Large model data enhancement method and device

Abstract

The invention relates to the technical field of large models, and particularly provides a large model data enhancement method and device, and the method comprises the following steps: S1, collecting original data through a sensor cluster, grouping the data through a Pearson's correlation coefficient, and for each group, determining a head sensor through employing a minimum spanning tree algorithm, and building a generation sequence; s2, starting a generator based on a long short-term memory (LSTM) network, and training data of a head sensor by using the LSTM generator; s3, starting a ridge regression generator, and training data of other sensors by using the ridge regression generator; and S4, generating corresponding data by using the model obtained by training according to a specified demand of a user, and increasing the data volume through interpolation. Compared with the prior art, the method has the advantages that inherent complex time behaviors in the industrial production process can be learned, the relation between sensors which is crucial to accurate simulation process conditions is kept, and it is ensured that the generated data mode has high fidelity while the original process characteristics are kept.

Description

Technical Field

[0001] The present invention relates to the technical field of large models, and specifically provides a large model data enhancement method and device. Background Art

[0002] Currently, large models are showing a development trend that uses basic large models as the technical foundation and industrial applications as the entry point. Compared to basic large models, industrial large models have advantages such as fewer parameters, higher degree of specialization, and stronger implementation. They can provide efficient solutions for technological breakthroughs, product innovation, and production changes in industrial verticals. The computational performance of industrial large models is highly dependent on the scale and quality of training data. In particular, in aspects such as model fine-tuning and search enhancement, they require the support of high-confidence and highly correlated industrial datasets. Although developers have proposed various methods for collecting industrial data using sensor clusters, many difficulties still exist in accurately replicating real-world production scenarios.

[0003] (1) There are many data noises and outliers; Industrial production environments are complex and ever-changing, and equipment operation generates various interferences and noises, resulting in a large amount of noise and outliers in the collected data. These noise and outliers can interfere with the learning of large models, reducing their accuracy and robustness.

[0004] (2) Missing and incomplete data; Data is often missing or incomplete due to equipment failures, sensor failures, communication interruptions, network delays, storage device failures, human errors, and imperfect data collection strategies. Missing data can prevent large models from fully learning the characteristics and patterns of the data, and can even cause them to malfunction.

[0005] (3) Data timeliness issues; Some data is highly time-sensitive, such as equipment operating status data, production process parameters, and product quality inspection data. If data is not updated in a timely manner, large models will make predictions and decisions based on outdated data, resulting in erroneous results.

[0006] In this context, how to fully utilize the actual data collected by sensor clusters to regenerate industrial data at each stage through the deep integration of data and mechanisms, and combine mechanisms such as expert knowledge and physical models to achieve dual-drive of data and knowledge, thereby improving the generalization and interpretability of generated data is an urgent problem to be solved in this field. Summary of the Invention

[0007] The present invention aims to address the deficiencies of the above-mentioned prior art and provides a highly practical large-scale model data enhancement method.

[0008] A further technical task of the present invention is to provide a large model data enhancement device with reasonable design, safety and applicability.

[0009] The technical solution adopted by the present invention to solve its technical problems is as follows: A large model data enhancement method has the following steps: S1. In the preprocessing stage, use a sensor cluster to collect raw data, group the data through the Pearson correlation coefficient, and for each group, use the minimum spanning tree algorithm to determine the head sensor and establish a generation sequence; S2. In the head sensor data training stage, start the generator based on the long short-term memory network LSTM, and use the LSTM generator to train the data of the head sensor; S3. In the non-head sensor data training stage, start the ridge regression generator, and use the ridge regression generator to train the data of other sensors; S4. In the data generation stage, according to the user-specified requirements, use the trained model to generate corresponding data, and increase the data volume through interpolation.

[0010] Further, in step S1, it further includes: S1-1. Use a sensor cluster to collect raw data, represent the time series data collected by each sensor in vector form. Assume that there are n sensors (from sensor 1 to sensor n) collecting the raw data set (from data 1 to data n, ), and these data are displayed in the Internet of Things data center in.csv format, and each data is represented as a vector with a length of m; S1-2. Calculate the Pearson correlation coefficient between any two vectors and , 1 ≤ i < j ≤ n, where represents the mean and variance of the vectors respectively, and finally form the Pearson matrix P; the value of the Pearson matrix P at the (i, j) position represents the Pearson correlation coefficient between the vectors and ; S1-3. Use the Pearson matrix P to calculate the Euclidean distance matrix D between n data, ; S1-4. Use multi-dimensional scaling analysis to obtain the coordinate matrix X corresponding to the reduced dimension of the raw data through the distance matrix D; S1-5. Based on the coordinate matrix X, set the threshold to 0.1, and cluster the data into different groups so that the data between different groups is almost uncorrelated; ​​S1-6. Use the Euclidean distance between each data as the weight to calculate the minimum spanning tree for each group. The sensor corresponding to the starting point of the minimum spanning tree is regarded as the head sensor, and the sensors corresponding to the non-starting points are regarded as non-head sensors.

[0011] Furthermore, in step S1-4, first, using the Euclidean distance matrix D, another new matrix B is calculated as follows, where each element Defined as: ; in, Indicates the matrix D The element at the position, and , , ; Then, the matrix B is decomposed to obtain its r largest eigenvalues and its corresponding eigenvector ; Finally, calculate the coordinate matrix X after dimensionality reduction, ,in, is a matrix containing r eigenvectors, is a diagonal matrix containing the r eigenvalues of B.

[0012] Furthermore, in step S1-6, the minimum spanning tree is established as follows: (1) Select a data from each group and initialize a tree; (2) Expand along an edge of the tree: Among the edges connecting the tree to the vertex that is not yet included, select the edge with the smallest weight and add it to the tree. The selection of the edge with the smallest weight means the smallest Euclidean distance. (3) Repeat the previous step until all vertices are included in the tree.

[0013] Furthermore, in step S2, it further includes: S2-1. To generate new head sensor data, it is necessary to use the previous time point data to predict the time point The data, the structure of the prediction function is defined as follows: ; in, Indicates a time point The predicted value of the function Used to generate predictions, the training goal is to optimize Parameters; S2-2, treat the data collected by the head sensor in a specific time period as a time series and construct the input matrix and the target matrix ; S2-3. Design LSTM neural network: Use Keras to design the LSTM network. The first layer of the LSTM network is the LSTM layer. The input matrix is set to (n−1)×1 to adapt to The specific structural requirements of the second layer are a fully connected layer with an output size of 1; Because our purpose is to use the time At the time (i.e., a total of n−1 data) to predict a single time point data; S2-4. Design an optimizer that combines first-order and second-order momentum to dynamically adjust the learning rate; S2-5. Train the generator based on the long short-term memory network LSTM. First, configure the optimizer to the generator. Then, train the training data set in batches according to the set number of training rounds EPOCH and batch size BATCH_SIZE. Specifically, for each training round, from 1 to EPOCH, the dataset and label sets Extract batch data of size BATCH_SIZE from the dataset to train the LSTM model and update its parameters until all batches of data are used for training and a full round of training is completed. Then it enters the next training round and repeats the above process until the preset number of training rounds EPOCH is reached. Finally, the trained and optimized LSTM generator is returned for the subsequent head sensor data generation task; S2-6. Use the trained LSTM generator to generate sensor data based on the input data. Specifically, use Take as input a new matrix to predict , the matrix represents the arrive Sensor data, matrix The dimension is (m−n)×1.

[0014] Furthermore, in step S2-2, the matrix and , using the sliding time window method is defined as follows: , , in, Indicates the first The value at a time point, therefore, The dimension of is (m−n)×(n−1), and The dimension is (m−n+1)×1.

[0015] Furthermore, in step S2-4, the first-order momentum and second-order momentum of each small step gradient are calculated, and deviation correction is performed, as follows: ; in, and are the estimates of the first-order momentum and second-order momentum of the gradient, and is an exponential decay rate, usually taking a value close to 1; According to the gradient of the loss function, the model parameters are gradually adjusted. The parameter adjustment rules are as follows: ; The default learning rate α=0.001, As can be seen from the expression, the updated step size calculation can be adaptively adjusted from the two perspectives of gradient mean and gradient square, rather than being directly determined by the current gradient.

[0016] Furthermore, in step S3, it further includes: S3-1. Design a scoring-based sensor ranking method within a group: In each group, for each sensor to generate data (target sensor), rank it based on its Pearson correlation with the sensors that have already generated data. The top five sensors with the highest Pearson correlation are selected as training features to generate data for the target sensor. If there are fewer than five available sensors, the sensor with the highest Pearson value is selected and polynomial feature expansion is performed as the training feature; S3-2. According to the sensor ranking method within the group based on the score, the non-head sensor (target sensor) to generate data Generate training feature matrix , using the generated and the corresponding Perform ridge regression training to obtain the ridge regression generator of the target sensor; S3-3, using the trained ridge regression generator to generate data of the target sensor .

[0017] Furthermore, in step S4, it includes: For data generated using LSTM and ridge regression, the data volume is expanded by interpolation by adjusting the timestamp granularity. The interpolation methods include linear interpolation, Newton interpolation, or Lagrange interpolation. The nth-order Lagrange interpolation polynomial is defined as: ; Among them, each are all n-degree polynomials, forming A set of bases.

[0018] A large model data enhancement device, comprising: at least one memory and at least one processor; The at least one memory is configured to store a machine-readable program; The at least one processor is configured to call the machine-readable program to execute a large model data enhancement method.

[0019] Compared with the prior art, the large model data enhancement method and device of the present invention have the following outstanding beneficial effects: This paper aims to replicate the data patterns of sensors at all key nodes in the industrial production process, capturing each key stage of the process. By utilizing a LSTM-based deep learning generator, it is able to learn the complex temporal behavior inherent in industrial production processes.

[0020] Furthermore, the present invention uses a Pearson matrix to automatically group sensor features and employs techniques such as minimum spanning tree (MST) and ridge regression to preserve the inter-sensor relationships that are crucial for accurately simulating process conditions. This ensures that the generated data model has high fidelity while maintaining the original process characteristics. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0022] Figure 1 It is a flowchart of a large model data enhancement method. DETAILED DESCRIPTION

[0023] In order to enable those skilled in the art to better understand the solutions of the present invention, the present invention will be further described in detail below in conjunction with specific embodiments. Obviously, the embodiments described are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0024] A best embodiment is given below: like Figure 1 As shown, a large model data enhancement method in this embodiment has the following steps: S1. Preprocessing stage including: S1-1. Use a sensor cluster to collect raw data, and represent the time series data collected by each sensor in vector form. Suppose there are n sensors (from sensor 1 to sensor n) collecting a raw data set (from data 1 to data n, ), and these data are presented in the IoT data center in.csv format. Represent each data as a vector of length m.

[0025] S1-2. Calculate the Pearson correlation coefficient between any two vectors and , where 1 ≤ i < j ≤ n. , where respectively represent the mean and variance of the vectors. Finally, form the Pearson matrix P. The value of the Pearson matrix P at the (i, j) position represents the Pearson correlation coefficient between the vectors and .

[0026] S1-3. Use the Pearson matrix P to calculate the Euclidean distance matrix D between n data .

[0027] S1-4. Through multi-dimensional scaling analysis, use the distance matrix D to obtain the corresponding coordinate matrix X after the raw data is dimensionally reduced. First, use the Euclidean distance matrix D to calculate another new matrix B in the following way, where each element is defined as: ; where represents the element at the position in the matrix D, and , , ; Subsequently, decompose the matrix B to obtain its r largest eigenvalues and their corresponding eigenvectors ; Finally, calculate the coordinate matrix X after dimensional reduction , where is a matrix containing r eigenvectors, is a diagonal matrix containing the r eigenvalues of B.

[0028] S1-5. Based on the coordinate matrix X, set the threshold to 0.1, and cluster the data into different groups so that the data between different groups is almost uncorrelated. The clustering method can use the DBSCAN algorithm.

[0029] Based on this threshold, DBSCAN will Clustered into k groups, the Pearson correlation coefficient between any two sensors from different groups is less than 0.1. However, when examining the DBSCAN results, we found that the Pearson correlation coefficient between two sensors in the same group can also be less than 0.1. This result is acceptable because, despite the low Pearson correlation coefficient values between some sensor pairs, there are still link paths through other sensors in the group with higher correlations. For example, consider sensors in the same group , the observed correlation may be , , , This indicates that there is a path within the group ( → → → ), where the Pearson correlation coefficient between each group of adjacent sensors exceeds 0.1.

[0030] S1-6. Use the Euclidean distance between each data as the weight to calculate the minimum spanning tree for each group. The sensor corresponding to the starting point of the minimum spanning tree is regarded as the head sensor, and the sensors corresponding to the non-starting points are regarded as non-head sensors.

[0031] For example, in the sequence → → → In the head sensor is identified as In this configuration, the The data is used to generate , from and The data is used to generate , and from 、 and The data is used to generate .

[0032] The steps to build a minimum spanning tree are as follows: 1) Randomly select a data from each group and initialize a tree.

[0033] 2) Expand along an edge of the tree: Among the edges connecting the tree to a vertex that is not yet included, select the edge with the smallest weight (i.e., the smallest Euclidean distance) and add it to the tree.

[0034] 3) Repeat the previous step until all vertices are included in the tree.

[0035] S2, head sensor data training stage; include: S2-1. To generate new head sensor data, it is necessary to use the previous time point data to predict the time point The data, the structure of the prediction function is defined as follows: ; in, Indicates a time point The predicted value of the function Used to generate predictions, the training goal is to optimize Parameters.

[0036] S2-2, treat the data collected by the head sensor in a specific time period as a time series and construct the input matrix and the target matrix The matrix and , using the sliding time window method is defined as follows: ; ; in, Indicates the first The value at each time point. Therefore, The dimension of is (m−n)×(n−1), and The dimension is (m−n+1)×1.

[0037] S2-3. Design LSTM neural network: Use Keras to design the LSTM network. The first layer of the LSTM network is the LSTM layer. The LSTM layer contains 128 units. The input matrix is set to (n−1)×1 to accommodate The second layer is a fully connected layer with an output size of 1.

[0038] The entire network has 66,689 parameters trained through LSTM forward and backward propagation. Deeper LSTM architectures can be explored to further improve simulation accuracy and performance metrics.

[0039] S2-4. Design an optimizer that combines first-order and second-order momentum to dynamically adjust the learning rate.

[0040] Calculate the first-order momentum and second-order momentum of each small step gradient and perform deviation correction as follows: ; in, and are estimates of the first-order momentum (mean) and second-order momentum (uncentered variance) of the gradient, and It is an exponential decay rate, usually close to 1. The default value can be set as follows , =0.999.

[0041] According to the gradient of the loss function, the model parameters are gradually adjusted. The parameter adjustment rules are as follows: ; The default learning rate α =0.001, As can be seen from the expression, the updated step size calculation can be adaptively adjusted from two perspectives: the gradient mean and the gradient square, rather than being directly determined by the current gradient.

[0042] S2-5. Train a generator based on a long short-term memory network (LSTM); First, configure the optimizer into the generator. Then, set the number of training rounds EPOCH=30 and the batch size BATCH_SIZE=100 to train the training dataset in batches.

[0043] Specifically, for each training round (from 1 to 30), the dataset and label sets We extract batches of data of size 100 from the dataset and use them to train the LSTM model and update its parameters until all batches of data are used for training, completing a full round of training. We then proceed to the next training round and repeat the above process until 30 rounds of training have been completed.

[0044] Finally, the trained and optimized LSTM generator is returned for subsequent head sensor data generation tasks.

[0045] S2-6, use the trained LSTM generator to generate sensor data based on the input data. Specifically, use Take as input a new matrix to predict , the matrix represents the arrive Sensor data, matrix The dimension is (m−n)×1.

[0046] S3, non-head sensor data training stage; include: S3-1. Design a scoring-based method for ranking sensors within a group: Within each group, sort each sensor to generate data (target sensor) based on its Pearson correlation with the sensors that have already generated data. Select the top five sensors with the highest Pearson correlations as training features to be used to generate data for the target sensor. If fewer than five sensors are available, select the sensor with the highest Pearson score and perform polynomial feature expansion on it as the training feature.

[0047] For example, Prim's algorithm is used to generate a minimum spanning tree sequence. → → → → → → , in for When generating data, available candidate features include , , , , and , we need to select the top 5 features based on Pearson correlation. generate When only a single feature is available, polynomial feature expansion is required. Specifically, if The data matrix is represented by x, then the expanded feature matrix is .

[0048] S3-2. According to the sensor ranking method within the group based on the score, the non-head sensor (target sensor) to generate data Generate training feature matrix , using the generated and the corresponding Perform ridge regression training to obtain the ridge regression generator of the target sensor.

[0049] S3-3, using the trained ridge regression generator to generate data of the target sensor .

[0050] S4, data generation stage; For data generated using LSTM and ridge regression, the data volume is expanded by interpolation by adjusting the timestamp granularity. Interpolation methods include linear interpolation, Newton interpolation, Lagrange interpolation, etc. The nth-order Lagrange interpolation polynomial is defined as: ; Among them, each are all n-degree polynomials, forming A set of bases.

[0051] For example, after generating data using LSTM and ridge regression, a dataset originally containing 1000 samples can be expanded to 5000 samples through Lagrangian interpolation.

[0052] Based on the above method, a large model data enhancement device in this embodiment includes: at least one memory and at least one processor; at least one memory for storing a machine-readable program; At least one processor is configured to call the machine-readable program to execute a large model data enhancement method.

[0053] The processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The processor may be a microprocessor or any conventional processor, etc.

[0054] The memory can be used to store computer programs and / or modules. The processor implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory and accessing the data stored in the memory. The memory can mainly include a program storage area and a data storage area. The program storage area can store the operating system and at least one application required for the function; the data storage area can store data generated based on the use of the terminal. In addition, the memory can also include high-speed random access memory and non-volatile memory, such as a hard disk, internal memory, a plug-in hard disk, a smart card (SMC), a secure digital (SD) card, a flash memory card, at least one disk storage device, a flash memory device, or other volatile solid-state memory devices.

[0055] The above-mentioned specific implementation methods are only specific cases of the present invention. The patent protection scope of the present invention includes but is not limited to the above-mentioned specific implementation methods. Any technical solutions that conform to the above-mentioned specific implementation methods of the present invention and any appropriate changes or substitutions made thereto by ordinary technicians in the relevant technical field shall fall within the patent protection scope of the present invention.

[0056] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A large model data enhancement method, characterized in that: The steps are as follows: S1, preprocessing stage uses the sensor cluster to collect raw data, groups the data by Pearson correlation coefficient, and for each group, uses the minimum spanning tree algorithm to determine the head sensor and establish the generation sequence; S2, the head sensor data training phase is to start the generator based on the long short-term memory network LSTM and use the LSTM generator to train the head sensor data; S3, the non-head sensor data training phase is to start the ridge regression generator and use the ridge regression generator to train the data of other sensors; S4, the data generation stage is to generate corresponding data based on the user's specified requirements using the trained model, and increase the data volume through interpolation.

2. A large model data enhancement method according to claim 1, characterized in that: In step S1, further comprising: S1-1. Use a sensor cluster to collect raw data and represent the time series data collected by each sensor as a vector. Assume that there are n sensors collecting raw data sets. These data are displayed in the IoT data center in .csv format, and each data is represented as a vector of length m. S1-2. Calculate any two vectors and , where 1 ≤ i < j ≤ n, and calculate the Pearson correlation coefficient between them. , where respectively represent the mean and variance of the vectors. Finally, form the Pearson matrix P. The value at the (i, j) position of the Pearson matrix P represents the Pearson correlation coefficient between the vectors and . S1-3, use the Pearson matrix P to calculate the Euclidean distance matrix D between n data, ; S1-4, using the distance matrix D through multidimensional scaling analysis to obtain the coordinate matrix X corresponding to the original data after dimensionality reduction; S1-5. Based on the coordinate matrix X, set the threshold to 0.1 and cluster the data into different groups so that the data between different groups are almost uncorrelated. S1-6. Use the Euclidean distance between each data as the weight to calculate the minimum spanning tree for each group. The sensor corresponding to the starting point of the minimum spanning tree is regarded as the head sensor, and the sensors corresponding to the non-starting points are regarded as non-head sensors.

3. A large model data enhancement method according to claim 2, characterized in that: In step S1-4, first, using the Euclidean distance matrix D, another new matrix B is calculated as follows, where each element Defined as: ； in, Indicates the matrix D The element at the position, and , , ; Then, the matrix B is decomposed to obtain its r largest eigenvalues and its corresponding eigenvector ; Finally, calculate the coordinate matrix X after dimensionality reduction, ,in, is a matrix containing r eigenvectors, is a diagonal matrix containing the r eigenvalues of B.

4. A large model data enhancement method according to claim 3, characterized in that: In step S1-6, the minimum spanning tree is established as follows: (1) Select a data from each group and initialize a tree; (2) Expand along an edge of the tree: Among the edges connecting the tree to the vertex that is not yet included, select the edge with the smallest weight and add it to the tree. The selection of the edge with the smallest weight means the smallest Euclidean distance. (3) Repeat the previous step until all vertices are included in the tree.

5. A large model data enhancement method according to claim 4, characterized in that: In step S2, further comprising: S2-1. To generate new head sensor data, it is necessary to use the previous time point data to predict the time point The data, the structure of the prediction function is defined as follows: ； in, Indicates a time point The predicted value of the function Used to generate predictions, the training goal is to optimize Parameters; S2-2, treat the data collected by the head sensor in a specific time period as a time series and construct the input matrix and the target matrix ; S2-3. Design LSTM neural network: Use Keras to design the LSTM network. The first layer of the LSTM network is the LSTM layer. The input matrix is set to (n−1)×1 to adapt to The specific structural requirements of the second layer are a fully connected layer with an output size of 1; S2-4. Design an optimizer that combines first-order and second-order momentum to dynamically adjust the learning rate; S2-5. Train the generator based on the long short-term memory network LSTM. First, configure the optimizer to the generator. Then, train the training data set in batches according to the set number of training rounds EPOCH and batch size BATCH_SIZE. Specifically, for each training round, from 1 to EPOCH), the dataset and label sets Extract batch data of size BATCH_SIZE from the dataset to train the LSTM model and update its parameters until all batches of data are used for training and a full round of training is completed. Then it enters the next training round and repeats the above process until the preset number of training rounds EPOCH is reached. Finally, the trained and optimized LSTM generator is returned for the subsequent head sensor data generation task; S2-6. Use the trained LSTM generator to generate sensor data based on the input data. Specifically, use Take as input a new matrix to predict , the matrix represents the arrive Sensor data, matrix The dimension is (m−n)×1.

6. A large model data enhancement method according to claim 5, characterized in that: In step S2-2, the matrix and , using the sliding time window method is defined as follows: ， ， in, Indicates the first The value at a time point, therefore, The dimension of is (m−n)×(n−1), and The dimension is (m−n+1)×1.

7. A large model data enhancement method according to claim 6, characterized in that: In step S2-4, the first-order momentum and second-order momentum of each small step gradient are calculated and the deviation correction is performed as follows: ； in, and are the estimates of the first-order momentum and second-order momentum of the gradient, and is an exponential decay rate, usually taking a value close to 1; According to the gradient of the loss function, the model parameters are gradually adjusted. The parameter adjustment rules are as follows: ； The default learning rate α=0.001, As can be seen from the expression, the updated step size calculation can be adaptively adjusted from the two perspectives of gradient mean and gradient square, rather than being directly determined by the current gradient.

8. A large model data enhancement method according to claim 7, characterized in that: In step S3, further comprising: S3-1. Design a scoring-based sensor ranking method within a group: In each group, for each sensor to generate data, rank it based on its Pearson correlation with the sensors that have already generated data. The top five sensors with the highest Pearson correlation are selected as training features to generate data for the target sensor. If there are fewer than five available sensors, the sensor with the highest Pearson value is selected and polynomial feature expansion is performed as the training feature; S3-2, according to the sensor sorting method within the group based on the score, for the non-head sensor to generate data Generate training feature matrix , using the generated and the corresponding Perform ridge regression training to obtain the ridge regression generator of the target sensor; S3-3, using the trained ridge regression generator to generate data of the target sensor .

9. A large model data enhancement method according to claim 8, characterized in that: In step S4, it includes: For data generated using LSTM and ridge regression, the data volume is expanded by interpolation by adjusting the timestamp granularity. The interpolation methods include linear interpolation, Newton interpolation, or Lagrange interpolation. The nth-order Lagrange interpolation polynomial is defined as: ； Among them, each are all n-degree polynomials, forming A set of bases.

10. A large model data enhancement device, characterized in that: include: at least one memory and at least one processor; The at least one memory is configured to store a machine-readable program; The at least one processor is configured to call the machine-readable program to execute the method according to any one of claims 1 to 9.

Images