An industrial missing data generation method and system for time-sharing power supply

Abstract

The application discloses a kind of industrial missing data generation methods and systems for time-sharing power supply, the method is preprocessed to fixed length containing missing industrial time series data, embedding representation, encoding side composite position coding, containing learnable period bias Transformer encoder, variational autoencoder module, decoding side composite position coding and containing learnable period bias Transformer decoder processing, reconstructs complete industrial data.The present application is fused to form composite position coding by time-sharing power supply period coding and absolute position coding, and introduces learnable period bias in multi-head attention, strengthens time series and working condition characteristics, can accurately and efficiently complete zinc electrolysis and other industrial scene missing data generation, strong stability, high adaptability.

Description

Technical Field

[0001] This application belongs to the field of industrial control technology, specifically relating to a method and system for generating missing industrial data for time-sharing power supply. Background Technology

[0002] Measurement data is a crucial tool for monitoring industrial processes and a vital foundation for their engineering development and scientific research. However, the on-site production environment of industrial processes is generally complex; for example, the on-site production environment of zinc electrolysis typically involves high-acid, concentrated fog, making it difficult to guarantee the long-term stable operation of measuring instruments. This results in varying degrees of data loss from sensors. Therefore, effectively filling in or generating missing industrial data is of great significance for industrial process monitoring and subsequent engineering and research tasks. Summary of the Invention

[0003] The purpose of this application is to provide a method and system for generating missing industrial data for time-of-use power supply, which can accurately complete the task of generating missing industrial data.

[0004] To achieve the above technical objectives, this application adopts the following technical solution: In a first aspect, this application provides a method for generating missing industrial data for time-of-use power supply, comprising: preprocessing missing industrial time-series data; inputting the preprocessed industrial time-series data into an initialized and trained missing data generation model; and outputting complete industrial data after processing by the missing data generation model. The missing data generation model includes an embedding representation module, an encoding-side composite positional encoding module, a Transformer encoder, a variational autoencoder module, a decoding-side composite positional encoding module, and a Transformer decoder. The embedding representation module takes preprocessed industrial time-series data with missing data as input and obtains the input embedding representation features through linear transformation. The encoding-side composite location encoding module performs composite location encoding on time series points; it combines the composite location encoding result with the input embedded representation feature to output encoded temporal features that fuse time-division location information. The Transformer encoder takes the encoded temporal features as input, extracts contextual features by incorporating a multi-head attention mechanism with learnable time-segment bias, and outputs encoder contextual features. The variational autoencoder module includes an encoding section and a decoding section; The encoding part takes the encoder context features as input, performs convolution operations and feature compression mapping to map the features to the latent space, and outputs latent representation features. Among them, the potential space is a pre-defined regularized feature space that satisfies the standard normal distribution and retains the key trends, time-of-use power supply characteristics, and operating mode of industrial data, in order to ensure that the generated data is stable, continuous, and conforms to the distribution law of industrial data. The decoding part takes the latent representation features as input, performs feature recovery mapping and transpose convolution operation (feature upsampling), and outputs the feature recovery result. The decoding-side composite position encoding module performs composite position encoding on time series points, combines the composite position encoding results with the feature recovery results, and outputs decoding-side features with position information. The Transformer decoder takes the encoder context features and the decoder-side features with position information as input, and reconstructs the data by incorporating an attention mechanism with learnable time-segment bias, ultimately outputting complete industrial data.

[0005] In one possible implementation, the method can be applied to the generation of missing industrial time-series data for the zinc electrolysis process, wherein the missing industrial time-series data is the time-series data of process parameters for the zinc electrolysis process.

[0006] In one possible implementation, the timing data of the process parameters of the zinc electrolysis process include DC current, DC voltage, fresh zinc concentration, waste zinc concentration, waste acid concentration, mixed zinc concentration, and mixed acid concentration.

[0007] In one possible implementation, the composite positional encoding includes: Based on the time-of-use power supply situation, a day is divided into off-peak, normal, and peak electricity consumption periods. The time-sharing power supply function is used to complete the time-segment mapping of time series points and generate time-segment mapping results. Based on the first learnable embedding representation matrix, the time period mapping result is encoded in time period segments to generate relative time period position encoding results; The sine-cosine position coding method is used to perform absolute position coding on time series points, generating absolute time position coding results; The relative time period location coding result and the absolute time period location coding result are concatenated along the feature dimension to obtain the composite location coding result.

[0008] In one possible implementation, the time-sharing power supply function is: ; in, Representing time series points The corresponding time period mapping results are as follows: 0 represents the off-peak electricity consumption period, 1 represents the normal electricity consumption period, and 2 represents the peak electricity consumption period. For time series points, the value range is [0,23], and the sequence length is 24; The one-hot encoding matrix of the time-period mapping result is represented as follows: , used to Convert to a feature representation that can be performed on matrices.

[0009] In one possible implementation, based on the first learnable embedding representation matrix, the time-segment mapping result is time-segmented and encoded to generate a relative time-segment position encoding result, calculated as follows: ; in, This is the result of the relative time period location encoding. For the first learnable embedding representation matrix, This represents the total feature dimension of the model.

[0010] In one possible implementation, a sine-cosine position coding method is used to perform absolute position coding on the time series points, generating an absolute time position coding result. The calculation expression is as follows: ; ; in, This is the result of encoding the absolute time position. Representing time series points The encoded value of the 2i-th feature dimension, i.e., the encoded value of the even-numbered feature dimensions; Representing time series points The encoded value of the (2i+1)th feature dimension, i.e., the encoded value of the odd-numbered feature dimensions; For time series points, Let i be the i-th feature dimension. , These are the sine and cosine functions, respectively.

[0011] In one possible implementation, the relative time period location encoding result and the absolute time period location encoding result are concatenated along the feature dimension to obtain the composite location encoding result, and the calculation expression is as follows: ; in, This is the result of composite position coding. This refers to the connection operation along the feature dimension.

[0012] In one possible implementation, the attention mechanism incorporating learnable time-segment bias includes multiple independent attention heads; Regarding the first Each attention head, the calculation process includes: Based on the The second learnable embedding representation matrix corresponding to each attention head right Encode to obtain the first The time-period encoding matrix corresponding to each attention head ; Time period encoding matrix Its transpose Perform matrix multiplication to obtain the first... The learnable time bias of each attention point ; Bias the learnable time period Join the The process of calculating the attention score of the first attention head yields the result. The calculation results of attention scores with bias in each attention head fusion period; The corresponding calculation expression is: ; in, For the first The time-segment encoding matrix of each attention head. For the first The second learnable embedding representation matrix of each attention head. For the first The learnable time bias of each attention point; Indicates the first The operation of calculating the attention score of each attention head. The first The query vector, key vector, and value vector of each attention head. For the first The feature dimensions of the attention head key vectors satisfy... , This is the activation function.

[0013] All are obtained by linear transformation and segmentation of the input features from the multi-head attention module (Transformer encoder / Transformer decoder).

[0014] Secondly, this application provides an industrial missing data generation system for time-of-use power supply, used to implement the above-mentioned industrial missing data generation method for time-of-use power supply; the system includes: The system includes a data preprocessing module and a missing data generation model module; The data preprocessing module is used to preprocess industrial time-series data containing missing information. The missing data generation model module is used to generate complete industrial data based on preprocessed industrial time-series data. The embedding representation module takes preprocessed industrial time-series data with missing data as input and obtains the input embedding representation features through linear transformation. The encoding-side composite location encoding module performs composite location encoding on time series points; it combines the composite location encoding result with the input embedded representation feature to output encoded temporal features that fuse time-division location information. The Transformer encoder takes the encoded temporal features as input, extracts contextual features by incorporating a multi-head attention mechanism with learnable time-segment bias, and outputs encoder contextual features. The variational autoencoder module includes an encoding section and a decoding section; The encoding part takes the encoder context features as input, performs convolution operations and feature compression mapping to map the features to the latent space, and outputs latent representation features. The decoding part takes the latent representation features as input, performs feature recovery mapping and transpose convolution operations, and outputs the feature recovery result. The decoding-side composite position encoding module performs composite position encoding on time series points, combines the composite position encoding results with the feature recovery results, and outputs decoding-side features with position information. The Transformer decoder takes the encoder context features and the decoder-side features with position information as input, and reconstructs the data by incorporating an attention mechanism with learnable time-segment bias, ultimately outputting complete industrial data.

[0015] Thirdly, this application provides an electronic device, including: a memory and a processor; The memory is used to store computer programs; The processor is used to invoke the computer program to execute the method described above.

[0016] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed on an electronic device, causes the electronic device to perform the method described above.

[0017] Fifthly, this application provides a computer program product, including a computer program that, when run on an electronic device, causes the electronic device to perform the method described above.

[0018] The specific implementation methods of the second to fifth aspects of this application can refer to the implementation methods of the first aspect, and will not be elaborated here.

[0019] Compared with the prior art, the advantages of this application are: To address the issue of missing industrial data during zinc electrolysis under time-of-use power supply, this application improves the existing Transformer position encoding mechanism by designing a learnable and updatable embedding representation matrix to encode the time-of-use power supply periods throughout the day. It also integrates the absolute position information from sine-cosine position encoding to form a composite position encoding method based on relative time period and absolute position. Simultaneously, a learnable and updatable embedding representation matrix is ​​designed within the Transformer's multi-head attention mechanism to calculate the learnable time-of-use power supply period bias, thereby effectively and accurately completing the task of generating missing industrial data. Attached Figure Description

[0020] Figure 1 This is a flowchart of a method in one embodiment of this application.

[0021] Figure 2 This is a visualization of attention weights in one embodiment of this application.

[0022] Figure 3 This is a schematic diagram of the generation result according to one embodiment of this application.

[0023] Figure 4 This is a schematic diagram illustrating the generation results of the comparison method (including the standard Transformer codec and the generative model of variational autoencoder). Detailed Implementation

[0024] The embodiments of this application are described in detail below. These embodiments are based on the technical solutions of this application and provide detailed implementation methods and specific operation processes to further explain the technical solutions of this application.

[0025] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this application described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0026] In this embodiment of the disclosure, the character " / " indicates that the objects before and after it are in an "or" relationship. For example, A / B means: A or B.

[0027] The term "and / or" describes an association between objects, indicating that three relationships can exist. For example, A and / or B means: A or B, or A and B.

[0028] In zinc electrolysis, switching of direct current (DC) often leads to significant changes in operating conditions. DC current is subject to "peak shaving and valley filling" based on time-of-use (TOU) pricing, meaning lower current is used during peak hours and higher current during off-peak hours. Operating conditions are similar within the same time period, resulting in relatively small fluctuations in related industrial data. Therefore, effectively incorporating TOU into the calculations of missing data generation models is of significant research importance.

[0029] Example 1: This embodiment uses the industrial missing data generation method for time-sharing power supply provided in this application to generate missing data for the zinc electrolysis process.

[0030] First, the input data is acquired and constructed. The input data includes DC current, DC voltage, concentration of fresh zinc solution, concentration of waste zinc solution, concentration of waste acid solution, concentration of mixed zinc solution, and concentration of mixed acid solution; the data sequence length is 24. The concentrations of waste zinc solution, waste acid solution, mixed zinc solution, and mixed acid solution contain missing data, with missing values ​​marked as 0. The DC current, DC voltage, and fresh zinc solution concentration do not contain missing data. Then, the data is preprocessed, such as using the max-min normalization method.

[0031] Then, the missing industrial time-series data is input into the trained missing data generation model to obtain the complete data generated by the model output.

[0032] The missing data generation model in this embodiment includes an embedding representation module, an encoding-side composite position encoding module, a Transformer encoder with learnable bias, a variational autoencoder module, a decoding-side composite position encoding module, and a Transformer decoder with learnable bias.

[0033] refer to Figure 1 As shown, the process of generating data from input industrial time-series data containing missing information using a missing data generation model includes: S1: Embedding representation of the input industrial time-series data containing missing data; S2: Perform composite location encoding on time series points, including: (1) Divide the day into different power supply periods according to the time-of-use power supply situation, and complete the time period mapping of the time series points through the time-of-use power supply function. The time-of-use power supply function is calculated as follows: ; in, These are time series points, ranging from [0, 23]. Representing time series points The corresponding time period mapping results are as follows: 0 represents the off-peak electricity consumption period, 1 represents the normal electricity consumption period, and 2 represents the peak electricity consumption period.

[0034] (2) Design the first learnable embedding representation matrix, encode the time period mapping results by time period, form the relative time period position code, and calculate the representation as follows: ; in, This is the result of the relative time period location encoding. The first learnable embedding representation matrix, This represents the total feature dimension of the model.

[0035] (3) Using the sine-cosine position coding method, the time series points are absolutely position coded to form the absolute time position code, which is calculated and expressed as: ; ; in, This is the result of encoding the absolute time position. Representing time series points The encoded value of the 2i-th feature dimension, i.e., the encoded value of the even-numbered feature dimensions; Representing time series points The encoded value of the (2i+1)th feature dimension, i.e., the encoded value of the odd-numbered feature dimensions; For time series points, Let i be the i-th feature dimension. , These are the sine and cosine functions, respectively.

[0036] (4) Concatenate the relative time period location coding result and the absolute time period location coding result along the feature dimension to obtain the composite location coding result. The calculation expression is: ; in, This is the result of composite position coding. This refers to the connection operation along the feature dimension.

[0037] S3: Combine the composite location encoding result with the input embedding representation feature (add element by element) to obtain the encoded temporal feature of fused time-division location information; S4: Input the encoded temporal features into the Transformer encoder with learnable bias. The Transformer encoder extracts contextual features by incorporating a multi-head attention mechanism with learnable temporal bias and outputs encoder contextual features. Among them, the attention mechanism that incorporates learnable time-segment bias includes multiple independent attention heads; Regarding the first Each attention head, the calculation process includes: Based on the The second learnable embedding representation matrix corresponding to each attention head right Encode to obtain the first The time-period encoding matrix corresponding to each attention head ; Time period encoding matrix Its transpose Perform matrix multiplication to obtain the first... The learnable time bias of each attention point ; Bias the learnable time period Join the The process of calculating the attention score of the first attention head yields the result. The calculation results of attention scores with bias in each attention head fusion period; The corresponding calculation expression is: ; in, For the first The time-segment encoding matrix of each attention head. For the first The second learnable embedding representation matrix of each attention head. For the first The learnable time bias of each attention point; Indicates the first The operation of calculating the attention score of each attention head. The first The query vector, key vector, and value vector of each attention head. For the first The feature dimensions of the attention head key vectors satisfy... , This is the activation function.

[0038] All are obtained by linear transformation and segmentation of the input features of the Transformer encoder.

[0039] S5: The encoding part of the variational autoencoder module performs convolution operations on the encoder context features with a kernel size of 24, and then maps the features to the latent space through feature compression mapping (low-dimensional mapping) to output the latent representation features; the decoding part of the variational autoencoder module performs feature recovery mapping (high-dimensional mapping) on ​​the latent representation features, and then performs transposed convolution operations with a kernel size of 24 to output the feature recovery result.

[0040] S6: Decoding-side composite position encoding module, performs composite position encoding on time series points, with the calculation formula as above; combines the composite position encoding result with the feature recovery result to output decoding-side features with position information.

[0041] S7: Transformer decoder, which takes the encoder context features and the decoder side features with position information as input, reconstructs the data by incorporating an attention mechanism with learnable time-segment bias, and finally outputs complete industrial data.

[0042] Among them, the The key and value vectors of each attention head can be obtained through linear transformation and segmentation of the encoder context features. The query vector of an attention head can be obtained by linear transformation and segmentation of the decoded side features with location information.

[0043] See Figure 2 The visualization results of the attention weights are shown, where yellow-green represents high weight and dark blue represents low weight. It can be seen that the model's attention weights are significantly higher within the same time-sharing power supply type, forming a clear weight heatmap that distinguishes between valleys, flat periods, and peak periods. This is because the method in this application performs a biased summation of the original standard attention distribution, indicating that the model effectively learns and utilizes time-sharing information. Figure 3 and Figure 4 The results of generating missing data for the zinc electrolysis process using the method of this application and the comparative method are shown respectively. The figure shows the two-dimensional plane dimensionality reduction result of the generated data output by the model using UMAP (Uniform Manifold Approximation and Projection) technology. Figure 3 and Figure 4 As can be seen from the comparison, the data generated by the method in this application has a higher degree of overlap with the real data, while the overlap is lower for the comparison method that does not incorporate time-of-use power supply information. The above visualization and comparison results demonstrate that incorporating time-of-use power supply information into the model effectively improves the generation effect of missing data.

[0044] Example 2: This embodiment provides an industrial missing data generation system for time-of-use power supply, used to implement the above-described industrial missing data generation method for time-of-use power supply; the system includes: The system includes a data preprocessing module and a missing data generation model module; The data preprocessing module is used to preprocess industrial time-series data containing missing information. The missing data generation model module is used to generate complete industrial data based on preprocessed industrial time-series data. The embedding representation module takes preprocessed industrial time-series data with missing data as input and obtains the input embedding representation features through linear transformation. The encoding-side composite location encoding module performs composite location encoding on time series points; it combines the composite location encoding result with the input embedded representation feature to output encoded temporal features that fuse time-division location information. The Transformer encoder takes the encoded temporal features as input, extracts contextual features by incorporating a multi-head attention mechanism with learnable time-segment bias, and outputs encoder contextual features. The variational autoencoder module includes an encoding section and a decoding section; The encoding part takes the encoder context features as input, performs convolution operations and feature compression mapping to map the features to the latent space, and outputs latent representation features. The decoding part takes the latent representation features as input, performs feature recovery mapping and transpose convolution operations, and outputs the feature recovery result. The decoding-side composite position encoding module performs composite position encoding on time series points, combines the composite position encoding results with the feature recovery results, and outputs decoding-side features with position information. The Transformer decoder takes the encoder context features and the decoder-side features with position information as input, and reconstructs the data by incorporating an attention mechanism with learnable time-segment bias, ultimately outputting complete industrial data.

[0045] Example 3: This embodiment provides an electronic device, including: a memory and a processor; The memory is used to store computer programs; The processor is configured to invoke the computer program to execute the method as described in Embodiment 1.

[0046] Example 4: This embodiment provides a computer-readable storage medium storing a computer program. When the computer program is run on an electronic device, it causes the electronic device to perform the method described in Embodiment 1.

[0047] Example 5: This embodiment provides a computer program product, including a computer program that, when run on an electronic device, causes the electronic device to perform the method described in Embodiment 1.

[0048] The specific implementation of the system, electronic device, computer-readable storage medium, and computer program product provided in this application can be referred to the specific embodiments of the above methods, and will not be repeated here.

[0049] Obviously, those skilled in the art should understand that the various units or steps of this application described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device, or fabricating them separately as individual integrated circuit modules, or fabricating multiple modules or steps into a single integrated circuit module. Thus, this application is not limited to any particular combination of hardware and software.

[0050] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for generating missing industrial data for time-of-use power supply, characterized in that, include: Preprocessing of industrial time-series data containing missing information; The preprocessed industrial time series data is input into the initialized and trained missing data generation model, and the missing data generation model outputs complete industrial data after processing. The missing industrial time-series data refers to the process parameter time-series data of the zinc electrolysis process, including DC current, DC voltage, fresh zinc concentration, waste zinc concentration, waste acid concentration, mixed zinc concentration, and mixed acid concentration. The missing data generation model includes: The embedding representation module takes preprocessed industrial time-series data containing missing data as input and obtains the input embedding representation features through linear transformation. The encoding-side composite location encoding module performs composite location encoding on time series points; it combines the composite location encoding result with the input embedded representation feature to output encoded time-series features that fuse time-sharing location information; wherein, the composite location encoding includes: dividing a day into off-peak, level, and peak electricity consumption periods according to the time-sharing power supply situation; completing the time-sharing point time-period mapping through the time-sharing power supply function to generate time-period mapping results; performing time-period encoding on the time-period mapping results based on the first learnable embedded representation matrix to generate relative time-period location encoding results; using the sine-cosine location encoding method to perform absolute location encoding on the time series points to generate absolute time-time location encoding results; and concatenating the relative time-period location encoding results and the absolute time-time location encoding results along the feature dimension to obtain the composite location encoding result; The Transformer encoder takes the encoded temporal features as input and extracts contextual features through a multi-head attention mechanism incorporating learnable time-segment biases, outputting encoder contextual features; wherein, the attention mechanism incorporating learnable time-segment biases includes multiple independent attention heads; for the first... The calculation process for the attention head includes: based on the first attention head... The second learnable embedding representation matrix corresponding to each attention head right Encode to obtain the first The time-period encoding matrix corresponding to each attention head ; Encode the time period matrix Its transpose Perform matrix multiplication to obtain the first... The learnable time bias of each attention point ; Bias the learnable time period Join the The process of calculating the attention score of the first attention head yields the result. The attention score is calculated based on the attention head fusion time period bias; the corresponding calculation expression is: ;in, For the first The time-segment encoding matrix of each attention head. For the first The second learnable embedding representation matrix of each attention head. For the first The learnable time bias of each attention point; Indicates the first The operation of calculating the attention score of each attention head. The first The query vector, key vector, and value vector of each attention head. For the first The feature dimensions of the attention head key vectors satisfy... , For activation function, This represents the total feature dimension of the model; All by the first The input features of the attention head are obtained by linear transformation; The variational autoencoder module includes an encoding section and a decoding section; The encoding part takes the encoder context features as input, performs convolution operations and feature compression mapping to map the features to the latent space, and outputs latent representation features. The decoding part takes the latent representation features as input, performs feature recovery mapping and transpose convolution operations, and outputs the feature recovery result. The decoding-side composite position encoding module performs composite position encoding on time series points, combines the composite position encoding results with the feature recovery results, and outputs decoding-side features with position information. The Transformer decoder takes the encoder context features and the decoder-side features with position information as input, and reconstructs the data by incorporating an attention mechanism with learnable time-segment bias, ultimately outputting complete industrial data.

2. The method for generating missing industrial data for time-sharing power supply according to claim 1, characterized in that, The time-sharing power supply function is: ； in, Representing time series points The corresponding time period mapping results are as follows: 0 represents the off-peak electricity consumption period, 1 represents the normal electricity consumption period, and 2 represents the peak electricity consumption period. For time series points, the value range is [0,23], and the sequence length is 24; The one-hot encoding matrix of the time-period mapping result is represented as follows , used to Convert to a feature representation that can be performed on matrices.

3. The method for generating missing industrial data for time-sharing power supply according to claim 2, characterized in that, Based on the first learnable embedding representation matrix, the time-segmented mapping results are encoded in time segments to generate relative time-segment position encoding results. The calculation expression is as follows: ； in, This is the result of the relative time period location encoding. is the first learnable embedding representation matrix.

4. The method for generating missing industrial data for time-sharing power supply according to claim 3, characterized in that, The sine-cosine position coding method is used to encode the absolute position of time series points, generating the absolute time position coding result. The calculation expression is as follows: ； ； in, This is the result of encoding the absolute time position. Representing time series points The encoded value of the 2i-th feature dimension, i.e., the encoded value of the even-numbered feature dimensions; Representing time series points The encoded value of the (2i+1)th feature dimension, i.e., the encoded value of the odd-numbered feature dimensions; For time series points, Let i be the i-th feature dimension. , These are the sine and cosine functions, respectively.

5. The method for generating missing industrial data for time-sharing power supply according to claim 1, characterized in that, The relative time period location encoding result and the absolute time period location encoding result are concatenated along the feature dimension to obtain the composite location encoding result. The calculation expression is as follows: ； in, This is the result of composite position coding. This refers to the connection operation along the feature dimension.

6. An industrial missing data generation system for time-sharing power supply, characterized in that, A method for generating missing industrial data for time-sharing power supply as described in any one of claims 1-5; the system comprises: The system includes a data preprocessing module and a missing data generation model module; The data preprocessing module is used to preprocess industrial time-series data containing missing information. The missing data generation model module is used to generate complete industrial data based on preprocessed industrial time-series data. The missing data generation model includes: The embedding representation module takes preprocessed industrial time-series data containing missing data as input and obtains the input embedding representation features through linear transformation. The encoding-side composite location encoding module performs composite location encoding on time series points; it combines the composite location encoding result with the input embedded representation feature to output encoded temporal features that fuse time-division location information. The Transformer encoder takes the encoded temporal features as input, extracts contextual features by incorporating a multi-head attention mechanism with learnable time-segment bias, and outputs encoder contextual features. The variational autoencoder module includes an encoding section and a decoding section; The encoding part takes the encoder context features as input, performs convolution operations and feature compression mapping to map the features to the latent space, and outputs latent representation features. The decoding part takes the latent representation features as input, performs feature recovery mapping and transpose convolution operations, and outputs the feature recovery result. The decoding-side composite position encoding module performs composite position encoding on time series points, combines the composite position encoding results with the feature recovery results, and outputs decoding-side features with position information. The Transformer decoder takes the encoder context features and the decoder-side features with position information as input, and reconstructs the data by incorporating an attention mechanism with learnable time-segment bias, ultimately outputting complete industrial data.

Images