Industrial time series AIGC basic model, electronic equipment and computer program products

Abstract

The embodiment of the present application provides an industrial time series AIGC basic model, electronic device and computer program product. The model includes: an encoding module for capturing the first relevant information of the original time series data in the frequency characteristic domain along the channel dimension, and down-sampling the first relevant information to obtain the encoded time series data, wherein the original time series data includes multiple types of original vibration data, and each channel collects one type of original vibration data; a decoding module for capturing the second relevant information of the encoded time series data along the time dimension, and up-sampling the second relevant information to obtain the decoded time series data; a synthesis module for synthesizing the decoded time series data according to the trend and time period and then splicing them to obtain the target time series data. The method is used to enhance the generation process of the target time series data, thereby generating high-fidelity target time series data.

Description

Technical Field

[0001] The present application relates to the field of time series technology, and in particular to an industrial time series AIGC basic model, electronic equipment and computer program product. Background Art

[0002] With the continuous development of the industrial Internet, industrial enterprises will collect a large amount of time series data and process it in real time during the production and operation process. Time series data is a data sequence recorded in chronological order.

[0003] Existing industrial time series data have the characteristics of poor data quality, diverse types, and high noise. When analyzing industrial time series data, existing models ignore the diversity of frequency variables. However, basic industrial time series information is contained in the frequency domain, such as the frequency of vibration signals. Although existing diffusion models can create time series, they mainly focus on the time domain and ignore the rich information in the frequency domain, resulting in low accuracy of extracted features and difficulty in generating high-fidelity time series data for prediction. Summary of the invention

[0004] The embodiments of the present application provide an industrial time series AIGC basic model, an electronic device, and a computer program product for processing diverse original time series data, while utilizing rich information in the time domain and frequency domain to enhance the generation process of target time series data, thereby generating high-fidelity target time series data.

[0005] An embodiment of the present application provides an industrial time series AIGC basic model, which is used to process multi-channel original time series data to obtain target time series data. The industrial time series AIGC basic model includes: an encoding module, which is used to capture first relevant information of the original time series data in the frequency characteristic domain along the channel dimension, and down-sample the first relevant information to obtain encoded time series data, wherein the original time series data includes multiple types of original vibration data, and each channel collects one type of original vibration data; a decoding module, which is used to capture second relevant information of the encoded time series data along the time dimension, and up-sample the second relevant information to obtain decoded time series data; a synthesis module, which is used to synthesize the decoded time series data according to trends and time periods and then splice them to obtain target time series data.

[0006] The industrial time series AIGC basic model, electronic device and computer program product provided in the embodiments of the present application can process diverse original time series data by generating target time series data from the channel dimension and the time dimension, and synthesizing them according to trends and time periods. At the same time, it can enhance the generation process of target time series data by utilizing rich information in the time domain and the frequency domain, thereby generating high-fidelity target time series data. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0008] Figure 1 Flow chart of the AIGC basic model of industrial time series provided for this application;

[0009] Figure 2 Schematic diagram of the structure of the AIGC basic model of industrial time series provided for this application Figure 1 ;

[0010] Figure 3 Schematic diagram of the structure of the AIGC basic model of industrial time series provided for this application Figure 2 ;

[0011] Figure 4 A schematic diagram of the structure of the encoding module provided for this application;

[0012] Figure 5 A schematic diagram of the structure of the decoding module provided in this application;

[0013] Figure 6 A schematic diagram of the structure of the mask processor provided in this application;

[0014] Figure 7 A schematic diagram of the structure of the synthesis module provided in this application;

[0015] Figure 8 Schematic diagram of the architecture of the AIGC basic model of industrial time series provided for this application;

[0016] Fig. 9 A schematic diagram of the architecture of the synthesis module provided for this application;

[0017] Fig.10 A schematic diagram of the structure of the electronic device provided in this application.

[0018] The above drawings have shown clear embodiments of the present application, which will be described in more detail later. These drawings and text descriptions are not intended to limit the scope of the present application in any way, but to illustrate the concept of the present application to those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0019] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The implementations described in the following exemplary embodiments do not represent all implementations consistent with the present application. Instead, they are merely examples of devices and methods consistent with some aspects of the present application as detailed in the appended claims.

[0020] Figure 1 A flow chart of the industrial time series generation method provided for this application is as follows: Figure 1 As shown, the industrial time series generation method of the present application includes:

[0021] Step S1: acquiring original time series data collected by industrial equipment through multiple channels, wherein the original time series data includes multiple types of original vibration data, and each channel collects one type of original vibration data;

[0022] In a possible implementation, the original vibration data may be vibration data of industrial equipment such as engines, bearings, and other machinery under various operating conditions. It is understood that the present application does not limit the specific source of the original time series data.

[0023] Step S2: Process the original time series data based on the industrial time series AIGC basic model to obtain target time series data, wherein the industrial time series AIGC basic model includes: an encoding module, which is used to capture the first relevant information of the original time series data in the frequency characteristic domain along the channel dimension, and down-sample the first relevant information to obtain encoded time series data; a decoding module, which is used to capture the second relevant information of the encoded time series data along the time dimension, and up-sample the second relevant information to obtain decoded time series data; a synthesis module, which is used to synthesize the decoded time series data according to trends and time periods, and then splice them to obtain target time series data.

[0024] Among them, Artificial Intelligence Generated Content (AIGC) refers to the technology based on artificial intelligence technical methods such as diffusion models, which generates relevant content with appropriate generalization capabilities by learning and identifying the distribution of existing industrial data. The meaning of "foundation" in the industrial time series AIGC basic model is that the model is a core framework, based on which diversified and multi-frequency target time series data can be generated, and it also has the ability to adapt to different industrial application scenarios.

[0025] The industrial time series AIGC basic model provided in this application generates target time series data from the channel dimension and time dimension, and synthesizes them according to trends and time periods. It can process diverse original time series data, and at the same time enhance the generation process of target time series data by utilizing rich information in the time domain and frequency domain, thereby generating high-fidelity target time series data.

[0026] The technical solution of the present application and how the technical solution of the present application solves the above-mentioned technical problems are described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of the present application will be described below in conjunction with the accompanying drawings.

[0027] Figure 2 Schematic diagram of the structure of the AIGC basic model of industrial time series provided for this application Figure 1 .like Figure 2 As shown, the industrial time series AIGC basic model 20 provided in this embodiment includes:

[0028] An encoding module 201 is used to capture first relevant information of the original time series data in a frequency characteristic domain along a channel dimension, and perform downsampling processing on the first relevant information to obtain encoded time series data;

[0029] A decoding module 202, configured to capture second related information of the encoded time series data along a time dimension, and perform upsampling processing on the second related information to obtain decoded time series data;

[0030] The synthesis module 203 is used to synthesize the decoded time series data according to trends and time periods and then splice them to obtain target time series data.

[0031] In a possible implementation, the number of encoding modules and decoding modules of the present application may be N, where N is a natural number. Figure 8 The schematic diagram of the architecture of the AIGC basic model of industrial time series provided for this application, Fig. 9 This is a schematic diagram of the architecture of the synthesis module provided in this application. In order to better describe the industrial time series AIGC basic model provided in this embodiment, the following will be combined with Figure 8 and Fig. 9 Provide specific instructions.

[0032] Figure 3 Schematic diagram of the structure of the AIGC basic model of industrial time series provided for this application Figure 2 .like Figure 3As shown, the industrial time series AIGC basic model 20 provided in this embodiment also includes a diffusion module, and the diffusion module includes a forward diffusion submodule 30, which is used to add preset noise to the original time series data according to a preset diffusion time step to obtain diffused original time series data.

[0033] The industrial time series AIGC base model of this application is a generative base model based on a diffusion probability architecture, designed for industrial time series generation. The diffusion module is able to estimate the distribution of training samples and extract effective data representations from them, thereby modeling complex patterns in time series and supporting a wide range of downstream tasks. In the context of modeling, diffusion is a process in which particles or information spread from high concentration areas to low concentration areas over time, eventually reaching an equilibrium state.

[0034] In a possible implementation, the diffusion process of the present application includes two main stages: forward diffusion and reverse diffusion. Forward diffusion is the stage of adding noise, also known as forward diffusion. Reverse diffusion is the stage of removing noise.

[0035] In the forward diffusion process, the original time series data Follow the distribution In each step of forward diffusion, Gaussian noise is added iteratively on the basis of the forward diffusion. The forward diffusion can be mathematically described by the iterative process of the following formula (1):

[0036] Formula (1)

[0037] in, Represents the noise data when the time step is t, that is, the original time series data has been diffused; is a parameter that controls the amount of noise added at each step, and N represents the normal distribution. At the end of the forward diffusion process, the original time series data is severely corrupted by noise and converted to a distribution that is closer to a standard Gaussian distribution.

[0038] In one possible implementation, Figure 3 As shown, the industrial time series AIGC basic model also includes an embedding module, and the embedding module includes:

[0039] A first embedding module 311 is used to perform latent variable embedding on the diffused original time series data by using one-dimensional convolution to obtain embedded time series data;

[0040] A second embedding module 312 is used to embed the discrete diffusion time steps using sinusoidal position embedding and a fully connected layer to obtain continuous time feature information;

[0041] The third embedding module 313 is used to embed the preset condition information using a masked fully connected layer to obtain vector condition information represented by a vector, wherein the condition information includes conditions for constraining the range of the embedded time series data.

[0042] Among them, latent variables refer to parameters that cannot be directly and accurately observed or that can be observed but still need to be synthesized through other methods. In order to process latent variables, diffusion time steps and conditional information to preserve the time relationship and conditional information, this application uses an embedding module to perform an embedding transformation for each input. Figure 8 In the embedded module, Figure 8 Convolutional embedding, time-step embedding, and conditional information embedding.

[0043] In one possible implementation, the first embedding module uses one-dimensional convolution to embed latent variables to obtain an embedded time series :

[0044] Formula (2)

[0045] Among them, X temb is the embedded time series data after embedding.

[0046] For the diffusion time step t, as shown in formulas (3) and (4), the second embedding module uses the sinusoidal position embedding PosEmbed(·) and the fully connected layer FC(·) to embed the discrete diffusion time step Convert to continuous time feature information :

[0047] Formula (3)

[0048] Formula (4)

[0049] in, represents time encoding, and GeLU represents the activation function.

[0050] For preset condition information , the third embedding module uses a masked FC network for embedding, converting the conditional information into a vector representation:

[0051] Formula (5)

[0052] Among them, X cemb The vector condition information is represented by a vector.

[0053] The industrial time series AIGC basic model of the present application adopts a Unet-like structure to generate target time series data. The main architecture of the industrial time series AIGC basic model includes an encoding module, a decoding module and a synthesis module, which aims to capture and reconstruct time series data.

[0054] Figure 4 This is a schematic diagram of the structure of the encoding module provided in this application. Figure 4 As shown, the encoding module 201 provided in this embodiment includes:

[0055] A first convolution component 40 is used to encode the original time series data through multiple encoder layers to obtain first intermediate data, where the first intermediate data includes the encoded intermediate time series data, continuous time feature information, and vector condition information;

[0056] A cross-channel learning component 41 is used to process the first intermediate data in a frequency characteristic domain along a channel dimension to obtain first relevant information;

[0057] The first sampling component 42 is used to downsample the first related information to obtain encoded time series data.

[0058] In one possible implementation, the first convolution component includes at least two convolution layers, and the convolution layer includes a plurality of encoder layers for encoding the entire embedded time series data. The continuous time feature information and the vector condition information can also be input to each convolution layer and encoded together with the embedded time series data to retain the diffusion step and condition information.

[0059] It is difficult for models in related technologies to learn multi-channel column correlations. Unlike continuous pixel values ​​in images, the dimensions in industrial time series data represent different variables, each with a unique meaning. These data involve multiple variables with complex interdependencies, making it difficult to learn the joint probability between these variables.

[0060] Therefore, it is crucial to consider the dependencies of channels when generating target time series data. This application proposes a cross-channel learning component that promotes communication between different channels by adopting a frequency cross-channel attention mechanism at each time step, thereby achieving learning of the correlations between multiple channels.

[0061] In a possible implementation, the cross-channel learning component 41 is Figure 8 The frequency cross-channel learner of Figure 4 As shown, the cross-channel learning component 41 includes:

[0062] A first frequency converter 411, configured to convert the first intermediate data into a frequency domain along a time step dimension by fast Fourier transform to obtain a first query vector;

[0063] A second frequency converter 412, configured to perform a fast Fourier transform on the product of the first intermediate data and a preset first key vector weight matrix to obtain a first key vector;

[0064] A third frequency converter 413, configured to perform a fast Fourier transform on the product of the first intermediate data and a preset first value vector weight matrix to obtain a first value vector;

[0065] A transposition module 414, configured to transpose the first query vector, the first key vector, and the first value vector, respectively, to obtain a transposed second query vector, a second key vector, and a second value vector;

[0066] The first capturer 415 is used to capture the first relevant information in the frequency characteristic domain along the channel dimension according to the second query vector, the second key vector and the second value vector.

[0067] In a possible implementation, the first intermediate data is converted to the frequency domain along the time step dimension by fast Fourier transform to obtain the first query vector. This process is also called a frequency conversion process. Specifically, the input first intermediate data is converted to the frequency domain along the time step dimension by fast Fourier transform FFT to capture the global frequency information, as shown in formula (6):

[0068] Formula (6)

[0069] in, represents the global frequency information, i.e., the first query vector, T is the number of time steps of each time series, and C is the number of channels; Represents the first intermediate data input to the i-th encoder layer. In this way, the industrial time series AIGC basic model can analyze periodic patterns and frequency components.

[0070] In one possible implementation, in the self-attention mechanism, the first query vector Q is used to ask questions, the first key vector K represents the features to be matched, and the first value vector V represents the corresponding information. In the cross-channel learning component, the first key vector and the first value vector are defined in the following way:

[0071] Formula (7)

[0072] Formula (8)

[0073] in, is the preset first key vector weight matrix, is the preset first value vector weight matrix, both of which are learnable weight matrices.

[0074] After conversion to the frequency domain, as shown in the series of formulas (9), the first query vector Q, the first key vector K and the first value vector V are transposed to obtain the transposed second query vector Qc, the second key vector Kc and the second value vector Vc to focus on the channel dimension:

[0075] Formula (9)

[0076] Then, multivariate correlation modeling is performed through the cross-channel learning component, and the frequency cross-channel attention mechanism is used to capture the dependencies between different variables in the frequency feature domain along the channel dimension:

[0077] Formula (10)

[0078] in, yes The dimension of the vector, Indicates that the softmax function is applied along the channel dimension to normalize the attention weights, so that the network can emphasize the influence of different channels and learn multi-channel correlations in the frequency characteristic domain. chan (Q c , K c , V c ) is the first relevant information.

[0079] In one possible implementation, Figure 4 As shown, the first sampling component 42 includes:

[0080] A first frequency inverter 421 is used to map the first related information back to the original time series domain through an inverse fast Fourier transform to obtain first time series data to be sampled;

[0081] The down-sampling subcomponent 422 is used to perform down-sampling processing on the first time series data to be sampled to obtain encoded time series data.

[0082] In a possible implementation, the first relevant information, namely the frequency attention vector By mapping the first relevant information back to the original time series domain through inverse fast Fourier transform, the first time series data to be sampled is obtained for further downsampling processing, as shown in the following formula (11):

[0083] Formula (11)

[0084] in, is the first time series data to be sampled.

[0085] The downsampling subcomponent may be a 1 / 2 downsampling block. The first time series data to be sampled is downsampled to obtain encoded time series data, as shown in the following formula (12):

[0086] Formula (12)

[0087] in, is the encoded time series data.

[0088] The decoding module of this application includes a cross-temporal learning component, which uses the Fourier domain and cross-temporal and spatial attention mechanisms to capture long-distance dependencies and complex patterns that are difficult to identify in the time domain.

[0089] Figure 5 This is a schematic diagram of the structure of the decoding module provided in this application. Figure 5 As shown, the decoding module 202 provided in this embodiment includes:

[0090] A second convolution component 50 is used to decode the encoded time series data through multiple decoder layers to obtain second intermediate data;

[0091] A cross-time learning component 51 is used to process the second intermediate data along the time dimension to obtain second relevant information of the encoded time series data;

[0092] The second sampling component 52 is used to perform up-sampling processing on the second related information to obtain decoded time series data.

[0093] In one possible implementation, the second convolutional component includes at least two convolutional layers, and the convolutional layer includes multiple decoder layers for decoding the entire encoded time series data. Each decoder layer may correspond to an encoder layer, and the decoding module can enhance details and restore the signal to its original size. An intermediate convolutional block may also be provided between the encoding module and the decoding module to act as a bridge to extract abstract features between the encoder and decoder stages.

[0094] In a possible implementation, the cross-time learning component 51 is Figure 8 The frequency cross-time learner of . Figure 5 As shown, the cross-time learning component 51 includes:

[0095] A fourth frequency converter 511, configured to convert the encoded time series data into a frequency domain along the time dimension by fast Fourier transform to obtain a third query vector;

[0096] A fifth frequency converter 512, configured to perform a fast Fourier transform on the product of the encoded time series data and a preset third key vector weight matrix to obtain a third key vector;

[0097] A sixth frequency converter 513, configured to perform a fast Fourier transform on the product of the encoded time series data and a preset third value vector weight matrix to obtain a third value vector;

[0098] The second capturer 514 is used to capture the second relevant information of the encoded time series data along the time dimension according to the third query vector, the third key vector and the third value vector.

[0099] In one possible implementation, in order to adaptively select important information across spatiotemporal dimensions, this application proposes a frequency-cross-spatial-temporal attention mechanism. This method is similar to the frequency-cross-channel attention mechanism, using FFT and linear transformation on the query matrix, key matrix, and value matrix, but does not transpose the attention matrix, but focuses on the time dimension.

[0100] In a possible implementation, the encoded time series data is converted to the frequency domain by fast Fourier transform along the time dimension to obtain a third query vector Q t , as shown in formula (13):

[0101] Formula (13)

[0102] in, represents the input of the i-th decoder layer, is the encoded time series data.

[0103] In one possible implementation, in the cross-time learning component, the third key vector and the third value vector are defined as follows:

[0104] Formula (14)

[0105] Formula (15)

[0106] in, is the preset third key vector weight matrix, is the preset third value vector weight matrix, both of which are learnable weight matrices.

[0107] Then, the second relevant information of the encoded time series data is captured along the time dimension, as shown in formula (16):

[0108] Formula (16)

[0109] Among them, the third query vector , the third key vector and the third value vector Represents the attention matrix across time in the frequency domain. Different from the frequency cross-channel attention mechanism, the frequency cross-time attention mechanism A softmax function is applied in the time dimension. This attention mechanism operates at different time steps, enabling the model of this application to capture long-range dependencies and dynamic patterns at different time steps.

[0110] In one possible implementation, Figure 5 As shown, the second sampling component 52 includes:

[0111] A mask processor 521, configured to perform adaptive high frequency mask processing on the second query vector to obtain a frequency enhancement parameter;

[0112] A second frequency inverter 522, configured to map the sum of the second related information and the frequency enhancement parameter back to the original time series domain through an inverse fast Fourier transform to obtain second time series data to be sampled;

[0113] The up-sampling subcomponent 523 is used to perform up-sampling processing on the second time series data to be sampled to obtain decoded time series data.

[0114] In a possible implementation, the mask processor adopts the adaptive high-frequency masking solution proposed in the present application, which allows the industrial time series AIGC basic model to dynamically adjust the filtering level and remove high-frequency noise components.

[0115] Figure 6 This is a schematic diagram of the structure of the mask processor provided in this application. Figure 6 As shown, the mask processor 521 provided in this embodiment includes:

[0116] A spectrum calculation component 61, configured to calculate a spectrum representing the intensity of each time series frequency according to the second query vector;

[0117] A filtering component 62, configured to normalize the spectrum and filter out frequency components below a preset training threshold to obtain a filtered spectrum;

[0118] The frequency enhancement component 63 is used to obtain frequency enhancement parameters according to the filtered spectrum and the preset adaptive weight.

[0119] In one possible implementation, in order to further improve the ability of the AIGC basic model of industrial time series to reduce noise while focusing on important frequency components, this application proposes an adaptive filtering mechanism. First, the spectrum representing the intensity of each time series frequency is calculated:

[0120] Formula (17)

[0121] Among them, X FFT That is, the third query vector, X SP The spectrum that characterizes the intensity of each time series frequency, also called the power spectrum, represents the intensity of each time series frequency.

[0122] In one possible implementation, the spectrum is normalized and frequency components below a preset training threshold are filtered out to obtain a filtered spectrum. For comparison, this threshold is back-propagated during training. Automatic optimization. Spectrum below threshold Frequencies are filtered out to reduce noise interference.

[0123] In one possible implementation, adaptive learnable weights are used to enhance the filtered frequency components. The overall process can be expressed as:

[0124] Formula (18)

[0125] Formula (19)

[0126] in, represents element-wise multiplication, Mask represents mask, and in formula (18), the frequency spectrum below the threshold is selectively filtered out, X filtered is the filtered spectrum. represents adaptive learnable weights that can selectively enhance important frequency components, X enhanced is the frequency enhancement parameter.

[0127] In a possible implementation, the second frequency inverter maps the sum of the second related information and the frequency enhancement parameter back to the original time series domain through an inverse fast Fourier transform (IFFT) to obtain the second time series data to be sampled. and the frequency attention vector obtained from the frequency cross-temporal attention mechanism Convert back to the time domain:

[0128] Formula (20)

[0129] in, is the second time series data to be sampled.

[0130] The upsampling subcomponent may be a 2x upsampling block. The second time series data to be sampled is upsampled to obtain decoded time series data, as shown in the following formula (21):

[0131] Formula (21)

[0132] in, Decoded time series data.

[0133] Models in related technologies have difficulty capturing complex temporal dynamic patterns. Industrial time series data are collected from physical equipment, reflecting the dynamic details of the equipment affected by the real environment and its performance over time. These data include long-term degradation processes and short-term dynamic changes, such as temperature and pressure fluctuations. These temporal dynamic changes make it difficult for diffusion models designed for static images to effectively synthesize industrial time series.

[0134] Therefore, in order to generate target time series data, the present application also provides a synthesis module, namely Figure 8 The contrast synthesis module can synthesize trend and seasonal information to reconstruct the time series, and perform denoising based on the initial input to finally generate high-fidelity time series data. By using a deep decomposition model architecture for contrast synthesis, the generated time series can synthesize both long-term degradation information and short-term detail information, promoting the accurate reproduction of the intrinsic temporal dynamics of the data during the generation process.

[0135] Figure 7 This is a schematic diagram of the structure of the synthesis module provided in this application. Figure 7 As shown, the synthesis module 203 provided in this embodiment includes:

[0136] A trend synthesis component 71, used to synthesize the decoded time series data according to the trend to obtain a first time series feature matrix;

[0137] A seasonal synthesis component 72, for synthesizing the decoded time series data according to a time period to obtain a second time series feature matrix;

[0138] A splicing component 73 is used to add the first time series feature matrix and the second time series feature matrix and then splice them with the original time series data to obtain spliced ​​time series data;

[0139] The reconstruction component 74 is used to perform deep processing and reconstruction on the spliced ​​time series data through a residual block composed of multiple layers of convolution and non-linear activation functions to obtain target time series data.

[0140] In one possible implementation, the trend synthesis component 71 is Fig. 9The present embodiment can use polynomial regression to perform trend synthesis. Polynomial regression allows the use of higher-order polynomials to model complex trends, so a method based on polynomial regression is used as a trend synthesis component to synthesize slowly changing long-term trend information.

[0141] Among them, the polynomial regression model can be a linear regression model Based on this, the input feature x is expanded into a polynomial basis matrix P, and the regression coefficients are replaced by the input data feature map Z. The polynomial basis is weighted so that the model can capture nonlinear trends.

[0142] Specifically, the polynomial regression model can be expressed as:

[0143] Formula (22)

[0144] Formula (23)

[0145] Formula (24)

[0146] in, is the bias, P is the polynomial basis space matrix, P i is the i-th row of the matrix P, and H is the preset normalization parameter.

[0147] By normalizing the time step t to , so that the features are on a unified scale, ensuring that the features have a relatively consistent range between different time steps, and avoiding the negative impact of too large or too small values ​​on model training.

[0148] In a possible implementation, the feature map Z is a local feature of the input data obtained by performing one-dimensional convolution and GELU nonlinear transformation on the input data X, that is, a first time series feature matrix, and the input data X is the original time series data.

[0149] In this polynomial regression model, the feature map Z acts as a feature matrix and the basis space P provides polynomial features. This combination actually applies weighting similar to regression coefficients based on the feature map. The model compares the predicted trend value with the actual trend value, calculates the loss and back-propagates, and continuously adjusts the convolution kernel weights and bias terms, so that the feature map Z can better capture the effective features of the input data and adjust the weighting parameters, thereby achieving optimal modeling of the trend of time series data.

[0150] In time series analysis, seasonality refers to recurring periodic patterns in the data. The seasonal composite component72 is Fig. 9In order to model these seasonal patterns, this application uses the Fourier series modeling method to decompose the complex time series into a series of simple sine and cosine functions. The basic structure includes convolutional layers and Fourier transform layers.

[0151] The convolution layer extracts the feature matrix S from the input time series data by performing a one-dimensional convolution operation on the input data X.

[0152] Formula (25)

[0153] In order to transform the output of the convolutional layer as weighted coefficients into Fourier space, the Fourier basis function is defined:

[0154] Formula (26)

[0155] Formula (27)

[0156] Formula (28)

[0157] Among them, P 1 and P 2 Represent the number of sine and cosine basis functions respectively. The Fourier basis matrix F consists of sine basis functions and cosine basis functions Merge to get.

[0158] Then, the feature matrix output by the convolutional layer is multiplied with the Fourier basis matrix as the weighted coefficients of different sine and cosine basis functions to obtain a similar periodic function, thus realizing the modeling of the seasonal component in the time series data:

[0159] Formula (29)

[0160] in, is the second time series feature matrix.

[0161] The seasonal synthesis component compares the predicted seasonal component with the actual seasonal component, calculates the loss and back-propagates, and continuously adjusts the convolution kernel weights and bias terms so that the feature matrix S can better capture the periodic characteristics of the input data, thereby adjusting the weighting coefficients of different frequency components to achieve optimal modeling of the seasonal component.

[0162] Without considering the noise information, relying solely on trend and seasonal components for prediction may result in differences between the restored time series and the original series. To address this problem, this application designs and introduces a comparative prediction function to enhance the model's ability in noise prediction.

[0163] In one possible implementation, the splicing component is Fig. 9 Specifically, the features processed by the trend synthesis component and the seasonal synthesis component are With the original input signal The splicing is performed as shown in formula (30). Among them, the feature processed by the trend synthesis component and the seasonal synthesis component is Z+V, and the original input signal r is the original time series data. The generated features are combined with the original signal, so that the industrial time series AIGC basic model can directly compare the difference between the synthesized time series and the original series.

[0164] Formula (30)

[0165] in, Indicates concatenating two feature tensors along a specific dimension.

[0166] Subsequently, the concatenated feature Xcat is processed by a multi-layer convolution operation W and a nonlinear activation function The residual block composed of θ is deeply processed and reconstructed to extract and enhance the noise component z in the feature.

[0167] Formula (31)

[0168] The concatenated features are input into the convolutional layer, which can map the high-dimensional abstract features to the target output dimension. For example, Fig. 9 The convolution layer is a one-dimensional convolution layer. Moreover, through the convolution operation, the industrial time series AIGC basic model converts the extracted features into actual noise prediction values ​​and makes the dimension of the output data match the dimension of the input data.

[0169] In one possible implementation, Figure 3 As shown, the diffusion module also includes:

[0170] The reverse diffusion submodule 32 is used to obtain the target noise of the target time series data, and perform denoising processing on the original time series data based on the target noise to obtain predicted time series data.

[0171] The back diffusion process aims to gradually denoise the data and recover the original time series data from the noisy observation data using the target noise.

[0172] The denoising model of the present application iteratively removes the added noise. The reverse diffusion process can be described by the following denoising steps:

[0173] Formula (32)

[0174] in, and is the mean and variance of the denoising model prediction at each step t. The denoising model with parameters is used to estimate these parameters through learning, thereby gradually removing the noise and reconstructing the original data to obtain the predicted time series data.

[0175] The electronic device provided in this embodiment can execute the method provided in the above method embodiment, and its implementation principle and technical effect are similar, so this embodiment will not be described in detail here.

[0176] Fig.10 This is a schematic diagram of the structure of the electronic device provided in this application. Fig.10 As shown, the electronic device 100 provided in this embodiment includes: at least one processor 1001 and a memory 1002. Optionally, the device 100 further includes a communication component 1003. The processor 1001, the memory 1002 and the communication component 1003 are connected via a bus 1004.

[0177] In a specific implementation process, at least one processor 1001 executes the computer-executable instructions stored in the memory 1002, so that at least one processor 1001 executes the above method.

[0178] The specific implementation process of the processor 1001 can be found in the above method embodiment, and its implementation principle and technical effect are similar, so this embodiment will not be repeated here.

[0179] In the above embodiments, it should be understood that the processor can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), etc. A general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in the invention can be directly implemented as a hardware processor, or can be implemented by a combination of hardware and software modules in the processor.

[0180] The memory may include a high-speed memory (Random Access Memory, RAM), and may also include a non-volatile memory (NVM), such as at least one disk storage.

[0181] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, the bus in the drawings of this application is not limited to only one bus or one type of bus.

[0182] The present application also provides a computer program product, including a computer program, which implements the above method when executed by a processor.

[0183] The present application also provides a computer-readable storage medium, in which computer-executable instructions are stored. When a processor executes the computer-executable instructions, the above method is implemented.

[0184] The above-mentioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk. The readable storage medium can be any available medium that can be accessed by a general or special-purpose computer.

[0185] An exemplary readable storage medium is coupled to a processor so that the processor can read information from the readable storage medium and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can be located in an application specific integrated circuit (Application Specific Integrated Circuits, referred to as: ASIC). Of course, the processor and the readable storage medium can also exist in the device as discrete components.

[0186] The division of units is only a logical function division, and there may be other divisions in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interface, device or unit, which can be electrical, mechanical or other forms.

[0187] The units described as segmented components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0188] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0189] If the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods of each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0190] Those skilled in the art can understand that all or part of the steps of implementing the above-mentioned method embodiments can be completed by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, the steps of the above-mentioned method embodiments are executed; and the aforementioned storage medium includes: ROM, RAM, disk or optical disk and other media that can store program codes.

[0191] Finally, it should be noted that those skilled in the art will readily conceive of other embodiments of the present invention after considering the specification and practicing the invention disclosed herein. The present invention is intended to cover any variations, uses or adaptations of the present invention, which follow the general principles of the present invention and include common knowledge or customary technical means in the art not disclosed by the present invention, are not limited to the precise structure described above and shown in the drawings, and may be modified and changed in various ways without departing from the scope thereof. The scope of the present invention is limited only by the appended claims.

Claims

1. An industrial time series AIGC basic model, characterized by: The industrial time series AIGC basic model is used to process the original time series data of multiple channels to obtain the target time series data. The industrial time series AIGC basic model includes: An encoding module, used for capturing first relevant information of original time series data in a frequency characteristic domain along a channel dimension, and performing downsampling processing on the first relevant information to obtain encoded time series data, wherein the original time series data includes multiple types of original vibration data, and each channel collects one type of original vibration data; A decoding module, used for capturing the second related information of the encoded time series data along the time dimension, and performing upsampling processing on the second related information to obtain decoded time series data; A synthesis module, used for synthesizing the decoded time series data according to trends and time periods and then splicing them to obtain target time series data; The encoding module comprises: A first convolution component is used to encode the original time series data through multiple encoder layers to obtain first intermediate data, where the first intermediate data includes the encoded intermediate time series data, continuous time feature information, and vector condition information; A cross-channel learning component, used to process the first intermediate data in a frequency characteristic domain along a channel dimension to obtain first relevant information; A first sampling component is used to perform downsampling processing on the first relevant information to obtain encoded time series data; The decoding module comprises: A second convolution component is used to decode the encoded time series data through multiple decoder layers to obtain second intermediate data; A cross-time learning component, used for processing the second intermediate data along the time dimension to obtain second relevant information of the encoded time series data; A second sampling component is used to perform up-sampling processing on the second related information to obtain decoded time series data; The synthesis module comprises: A trend synthesis component, used for synthesizing the decoded time series data according to the trend to obtain a first time series feature matrix; A seasonal synthesis component, used for synthesizing the decoded time series data according to time periods to obtain a second time series feature matrix; A splicing component, used for adding the first time series feature matrix and the second time series feature matrix and splicing them with the original time series data to obtain spliced ​​time series data; The reconstruction component is used to perform deep processing and reconstruction on the spliced ​​time series data through a residual block composed of multi-layer convolution and non-linear activation functions to obtain target time series data.

2. The AIGC basic model of industrial time series according to claim 1 is characterized in that: The cross-channel learning component includes: A first frequency converter, used for converting the first intermediate data into a frequency domain along a time step dimension by fast Fourier transform to obtain a first query vector; a second frequency converter, configured to perform a fast Fourier transform on the product of the first intermediate data and a preset first key vector weight matrix to obtain a first key vector; a third frequency converter, configured to perform a fast Fourier transform on the product of the first intermediate data and a preset first value vector weight matrix to obtain a first value vector; a transposer, configured to transpose the first query vector, the first key vector and the first value vector respectively to obtain a transposed second query vector, a second key vector and a second value vector; A first capturer is used to capture first relevant information in a frequency characteristic domain along a channel dimension according to the second query vector, the second key vector, and the second value vector.

3. The industrial time series AIGC basic model according to claim 2 is characterized in that: The first sampling component comprises: A first frequency inverter, used for mapping the first relevant information back to the original time series domain through inverse fast Fourier transform to obtain first time series data to be sampled; The down-sampling subcomponent is used to perform down-sampling processing on the first time series data to be sampled to obtain encoded time series data.

4. The AIGC basic model for industrial time series according to claim 1 is characterized in that: The cross-time learning component includes: a fourth frequency converter, used for converting the encoded time series data into a frequency domain along the time dimension by fast Fourier transform to obtain a third query vector; a fifth frequency converter, configured to perform a fast Fourier transform on the product of the encoded time series data and a preset third key vector weight matrix to obtain a third key vector; a sixth frequency converter, configured to perform a fast Fourier transform on the product of the encoded time series data and a preset third value vector weight matrix to obtain a third value vector; A second capturer is used to capture second relevant information of the encoded time series data along the time dimension according to the third query vector, the third key vector and the third value vector.

5. The AIGC basic model of industrial time series according to claim 4 is characterized in that: The second sampling component comprises: A mask processor is used to perform adaptive high-frequency mask processing on the third query vector to obtain a frequency enhancement parameter, wherein the mask processor includes: a spectrum calculation component, used to calculate a spectrum representing the intensity of each time series frequency according to the third query vector; a filtering component, used to normalize the spectrum and filter out frequency components below a preset training threshold to obtain a filtered spectrum; a frequency enhancement component, used to obtain a frequency enhancement parameter according to the filtered spectrum and a preset adaptive weight; A second frequency inverter, used for mapping the sum of the second related information and the frequency enhancement parameter back to the original time series domain through an inverse fast Fourier transform to obtain second time series data to be sampled; The up-sampling subcomponent is used to perform up-sampling processing on the second time series data to be sampled to obtain decoded time series data.

6. An electronic device, characterized in that: include: Memory, processor; The memory stores computer-executable instructions; The processor executes the computer-executable instructions stored in the memory, so that the processor executes the industrial time series AIGC basic model as described in any one of claims 1-5.

7. A computer program product, characterized in that It includes a computer program, which, when executed by a processor, implements the industrial time series AIGC basic model described in any one of claims 1 to 5.

Images