Time series data generation method, storage medium, electronic device, and computer program

Abstract

The present application discloses a method for generating time series data, a storage medium, an electronic device, and a computer program. The method includes: processing task data corresponding to a machine device through a device cognitive big model to obtain a description text corresponding to the task data; inputting the description text into a data generation big model connected to the device cognitive big model, controlling the data generation big model to determine the coarse-grained features of the time series data to be generated in the text field, and when obtaining reference time series data corresponding to the time series data to be generated, inputting the reference time series data into the data generation big model, controlling the data generation big model to determine the fine-grained features corresponding to the time series data to be generated in the time series field, where the reference time series data is historical time series data matching the machine device; regulating the time series data output by the data generation big model in real time according to the coarse-grained features and the fine-grained features to obtain the target time series data corresponding to the task data.

Description

Technical Field

[0001] The present application relates to the field of big data, and specifically, to a method for generating time series data, a storage medium, an electronic device, and a computer program. Background Art

[0002] The rise and development of the Industrial Internet has led to a continuous increase in the amount of time series data generated by machinery and equipment, creating an increasingly urgent demand for high-quality time series data resources. Industrial enterprises hope to optimize production processes and improve efficiency by analyzing and mining time series data. However, obtaining high-quality, authentic time series data is difficult and the amount of data available is limited. To address these challenges, the rapid development of generative artificial intelligence (AI) technology has provided powerful tools and methods for generating time series data for machinery and equipment. Generative models can learn the underlying distribution of existing time series data and generate new, diverse, and authentic samples. This expands datasets, increases data diversity, and improves the performance of machine learning algorithms. In the machinery and equipment sector, generating large amounts of high-quality time series data through generative models can help optimize equipment performance and improve applications such as fault diagnosis and predictive maintenance. Generative models can also support system simulation and virtual testing. By generating data with time series characteristics, simulation platforms can be built in real-world scenarios to test and verify the performance, stability, and feasibility of machinery and equipment, reducing the cost and risk of actual testing and increasing efficiency and speed.

[0003] However, the diverse machines and devices in the Industrial Internet scenario have different physical characteristics and behavioral patterns. These devices will generate time series data with different value types, frequencies, periods, and other attributes under different working conditions. Although current time series data generation models based on RNN (Recurrent Neural Network, RNN for short) and GAN (Generative Adversial Network, GAN for short) can simply simulate the time series data of a certain type of machine equipment, they cannot capture the long-term correlations and complex dependencies in complex time series data. The quality of the generated results is usually not high, and it is impossible to achieve precise control over the generated content. In addition, there is currently a lack of sufficient time series data training sets to directly train a large general time series data generation model.

[0004] Regarding the related technologies, the existing technologies for the current time series data generation model have problems such as low generation quality and difficulty in control, and no effective solutions have been proposed yet. Summary of the Invention

[0005] The embodiments of the present application provide a method for generating time series data, a storage medium, an electronic device, and a computer program to at least solve the problems in the related art, such as the low quality of generated results and the difficulty in controlling the current time series data generation model in the existing technology.

[0006] According to one embodiment of the present application, a method for generating time series data is provided, comprising: processing task data corresponding to a machine device through a device cognitive big model to obtain a description text corresponding to the task data, wherein the task data is used to describe the data features of the time series data to be generated by the machine device, the device cognitive big model is a model for generating the description text established by combining a self-training mechanism with a natural language big model, and the description text corresponds to the time series data generated by the machine device; inputting the description text into a data generation big model connected to the device cognitive big model, controlling the data generation big model to determine the coarse-grained features of the time series data to be generated in the text field, wherein the The data generation large model is used to determine coarse-grained features; when reference time series data is obtained, the reference time series data is input into the data generation large model, and the data generation large model is controlled to determine the fine-grained features corresponding to the time series data to be generated in the time series field, wherein the data generation large model is also used to determine fine-grained features, and the reference time series data is historical time series data that matches the machine equipment; according to the coarse-grained features and the fine-grained features, the time series data output by the data generation large model in real time is regulated to obtain the target time series data corresponding to the task data, wherein the data generation large model is also used to generate corresponding time series data based on the coarse-grained features and the fine-grained features.

[0007] According to one embodiment of the present application, a device for generating time series data is provided, comprising: a processing module for processing task data corresponding to a machine device through a device cognitive big model to obtain a description text corresponding to the task data, wherein the task data is used to describe the data features of the time series data to be generated by the machine device, the device cognitive big model is a model for generating the description text established by combining a self-training mechanism with a natural language big model, and the description text corresponds to the time series data generated by the machine device; a first determination module for inputting the description text into a data generation big model connected to the device cognitive big model, and controlling the data generation big model to determine the coarse-grained features of the time series data to be generated in the text field, wherein the The data generation large model is used to determine coarse-grained features; the second determination module is used to input the reference time series data into the data generation large model when the reference time series data is obtained, and control the data generation large model to determine the fine-grained features corresponding to the time series data to be generated in the time series field, wherein the data generation large model is also used to determine fine-grained features, and the reference time series data is historical time series data matching the machine equipment; the control module is used to control the time series data output in real time by the data generation large model according to the coarse-grained features and the fine-grained features, and obtain the target time series data corresponding to the task data, wherein the data generation large model is also used to generate corresponding time series data based on the coarse-grained features and the fine-grained features.

[0008] According to another aspect of an embodiment of the present application, a computer-readable storage medium is further provided, in which a computer program is stored, wherein the computer program is configured to execute the above-mentioned method for generating time series data when running.

[0009] According to another aspect of an embodiment of the present application, an electronic device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the method for generating time series data through the computer program.

[0010] According to another embodiment of the present application, a computer program product is provided, including a computer program, which implements the steps of any of the above method embodiments when executed by a processor.

[0011] In an embodiment of the present application, the task description information corresponding to the time series data to be generated is first processed by the device cognitive large model to obtain text data; then the coarse-grained features of the time series data output by the large model in real time in the text field are generated according to the text data control data, and at the same time, when the reference time series data corresponding to the time series data to be generated is obtained, the fine-grained features corresponding to the time series data output by the large model in real time in the time series field are generated by reference to the time series control data, thereby obtaining the target time series data that satisfies both the coarse-grained features and the fine-grained features, and determining the target time series data as the time series data corresponding to the task data Data; adopting the above scheme, coarse-grained descriptive text is used to guide the time series data to be generated, ensuring that the generated results have the expected abstract features (such as the trend of time series data, the periodicity of time series data, etc.), and then, fine-grained time series guidance is used to ensure that the generated results have consistent specific features (such as the amplitude of time series data, the mean of time series data, the variance of time series data, etc.), which improves the quality of the time series data journey and the controllability of the generation, thereby solving the problem in related technologies that the existing technology has low generation quality and difficult to control for the current time series data generation model. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 This is a hardware structure block diagram of a computer terminal for an optional method for generating time series data according to an embodiment of the present application;

[0013] Figure 2 This is a flow chart of an optional method for generating time series data according to an embodiment of the present application;

[0014] Figure 3 This is a schematic diagram of an optional process for constructing a large device cognitive model according to an embodiment of the present application;

[0015] Figure 4 This is a schematic diagram of an optional cross-domain knowledge mapping based on a conditional diffusion model in the training phase of an embodiment of the present application;

[0016] Figure 5 This is an optional schematic diagram of cross-domain knowledge mapping based on the conditional diffusion model in the reasoning stage of an embodiment of the present application;

[0017] Figure 6 This is an optional schematic diagram of generating time series data based on two-stage optimization in an embodiment of the present application;

[0018] Figure 7 It is an optional time series data generation architecture based on cross-domain multi-model collaboration in an embodiment of the present application;

[0019] Figure 8 This is an optional flowchart of an embodiment of the present application for generating time series data based on a time series data generation architecture with cross-domain multi-model collaboration;

[0020] Figure 9 This is an optional schematic diagram of data flow in the process of generating time series data in an embodiment of the present application;

[0021] Figure 10 This is a structural block diagram of a device for generating time series data according to an embodiment of the present application. DETAILED DESCRIPTION

[0022] In order to enable those skilled in the art to better understand the present invention, the following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of this application.

[0023] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequential order. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in a sequence other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0024] The method embodiments provided in the embodiments of the present application can be executed on a computer terminal or a similar computing system. Taking running on a computer terminal as an example, Figure 1 This is a hardware structure block diagram of a computer terminal for a method of generating time series data according to an embodiment of the present application. Figure 1 As shown, the computer terminal may include one or more ( Figure 1 Only one is shown) a processor 102 (the processor 102 may include but is not limited to a processing system such as a microprocessor MCU or a programmable logic device FPGA) and a memory 104 for storing data. In an exemplary embodiment, the computer terminal may further include a transmission device 106 and an input / output device 108 for communication functions. It will be understood by those skilled in the art that Figure 1 The structure shown is only for illustration and does not limit the structure of the above-mentioned computer terminal. For example, the computer terminal may also include Figure 1 More or fewer components than shown, or with Figure 1 Equivalent functions or comparisons shown Figure 1 Shown are different configurations with more functionality.

[0025] The memory 104 can be used to store computer programs, for example, software programs and modules of application software, such as the computer program corresponding to the method for generating time series data in the embodiment of the present application. The processor 102 executes various functional applications and data processing by running the computer program stored in the memory 104, that is, implementing the above method. The memory 104 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage systems, flash memory, or other non-volatile solid-state memory. In some instances, the memory 104 may further include a memory remotely located relative to the processor 102, and these remote memories can be connected to the computer terminal via a network. Examples of the above-mentioned network include but are not limited to the Internet, an intranet, a local area network, a mobile communication network, and a combination thereof.

[0026] Transmission device 106 is used to receive or transmit data via a network. A specific example of the aforementioned network may include a wireless network provided by a communications provider of a computer terminal. In one embodiment, transmission device 106 includes a network interface controller (NIC), which can be connected to other network devices via a base station to enable communication with the Internet.

[0027] In this embodiment, a method for generating time series data is provided, which is applied to the above-mentioned computer terminal. Figure 2 This is a flowchart of an optional method for generating time series data according to an embodiment of the present application, which includes the following steps:

[0028] Step S202: Process the task data corresponding to the machine device through the device cognitive macro model to obtain a description text corresponding to the task data, wherein the task data is used to describe the data characteristics of the time series data to be generated by the machine device. The device cognitive macro model is a model established through a self-training mechanism combined with a natural language macro model to generate the description text, and the description text corresponds to the time series data generated by the machine device.

[0029] Optionally, before executing the above step S202, in order to ensure the accuracy of the device cognitive model in recognizing the task description information, it is necessary to construct the device cognitive model in the following way: Figure 3 This is a schematic diagram of an optional process for constructing a large device cognitive model according to an embodiment of the present application; specifically, the following steps are included:

[0030] Step S21, selection step: determining the natural language large model according to preset conditions;

[0031] Step S22, labeling step: When the preset seed data corresponding to the machine device is obtained, target data similar to the preset seed data is determined from the unlabeled dataset, and the target data is labeled using the natural language large model to obtain multiple pairs of target data corresponding to the machine device, wherein the preset seed data includes at least one pair of data, each data pair consisting of device information of a machine device and reference description text describing time series data generated by the machine device, and the unlabeled dataset is text associated with the machine device intercepted from the network;

[0032] Step S23, screening step: selecting training data from the plurality of target data pairs according to preset screening rules;

[0033] Step S24, training step: using the training data and the preset seed data to perform model training on the natural language large model to obtain a target training model corresponding to the natural language large model;

[0034] Step S25, looping through the labeling step, the screening step, and the training step, and determining the target training model that completes each round of training as the initial model of the natural language large model in the next round of model training.

[0035] Step S26: Acquire a multiple training model obtained by training the natural language large model for the Nth time, wherein N is an integer greater than 1, and the multiple training model is a target training model obtained by completing the model training for the natural language large model for the Nth time;

[0036] Step S27: When the recognition accuracy of the multiple training models for the machine device meets a preset expected value, the multiple training models are determined as the device recognition macro model corresponding to the machine device.

[0037] As an optional implementation method, considering that the current industrial scenarios lack large-scale data sets to directly train general time series data generation big models, the time series data generation problem is decomposed into building a device cognitive big model (solving the problem of what to generate) and a time series data generation big model (solving the problem of how to generate), and a more general and accurate time series data generation capability is constructed through model integration. Specifically, first, a device cognitive big model oriented to the physical characteristics and behavioral mechanisms of machine equipment is established to summarize and understand the operating mechanisms and behavioral performances of different machine equipment in specific scenarios. On this basis, the sequence modeling capabilities and multi-sequence relationship modeling of the big model are used to improve the data generation capability, and the domain knowledge generated by the cognitive big model is integrated into the time series data generation process, and the time series generation big model is trained by combining prior knowledge and data-driven methods.

[0038] In order to build a cognitive big model for industrial Internet machines and equipment, we can use the current mainstream natural language big models, such as ChatGLM, LLaMA, etc., as the base model, and use LoRA (Low-Rank Adaptation, low-order adaptation, referred to as LoRA) and other methods to fine-tune the model. Considering the current lack of data sets to support the training of the above-mentioned device cognitive big models, the optional embodiment of this application proposes a training method of a self-training mechanism, which uses the text summarization extraction ability of the language big model to construct a training data set, and iteratively screens out high-quality data sets to train the device cognitive big model. Specifically, the self-training method starts with a basic language big model, for example, ChatGPT as a seed model, as well as a small amount of seed data and a large amount of unlabeled data sets.

[0039] Specifically, the training process of the device recognition model is as follows:

[0040] In step one, the basic language model is fine-tuned on seed data to learn how to describe time series features for different devices, generating training data pairs for subsequent tuning of the device recognition model. Then, because the basic language model inherently possesses advanced capabilities for text summarization, dialogue generation, and translation, it can automatically filter and summarize massive unlabeled datasets and generate device-time series feature description pairs in the aforementioned format.

[0041] In step 2, to select high-quality sample data for training the device recognition model, the previously trained model is used to filter the generated machine device-time series data feature description pairs during the current round of device recognition model training. This results in high-quality sample data. This sample data is then used together with the seed data to train the device recognition model.

[0042] Step 3: Iterate the above steps until a large device recognition model that meets the expectations is trained.

[0043] Optionally, the above-mentioned seed data can be a machine equipment-time series data feature description data pair, for example, PLC equipment-PLC time series data has characteristics such as periodicity, discreteness, multivariateness, abnormality and time consumption; the above-mentioned unlabeled data set can be a network corpus, etc. It should be noted that unlabeled data is often a large number of documents of various forms, including various topics related to industrial Internet equipment. The above is only an example and does not limit this application.

[0044] It should be noted that the device recognition model has the ability to recognize different machines and devices. During the time series data generation process, the device recognition model first generates text about the time series characteristics of the current machine and device, and inputs this text as prior knowledge into the time series data model. This provides support for process control, attribute identification, result analysis, and optimization during the machine and device time series data generation process.

[0045] Step S204: input the description text into a data generation model connected to the device cognitive model, and control the data generation model to determine coarse-grained features of the time series data to be generated in the text field, wherein the data generation model is used to determine the coarse-grained features;

[0046] Step S206, when the reference time series data is obtained, the reference time series data is input into the data generation big model, and the data generation big model is controlled to determine the fine-grained features corresponding to the time series data to be generated in the time series field, wherein the data generation big model is also used to determine the fine-grained features, and the reference time series data is the historical time series data that matches the machine equipment; it can be understood that by obtaining the historical time series data of the machine equipment as a reference, the time series data that meets the machine equipment can be output in real time through the data generation big model, thereby improving the accuracy of time series data generation. Optionally, the historical time series data is determined based on the existing or historically generated time series data of the machine equipment.

[0047] As an optional implementation, before inputting the descriptive text into the data generation big model connected to the device cognitive big model, the method also includes: establishing a cross-domain knowledge mapping between the device cognitive big model and the data generation big model; wherein, the establishment of the cross-domain knowledge mapping between the device cognitive big model and the data generation big model includes: obtaining reference text data for training the device cognitive big model, and obtaining reference time series data for training the data generation big model, wherein the reference text data corresponds to the reference time series data; using the first type of representation model corresponding to the text field to extract knowledge representation from the reference text data to obtain a first representation; using the second type of representation model corresponding to the time series field to extract knowledge representation from the reference time series data to obtain a second representation; associating the first representation and the second representation according to a preset conditional diffusion model to obtain a target mapping relationship, and determining the mapping result of the cross-domain knowledge mapping corresponding to the data generation big model and the device cognitive big model based on the target mapping relationship.

[0048] In an exemplary embodiment, the first representation and the second representation are associated according to a preset conditional diffusion model to obtain a target mapping relationship, including: processing the first representation and the second representation respectively according to the preset conditional diffusion model, wherein the conditional diffusion model at least includes: an unconditional generation sub-model and a conditional generation sub-model, and the output result of the unconditional generation sub-model is allowed to serve as a guiding signal for the conditional generation sub-model; using the first representation and the reference text data to train the unconditional generation sub-model to obtain a first sub-model; using the second representation and the reference time series data to train the conditional generation model to obtain a second sub-model; when the first representation is set as the basic condition for calling the second representation, determining the target association relationship according to the first output result corresponding to the first sub-model and the second output result corresponding to the second sub-model; and determining the target mapping relationship between the first representation and the second representation according to the target association relationship.

[0049] It is understandable that in order to introduce the output of the device cognitive large model into the time series generation large model, it is necessary to solve the heterogeneous mapping and fusion problems of knowledge in different fields. To this end, in the optional embodiments of the present application, representation models of different fields are used to extract knowledge representations of text data and time series data respectively. For example, BERT is used to perform representation learning on text data and autoencoders are trained to perform representation learning on time series data. After obtaining the knowledge representations of different fields, the present application uses the conditional diffusion model (DM Diffusion Model, diffusion model, referred to as Classifier-FreeDM) to use the knowledge representation of the text field as a guide to control the generation of knowledge representation of the time series field, thereby realizing the mapping learning of knowledge in different fields.

[0050] Optionally, cross-domain knowledge mapping based on the conditional diffusion model can be performed as follows: Figure 4 As shown, Figure 4 This is an optional schematic diagram of cross-domain knowledge mapping based on the conditional diffusion model in the training phase of an embodiment of the present application: wherein the training phase includes: a text domain phase 42, a time series domain phase 44, and a conditional diffusion model 46;

[0051] The text domain stage 42 processes the text data x corresponding to the training data pair through an expert model f(x) trained in the text domain; optionally, the above f(x) can be a classification model, a generation model, etc.

[0052] Specifically, the text domain stage 42 may perform the following steps:

[0053] Step 1: Determine the composition of f(x), i.e., f(x)=Dx(Ex(x)), where Dx is the sub-model of the front part of the expert model that encodes the text data, and Ex is the sub-model of the back part of the expert model that decodes the text data;

[0054] Step 2: Use the Ex and Dx sub-models in the expert model to perform dimensionality reduction processing on the text data x to obtain extracted data that can extract the knowledge representation corresponding to the text data x;

[0055] Step 3: Extract the knowledge representation corresponding to the text data x and obtain the knowledge representation Zx corresponding to the text field; Zx = Ex(x) represents the knowledge representation about x learned by f(x).

[0056] That is, after determining the corresponding expert model in the text domain stage, for any text, the expert model f(x) can be used to quickly determine the knowledge representation Zx corresponding to the text.

[0057] The text domain stage 42 processes the text data x corresponding to the training data pair through an expert model f(x) trained in the text domain; optionally, the above f(x) can be a classification model, a generation model, etc.

[0058] Specifically, the time series domain stage 44 processes the time series data y corresponding to the training data pair through the expert model g(y) trained in the time series domain; optionally, the above g(y) can be a classification model, a generation model, etc.

[0059] Specifically, the timing domain stage 44 may perform the following steps:

[0060] Step 1: Determine the composition of g(y), i.e., g(y)=Dy(Ey(y)), where Dy is the sub-model of the front part of the expert model that encodes the text data, and Ey is the sub-model of the back part of the expert model that decodes the text data;

[0061] Step 2: Use the Dy and Ey sub-models in the expert model to perform dimensionality reduction on the time series data y to obtain extracted data that can be used to extract the knowledge representation corresponding to the time series data y;

[0062] Step 3: Extract the knowledge representation corresponding to the time series data y and obtain the knowledge representation Zy corresponding to the time series domain; Zy = Ey(y) represents the knowledge representation about y learned by f(y).

[0063] That is, after determining the corresponding expert model in the time series domain, for any time series data, the expert model g(y) can be used to quickly determine the knowledge representation Zy corresponding to the time series data.

[0064] It should be noted that, since Zx and Zy are usually heterogeneous and misaligned, this application uses a Classifier-Free conditional diffusion model to achieve the mapping between Zx and Zy to avoid the above-mentioned heterogeneous and misaligned situation. Specifically, in the conditional diffusion model stage 46, the knowledge representation Zx is set as the text guide of the knowledge representation Zy during the training process, thereby achieving the alignment of Zx and Zy.

[0065] It can be understood that in the forward diffusion process, for any data x sampled from Zx t , continuously adding noise to the training data, where β t is a constant that controls the noise step size. t |x t-1 ) is the distribution of added noise, is the standard Gaussian distribution corresponding to the noise, and I is the Gaussian noise range parameter. Then, in the reverse denoising process, the required data samples are constructed from the noise by reverse learning.

[0066] p θ，ξ (x t-1 |x t )=Zp θ (x t-1 |x t )p ξ (Z x |x t-1 );

[0067]

[0068] Among them, Z is a constant used for normalization, p ξ (Z x |x t-1 ) represents the intermediate generation result, x t-1 Belong to Z x The probability of reverse denoising learning is to θ and variance Make predictions, θ is the model parameter of the conditional diffusion model, Pθ is the distribution of the predicted noise to be added, and p θ，ξ (x t-1 |x t ) is the reverse process corresponding to the forward diffusion process. Indicates that it contains the mean μ θ and variance Gaussian distribution.

[0069] Compared to Classifer Gudiance DM, which requires an additional classifier based on noise data, Classifier-Free DM replaces the explicit classifier with an implicit classifier, eliminating the need to directly calculate the explicit classifier and its gradient. According to the Bayesian formula, the diffusion score update formula of Classifier-Free DM can be expressed as:

[0070]

[0071] From the above, we can see that the new generation process no longer depends on the displayed classifier, where ω represents the weight of the guidance signal. is the gradient of the implicit classifier, is the updated diffusion score, ∈ θ (x t ,t,Z x ) is the diffusion score before update corresponding to the conditional generation case, α is a constant, ∈ θ (x t , t) is the diffusion score before update corresponding to the unconditional generation. That is, in the sampling stage, the noise predicted under conditional control and the noise predicted under unconditional control are combined to determine the final noise.

[0072] In addition, in the training process of the conditional diffusion model, in order to enable the conditional diffusion model to achieve more comprehensive mapping learning of knowledge in different fields, in actual applications, this application can also automatically construct training data pairs by utilizing existing algorithm tools, that is, construct training data pairs that generate time series data-text. Optionally, mainstream statistical algorithms, time series algorithms, machine learning algorithms, etc. are used to perform period analysis, amplitude calculation, anomaly detection, etc. on the time series data, and the analysis results are expressed in natural language. For example, by performing period detection, trend analysis, anomaly detection, etc. on the given time series data, the following text description is automatically generated: "This time series data is a square wave signal with a duration of 180 minutes, containing 3600 sampling points, with a period of 20 minutes, a duty cycle of 40%, a maximum value of 2, and a minimum value of -1", "This time series data is a discrete data with a duration of 30 days, showing an oscillating upward trend, gradually rising from 12.6 to 48.9".

[0073] The above method allows for textual description of any generated time series data. Conditional diffusion models are trained based on the obtained training data pairs. Since the current Classifier-Free DM conditional diffusion model can be initially viewed as consisting of two submodels: an unconditional generative submodel and a conditional generative submodel, these two models can be represented by the same parameter model. During training, it is sufficient to simply set the conditions to null with a certain probability. Specifically, during training, an empty marker symbol can be added to the parameter model as a category identifier. When an empty marker symbol appears in both training data pairs, it indicates that the training process is for the unconditional generative submodel; otherwise, it indicates that the training process is for the conditional generative submodel.

[0074] During the inference phase, after determining the new text data corresponding to the text domain, linear extrapolation can be performed through a conditional diffusion model containing a conditional generation sub-model and an unconditional generation sub-model to obtain the time series data that the new text data should correspond to in the time series domain.

[0075] Specifically, Figure 5 This is an optional schematic diagram of cross-domain knowledge mapping based on the conditional diffusion model in the reasoning phase of an embodiment of the present application. Specifically, knowledge mapping from the text domain to the time series domain includes the following steps:

[0076] Step S502: For any text data x, use the expert model f corresponding to the text domain stage to obtain the knowledge representation Zx.

[0077] Step S504: Use Zx as a guidance signal for completing the training of the conditional diffusion model, and instruct the conditional diffusion model to generate a corresponding knowledge representation Zy in the time series domain.

[0078] Step S506: Use the reconstruction model to decode Zy in the time series field to obtain corresponding time series data, and determine the time series data as the time series data corresponding to the above-mentioned arbitrary text data x.

[0079] In summary, a cross-domain knowledge mapping method is established based on the conditional diffusion model to achieve the association between the time series domain and the text domain. Existing representation models in different domains are used to extract knowledge representations for text data and time series data respectively. Based on the conditional diffusion model, the knowledge representation of the text domain is used as a guide to control the generation of knowledge representation in the time series domain. The generated knowledge representation in the time series domain is then decoded to generate the corresponding time series data, thereby achieving mapping learning of knowledge in different domains. In addition, the knowledge acquired in the text domain is introduced into the time series domain. By combining prior knowledge and data-driven methods, the efficiency of model training is improved, the problem of missing large-scale labeled time series data training sets is alleviated, and the effect of improving cross-domain knowledge fusion is achieved.

[0080] In an exemplary embodiment, after determining the mapping results of the cross-domain knowledge mapping corresponding to the data generation big model and the device cognitive big model based on the target mapping relationship, the method further includes: establishing a generation process for generating the time series data through the data generation big model and the device cognitive big model according to the mapping results; wherein, the generation process includes: a first step, determining the first requirement for the data generation big model to generate the time series data to be generated by the descriptive text output by the device cognitive big model, and determining the first target representation corresponding to the time series data output in real time in the text field and the target coarse-grained features included in the first target representation according to the descriptive text; a second step, when the reference time series data is input into the data generation big model , determine the second target representation corresponding to the time series data output in real time in the time series field and the target fine-grained features included in the second target representation according to the reference time series data, and determine the first target representation as the guidance signal of the second target representation according to the mapping result, and obtain the second requirement of the data generation large model to generate the time series data to be generated, wherein the second target representation is a representation with the same features as some of the features in the first target representation; the third step, when the data generation large model generates the corresponding knowledge representation based on the target fine-grained features and the target coarse-grained features, the knowledge representation is decoded by the expert model existing in the data generation large model, and the time series data output in real time by the data generation large model is determined according to the decoding result.

[0081] In simple terms, the above three steps can be expressed through a two-stage optimization process when applied to the above-mentioned device cognitive model and data generation model. Specifically, due to the problem of missing information when describing data in text form, the above-mentioned DM generation process based on coarse-grained text guidance can ensure that the abstract features of the generated results, such as trends, periodicity, etc., are consistent with the expected expression, but it cannot guarantee that the generated time series data has the expected fine features, such as single-point sampling values, local changes, etc. Moreover, for randomly sampled noise inputs, the generation results of DM may be different data with the same abstract features but different real values. To this end, the present application uses a time series data generation process based on two-stage optimization, Figure 6 This is an optional schematic diagram of generating time series data based on two-stage optimization in an embodiment of the present application;

[0082] Optionally, in stage one, the above text is used as a guide signal to ensure that the generated time series data has consistent abstract features with the specific machine equipment. In stage two, the existing or historically generated time series data is used as a time series fine-grained guide. In the process of generating long-term time series data, based on the cyclic generation strategy, the generation results of the previous round are determined as the time series fine-grained guide for the current generation round, and data generation is achieved through iteration. On the one hand, it can solve the problem of generating long-term data, and on the other hand, it can ensure that the generated results have consistent statistical characteristics, such as amplitude, mean, variance, etc.

[0083] In summary, by introducing a two-stage guidance optimization mechanism, on the one hand, coarse-grained text guidance is used to ensure that the generated results have the expected abstract features, and on the other hand, fine-grained time series guidance is used to ensure that the generated results have consistent specific features. This can simultaneously improve the quality and controllability of the generated results, effectively improving the quality and controllability of the generated time series data.

[0084] Step S208, regulating the time series data output in real time by the data generation large model according to the coarse-grained features and the fine-grained features, and obtaining target time series data corresponding to the task data, wherein the data generation large model is also used to generate corresponding time series data based on the coarse-grained features and the fine-grained features.

[0085] Optionally, the data period corresponding to the target time series data required in the actual scenario is different. When the task data indicates that long-period data needs to be generated, the cyclic generation strategy of the data generation large model is started, wherein the cyclic generation strategy is used to indicate that when the long-period data is not completely generated, the target time series data generated in the previous round of the data generation large model is determined as the reference time series data for the current generation round, and new fine-grained features corresponding to the time series field are obtained. The time series data output in real time is iteratively generated according to the coarse-grained features and the new fine-grained features to obtain iterative time series data, so as to determine the long-period data according to the iterative time series data; the iterative time series data generated by the data generation large model executing the cyclic generation strategy is obtained, and the data period length corresponding to the iterative time series data is counted; when the data period length is greater than or equal to the target data period length corresponding to the long-period data, the current iterative time series data output in real time by the data generation large model is determined as the target time series data corresponding to the task data.

[0086] Through the above steps, the task description information corresponding to the time series data to be generated is first processed by the device cognitive large model to obtain text data; then the coarse-grained features of the time series data output by the large model in real time in the text field are generated according to the text data control data. At the same time, when the reference time series data corresponding to the time series data to be generated is obtained, the fine-grained features corresponding to the time series data output by the large model in real time in the time series field are generated by the reference time series control data, thereby obtaining the target time series data that satisfies both the coarse-grained features and the fine-grained features, and determining the target time series data as the time series data corresponding to the task data. According to the above scheme, coarse-grained descriptive text is used to guide the time series data to be generated, ensuring that the generated results have the expected abstract features (such as the trend of time series data, the periodicity of time series data, etc.). Then, fine-grained time series guidance is used to ensure that the generated results have consistent specific features (such as the amplitude of time series data, the mean of time series data, the variance of time series data, etc.), which improves the quality of the time series data process and the controllability of the generation, thereby solving the problems in related technologies that the existing technology has low generation quality and difficult to control for the current time series data generation model.

[0087] Simply put, through the above-mentioned time series data generation method, the general time series data generation problem is decomposed into building a large device recognition model (solving the problem of what to generate) and a large time series data generation model (solving the problem of how to generate). The former is responsible for recognizing the characteristics of the time series data of machine equipment (text domain), and the latter integrates the domain knowledge of the former into the time series data generation process. Through model integration, a more general time series data generation method is constructed, thereby enabling the generation of general time series data for machine equipment.

[0088] In an exemplary embodiment, task data corresponding to a machine device is processed by a device cognitive big model to obtain a description text corresponding to the task data, including: determining the parsing result of the device cognitive big model on the task data; expressing the model code corresponding to the parsing result in natural language to obtain the description text corresponding to the task data.

[0089] It should be noted that the model optimization method (i.e. the above-mentioned preset training method) can be a stochastic gradient descent method or other optimization algorithms; this application does not impose any restrictions on this.

[0090] Obviously, the embodiments described above are only part of the embodiments of this application, not all of them. In order to better understand the above-mentioned method for generating time series data, the above process is described below in conjunction with the embodiments, but it is not intended to limit the technical solutions of the embodiments of this application. Specifically:

[0091] A GAN-based time series data generation model has been proposed. This model uses a generative adversarial approach to construct a generator with recurrent units, enabling the generation of complex sequential time series data. Traditional GAN ​​models fail to capture the complex associations between temporal features and their associated (invariant) attributes, as well as the long-term correlations within the time series, when modeling time series. This solution uses two networks—a multi-layer perceptron (MLP) and a recurrent network—to capture temporal dependencies. At the same time, it separates attribution generation to better capture the correlations between the time series and its attributes (such as user age, location, and gender). This solution also has the ability to generate data features conditioned on data attributes. Its conditional generative architecture provides the flexibility to modify and conditionally manipulate attribute distributions, helping to conceal attribute distributions and enhance privacy. Furthermore, to address the imbalanced distribution of training data during GAN training, this solution proposes an automatic normalization scheme. This involves normalizing data features before training and adding the minimum and maximum ranges of the features as two additional attributes for each sample. This involves adding a normalization factor to the generator to constrain the feature range, accelerating GAN convergence and preventing mode collapse. The above scheme can effectively expand the GAN-based time series data generation model to capture dependencies and correlations within long time series, thereby generating complex time series data.

[0092] However, while the aforementioned solution uses GANs to construct a time series generation model, capable of generating high-quality sequential time series data, it is limited by the generation capabilities of GANs and the long-term forgetting flaw of RNNs, making it ineffective for generating highly complex, long-duration time series data. Furthermore, the aforementioned solution lacks universal time series data generation capabilities. Generating different types of time series data requires different training datasets for model training. Therefore, it is only suitable for applications with limited scale and fixed tasks, and is not suitable for scenarios requiring the generation of complex, diverse time series data.

[0093] Another prior art strategy proposes a time series forecasting strategy based on a large model. This is the first exploration of large models in time series forecasting, leveraging large models to enhance multimodal data processing and interpretability in financial forecasting scenarios. This solution primarily consists of four components: time series processing, textual information introduction, prompt design, and large model fine-tuning using instruction tuning. First, the time series is discretized and converted into symbolic representations for easier input into the large model. Next, GPT4 is used to generate information, including a company overview, textual information such as strengths and weaknesses that may influence stock prices, as well as summary generation and keyword extraction of news information for subsequent input. Next, predictions are made based on prompts. Building on this foundation, this solution further incorporates instruction tuning, constructing 30k weekly forecasts and 7k monthly forecasts using the aforementioned method. Instruction tuning is performed using the open-source LLaMa model. Implementing this solution enables effective forecasting of time series data.

[0094] However, the above solution can only use large models to capture the dependencies or associations of complex time series data and improve time series prediction performance by introducing multimodal information. Furthermore, the above solution simply converts the time series prediction task into a text task, which can only solve coarse-grained simple prediction problems and cannot guarantee optimal prediction results. Furthermore, the above solution learns the associations within time series data (Instruction Tuning) through text learning, which is essentially still text learning and cannot truly learn the regular characteristics of time series data.

[0095] To address these challenges, it is necessary to build a time series data generation model for industrial Internet machines and devices. First, we need to research controllable, general-purpose time series data generation models that can capture the complex temporal correlations and long-term dependencies in time series data and intelligently generate time series data with desired attributes based on the physical and behavioral characteristics of the machines and devices. Traditional time series data generation models based on RNNs, GANs, and other techniques are generally unable to model the dependencies in complex time series data and require training separate generation models for different machines and devices, making them less universal. To this end, we must first establish a multi-faceted machine knowledge framework that can understand the physical characteristics, operating mechanisms, and behavioral performance of different machines and devices. Based on this, we can achieve high-quality generation of time series data with arbitrary attributes. Second, we need to address the shortage of training datasets for time series data. Directly training large, general-purpose generative models, such as ChatGPT and LLaMA, using large amounts of annotated time series data is unrealistic. Therefore, we need to develop efficient model learning strategies that incorporate knowledge from other domains into the time series generation task, addressing the shortage of training data through a combined prior knowledge and data-driven approach.

[0096] Therefore, in the process of building a general time series generation model for industrial Internet machines and equipment, when training data is limited, it is necessary to integrate the knowledge trained in the text field into the time series data generation process through cross-domain multi-model collaboration to solve the problem of lack of training data and improve the quality of the generated results and the controllability of the generation process. This is the core problem to be solved in time series data generation based on cross-domain multi-model collaboration.

[0097] Based on the above problems, this application provides an optional embodiment, such as Figure 7 As shown, Figure 7 It is an optional time series data generation architecture based on cross-domain multi-model collaboration in the embodiment of the present application. Based on the ideas of task decomposition and multi-model collaboration, it decomposes the complex general time series data generation problem into building a large device cognitive model (solving the problem of what to generate) and a large data generation model (solving the problem of how to generate). It uses the sequence modeling capabilities and multi-sequence relationship modeling of the large model to improve the data generation capability, and integrates domain knowledge into the time series data generation process, which can generate high-quality time series data with specified attributes. At the same time, building a cross-domain multi-model collaboration mechanism can integrate multiple models into a whole, using their respective advantages and expertise to form a set of collaborative strategies that efficiently and flexibly integrate different vertical categories of knowledge, which is referenceable and replicable in other application scenarios.

[0098] Optionally, the above-mentioned time series data generation architecture based on cross-domain multi-model collaboration mainly includes the following three key parts:

[0099] Part I: Task Input 72, mainly used to input complex general time series data generation problems for machine equipment;

[0100] The second part: the device recognition model 74, which is mainly used to perform task analysis, device recognition, and timing feature determination on the task input, and obtain information about the content to be generated corresponding to the task input;

[0101] The third part: the data generation model 76 is mainly used to determine the domain knowledge based on the information of the content to be generated corresponding to the task input, and to generate abstract features and specific features according to the preset loop generation strategy, and output the final generated content as a result.

[0102] As an optional implementation method, in actual applications, the time series data generation process using the above-mentioned time series data generation architecture based on cross-domain multi-model collaboration is as follows: Figure 8 This is a flowchart of an optional time series data generation architecture based on cross-domain multi-model collaboration in an embodiment of the present application, which specifically includes the following steps:

[0103] Step S801: Determine the task input into the device cognitive model under the time series data generation architecture;

[0104] Optionally, describe the current time series data generation task, including the machine equipment, working status, user expectations, etc. For example, "Please generate the temperature changes of a temperature sensor in a home for about 3 months." If necessary, provide a time series data sample as a reference.

[0105] Step S802: parse the task input using a device cognitive model trained with domain data to obtain a text description containing domain knowledge;

[0106] It should be noted that the trained device cognitive model has the ability to recognize the characteristics of time series data of different machines and equipment, and can generate corresponding text description sections according to the input. The text description can include text information such as text output and numerical representation.

[0107] Optionally, for example, the text description of the time series data feature corresponding to the above task input is as follows: "The temperature of a home room measured by a temperature sensor usually fluctuates slightly during the day, and the long-term trend is constant temperature."

[0108] Step S803: Input the text description generated by the device cognitive model into the data generation model. The data generation model starts with initial noise and uses the above text description as a text field guidance signal to control the generation of time series data with expected abstract features, such as trends, periodicity, etc., and corresponds to the result output of the data generation model.

[0109] Optionally, in order to more clearly demonstrate the above time series data generation process, Figure 9 Provides a visual display of data flow, Figure 9 This is an optional data flow diagram in the process of generating time series data in an embodiment of the present application, which mainly includes: after the task is input into the device cognitive big model, the domain knowledge of the task in the text field is output through the internal conversion of the device cognitive big model, and the domain knowledge is further input into the data generation big model, thereby obtaining a result output with time series data corresponding to the task input, that is, through the mutual cooperation between the device cognitive big model and the data generation big model, the time series data that meets the task input is obtained, thereby realizing high-quality generation of time series data.

[0110] It should be noted that while the data generation model generates the corresponding time series data in real time, the reference time series data provided by the user is input into the data generation model, and the reference time series data is used as the time series field guidance signal to control the generated time series data to have the expected specific characteristics, such as amplitude, mean, variance, etc. Optionally, in the process of generating time series data for a long period of time, the time series data generation results of the previous round are used as the time series guidance signal of the current generation round, and data generation for a long period of time is achieved through iteration. In addition, in order to ensure the validity of the generated time series data, the generated time series data can also be tested, including accuracy, completeness, consistency, smoothness and other test contents.

[0111] It should be noted that the aforementioned time series data generation architecture based on cross-domain multi-model collaboration supports the high-quality, controlled generation of time series data for heterogeneous machines and devices. To address the current lack of large-scale annotated time series data training sets, this approach abandons the practice of directly training a general-purpose time series data generation model. Instead, it decomposes the complex time series data generation problem into two sub-problems: machine device cognition and time series data generation. By leveraging existing general-purpose language models to filter unlabeled web corpora, data pairs in a specific format are generated as training samples. High-quality training samples are then iteratively selected and used to train the cognition model. Based on the cognition of the physical properties and behavioral characteristics of machines and devices, the data generation model is trained using a combination of prior knowledge and data-driven methods, effectively addressing the lack of time series data training sets. Furthermore, by constructing large models tailored to different objectives and designing an interactive collaboration mechanism between these models, the domain knowledge of the device cognition model is integrated into the time series data generation process, enabling the high-quality generation of time series data with arbitrary attributes.

[0112] Optionally, by implementing the time series data generation process using the aforementioned cross-domain, multi-model collaborative time series data generation architecture, the complex problem of generating general time series data for machines and devices can be broken down into two steps: determining what to generate using a device cognitive macromodel, and determining how to generate data using a data generation macromodel. This allows for the establishment of a device cognitive macromodel oriented towards the physical characteristics and behavioral mechanisms of machines and devices, summarizing and understanding the time series data characteristics of different machines and devices. Building on this foundation, a data generation macromodel can be constructed to capture the complex relationships within time series data and integrate domain knowledge into the time series data generation process, improving the generation effectiveness and versatility of the generation model.

[0113] In addition, in the process of cross-modal model collaboration, in order to solve the problem of alignment and fusion of knowledge in different fields, this application first uses the existing representation models (Learning Representation Models) in different fields to extract the knowledge representations of different modalities respectively. On this basis, based on the conditional diffusion model (Conditional Diffusion Models), one field knowledge representation is used as a guide to control the generation of another field knowledge representation, thereby performing mapping learning of knowledge in different fields. At the same time, in order to solve the above-mentioned training data labeling problem, this application uses existing tools to automatically construct training data pairs for model training. For example, use existing statistical algorithms, time series algorithms, machine learning algorithms, etc. to perform period analysis, amplitude calculation, anomaly detection, etc. on time series data, and express the analysis results in natural language to generate time series data-text training data pairs.

[0114] Simply put, in the process of knowledge mapping between the text domain and the time series domain, the Classfier-Free diffusion model uses the knowledge representation of the text domain as a guide to control the generation of knowledge representation in the time series domain. The generated knowledge representation in the time series domain is then decoded to generate the corresponding time series data, which can achieve efficient mapping of knowledge in different domains.

[0115] In addition, this application proposes a two-stage optimized generation process. On the one hand, it is necessary to ensure that the currently generated time series data meets the semantic requirements of the semantic large model, and on the other hand, it is necessary to ensure the consistency of the statistical characteristics of the generated data. In addition, to solve the problem of limited step size in the time series data generation process, this application uses a cyclic mechanism generation strategy, using the generation results of the previous round as additional input for the current generation round, and achieving data generation over a long period of time through iteration.

[0116] Through the above two-stage optimized generation process, during the time series data generation process, first, coarse-grained text guidance is used to ensure that the generation results have the expected abstract features. Then, fine-grained time series guidance is used to ensure that the generation results have consistent specific features. This can simultaneously improve the quality and controllability of the generation results.

[0117] In summary, the solutions in the optional embodiments of this application can, on the one hand, enable the generation of time series data for multiple machine devices, improving the versatility of the generated models through multi-model collaboration. On the other hand, through cross-domain knowledge collaboration and conditional generation mechanisms, the quality and controllability of time series data generation can be improved.

[0118] Through the description of the above implementation methods, those skilled in the art can clearly understand that the method according to the above embodiment can be implemented by means of software plus the necessary general hardware platform, and of course it can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes a number of instructions for enabling a terminal device (which can be a mobile phone, computer, server, or network device, etc.) to execute the methods of each embodiment of the present application.

[0119] Figure 10 is a structural block diagram of a device for generating time series data according to an embodiment of the present application; Figure 10 As shown, including:

[0120] Processing module 92 is configured to process task data corresponding to a machine device through a device cognitive macromodel to obtain a description text corresponding to the task data, wherein the task data is used to describe the data characteristics of the time series data to be generated by the machine device. The device cognitive macromodel is a model established through a self-training mechanism combined with a natural language macromodel to generate the description text, and the description text corresponds to the time series data generated by the machine device.

[0121] A first determination module 94 is configured to input the description text into a data generation model connected to the device cognitive model, and control the data generation model to determine coarse-grained features of the time series data to be generated in the text domain, wherein the data generation model is configured to determine coarse-grained features;

[0122] A first determination module 96 is configured to, upon obtaining reference time series data, input the reference time series data into the data generation macro model, and control the data generation macro model to determine fine-grained features corresponding to the time series data to be generated in the time series domain. The data generation macro model is further configured to determine fine-grained features, and the reference time series data is historical time series data that matches the machine device.

[0123] The control module 98 is used to control the time series data output in real time by the data generation large model according to the coarse-grained features and the fine-grained features, and obtain the target time series data corresponding to the task data, wherein the data generation large model is also used to generate corresponding time series data based on the coarse-grained features and the fine-grained features.

[0124] Through the above-mentioned device, the task description information corresponding to the time series data to be generated is first processed by the device recognition large model to obtain text data; then the coarse-grained features of the time series data output by the large model in real time in the text field are generated according to the text data control data. At the same time, when the reference time series data corresponding to the time series data to be generated is obtained, the fine-grained features corresponding to the time series data output by the large model in real time in the time series field are generated by the reference time series control data, thereby obtaining the target time series data that satisfies both the coarse-grained features and the fine-grained features, and determining the target time series data as the time series data corresponding to the task data. According to the above scheme, coarse-grained descriptive text is used to guide the time series data to be generated, ensuring that the generated results have the expected abstract features (such as the trend of time series data, the periodicity of time series data, etc.). Then, fine-grained time series guidance is used to ensure that the generated results have consistent specific features (such as the amplitude of time series data, the mean of time series data, the variance of time series data, etc.), which improves the quality of the time series data process and the controllability of the generation, thereby solving the problems in related technologies that the existing technology has low generation quality and difficult to control for the current time series data generation model.

[0125] In the embodiments of the present application, the term "module" or "unit" refers to a computer program or a part of a computer program that has a predetermined function and works together with other related parts to achieve a predetermined goal, and can be implemented in whole or in part by using software, hardware (such as processing circuits or memories) or a combination thereof. Similarly, a processor (or multiple processors or memories) can be used to implement one or more modules or units. In addition, each module or unit can be part of an overall module or unit that includes the function of the module or unit.

[0126] It should be noted that for the aforementioned method embodiments, for simplicity of description, they are all expressed as a series of action combinations. However, those skilled in the art should be aware that the present invention is not limited by the order of the actions described, because according to the present invention, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present invention.

[0127] An embodiment of the present application further provides a computer-readable storage medium, in which a computer program is stored. The computer program is configured to execute the steps of any of the above method embodiments when run.

[0128] In an exemplary embodiment, the computer-readable storage medium may include, but is not limited to, various media that can store computer programs, such as a USB flash drive, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk, or an optical disk.

[0129] An embodiment of the present application further provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute the steps in any one of the above method embodiments.

[0130] In an exemplary embodiment, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor, and the input / output device is connected to the processor.

[0131] Optionally, specific examples in this embodiment may refer to the examples described in the above embodiments and optional implementation modes, and this embodiment will not be described in detail here.

[0132] An embodiment of the present application further provides a computer program product, which includes a computer program. When the computer program is executed by a processor, the steps in any one of the above method embodiments are implemented.

[0133] An embodiment of the present application further provides another computer program product, comprising a non-volatile computer-readable storage medium, wherein the non-volatile computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of any of the above method embodiments are implemented.

[0134] An embodiment of the present application further provides a computer program, which includes computer instructions stored in a computer-readable storage medium; a processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device performs the steps of any of the above method embodiments.

[0135] Obviously, those skilled in the art should understand that the modules or steps of the present application described above can be implemented using a general-purpose computing system, they can be concentrated on a single computing system, or distributed across a network composed of multiple computing systems. Alternatively, they can be implemented using program code executable by the computing system, so that they can be stored in a storage system and executed by the computing system. In some cases, the steps shown or described can be performed in a different order than herein, or they can be fabricated into separate integrated circuit modules, or multiple modules or steps can be fabricated into a single integrated circuit module for implementation. Thus, the present application is not limited to any specific combination of hardware and software.

[0136] The above description is merely a preferred embodiment of the present application and is not intended to limit the present application. Various modifications and variations are possible for those skilled in the art. Any modifications, equivalent substitutions, improvements, etc. made within the principles of the present application shall be included within the scope of protection of the present application.

Claims

1. A method for generating time series data, characterized in that: include: Processing task data corresponding to a machine device through a device cognitive macromodel to obtain a description text corresponding to the task data, wherein the task data is used to describe the data features of the time series data to be generated by the machine device. The device cognitive macromodel is a model established through a self-training mechanism combined with a natural language macromodel for generating the description text, and the description text corresponds to the time series data generated by the machine device. Inputting the description text into a data generation model connected to the device cognitive model, and controlling the data generation model to determine coarse-grained features of the time series data to be generated in the text field, wherein the data generation model is used to determine the coarse-grained features; When reference time series data is obtained, the reference time series data is input into the data generation model, and the data generation model is controlled to determine fine-grained features corresponding to the time series data to be generated in the time series field, wherein the data generation model is also used to determine fine-grained features, and the reference time series data is historical time series data that matches the machine equipment; The time series data output in real time by the data generation large model is regulated according to the coarse-grained features and the fine-grained features to obtain target time series data corresponding to the task data, wherein the data generation large model is also used to generate corresponding time series data based on the coarse-grained features and the fine-grained features.

2. The method for generating time series data according to claim 1, wherein: Before processing the task data corresponding to the machine equipment by the device cognitive model to obtain a description text corresponding to the task data, the method further includes: Selection step: determining the natural language large model according to preset conditions; Labeling step: When the preset seed data corresponding to the machine device is obtained, target data similar to the preset seed data is determined from the unlabeled dataset, and the target data is labeled using the natural language large model to obtain multiple pairs of target data corresponding to the machine device, wherein the preset seed data includes at least one pair of data, each data pair consisting of device information of a machine device and reference description text describing the time series data generated by the machine device, and the unlabeled dataset is text associated with the machine device intercepted from the network; Screening step: selecting training data from the plurality of target data pairs according to preset screening rules; Training step: using the training data and the preset seed data to perform model training on the natural language large model to obtain a target training model corresponding to the natural language large model; cyclically executing the labeling step, the screening step, and the training step, and determining the target training model that completes each round of training as the initial model of the natural language large model in the next round of model training; Obtain a multiple training model obtained by training the large natural language model for the Nth time, wherein N is an integer greater than 1, and the multiple training model is a target training model obtained by completing the model training for the large natural language model for the Nth time; When the recognition accuracy of the multiple training models for the machine device meets a preset expected value, the multiple training models are determined as the device recognition macro model corresponding to the machine device.

3. The method for generating time series data according to claim 1, wherein: Before inputting the description text into the data generation model connected to the device cognitive model, the method further includes: Establishing a cross-domain knowledge mapping between the device cognitive model and the data generation model; The step of establishing a cross-domain knowledge mapping between the device cognitive model and the data generation model includes: Acquire reference text data for training the device recognition large model, and acquire reference time series data for training the data generation large model, wherein the reference text data corresponds to the reference time series data; Using a first type of representation model corresponding to the text field to extract knowledge representation from the reference text data to obtain a first representation; Using a second type of representation model corresponding to the time series field to extract knowledge representation from the reference time series data to obtain a second representation; The first representation and the second representation are associated according to a preset conditional diffusion model to obtain a target mapping relationship, and a mapping result of the cross-domain knowledge mapping corresponding to the data generation big model and the device cognition big model is determined based on the target mapping relationship.

4. The method for generating time series data according to claim 3, wherein: Associating the first representation and the second representation according to a preset conditional diffusion model to obtain a target mapping relationship includes: The first representation and the second representation are processed respectively according to the preset conditional diffusion model, wherein the conditional diffusion model includes at least: an unconditional generation sub-model and a conditional generation sub-model, and an output result of the unconditional generation sub-model is allowed to serve as a guidance signal of the conditional generation sub-model; Using the first representation and the reference text data to train the unconditional generative sub-model to obtain a first sub-model; Training the conditional generative model using the second representation and the reference time series data to obtain a second sub-model; In a case where the first representation is set as a basic condition for calling the second representation, determining a target association relationship according to a first output result corresponding to the first sub-model and a second output result corresponding to the second sub-model; A target mapping relationship between the first representation and the second representation is determined according to the target association relationship.

5. The method for generating time series data according to claim 3, wherein: After determining a mapping result of the cross-domain knowledge mapping corresponding to the data generation large model and the device cognition large model based on the target mapping relationship, the method further includes: Establishing a generation process for generating the time series data through the data generation model and the device recognition model according to the mapping result; The generation process includes: The first step is to determine the first requirement of the data generation model to generate the time series data to be generated by the description text output by the device cognitive model, and determine the first target representation corresponding to the time series data output in real time in the text domain and the target coarse-grained features included in the first target representation based on the description text; In the second step, when the reference time series data is input into the data generation macro model, a second target representation corresponding to the time series data output in real time in the time series domain and target fine-grained features included in the second target representation are determined based on the reference time series data, and the first target representation is determined as a guiding signal for the second target representation based on the mapping result, thereby obtaining a second requirement for the data generation macro model to generate the time series data to be generated, wherein the second target representation is a representation with some features identical to those in the first target representation; In the third step, when the data generation large model generates corresponding knowledge representation based on the target fine-grained features and the target coarse-grained features, the knowledge representation is decoded through the expert model existing in the data generation large model, and the time series data output in real time by the data generation large model is determined according to the decoding result.

6. The method for generating time series data according to claim 1, wherein: The method further comprises: In the case where the task data indicates that long-term data needs to be generated, the cyclic generation strategy of the data generation large model is started, wherein the cyclic generation strategy is used to indicate that when the long-term data is not completely generated, the target time series data generated in the previous round of the data generation large model is determined as the reference time series data for the current generation round, and new fine-grained features corresponding to the time series field are obtained. The time series data output in real time is iteratively generated based on the coarse-grained features and the new fine-grained features to obtain iterative time series data, so as to determine the long-term data based on the iterative time series data; Acquire the iterative time series data generated by the data generation macro model executing the cyclic generation strategy, and count the data period length corresponding to the iterative time series data; When the data period length is greater than or equal to the target data period length corresponding to the long period data, the current iterative time series data output in real time by the data generation large model is determined as the target time series data corresponding to the task data.

7. The method for generating time series data according to claim 1, wherein: The task data corresponding to the machine equipment is processed by the device cognitive model to obtain a description text corresponding to the task data, including: Determine the parsing result of the device cognitive model on the task data; The model code corresponding to the analysis result is expressed in natural language to obtain the description text corresponding to the task data.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored program, wherein the program executes the steps of the method according to any one of claims 1 to 7 when executed.

9. An electronic device comprising a memory and a processor, characterized in that: A computer program is stored in the memory, and the processor is configured to execute the steps of the method according to any one of claims 1 to 7 through the computer program.

10. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images