic diagram of a process of acquiring a grain storage temperature prediction model provided by an embodiment of the present invention, wherein the process of acquiring the grain storage temperature prediction model includes:
[0057] S101. After preprocessing the grain storage data, the training set and the training set are divided;
[0058] Wherein, the training set and the training set divided after preprocessing the grain storage data include:
[0059] Removing abnormal points of the grain storage data and supplementing missed points of the grain storage data;
[0060] Dividing the grain storage data into a training set and a test set according to a timeline;
[0061] The stored grain data in the training set and the test set are normalized respectively.
[0062] The grain storage data includes relevant grain storage temperature and humidity data, which are collected by time. After removing the abnormal points of the grain storage data and supplementing the missed points of the grain storage data, the first row corresponds to the first day's data, the second row corresponds to the second day's data, and so on; the grain storage data is set to 720 days of data according to the timeline, which can be divided into the first 690 lines for training and the last 30 lines for testing. It can also be directly divided into training sets and test sets according to the required ratio, such as 23:1.
[0063] The grain storage data includes: obtaining the temperature of the grain stored in the granary, the humidity of the grain stored in the granary, the temperature inside the granary, the humidity inside the granary, and the temperature outside the granary, the humidity outside the granary;
[0064] Interpolation method can be used to supplement the missed points of the grain storage data.
[0065] The normalization method is the range normalization method.
[0066] S102. Input the grain storage data in the training set into a feature fusion module to obtain a feature fusion result, perform feature mapping on the feature fusion result to obtain a feature node, perform enhancement transformation on the feature node to obtain an enhanced node, process the enhanced node through a multi-head self-attention module to generate a new enhanced node; perform node splicing on the new enhanced node and the feature node to obtain a splicing matrix, and solve the connection weight according to the splicing matrix and the output layer;
[0067] S103. The test set verifies the prediction effect of the grain storage temperature prediction model using the connection weights, and uses the verification result to determine whether to retrain the grain storage temperature prediction model.
[0068] The test set verifies the prediction effect of the grain storage temperature prediction model using the connection weights, and uses the verification result to determine whether to retrain the grain storage temperature prediction model, including:
[0069] Obtain the predicted value of grain storage temperature and the actual value of grain storage temperature;
[0070] The error value between the predicted value of the grain storage temperature and the true value of the grain storage temperature is determined by using a mean absolute error function, a root mean square error function and a mean absolute percentage error function;
[0071] Whether to retrain the grain storage temperature prediction model is determined based on the error value.
[0072] Determining whether to retrain the stored grain temperature prediction model according to the error value specifically includes:
[0073] Determine whether the error value is within a preset error allowable range;
[0074] If not, the grain storage temperature prediction model is retrained.
[0075] The preset error allowable range can be set independently.
[0076] Figure 2 : is a flow chart of a test process of a grain storage temperature prediction model provided by an embodiment of the present invention, wherein the test process includes:
[0077] S201. Inputting the grain storage data in the training set into a feature fusion module to obtain a feature fusion result, and performing feature mapping on the feature fusion result to obtain a feature node;
[0078] Wherein, performing feature mapping on the feature fusion result to obtain feature nodes includes:
[0079]
[0080] Among them, the θ i is a linear or nonlinear activation function, and stated are randomly generated weights and biases, n is the number of feature nodes, X is the feature fusion result, and Z i is the characteristic node.
[0081] S202. Performing an enhancement transformation on the feature node to obtain an enhanced node, and processing the enhanced node through a multi-head self-attention module to generate a new enhanced node;
[0082] The step of performing an enhancement transformation on the characteristic node to obtain an enhanced node includes:
[0083]
[0084] Among them, the is a nonlinear activation function, and stated are randomly generated weights and biases, m is the number of feature nodes, and Z n is the characteristic node set, the H j is the enhanced node.
[0085] Figure 3 is a schematic diagram of a multi-head self-attention module provided in an embodiment of the present invention, and processing the enhanced node by the multi-head self-attention module to generate a new enhanced node includes:
[0086]
[0087] MSA(H m ,H m ,H m )
[0088] =Concat(Self_attention1,…,Self_attention h )W'
[0089]
[0090] Among them, the H m is the set of enhanced nodes, the MSA is the multi-head self-attention mechanism function, the Norm is the normalization function, the Tanh is the tanh nonlinear activation function, and H m’ is a new enhanced node, the Self_attention is a self-attention mechanism function, the Concat is a concatenation function that aggregates multiple self-attentions together, and the softmax is a softmax activation function. For the H m The transposed matrix of , m is the dimension of the input enhancement node.
[0091] S203. Node splicing of the new enhanced node and the characteristic node to obtain a splicing matrix;
[0092] The step of performing node splicing on the new enhanced node and the characteristic node to obtain a splicing matrix includes:
[0093] Y=[Z1,Z2,…Z n |H'1,H'2,…H' m ]W=[Z n |H m’ ]W=AW
[0094] The [Z1, Z2, ... Z n ] is the characteristic node set, and HH'1, H'2, ... H' m ] is the new enhanced node set, and the [Z n |H m’ ] That is, A is the concatenation matrix, and W is the connection weight.
[0095] S204. Solve the connection weight according to the concatenation matrix and the output layer.
[0096] Wherein, solving the connection weight according to the splicing matrix and the output layer includes:
[0097] W=A + Y
[0098]
[0099] Wherein, λ is the regularization coefficient, I is the unit matrix, Y is the true value matrix of grain storage temperature, and A T is the transposed matrix of A, + is the pseudo-inverse matrix of A.
[0100] The number of rows in the grain storage temperature true value matrix is strictly equal to the number of rows in the input grain storage temperature data. The number of columns can be adjusted according to the data construction requirements (depending on which setting will have higher prediction accuracy), but each column must be consistent, such as [known row, 1 column] or [known row, 30 columns], etc.
[0101] Figure 4 The present invention provides a flowchart of inputting the grain storage data in the training set into a feature fusion module to obtain a feature fusion result, wherein the grain storage data in the training set is input into a feature fusion module to obtain a feature fusion result, including:
[0102] S401. Obtaining grain storage temperature data weight information characteristics and grain storage humidity data weight information characteristics corresponding to grain storage temperature data and grain storage humidity data, respectively, wherein the grain storage data includes: the grain storage temperature data and the grain storage humidity data;
[0103] The process of acquiring the grain storage temperature weight information feature includes:
[0104] The grain storage temperature data is sliced according to the timeline, and the grain storage temperature data after slicing is globally averaged pooled to obtain a global average pooling result. After the global average pooling result passes through the first fully connected layer, a rule activation function is used to obtain an activation result. The activation result passes through the second fully connected layer and is used as an input of a sigmoid function to obtain a weight vector. The weight vector is multiplied by the grain storage temperature data before slicing to obtain weight information features.
[0105] The process of obtaining the weight information feature can be expressed as follows:
[0106]
[0107] Wherein, X is the grain storage temperature data before slicing, AvgPool is the global average pooling, Fc1 is the first fully connected layer, relu is the rule activation function, Fc2 is the second fully connected layer, sigmoid is the sigmoid function, is multiplication, and F(X) is the weight information feature.
[0108] When X=X T That is, when the grain storage temperature data is used, what is obtained is the grain storage temperature weight information feature F(X T ).
[0109] When X=X H That is, when the grain humidity data is stored, what is obtained is the grain humidity weight information feature F(X H ).
[0110] The process of acquiring the grain storage humidity weight information feature is the same as the process of acquiring the grain storage temperature weight information feature, and it is only necessary to replace the grain storage temperature data with the grain storage humidity data.
[0111] The grain storage temperature data is sliced according to the timeline. The grain storage data is also sliced by row corresponding to the number of days, that is, the first row corresponds to the grain storage data of the first day, the second row corresponds to the grain storage data of the second day, and so on. The grain storage data is set to 10 for the 600-day data dimension according to the timeline, and can be directly divided according to the required ratio (such as the required number of slices is 20), that is, 600 divided by 20 equals 30, 30 small blocks, each of which is a (row, column) = (30, 10) (that is, equal slicing, but the data dimension remains unchanged), and the division needs to meet the timeline relationship condition. In this way, the time relationship of the data can be considered.
[0112] S402. Feature splicing of the grain storage temperature data weight information feature and the grain storage humidity data weight information feature to obtain a splicing result;
[0113] The process of feature splicing can be expressed as follows:
[0114] X TH =Concat(F(X T ),F(X H ))
[0115] Among them, X TH is the concatenation result, and Concat is the feature concatenation function.
[0116] S403. Processing the splicing results to obtain the grain storage temperature and humidity data to be integrated;
[0117] The splicing result is processed to obtain the stored grain temperature and humidity data to be fused, including:
[0118] The splicing result is sliced according to the timeline, and the splicing result after slicing is subjected to global average pooling and global maximum pooling respectively to obtain a global average pooling result and a global maximum pooling result, and the global average pooling result is added to the global maximum pooling result to obtain an addition result, and the addition result is used as the input of the sigmoid function, and the output of the sigmoid function passes through the third fully connected layer, and then uses the rule activation function to obtain an activation result, and the activation result passes through the fourth fully connected layer and is again used as the input of the sigmoid function to obtain a weight vector, and the weight vector is multiplied by the splicing result before slicing to obtain the stored grain temperature and humidity data to be fused.
[0119] The process of acquiring the stored grain temperature and humidity data to be fused can be expressed as follows:
[0120]
[0121] X TH is the splicing result before slicing, AvgPool is the global average pooling, MaxPool is the global maximum pooling, Fc3 is the third fully connected layer, relu is the rule activation function, Fc4 is the fourth fully connected layer, sigmoid is the sigmoid function, For multiplication, is addition, and X′ is the stored grain temperature and humidity data to be integrated.
[0122] S404. Add the stored grain temperature and humidity data to be fused and the stored grain data when input into the feature fusion module to obtain output data of the feature fusion module.
[0123] The process of obtaining the output data of the feature fusion module can be expressed as follows:
[0124]
[0125] Among them, X' is the stored grain temperature and humidity data to be fused, and X is the stored grain data when input into the feature fusion module. is addition, X * The output data is the feature fusion result.
[0126] Figure 5 The trained grain storage temperature prediction model was subjected to a 30-day processing temperature prediction experiment. The blue line is the actual observed value (the true value of the grain storage temperature data), the green line is the grain storage temperature prediction value of the trained width learning model, and the red line is the grain storage temperature prediction value of the trained improved width learning model. The MAPE (mean absolute percentage error) between the grain storage temperature prediction value of the improved width learning model and the corresponding grain storage temperature true value is 1.12%, MAE (mean absolute error) is 0.31, and RMSE (root mean square error) is 0.33. Compared with the MAPE of 3.72%, MAE of 0.84, and RMSE of 0.86 between the grain storage temperature prediction value of the width learning model and the corresponding grain storage temperature true value, the overall error is reduced by more than half. It can be seen that the accuracy of the improved grain storage temperature prediction model has been improved. At the same time, the training time does not exceed 10 seconds. The grain storage temperature prediction model can accurately predict the change of grain storage temperature through rapid training.
[0127] Experimental equipment configuration: Windows 10, 64-bit operating system, Intel(R) core(TM) i7-10700 CPU, 64GB RAM, Python 3.10 and Pytorch 2.0 operating environment.
[0128] In the acquisition process, i.e., the training process, of the grain storage temperature prediction model of the present invention, firstly, the abnormal data of the grain storage data is removed and the missing data is supplemented to enhance the trainability of the data, and the grain storage data is divided into a training set and a test set. For the training set, firstly, the grain storage temperature data and the grain storage humidity data are effectively fused by using a feature fusion module, and secondly, the time correlation between the enhanced nodes is extracted by using a multi-head self-attention module, and finally, the mapped feature nodes are spliced with the new enhanced nodes, and the connection weights of the grain storage temperature prediction model are solved according to the splicing matrix and the output layer to obtain a better grain storage temperature prediction result, and the grain storage temperature is quickly predicted using the model. The grain storage temperature prediction model of the present invention has a fast training speed and a low updating cost, and provides a very feasible method for predicting grain storage temperature.
[0129] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and substitutions can be made without departing from the technical principles of the present invention. These improvements and substitutions should also be regarded as the scope of protection of the present invention.
Claims
1. A method for predicting stored grain temperature based on an improved width learning model, characterized in that: include: Collecting grain storage data, inputting the grain storage data into a preset grain storage temperature prediction model, and the grain storage temperature prediction model outputs a grain storage temperature prediction result; The grain storage temperature prediction model is an improved width learning model. The acquisition process of the grain storage temperature prediction model includes: pre-processing the grain storage data to divide it into a training set and a test set; inputting the grain storage data in the training set into a feature fusion module to obtain a feature fusion result, performing feature mapping on the feature fusion result to obtain a feature node, performing enhancement transformation on the feature node to obtain an enhanced node, and processing the enhanced node through a multi-head self-attention module to generate a new enhanced node; performing node splicing on the new enhanced node and the feature node to obtain a splicing matrix, and solving the connection weight according to the splicing matrix and the output layer, wherein the training set and the training set divided after pre-processing the grain storage data include: removing abnormal points of the grain storage data and supplementing missed points of the grain storage data; dividing the grain storage data into a training set and a test set according to the timeline; normalizing the grain storage data in the training set and the test set respectively; The test set verifies the prediction effect of the grain storage temperature prediction model using the connection weights, and uses the verification result to determine whether to retrain the grain storage temperature prediction model.
2. The method for predicting stored grain temperature based on the improved width learning model according to claim 1, characterized in that: Inputting the stored grain data in the training set into a feature fusion module to obtain a feature fusion result includes: Obtaining grain storage temperature data weight information features and grain storage humidity data weight information features corresponding to grain storage temperature data and grain storage humidity data, respectively, wherein the grain storage data includes: the grain storage temperature data and the grain storage humidity data; The weight information feature of the grain storage temperature data and the weight information feature of the grain storage humidity data are feature spliced to obtain a splicing result; The splicing results are processed to obtain the grain storage temperature and humidity data to be fused; The grain storage temperature and humidity data to be fused are added to the grain storage data when input into the feature fusion module to obtain the output data of the feature fusion module, i.e., the feature fusion result.
3. The method for predicting stored grain temperature based on the improved width learning model according to claim 2, characterized in that: The process of acquiring the grain storage temperature weight information feature includes: The grain storage temperature data is sliced according to the timeline, and the grain storage temperature data after slicing is globally averaged pooled to obtain a global average pooling result. After the global average pooling result passes through the first fully connected layer, a rule activation function is used to obtain an activation result. The activation result passes through the second fully connected layer and is used as an input of a sigmoid function to obtain a weight vector. The weight vector is multiplied by the grain storage temperature data before slicing to obtain weight information features.
4. The method for predicting stored grain temperature based on the improved width learning model according to claim 2, characterized in that: After processing the splicing results, the stored grain temperature and humidity data to be fused include: The splicing result is sliced according to the timeline, and the splicing result after slicing is subjected to global average pooling and global maximum pooling respectively to obtain a global average pooling result and a global maximum pooling result, and the global average pooling result is added to the global maximum pooling result to obtain an addition result, and the addition result is used as the input of the sigmoid function, and the output of the sigmoid function passes through the third fully connected layer, and then uses the rule activation function to obtain an activation result, and the activation result passes through the fourth fully connected layer and is again used as the input of the sigmoid function to obtain a weight vector, and the weight vector is multiplied by the splicing result before slicing to obtain the stored grain temperature and humidity data to be fused.
5. The method for predicting stored grain temperature based on the improved width learning model according to claim 1, characterized in that: Performing feature mapping on the feature fusion result to obtain feature nodes includes: Among them, the is a linear or nonlinear activation function, and stated are randomly generated weights and biases, n is the number of feature nodes, X is the feature fusion result, and Z i is the characteristic node.
6. The method for predicting stored grain temperature based on the improved width learning model according to claim 1, characterized in that: Performing an enhancement transformation on the characteristic node to obtain an enhanced node includes: Among them, the is a nonlinear activation function, and stated are randomly generated weights and biases, m is the number of feature nodes, and Z n is the characteristic node set, the H j is the enhanced node.
7. The method for predicting stored grain temperature based on the improved width learning model according to claim 1, characterized in that: Processing the enhanced node by a multi-head self-attention module to generate a new enhanced node includes: Among them, the H m is the set of enhanced nodes, the MSA is the multi-head self-attention mechanism function, the Norm is the normalization function, the Tanh is the tanh nonlinear activation function, and H m’ For the newly enhanced node, the Self attention is a self-attention mechanism function, Concat is a concatenation function that aggregates multiple self-attentions together, and softmax is a softmax activation function. For the H m The transposed matrix of , m is the input augmented node dimension.
8. The method for predicting stored grain temperature based on the improved width learning model according to claim 1, characterized in that: Performing node splicing on the new enhanced node and the characteristic node to obtain a splicing matrix includes: Y=[Z1,Z2,…Z n [H'1,H'2,…H' m ]W=[Z n |H m’ ]W=AW The [Z1, Z2, ... Z n ] is the characteristic node set, and [H'1,H' 2, …H' m ] is the new enhanced node set, and the [Z n |H m’ ] That is, A is the concatenation matrix, and W is the connection weight.
9. The method for predicting stored grain temperature based on the improved width learning model according to claim 1, characterized in that: Solving the connection weights based on the concatenation matrix and the output layer includes: W=A J Y Wherein, λ is the regularization coefficient, I is the unit matrix, Y is the true value matrix of grain storage temperature, and A 8 is the transposed matrix of A, J is the pseudo-inverse matrix of A.
10. The method for predicting stored grain temperature based on improved width learning model according to claim 1, characterized in that: The test set verifies the prediction effect of the grain storage temperature prediction model using the connection weights, and verifies the result to determine whether to retrain the grain storage temperature prediction model, including: Obtain the predicted value of grain storage temperature and the actual value of grain storage temperature; The error value between the predicted value of the grain storage temperature and the true value of the grain storage temperature is determined by using a mean absolute error function, a root mean square error function and a mean absolute percentage error function; Whether to retrain the grain storage temperature prediction model is determined based on the error value.