An experimental process intelligent error correction method and device

Abstract

The present application relates to the technical field of machine vision, in particular to an experimental process intelligent error correction method and device. The method comprises: dynamically evaluating the importance of each channel of multi-channel time series data and performing dimension reordering; inputting the reordered data into an encoder to obtain first and second latent vectors of adjacent time steps, and constructing a spliced latent vector that fuses time series and causal information; inputting the spliced vector into a forward and backward decoder to calculate a reconstruction inconsistency score; when the score exceeds a threshold value, locking the first latent vector, optimizing the gradient of the second latent vector to minimize inconsistency, obtaining a correction vector and generating error correction data. The scheme of the present application improves the recognition accuracy of complex anomalies, enhances the ability to distinguish between normal changes and errors, and improves the reliability of error correction results.

Description

Technical Field

[0001] This invention relates to the field of machine vision technology. More specifically, this invention relates to an intelligent error correction method and apparatus for experimental processes. Background Technology

[0002] In complex scientific experiments and industrial production processes, abnormal or erroneous values ​​often appear in process data due to factors such as sensor failure, sudden environmental changes, human error, or transmission interference. Therefore, intelligent and automatic anomaly detection and correction of time-series data is crucial for improving system efficiency and ensuring operational safety. In existing technologies, error correction methods based on statistical models, such as the 3σ criterion and moving average methods, are widely used. However, these methods typically make strong prior assumptions about the probability distribution of the data and struggle to effectively capture complex nonlinear relationships between different variables when processing high-dimensional data.

[0003] To address the limitations of the aforementioned statistical models, improved techniques based on deep learning have been developed in recent years. These techniques employ models such as autoencoders, recurrent neural networks, or long short-term memory networks to learn the distribution patterns of data within normal patterns and utilize reconstruction errors to identify and locate outliers. This approach demonstrates stronger nonlinear feature detection capabilities and is better suited to high-dimensional and complex data. By learning normal behavior patterns in a data-driven manner, these methods provide a more effective approach for anomaly detection in time-series data.

[0004] However, the aforementioned deep learning-based improvements still have significant shortcomings in practical applications. First, when processing multi-channel data, most models fail to fully consider the dynamic correlations between channels, often treating all channels equally or assuming static correlations. This limits the model's ability to identify complex anomalies arising from specific channel combinations. Second, in modeling temporal dependencies, existing models focus only on information compression at the current time point during encoding, ignoring the potential temporal evolution and causal relationships between adjacent time steps. This makes it difficult for the model to accurately distinguish between normal system state transitions and genuine anomalous mutations. Furthermore, in the error correction stage, existing methods often employ a simple repair strategy of directly replacing model reconstruction values. This approach lacks an optimization process oriented towards restoring data consistency. When anomalies have contaminated the model's latent vectors, the reconstruction result itself may be unreliable, leading to the corrected data sequence not fully conforming to the system's true physical laws. Summary of the Invention

[0005] The purpose of this invention is to propose an intelligent error correction method and device for experimental processes, in order to solve the problem of poor reliability and fidelity of error correction results in the prior art; to this end, this invention provides solutions in the following two aspects.

[0006] In a first aspect, the present application provides an intelligent error correction method for an experimental process, comprising the following steps:

[0007] The method comprises the following steps: obtaining multi-channel time series data of an experimental process to be corrected; dynamically evaluating the importance of each channel and reordering the channel dimension for the data in a time window to obtain channel rearrangement data; inputting the channel rearrangement data into an encoder; obtaining a first latent vector and a second latent vector for adjacent first and second time steps, respectively, and constructing a spliced latent vector that fuses timing and causal information based on the first and second latent vectors; inputting the spliced latent vector into a forward decoder and a backward decoder, respectively, to obtain forward reconstruction data and backward reconstruction data for the second time step, and calculating a reconstruction inconsistency score for the current time window based on the norm difference of the reconstruction residuals formed by the forward and backward reconstruction data and the original data of the second time step; when the reconstruction inconsistency score exceeds a preset threshold, locking the first latent vector, performing gradient optimization on the second latent vector to obtain a corrected second latent vector, and reconstructing the spliced latent vector using the locked first latent vector and the corrected second latent vector, and decoding the reconstructed spliced latent vector through the forward decoder to generate corrected experimental process data.

[0008] Preferably, the step of dynamically evaluating the importance of each channel comprises generating a channel attention mask through a neural network to represent the importance of each channel according to the local statistical characteristics of the data in the time window and the reconstruction inconsistency score of the previous time window.

[0009] Preferably, the step of reordering the channel dimension comprises generating a channel permutation matrix by adjusting the relevance measure between channels using the channel attention mask, and reordering the channel dimension of the multi-channel time series data using the channel permutation matrix to obtain channel rearrangement data.

[0010] Preferably, the step of constructing a spliced latent vector that fuses timing and causal information comprises calculating the Hadamard product of the first and second latent vectors as a causal component, calculating the vector difference between the second and first latent vectors as a timing component, and splicing the first latent vector, causal component, and timing component to construct a spliced latent vector.

[0011] Preferably, the step of generating a channel attention mask by the neural network comprises: calculating the mean and standard deviation of the time series data of each channel in the time window as local statistical features; concatenating the mean and standard deviation of all channels into a statistical feature vector, and concatenating the statistical feature vector with the reconstruction inconsistency score of the previous time window to form a fusion feature vector; inputting the fusion feature vector into a double-layer multilayer perception, and outputting a channel attention mask with the same dimension as the number of channels through a Sigmoid activation function.

[0012] Preferably, the step of generating a channel replacement matrix by adjusting the relevance measure between channels using the channel attention mask comprises: calculating the Pearson correlation coefficient between all channels in the multichannel time series data to form an initial relevance matrix; performing Hadamard product operation on the initial relevance matrix after broadcasting the channel attention mask to obtain an adjusted relevance matrix; based on the adjusted relevance matrix, modeling the channel reordering problem as a maximum weight matching problem, and solving it by using the Hungarian algorithm to obtain the optimal channel replacement matrix.

[0013] Preferably, the step of concatenating the first latent vector, the causal component and the time series component to construct a concatenated latent vector comprises:

[0014] Let the first latent vector be , and the second latent vector be ; calculate the causal component , wherein is a Hadamard product operator; calculate the time series component ; sequentially concatenate , and in the feature dimension to obtain a time series causal concatenated latent vector.

[0015] Preferably, the step of calculating the reconstruction inconsistency score of the current time window based on the norm difference of the reconstruction residuals formed by the forward reconstruction data and the backward reconstruction data and the original data of the second time step comprises:

[0016] Let the original data of the second time step be , the forward reconstruction data be , and the backward reconstruction data be ; calculate the forward reconstruction residual , and calculate the norm ; calculate the backward reconstruction residual , and calculate the norm ; the reconstruction inconsistency score of the current time window is the absolute difference of the two norms, that is .

[0017] Preferably, the step of optimizing the second latent vector with the optimization objective of minimizing the reconstruction inconsistency score includes: setting the initial value of the second latent vector to be corrected as the original second latent vector; using an Adam optimizer with a learning rate of 0.001, taking the reconstruction inconsistency score as a loss function, and performing 50 times of iterative updates on the second latent vector to be corrected; and taking the second latent vector after the iterative update as the corrected second latent vector.

[0018] In a second aspect, an experimental process intelligent error correction device comprises:

[0019] a processor; and a memory storing computer instructions for experimental process intelligent error correction, which, when executed by the processor, causes the system to perform the experimental process intelligent error correction method described above.

[0020] The beneficial effects of the present application are: by fusing local statistical features and historical reconstruction information to generate a channel permutation matrix, the dimensions of multi-channel data are reordered, so that the model can detect the complex correlation between channels in different time windows, and the recognition accuracy of complex anomalies is improved; by constructing a time causal splicing latent vector, the latent information of adjacent time steps is decomposed into causal components and time components, and the ability of the model to distinguish between normal state changes and real data errors is enhanced; in the error correction stage, the correct latent vector of the previous moment is locked, and the latent vector of the current moment is optimized in the latent space with the objective of minimizing the forward and backward reconstruction inconsistency, the error correction method ensures that the corrected data points maintain high consistency with the context data in the time sequence logic, and improves the reliability and fidelity of the error correction result. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 The step flowchart of the experimental process intelligent error correction method in the embodiment is schematically shown;

[0022] Figure 2 The structural block diagram of the experimental process intelligent error correction device in the embodiment is schematically shown. DETAILED DESCRIPTION

[0023] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application.

[0024] As Figure 1 shown, the experimental process intelligent error correction method in the embodiment includes the following steps:

[0025] Step S1: Obtain multi-channel time-series data of the experimental process to be corrected. For the data within the time window, dynamically evaluate the importance of each channel and reorder the channel dimensions to obtain channel reordering data.

[0026] The step of dynamically evaluating the importance of each channel includes: generating a channel attention mask through a neural network based on the local statistical characteristics of the data within the time window and the reconstruction inconsistency score of the previous time window, in order to characterize the importance of each channel.

[0027] The steps for reordering the channel dimensions include: using the channel attention mask to adjust the correlation metric between channels to generate a channel permutation matrix, and using the channel permutation matrix to reorder the channel dimensions of the multi-channel time series data to obtain the channel-reordered data.

[0028] Specifically, multiple sensors deployed on the chemical reactor collect multi-dimensional process parameters, such as temperature, pressure, stirring speed, and feed flow rate, forming an N-row, M-column time-series data matrix, where N is the total number of time steps and M is the number of channels or variables. The data matrix is ​​then divided into a series of window data segments using a sliding time window of length T.

[0029] The mean, variance, kurtosis, and skewness of each channel data within the current time window are calculated as local statistical feature vectors. These feature vectors are concatenated with the reconstruction inconsistency score scalar value calculated in the previous time window and input into a multilayer perceptron network, which outputs a channel attention mask vector of length M. The Pearson correlation coefficient matrix among all channels within the current window is calculated as an initial correlation measure. The channel attention mask vector is outer-productted with its transpose to obtain an attention weight matrix, which is then element-wise multiplied with the Pearson correlation coefficient matrix to obtain an adjusted correlation matrix. Based on the adjusted correlation matrix, a greedy algorithm is used to determine the channel importance order. For example, the channel with the highest total correlation is selected as the first channel, then the remaining channels with the highest correlation to the first channel are selected as the second channel, and so on, generating an M-order channel permutation matrix. This permutation matrix is ​​then right-multiplied by the data of the current time window to complete the channel dimension reordering.

[0030] In step S1 above, by introducing dynamic evaluation and channel reordering mechanisms, the limitations of traditional methods that treat all data channels equally and assume that their correlation remains static are overcome. This enables the model to adaptively focus on the most important information channels according to the specific stage of the current experimental process. By optimizing the data input order, the subsequent network's ability to capture complex temporal dependencies among multiple variables is enhanced, thereby significantly improving the system's accuracy and dynamic adaptability in identifying complex and hidden anomalies under varying conditions.

[0031] In an optional embodiment, the step of generating a channel attention mask by the neural network comprises: calculating the mean and standard deviation of the time series data of each channel in the time window as local statistical features; concatenating the mean and standard deviation of all channels into a statistical feature vector, and concatenating the statistical feature vector with the reconstruction inconsistency score of the previous time window to form a fusion feature vector; inputting the fusion feature vector into a double-layer multilayer perceptron, and outputting a channel attention mask with the same dimension as the number of channels through a Sigmoid activation function.

[0032] Specifically, assuming that the multi-channel time series data in the current time window contains 10 channels, each channel has 100 data points. The mean and standard deviation of the data of the 10 channels are calculated independently, for example, the mean of channel 1 is 5.2 and the standard deviation is 1.1, the mean of channel 2 is 6.1 and the standard deviation is 1.3, and so on, obtaining 10 means and 10 standard deviations.

[0033] The 10 means and 10 standard deviations obtained above are concatenated into a 20-dimensional statistical feature vector. Assuming that the reconstruction inconsistency score calculated in the previous time window is 0.8, the reconstruction inconsistency score is concatenated to the end of the 20-dimensional vector to form a 21-dimensional fusion feature vector. The 21-dimensional fusion feature vector is input as an input into a pre-trained double-layer multilayer perceptron, for example, the first layer maps 21 dimensions to 16 dimensions, the second layer maps 16 dimensions to 10 dimensions, and the last layer outputs a 10-dimensional vector through a Sigmoid activation function, each element value of which is between 0 and 1. The vector is the channel attention mask, representing the importance weight of each channel.

[0034] In an optional embodiment, the step of generating a channel attention mask by the neural network comprises: calculating the mean and standard deviation of the time series data of each channel in the time window as local statistical features; concatenating the mean and standard deviation of all channels into a statistical feature vector, and concatenating the statistical feature vector with the reconstruction inconsistency score of the previous time window to form a fusion feature vector; inputting the fusion feature vector into a double-layer multilayer perceptron, and outputting a channel attention mask with the same dimension as the number of channels through a Sigmoid activation function.

[0035] Specifically, taking a 10-channel time series data as an example, the Pearson correlation coefficient between any two channels is calculated, for example, the correlation coefficient between channel 1 and channel 2 is 0.9, the correlation coefficient between channel 1 and channel 3 is 0.2, and so on, thereby constructing a symmetric matrix of the initial correlation matrix, the diagonal elements are 1.

[0036] Take the 10-dimensional channel attention mask generated in the previous step and broadcast it to a matrix, where each row is the mask vector. Perform a Hadamard product, i.e., element-wise multiplication, between the broadcasted mask matrix and the initial relevance matrix. This operation adjusts the relevance between channels according to their importance, for example, the relevance of unimportant channels is weakened accordingly, resulting in an adjusted relevance matrix. Treat the channel reordering problem as a maximum weight matching problem in graph theory, where channels are nodes and the values in the adjusted relevance matrix are edge weights. Use the Hungarian algorithm to solve this problem and obtain a channel arrangement that maximizes the total relevance weight, which is represented in the form of a permutation matrix.

[0037] In step S2, input the channel rearrangement data into the encoder, obtain a first latent vector and a second latent vector for adjacent first and second time steps, respectively, and construct a spliced latent vector that integrates temporal and causal information based on the first and second latent vectors.

[0038] The step of reordering the channel dimension includes: generating a channel permutation matrix by adjusting the relevance measure between channels using the channel attention mask, and reordering the channel dimension of the multi-channel time series data using the channel permutation matrix to obtain channel rearrangement data.

[0039] Specifically, the encoder is a network structure containing multiple layers of gated recurrent units (GRU), and the channel-arranged time window data sequence is input into the encoder. The hidden state output by the encoder when processing time step t-1 in the sequence is the first latent vector, and the hidden state output when processing time step t is the second latent vector.

[0040] Perform element-wise multiplication on the first and second latent vectors to obtain a causal component, and perform element-wise subtraction on the second and first latent vectors to obtain a temporal component. Concatenate the first latent vector, the causal component vector, and the temporal component vector in the dimension to form a temporal and causal spliced latent vector with three times the original latent vector dimension.

[0041] In the above step S2, a spliced latent vector that integrates previous state, temporal change, and causal correlation is constructed, providing the model with a dynamic perspective on how the system state evolves. This decomposition of the time evolution process enables the model to deeply understand the internal logic and reasonable rate of system dynamic changes, thereby greatly enhancing its ability to distinguish between normal state fluctuations and real data anomalies, providing a more information-rich and logically rigorous basis for subsequent accurate anomaly detection and error correction.

[0042] In an optional embodiment, the first latent vector, the causal component and the temporal component are spliced, and the step of constructing the spliced latent vector comprises:

[0043] Let the first latent vector be , and the second latent vector be ; calculate the causal component , wherein is a Hadamard product operator; calculate the temporal component ; sequentially splice , and in the feature dimension to obtain the temporal causal spliced latent vector.

[0044] Specifically, it is assumed that the encoder maps the data of each time step to a 64-dimensional latent vector. Let the latent vector of the first time step be , which is a 64-dimensional vector, and the latent vector of the second time step be , which is another 64-dimensional vector.

[0045] Subsequently, the causal component is calculated by performing Hadamard product, i.e., multiplying corresponding elements, on and to obtain a new 64-dimensional vector. The temporal component is calculated by performing corresponding element subtraction on and to obtain another 64-dimensional vector. In the feature dimension, the 64-dimensional vector, the 64-dimensional vector and the 64-dimensional vector are sequentially connected to form a temporal causal spliced latent vector with a total dimension of 192.

[0046] Step S3, input the spliced latent vector into the forward decoder and the backward decoder respectively to obtain the forward reconstruction data and the backward reconstruction data of the second time step, and calculate the reconstruction inconsistency score of the current time window based on the norm difference of the reconstruction residuals formed by the forward reconstruction data and the backward reconstruction data and the original data of the second time step.

[0047] Specifically, the forward decoder and the backward decoder are both GRU networks symmetrical to the structure of the encoder; the temporal causal spliced latent vector is input as the initial hidden state into the two decoders respectively to generate the reconstruction results of the data of the second time step, i.e., the forward reconstruction data and the backward reconstruction data; the difference between the forward reconstruction data and the original data is calculated to obtain the forward reconstruction residual, and the difference between the backward reconstruction data and the original data is calculated to obtain the backward reconstruction residual; the norms of the forward and backward reconstruction residual vectors norms of the forward and backward reconstruction residual vectors, and taking the absolute difference between the two norm values as the reconstruction inconsistency score of the current time window.

[0048] In step S3, by calculating the reconstruction inconsistency scores between the forward and backward reconstruction results and the true values, time logic consistency is introduced as a higher-dimensional evaluation standard, making the model more sensitive to capture instantaneous abnormalities that destroy the time sequence continuity, while having stronger robustness to the normal drastic changes of the system, thereby greatly improving the accuracy and reliability of the anomaly detection, effectively reducing the false positive rate and the false negative rate.

[0049] In an optional embodiment, the reconstruction inconsistency score of the current time window is calculated based on the norm difference of the reconstruction residuals formed by the forward reconstruction data and the backward reconstruction data with the original data of the second time step, including:

[0050] Let the original data of the second time step be , the forward reconstruction data be , and the backward reconstruction data be ; calculate the forward reconstruction residual , and calculate the norm ; calculate the backward reconstruction residual , and calculate the norm ; the reconstruction inconsistency score of the current time window is calculated as the absolute difference between the two norms, i.e. .

[0051] For example, assume that in the second time step, the original data is a 100-dimensional vector. Based on the information of the first time step, the forward reconstruction data predicted by the forward model is also a 100-dimensional vector. Similarly, based on the information of the third time step, the backward reconstruction data predicted by the backward model is also a 100-dimensional vector.

[0052] Calculate the forward reconstruction residual , i.e., the element-wise difference between and , to obtain a 100-dimensional residual vector. Then calculate the norm of the residual vector, i.e., the square root of the sum of the squares of all elements, to obtain a scalar value. In the same way, calculate the backward reconstruction residual , i.e., the difference between and , and calculate the norm. The reconstruction inconsistency score of the current time window is the absolute difference of the norm values of the two .

[0053] Step S4, when the reconstruction inconsistency score exceeds the preset threshold, locking the first latent vector, and performing gradient optimization on the second latent vector with the optimization objective of minimizing the reconstruction inconsistency score to obtain a corrected second latent vector, and reconstructing a splicing latent vector by using the locked first latent vector and the corrected second latent vector, and decoding the reconstructed splicing latent vector through the forward decoder to generate the error-corrected experimental process data.

[0054] Specifically, the preset threshold is determined by calculating the reconstruction inconsistency scores of all time windows on a validation data set not containing anomalies, and taking the 99th percentile; when the score exceeds the limit, the first latent vector is regarded as a fixed constant and does not participate in gradient calculation; the reconstruction inconsistency score is a loss function, and an Adam optimizer is used to calculate the gradient of the loss function for each element of the second latent vector and perform iterative updating in the opposite direction of the gradient, for example, 20 times of iterative updating, a learning rate of 1 / 1000, until the loss function value converges or the upper limit of the number of iterations is reached, at which time the obtained second latent vector is the corrected second latent vector.

[0055] The first latent vector that remains unchanged and the corrected second latent vector are used to recalculate the Hadamard product and the vector difference according to the above steps, and are spliced into a new time-series causal splicing latent vector; the new splicing latent vector is input into the forward decoder for one forward propagation calculation, and the data points output by the decoder are the error correction results of the second time step original abnormal data, which are used to replace the error values in the original data stream.

[0056] In the above step S4, the correct latent vector of the previous time is locked as an unshakable historical reference, and the logical consistency of the recovered data in the time series (i.e., minimizing the reconstruction inconsistency score) is taken as the explicit target to iteratively optimize the latent vector at the abnormal time. This method ensures that the generated error correction data is not simply numerical filling, but a high-fidelity reconstruction result that fundamentally conforms to the inherent dynamic law of the system, thereby greatly improving the reliability and time series continuity of the error-corrected data.

[0057] In an optional embodiment, the step of obtaining the preset threshold is: using a segment of normal experimental process data not containing anomalies to calculate the reconstruction inconsistency score of each time window to form a score distribution; calculating the mean and the standard deviation of the score distribution; and setting the preset threshold as .

[0058] In an optional embodiment, the step of performing gradient optimization on the second latent vector with the goal of minimizing the reconstruction inconsistency score to obtain the corrected second latent vector includes:

[0059] The initial value of the second potential vector to be corrected is set to the original second potential vector; the Adam optimizer with a learning rate of 0.001 is used to reconstruct the inconsistency score as the loss function, and the second potential vector to be corrected is updated 50 times; the second potential vector after the iteration update is used as the corrected second potential vector.

[0060] Specifically, when the reconstruction inconsistency score of a time window exceeds a preset threshold, a correction procedure is initiated. This involves retrieving the second latent vector generated by the encoder corresponding to the abnormal time window. Assume this is the starting point for optimization. It is a 64-dimensional vector. An iterative optimization loop is entered. In this loop, the reconstruction inconsistency score is defined as the loss function. The Adam optimizer is used, with a learning rate of 0.001. In each iteration, the loss function is calculated with respect to the latent vector. The gradient of each element is calculated, and then the Adam optimizer updates the algorithm based on this gradient information. The value of makes the updated value This can produce a smaller inconsistency score. Repeat the above process 50 times. For example, after the first iteration, Fine-tuning reduced the reconstruction inconsistency score from 0.9 to 0.85; after 50 iterations, the reconstruction inconsistency score may decrease to a near-normal level of 0.3. The optimization process stops after 50 iterations, and the resulting 64-dimensional latent vector is the corrected second latent vector.

[0061] This invention also provides an intelligent error correction system for experimental processes. For example... Figure 2 As shown, the system includes a processor and a memory, the memory storing computer program instructions, which, when executed by the processor, implement the intelligent error correction method for the experimental process according to the present invention.

[0062] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and therefore will not be described in detail here.

[0063] In this description, the term "application" also means any computer program product storing such a program for use with or in connection with a computer system, apparatus or device. The program can be stored on any apparatus-readable medium, for example, but not limited to, any volatile memory or non-volatile memory. In this description, the term "memory" also means any computer program product storing such a program for use with or in connection with a computer system, apparatus or device. The program can be stored on any apparatus-readable medium, for example, but not limited to, any volatile memory or non-volatile memory. In this description, the term "computer-readable medium" means any tangible medium that stores, communicates, or otherwise provides data that can be used by an instruction execution system, apparatus or device. The computer-readable medium can be any suitable magnetic storage medium or magneto-optical storage medium, such as, for example, resistive random access memory (RRAM), dynamic random access memory (DRAM), static random access memory (SRAM), enhanced dynamic random access memory (EDRAM), high-bandwidth memory (HBM), hybrid memory cube (HMC), and the like, or any other medium that can be used to store the desired information and that can be accessed by an application, module, or both. Any such computer storage media can be part of the device or accessible or connectable thereto. Any application or module described in this description can be implemented by computer-readable / executable instructions stored or otherwise held by such computer-readable media.

[0064] In the description of the present description, the meaning of "a plurality of" is at least two, for example, two, three or more, and the like, unless otherwise explicitly specified.

[0065] Although the present description has shown and described a number of embodiments of the present application, it will be apparent to those skilled in the art that many modifications, variations, and alternatives to the embodiments described herein can be made in light of the teachings herein.

Claims

1. An intelligent error correction method for an experimental process, characterized in that, Includes the following steps: Acquire multi-channel time-series data of the experiment process to be corrected, dynamically evaluate the importance of each channel for the data within the time window, and reorder the channel dimensions to obtain channel reordering data; The channel rearranged data is input into the encoder. For adjacent first and second time steps, a first latent vector and a second latent vector are obtained respectively. Based on the first and second latent vectors, a spliced ​​latent vector that integrates temporal and causal information is constructed. The concatenated latent vectors are input into the forward decoder and the backward decoder respectively to obtain the forward reconstruction data and the backward reconstruction data of the second time step. Based on the norm difference of the reconstruction residuals formed by the forward reconstruction data and the backward reconstruction data and the original data of the second time step respectively, the reconstruction inconsistency score of the current time window is calculated. When the reconstruction inconsistency score exceeds a preset threshold, the first latent vector is locked, and the second latent vector is subjected to gradient optimization with the goal of minimizing the reconstruction inconsistency score to obtain the corrected second latent vector. The locked first latent vector and the corrected second latent vector are used to reconstruct the spliced ​​latent vector, and the reconstructed spliced ​​latent vector is decoded by the forward decoder to generate the error-corrected experimental process data.

2. The intelligent error correction method for the experimental process according to claim 1, characterized in that, The step of dynamically evaluating the importance of each channel includes: generating a channel attention mask through a neural network based on the local statistical characteristics of the data within the time window and the reconstruction inconsistency score of the previous time window, in order to characterize the importance of each channel.

3. The intelligent error correction method for the experimental process according to claim 2, characterized in that, The step of reordering the channel dimensions includes: using the channel attention mask to adjust the correlation metric between channels to generate a channel permutation matrix, and using the channel permutation matrix to reorder the channel dimensions of the multi-channel time series data to obtain the channel-reordered data.

4. The intelligent error correction method for the experimental process according to claim 1, characterized in that, The steps for constructing a concatenated latent vector that integrates temporal and causal information include: calculating the Hadamard product of the first latent vector and the second latent vector as the causal component, calculating the vector difference between the second latent vector and the first latent vector as the temporal component, and concatenating the first latent vector, the causal component, and the temporal component to construct the concatenated latent vector.

5. The intelligent error correction method for the experimental process according to claim 2, characterized in that, The step of generating a channel attention mask using a neural network includes: For the time series data of each channel within the time window, the mean and standard deviation are calculated as local statistical features; The mean and standard deviation of all channels are concatenated into a statistical feature vector, and the statistical feature vector is concatenated with the reconstruction inconsistency score of the previous time window to form a fusion feature vector. The fused feature vector is input into a two-layer multilayer perceptron, and a channel attention mask with the same dimension as the number of channels is output through the Sigmoid activation function.

6. The intelligent error correction method for the experimental process according to claim 3, characterized in that, The step of adjusting the correlation metric between channels using the channel attention mask to generate the channel permutation matrix includes: Calculate the Pearson correlation coefficients between all pairs of channels in the multi-channel time series data to form an initial correlation matrix; After broadcasting the channel attention mask, perform a Hadamard product operation with the initial correlation matrix to obtain the adjusted correlation matrix; Based on the adjusted correlation matrix, the channel reordering problem is modeled as a maximum weight matching problem, and the Hungarian algorithm is used to solve it to obtain the optimal channel permutation matrix.

7. The intelligent error correction method for the experimental process according to claim 4, characterized in that, The step of concatenating the first latent vector, the causal component, and the temporal component to construct the concatenated latent vector includes: Let the first potential vector be The second potential vector is ; Calculate causal components ,in, This is the Hadamard product operator; Calculate time-series components ; In terms of feature dimension, , and By sequentially concatenating the vectors, we obtain the temporal causal concatenation potential vector.

8. The intelligent error correction method for the experimental process according to claim 1, characterized in that, The calculation of the reconstruction inconsistency score for the current time window, based on the norm difference between the reconstruction residuals formed by the forward and backward reconstruction data and the original data at the second time step, includes: Let the original data of the second time step be Forward reconstruction data is The backward reconstruction data is ; Calculate the forward reconstruction residual and calculate norm ; Calculate the backward reconstruction residual and calculate norm ; Reconstruction inconsistency score for the current time window Calculated as the absolute difference between the two norms, i.e. .

9. The intelligent error correction method for the experimental process according to claim 1, characterized in that, The step of performing gradient optimization on the second latent vector with the optimization objective of minimizing the reconstruction inconsistency score to obtain the corrected second latent vector includes: Set the initial value of the second potential vector to be corrected to the original second potential vector; The Adam optimizer with a learning rate of 0.001 is used to reconstruct the inconsistency score as the loss function, and the second potential vector to be corrected is updated 50 times. The second potential vector after the iterative update is completed is used as the corrected second potential vector.

10. An intelligent error correction device for an experimental process, characterized in that, include: processor; A memory storing computer instructions for intelligent error correction during the experimental process, which, when executed by the processor, cause the system to perform the intelligent error correction method for the experimental process according to any one of claims 1-9.

Images