Task-oriented multi-source data generation method, device and equipment for industrial time series large model

Abstract

The task-oriented multi-source data generation method for an industrial time sequence large model provided in the application relates to the fields of artificial intelligence and industrial time sequence data processing. The method first acquires original time sequence samples of multiple sensors of an industrial device, a task type, and a corresponding condition label, then generates noise-free estimation samples by using a condition diffusion generation model containing a diffusion denoising backbone and a mixed expert routing network, inputs the noise-free estimation samples into a frozen task evaluator, obtains output task evaluation signals, and adjusts routing probabilities of the mixed expert routing network in the condition diffusion generation model to generate industrial multi-source time sequence samples that meet condition constraints and task requirements. The method provided in the application achieves the technical effect of considering both the distribution authenticity and task utility of industrial multi-source time sequence samples.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and industrial time series data processing, and in particular to a task-oriented multi-source data generation method, apparatus and equipment for large industrial time series models. Background Technology

[0002] In the industrial sector, generative adversarial networks, variational autoencoders, autoregressive models, or diffusion models are commonly used to model and generate samples for multivariate time series data to serve tasks such as fault diagnosis, lifespan prediction, and anomaly detection.

[0003] However, existing solutions often focus on distribution fitting or statistical similarity. The actual utility of the generated samples for downstream classification, regression and other tasks usually needs to be verified separately after training, making it difficult to achieve consistent optimization during the generation stage. At the same time, they do not adequately characterize the channel differences, local dynamics and different working conditions in multi-source time series, which can easily affect the generation quality, adaptability and generalization stability.

[0004] Therefore, how to balance distribution accuracy and task utility in multi-source time series generation has become a technical problem that needs to be solved. Summary of the Invention

[0005] This application provides a task-oriented multi-source data generation method, apparatus, and device for large-scale industrial time-series models to address the aforementioned technical problems. This solution targets multi-sensor time-series generation scenarios for industrial equipment, collaboratively incorporating target task requirements and the inherent conditional attributes of industrial data into the generation process. While ensuring that the generated samples accurately reflect industrial operating conditions and label constraints, it enables the generated results to be specifically optimized for target tasks such as fault diagnosis, lifespan prediction, and anomaly detection, thereby achieving a technical effect that balances the distributional accuracy of industrial multi-source time series with task utility.

[0006] Firstly, this application provides a task-oriented multi-source data generation method for large-scale industrial time-series models, including:

[0007] Acquire raw time-series samples collected by multiple sensors in industrial equipment and the task type of the target task;

[0008] Based on the original time series samples, determine the corresponding condition labels, which include at least one of the following: fault category label, operating condition label, equipment health status label, remaining service life range label, and continuous remaining service life label.

[0009] A pre-trained conditional diffusion generation model is adopted. Based on the original time series samples, task type and conditional label, conditional diffusion generation processing is performed to obtain the corresponding noiseless estimation samples. The conditional diffusion generation model is a neural network model used to learn the distribution characteristics of real industrial time series and generate new industrial time series. The conditional diffusion generation model includes a diffusion denoising backbone network. The feedforward sub-layer of the diffusion denoising backbone network is set with a hybrid expert routing network. The hybrid expert routing network includes multiple expert sub-networks and a routing controller. The routing controller is used to calculate the routing probability of each expert sub-network and select the expert sub-network to be activated.

[0010] The noiseless estimation samples are input into the pre-built frozen task evaluator to obtain the corresponding task evaluation signal. The frozen task evaluator is a task evaluation device used to output the performance score corresponding to the target task based on the noiseless estimation samples.

[0011] Based on the task evaluation signal, the routing probability of the route controller in the conditional diffusion generation model is adjusted to obtain the adjusted routing probability;

[0012] Based on the adjusted routing probabilities, a conditional diffusion generation model is used to generate target industrial multi-source time series samples that meet the conditional labels and task types.

[0013] Secondly, this application provides a task-oriented multi-source data generation device for large-scale industrial time-series models, comprising:

[0014] The acquisition module is used to acquire raw time series samples collected by multiple sensors of industrial equipment and the task type of the target task;

[0015] The determination module is used to determine the corresponding condition labels based on the original time series samples. The condition labels include at least one of the following: fault category label, operating condition label, equipment health status label, remaining service life range label, and continuous remaining service life label.

[0016] The first processing module is used to perform conditional diffusion generation processing based on the original time series samples, task type and condition label using a pre-trained conditional diffusion generation model to obtain the corresponding noiseless estimation samples. The conditional diffusion generation model is a neural network model used to learn the distribution characteristics of real industrial time series and generate new industrial time series. The conditional diffusion generation model includes a diffusion denoising backbone network. The feedforward sublayer of the diffusion denoising backbone network is equipped with a hybrid expert routing network. The hybrid expert routing network includes multiple expert subnetworks and a routing controller. The routing controller is used to calculate the routing probability of each expert subnetwork and select the expert subnetwork to be activated.

[0017] The second processing module is used to input the noiseless estimation samples into the pre-built frozen task evaluator to obtain the corresponding task evaluation signal. The frozen task evaluator is a task evaluation device used to output the performance score corresponding to the target task based on the noiseless estimation samples.

[0018] The adjustment module is used to adjust the routing probability of the route controller in the conditional diffusion generation model based on the task evaluation signal, so as to obtain the adjusted routing probability.

[0019] The generation module is used to generate target industrial multi-source time series samples that meet the conditional labels and task types based on the adjusted routing probabilities and through a conditional diffusion generation model.

[0020] Thirdly, this application provides a task-oriented multi-source data generation device for industrial time-series large models, including: a memory and a processor;

[0021] The memory stores the instructions that the computer executes;

[0022] The processor executes computer execution instructions stored in memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.

[0023] Fourthly, this application provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible embodiments of the first aspect.

[0024] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the first aspect and / or various possible implementations of the first aspect.

[0025] This application provides a task-oriented multi-source data generation method, apparatus, and equipment for large-scale industrial time-series models. It acquires raw time-series samples collected by multiple sensors in industrial equipment and the task type of the target task, determines the conditional labels corresponding to the raw time-series samples, and utilizes a pre-trained conditional diffusion generation model containing a hybrid expert routing network set in the feedforward sublayer of the diffusion denoising backbone network. Based on the raw time-series samples, task type, and conditional labels, conditional diffusion generation is performed. Furthermore, the routing probability of the routing controller in the hybrid expert routing network is adjusted by combining the task evaluation signal output by the frozen task evaluator. This process simultaneously considers learning the distribution characteristics of real industrial time series, preserving conditional constraints, and optimizing task performance. Ultimately, it generates target industrial multi-source time-series samples that better meet the requirements of conditional labels and task types, are more adaptable to downstream classification or regression tasks, and more fully preserve the differences in multi-source time-series channels, local dynamics, and different operating conditions. The overall process achieves the technical effect of balancing the distribution authenticity of industrial multi-source time series and the utility of the task. Attached Figure Description

[0026] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0027] Figure 1 This application provides a schematic diagram of an application data processing system architecture.

[0028] Figure 2 A flowchart illustrating the task-oriented multi-source data generation method for large-scale industrial time-series models provided in this application embodiment. Figure 1 ;

[0029] Figure 3 A flowchart illustrating the task-oriented multi-source data generation method for large-scale industrial time-series models provided in this application embodiment. Figure 2 ;

[0030] Figure 4 A flowchart illustrating the task-oriented multi-source data generation method for large-scale industrial time-series models provided in this application embodiment. Figure 3 ;

[0031] Figure 5 A flowchart illustrating the task-oriented multi-source data generation method for large-scale industrial time-series models provided in this application embodiment. Figure 4 ;

[0032] Figure 6 A flowchart illustrating the task-oriented multi-source data generation method for large-scale industrial time-series models provided in this application embodiment. Figure 5 ;

[0033] Figure 7 A schematic diagram of the structure of a task-oriented multi-source data generation device for industrial time-series large models provided in this application embodiment;

[0034] Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0035] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0036] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0037] Industrial multi-source time series generation technology is mainly applied to scenarios such as equipment health management, fault diagnosis, remaining service life prediction, and anomaly detection. In systems such as aero-engines, wind turbines, and industrial production lines, multiple sensors typically collect raw time series samples such as temperature, vibration, pressure, and current, which are then combined with upper-level analysis models to complete state identification, trend prediction, and operation and maintenance decisions.

[0038] In existing technologies, generative adversarial networks, variational autoencoders, autoregressive models, or diffusion models are commonly used to model multivariate time series data. These methods generate new time series samples by learning the statistical distribution of historical samples to supplement fault samples, enhance training data, or simulate operational data under different conditions. Such approaches typically prioritize the similarity in distribution between generated and real samples as the primary optimization objective, and then use the generated samples for training and validating classification, regression, or detection models.

[0039] However, the above-mentioned schemes often focus on distribution fitting or morphological similarity, and the generation process lacks constraints directly related to the performance of the target task. As a result, although the generated samples are statistically close to the real data, they may not be able to effectively improve the accuracy of fault classification, lifetime prediction accuracy, or anomaly identification. At the same time, the physical semantics, local dynamic changes, and pattern differences under different operating conditions corresponding to different sensor channels in multi-source time series data are often difficult to fully reflect in the unified generation process.

[0040] Furthermore, when equipment is in different health states, fault categories, or life stages, the sample distribution often shows a significant shift. If the generated model lacks the ability to collaboratively model task types and condition information, it is prone to problems such as insufficient adaptability of generated samples, unstable generalization across operating conditions, and limited support for downstream models, thereby restricting its practical application value in industrial scenarios.

[0041] How to balance distribution accuracy and task utility in multi-source time series generation has become a technical problem that needs to be solved.

[0042] To address the aforementioned issues, this application provides a task-oriented multi-source data generation method, apparatus, and equipment for large-scale industrial time-series models. By acquiring raw time-series samples collected by multiple sensors of industrial equipment and the task type of the target task, conditional labels corresponding to the raw time-series samples are determined. A pre-trained conditional diffusion generation model, including a hybrid expert routing network set in the feedforward sublayer of the diffusion denoising backbone network, is used to perform conditional diffusion generation based on the raw time-series samples, task type, and conditional labels. The routing probability of the routing controller in the hybrid expert routing network is adjusted by combining the task evaluation signal output by the frozen task evaluator. This process simultaneously considers learning the distribution characteristics of real industrial time series, preserving conditional constraints, and optimizing task performance. This results in target industrial multi-source time-series samples that better meet the requirements of conditional labels and task types, are more adaptable to downstream classification or regression tasks, and more fully retain the differences in multi-source time-series channels, local dynamics, and different operating conditions. The overall process achieves the technical effect of balancing the distribution authenticity of industrial multi-source time series and the utility of the task.

[0043] In terms of application architecture, this method can be deployed in an industrial intelligent analysis system consisting of a data acquisition layer, a generation model layer, and a downstream task layer. The data acquisition layer is responsible for acquiring time-series data from multiple sensors of the equipment, the generation model layer performs conditional diffusion generation, task evaluation, and routing probability adjustment, and the downstream task layer performs classification, regression, or detection applications on the generated samples.

[0044] By adopting the above technical approach, the generation process can be more closely linked to the requirements of the target task, thereby improving the usability of the generated samples in actual industrial tasks.

[0045] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0046] Figure 1This is a schematic diagram of an application data processing system architecture provided in an embodiment of this application. The application data processing system is a computer device. Figure 1 As shown, the above architecture includes at least one of a data acquisition device 101, a processing device 102, and a display device 103.

[0047] It is understood that the structures illustrated in the embodiments of this application do not constitute a specific limitation on the architecture of the application data processing system. In other feasible embodiments of this application, the above architecture may include more or fewer components than illustrated, or combine some components, or split some components, or arrange different components, which can be determined according to the actual application scenario and is not limited here. Figure 1 The components shown can be implemented in hardware, software, or a combination of both.

[0048] In the specific implementation process, the data acquisition device 101 may include an input / output interface or a communication interface, and the data acquisition device 101 can be connected to the processing device through the input / output interface or the communication interface.

[0049] Processing device 102 can first acquire the original time series samples of industrial equipment and the target task type to determine the corresponding condition labels. A pre-trained conditional diffusion generation model containing a hybrid expert routing network is used to perform conditional diffusion generation processing to obtain noiseless estimated samples. These noiseless estimated samples are then input into a frozen task evaluator to obtain task evaluation signals. Based on these signals, the routing probability of the routing controller is adjusted, ultimately generating target industrial multi-source time series samples that meet the condition labels and task types.

[0050] The display device 103 can also be a touch screen or the screen of a terminal device, used to receive user commands while displaying the above-mentioned content, so as to realize interaction with the user.

[0051] It should be understood that the aforementioned processing device can be implemented by a processor reading instructions from memory and executing those instructions, or it can be implemented by a chip circuit.

[0052] Furthermore, the network architecture and business scenarios described in the embodiments of this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided in the embodiments of this application. As those skilled in the art will know, with the evolution of network architecture and the emergence of new business scenarios, the technical solutions provided in the embodiments of this application are also applicable to similar technical problems.

[0053] Figure 2 A flowchart illustrating the task-oriented multi-source data generation method for large-scale industrial time-series models provided in this application embodiment. Figure 1 ,like Figure 2As shown in this embodiment, the task-oriented multi-source data generation method for industrial time-series large models includes:

[0054] S201. Obtain the raw time series samples collected by multiple sensors of industrial equipment and the task type of the target task.

[0055] In this embodiment, the original time-series samples serve as the data foundation for subsequent generation and processing, characterizing the observation sequences acquired synchronously or quasi-synchronously by multiple sensors during the actual operation of industrial equipment. The task type of the target task characterizes the downstream industrial analysis direction served by the generated samples and participates in the subsequent generation process as a task-oriented constraint.

[0056] For example, the execution entity can be deployed on the generative model layer server in an industrial intelligent analysis system, or on an edge computing node close to the device. The data acquisition layer collects continuous monitoring data from multiple sensors on the industrial equipment and aggregates it. For the acquired multi-source channel data, necessary time alignment and preprocessing can be performed to form raw time series samples that can be used for subsequent processing.

[0057] In one possible embodiment, the task type is provided synchronously by the upstream task configuration module and a sample-level association is established with the original time series samples. The task type can correspond to one of fault classification, anomaly detection, health status identification, remaining useful life interval prediction, or continuous remaining useful life regression.

[0058] Based on the above processing, the system has completed the joint reception of real industrial multi-source time series observation data and their target task semantics, so that the subsequent generation process is based on a clear task orientation, rather than just a targetless distribution reproduction of the input data.

[0059] S202. Determine the corresponding condition labels based on the original time series samples.

[0060] In this embodiment, the condition labels include at least one of the following: fault category label, operating condition label, equipment health status label, remaining service life interval label, and continuous remaining service life label. The condition labels are used to describe the industrial state information corresponding to the original time series sample and to provide scenario constraints and state constraints for subsequent condition diffusion generation processing. Among them, the fault category label is used to identify the fault type corresponding to the sample, the operating condition label is used to identify the load, speed, power, ambient temperature or process stage of the equipment when the sample is collected, the equipment health status label is used to characterize the state such as healthy, slightly degraded, severely degraded or faulty, the remaining service life interval label is used to give the service life interval, and the continuous remaining service life label is used to give the service life information in continuous numerical form.

[0061] In practice, the implementing entity selects the label determination method based on the sample source. For industrial systems with established operation and maintenance databases or test bench databases, the corresponding labels can be directly read from sample metadata, fault record tables, maintenance record tables, operating condition record tables, and lifespan label tables, and aligned according to the sample time window and the label time range. For online data acquisition scenarios, if some labels are not directly stored, a reverse lookup is performed based on the device identifier, sampling time, operating condition information, and device status information corresponding to the original time series sample to obtain labels consistent with the time window.

[0062] In one possible embodiment, the condition label consists of a single label, or it can consist of multiple labels combined.

[0063] Specifically, for aircraft engine degradation data, operating condition labels and continuous remaining service life labels can be used together as input conditions; for wind turbine fault diagnosis scenarios, fault category labels and equipment health status labels can be used simultaneously; for industrial production line anomaly detection scenarios, operating condition labels corresponding to the process stage can be used as input conditions.

[0064] Through this step, the industrial state corresponding to the sample is transformed into a computable condition constraint, and the subsequently generated time series can maintain a correspondence with the input state in terms of fault type, operating mode, health level, or life trend.

[0065] S203. A pre-trained conditional diffusion generation model is used to perform conditional diffusion generation processing based on the original time series samples, task type, and conditional labels to obtain the corresponding noiseless estimated samples.

[0066] In this embodiment, the conditional diffusion generation model is a neural network model used to learn the distribution characteristics of real industrial time series and generate new industrial time series. The conditional diffusion generation model includes a diffusion denoising backbone network. The feedforward sublayer of the diffusion denoising backbone network is equipped with a hybrid expert routing network. The hybrid expert routing network includes multiple expert subnetworks and a routing controller. The routing controller is used to calculate the routing probability of each expert subnetwork and select the expert subnetwork to be activated.

[0067] The pre-trained conditional diffusion generative model is the core model of the generation process. It is used to learn the distribution of real industrial multi-source time series under conditional constraints and output noise-free estimated samples in the reverse denoising process.

[0068] The diffusion denoising backbone network can employ a temporal transformer suitable for time series modeling, a temporal convolutional network combined with attention structure, or a sequence U-shaped network with residual blocks.

[0069] The hybrid expert routing network set up in the feedforward sublayer is used to enable differentiated computation paths for different sample patterns, task types, and condition labels.

[0070] Multiple expert subnetworks are deployed in parallel. Each expert subnetwork performs a specific feature transformation. The routing controller calculates the routing probability of each expert based on the current input features and condition information, and selects the expert subnetworks to be activated to participate in the calculation.

[0071] In one possible embodiment, task type and condition labels can also be used to adjust the denoising response of the backbone network so that the network exhibits different denoising characteristics under different tasks and conditions.

[0072] The conditional diffusion generation model has completed parameter learning based on large-scale real industrial time series samples during the pre-training stage. Therefore, in this step, it can perform conditional diffusion generation based on the input samples and conditional information.

[0073] Optionally, the noiseless estimated sample can be the estimation result of the original noiseless sequence by the current reverse denoising step, or it can be the final noiseless representation obtained after the complete reverse process.

[0074] If the standard noise prediction method is used, the backbone network outputs a noise estimate, and then calculates a noise-free estimated sample based on the diffusion inversion process; if the direct sample prediction method is used, the network directly outputs a noise-free estimated sample.

[0075] By introducing a hybrid expert routing network into the feedforward sublayer, the conditional diffusion generative model enables different experts to model different types of industrial time series features. The routing controller then selects the corresponding expert routing network combination based on the current target task and condition label, so that the noiseless estimation sample maintains the statistical structure of the real industrial time series and reflects the differentiated expression for the target task.

[0076] In one possible embodiment, a multi-scale noise modeling mechanism is further introduced during the conditional diffusion generation process.

[0077] Specifically, the time series is decomposed into three scales: short-term, medium-term, and long-term. Independent diffusion noise addition and denoising processes are performed on each scale to generate noise sequences at the corresponding scales. In the denoising network, a cross-scale attention mechanism is used to fuse the multi-scale noise sequences, generating the final noise-free estimation sample. This mechanism is dynamically optimized through reward feedback from the task evaluator. For example, in the Remaining Useful Life (RUL) prediction task, long-term scale noise modeling can enhance the ability to capture degradation trends.

[0078] S204. Input the noiseless estimation sample into the pre-built frozen task evaluator to obtain the corresponding task evaluation signal.

[0079] In this embodiment, the frozen task evaluator is a task evaluation device used to output the performance score corresponding to the target task based on the noiseless estimation sample.

[0080] The frozen task evaluator is used to quantitatively evaluate the usability of the noiseless estimated samples on the target task and outputs a task evaluation signal as the basis for subsequent route probability adjustment. The parameters of the frozen task evaluator remain fixed during the execution phase of this method and do not participate in the parameter updates of the conditional diffusion generative model. This ensures that the evaluation results only reflect the degree to which the samples support the target task and are not coupled with the parameters of the generative model.

[0081] Specifically, corresponding task evaluators can be pre-built for different task types to output performance scores corresponding to the target task. To ensure consistency between the evaluator and the generator interface, data transformation consistent with the evaluator's processing requirements can be performed before inputting noiseless estimation samples.

[0082] Based on this step, whether the generated samples are suitable for the target task no longer depends on the indirect reflection of the loss within the generation model, but is directly quantified by a separate task evaluation module.

[0083] S205. Based on the task evaluation signal, adjust the routing probability of the routing controller in the conditional diffusion generation model to obtain the adjusted routing probability.

[0084] In this embodiment, the routing probability represents the likelihood of each expert subnetwork being selected under the current input conditions, and the task evaluation signal is used to drive the routing controller to update the expert combination that is more suitable for the target task.

[0085] For example, in the hybrid expert routing network of each layer in the diffusion denoising backbone network, the original routing probability distribution of that layer when generating noiseless estimation samples is first obtained, and then the routing probability is adjusted according to the task evaluation signal to enhance the expert selection tendency that is more favorable to the target task. For batch samples, the routing probability can also be comprehensively adjusted by combining the task evaluation results corresponding to multiple samples to reduce the impact of single sample fluctuations on route updates.

[0086] Through this step, the expert selection in the conditional diffusion generative model is not only determined by the distributed modeling process, but also receives feedback from the performance of the target task, thereby adjusting the routing path to a computational path more consistent with the target task.

[0087] S206. Based on the adjusted routing probabilities, generate target industrial multi-source time series samples that meet the conditional labels and task types through a conditional diffusion generation model.

[0088] In this embodiment, the adjusted routing probability already reflects the correction result of the task evaluation signal on the expert selection strategy. Therefore, in the subsequent reverse denoising process, the conditional diffusion generation model generates target industrial multi-source time series samples according to the new expert activation combination.

[0089] For example, one can start from a random noise sequence and perform a backdiffusion process given a task type and conditional labels; or one can perform conditional sampling with the original time series sample as a reference to generate a time series sample that is semantically close to the original sample but has new variants.

[0090] In the denoising time step, the diffusion denoising backbone network receives the current noisy sample and condition information related to the task type and condition label, and selects the corresponding expert subnetwork in the feedforward sublayer according to the adjusted routing probability.

[0091] After denoising, the target industrial multi-source time series sample is output. This target sample maintains the same data structure as the original time series sample, thus it can be directly used for subsequent industrial task model training, validation, or simulation.

[0092] In one possible embodiment, if the condition label includes a fault category label and an operating condition label, the generated sample simultaneously reflects the corresponding fault mode and operating condition characteristics; if the condition label includes a remaining service life interval label or a continuous remaining service life label, the generated sample reflects a degradation trend consistent with the service life stage; if the task type points to anomaly detection, the generated sample satisfies the condition label constraints while retaining the abnormal fragment features that can be identified by the detection model.

[0093] After generation, necessary post-processing and storage can be performed on the target industrial multi-source time series samples for use by downstream classification, regression or detection models.

[0094] Based on the above analysis, it can be seen that the adjusted routing probability ultimately acts on the core computational path of conditional diffusion generation, so that the generation process is simultaneously controlled by real distribution learning, conditional label constraints and target task feedback. The generated target industrial multi-source time series samples are consistent with the input constraints in terms of industrial state semantics and task adaptation direction.

[0095] This application provides a task-oriented multi-source data generation method for large-scale industrial time-series models. It acquires raw time-series samples collected by multiple sensors in industrial equipment and the task type of the target task, determines the conditional labels corresponding to the raw time-series samples, and utilizes a pre-trained conditional diffusion generation model containing a hybrid expert routing network set in the feedforward sublayer of the diffusion denoising backbone network. Based on the raw time-series samples, task type, and conditional labels, conditional diffusion generation is performed. Furthermore, the routing probability of the routing controller in the hybrid expert routing network is adjusted by freezing the task evaluation signal output by the task evaluator. This method simultaneously considers learning the distribution characteristics of real industrial time series, preserving conditional constraints, and optimizing task performance during the generation process. This results in target industrial multi-source time-series samples that better meet the requirements of conditional labels and task types, are more adaptable to downstream classification or regression tasks, and more fully retain the differences in multi-source time-series channels, local dynamics, and different operating conditions. The overall process achieves the technical effect of balancing the distribution authenticity of industrial multi-source time series and the utility of the task.

[0096] Figure 3 A flowchart illustrating the task-oriented multi-source data generation method for large-scale industrial time-series models provided in this application embodiment. Figure 2 ,like Figure 3 As shown, this embodiment, based on the above embodiments, provides a detailed explanation of the specific process for obtaining noise-free estimation samples, including:

[0097] S301. Divide the original time series sample into multiple time segments of fixed length along the time axis.

[0098] In this embodiment, the original time series sample contains multiple sensor data acquisition channels.

[0099] For example, the original time series samples can be synchronously acquired by multiple sensor channels in an aircraft engine, wind turbine, or industrial production line. After sampling, a two-dimensional sequence containing multi-channel data such as temperature, vibration, pressure, or current is formed, and its mathematical representation is... for:

[0100]

[0101] in, Number of sensor channels This represents the total time step.

[0102] The system performs non-overlapping segmentation of the sequence along the time axis by a preset length P, resulting in... The local sequence of the c-th channel in the n-th time segment is as follows:

[0103]

[0104] in, Let P be the local sequence of the t-th diffusion step, the c-th channel, and the n-th time segment, where t is the time step of the diffusion process, c is the sensor channel index, n is the time segment index, P is the length of a single time segment (number of time steps), and N is the total number of time segments.

[0105] The segment length can be set to 32, 64, or 128 time steps depending on the sampling frequency and task requirements, so that each time segment contains multi-channel observations within the same continuous time window. The core purpose of this segmentation operation is to reduce the computational complexity of long sequence processing, while ensuring that each segment contains sufficient equipment operating status information, which meets the adaptation requirements of the diffusion model for the length of the input sequence.

[0106] S302. Perform independent mapping processing on the time segment corresponding to each sensor data acquisition channel to obtain the subspace representation corresponding to each sensor data acquisition channel.

[0107] For example, the system configures a dedicated independent mapping layer for each sensor channel, mapping channel segments with different physical meanings to a subspace of a unified dimension, where the subspace representation of the c-th channel in the n-th time segment is... The mathematical expression is:

[0108]

[0109] in, The mapping function for the c-th channel can be implemented using linear projection, one-dimensional convolutional projection, or a two-layer small feedforward network. For the subspace dimension, all channels They have the same value.

[0110] Sensor data with different physical meanings, dimensions, and sampling precisions are transformed into a unified feature space through a dedicated mapping layer. This preserves the differences in physical characteristics of each channel while eliminating feature interference caused by inconsistencies in dimensions. This design differs from traditional full-channel unified mapping and effectively avoids the problem of weak signal channels being overwhelmed by strong signal channels.

[0111] S303. The subspace representations of each sensor data acquisition channel within the same time segment are spliced ​​together and time position encoding is added to generate a time segment-level feature sequence.

[0112] For example, the splicing operation is performed along the channel dimension, splicing c channels within the same time segment. dimensional subspace representation splicing is The feature vectors of the segments are then superimposed with sine and cosine position codes that correspond one-to-one with the segment indices. To form the nth time-segment level token, parameters can be used. This indicates that its corresponding expression is:

[0113]

[0114] In summary, the complete token sequence corresponding to the entire input window is:

[0115]

[0116] Where Z is the complete time-segment level feature sequence, and d is the total dimension of a single time-segment level token.

[0117] Since the attention mechanism itself lacks temporal position awareness, adding temporal position encoding can explicitly express the chronological order and time interval relationship between different time segments. The resulting time segment-level feature sequence can be directly input into the diffusion denoising backbone network for subsequent processing.

[0118] S304. Convert the time-segment level feature sequence into a tensor containing a channel subspace structure, which is composed of the subspace representation corresponding to each sensor data acquisition channel.

[0119] For example, the above conversion operation will concatenate the results. The eigenvectors of a dimensional segment are reconstructed into a structured tensor, the mathematical representation of which is:

[0120]

[0121] in, Let B be a four-dimensional tensor containing the channel subspace structure, and let B be the training batch size.

[0122] This structured tensor provides clear channel boundaries for subsequent channel-dimensional attention operations, enabling attention weights to be applied precisely to each individual sensor channel feature and avoiding confusion between different channel features.

[0123] S305. In the time dimension, attention aggregation is performed on the features of each sensor data acquisition channel to obtain the time aggregation representation of each sensor data acquisition channel; in the channel dimension, attention weighting is performed on the time aggregation representation of each sensor data acquisition channel to obtain the global running state vector.

[0124] For example, firstly, temporal attention aggregation is performed: Temporal attention weights are calculated independently for each sensor channel. The weight values ​​reflect the importance of different time segments in representing the current device state. The calculation formula is as follows:

[0125]

[0126] in, Let b be the temporal attention weights for the b-th sample, the n-th segment, and the c-th channel. For time-attention, there is a learnable weight vector. Let m be the subspace representation of the b-th sample, the n-th segment, and the c-th channel, where m is the summation index of the time segments.

[0127] The multi-time segment features of each channel are compressed into a single temporal aggregate representation by weighted summation:

[0128]

[0129] in, This is the time-aggregated representation of the b-th sample and the c-th channel.

[0130] Then, channel-level attention weighting is performed: the channel attention weight for each channel is calculated. The weight value reflects the contribution of different sensor channels to the global device state. The calculation formula is as follows:

[0131]

[0132] in, Let b be the channel attention weights for the b-th sample and c-th channel. Let be the learnable weight vector for channel attention, and j be the summation index of the sensor channels.

[0133] Multi-channel features are compressed into a single global running state vector through weighted summation:

[0134]

[0135] in, Let be the global running state vector of the b-th sample.

[0136] The aforementioned two-dimensional attention mechanism can be implemented by a multi-head self-attention layer, a layer normalization layer, and residual connections. The network dimension can be flexibly adjusted according to the number of sensor channels and the length of time segments, and this application does not impose specific limitations.

[0137] S306. Generate a global modulation vector based on the global operating state vector and the time aggregation representation of each sensor data acquisition channel.

[0138] For example, the global runtime state vector is concatenated with the time aggregation representation of each channel, and the final global modulation vector is generated through linear transformation and sigmoid activation. The mathematical expression is as follows:

[0139]

[0140] in, Let b be the global modulation vector of the b-th sample. It is the Sigmoid activation function. The learnable weight matrix for global modulation. For feature splicing operations, This is a learnable bias vector for global modulation.

[0141] This modulation vector simultaneously integrates the local temporal dynamic features of each channel and the global operating state features of the device, which can provide global context constraints for the feature transformation of the subsequent feedforward sub-layer, making the generated features more consistent with the actual operating state of the device.

[0142] S307. Input the global modulation vector into the feedforward sub-layer of the diffusion denoising backbone network to obtain the intermediate output of the feedforward sub-layer.

[0143] In this embodiment, the intermediate output of the feedforward sub-layer is a high-dimensional feature vector containing the temporal variation characteristics, inter-channel correlation characteristics, and global operating status characteristics of multi-sensor data from industrial equipment.

[0144] For example, the feedforward sublayer of the diffusion denoising backbone network adopts the standard Transformer structure, containing two linear transformation layers and a Gaussian Error Linear Unit (GELU) activation function. A hybrid expert routing network is embedded between the two linear layers. After the global modulation vector is input to the first linear layer for preliminary feature transformation, the intermediate output of the feedforward sublayer is obtained. The basic diffusion training loss in this embodiment uses the noise prediction mean square error to constrain the distribution fidelity of the generated samples.

[0145]

[0146] in, To reduce noise loss during diffusion, For mathematical expectation, It is real Gaussian noise. For the noise prediction function of the diffusion denoising backbone network, For the noise sample at the t-th diffusion step, For the conditional labels corresponding to the samples, For downstream task semantics (such as classification, regression, anomaly detection). It is the square of the L2 norm.

[0147] The intermediate output of the feedforward sublayer retains the temporal change patterns extracted by the time dimension attention, the inter-sensor correlations extracted by the channel dimension attention, and the overall health information of the device carried by the global operating state vector, providing comprehensive feature basis for expert selection in the hybrid expert routing network.

[0148] S308. Input the intermediate output of the feedforward sublayer into each expert subnetwork of the hybrid expert routing network to obtain noiseless estimation samples.

[0149] For example, the hybrid expert routing network first calculates the routing probability of each expert based on the intermediate output of the feedforward sublayer and the task mode vector. A predetermined number of experts with the highest routing probabilities are selected to form an activation set. The intermediate output is then input into the activation experts for feature transformation. The outputs of each expert are weighted and summed before being returned to the feedforward sublayer for residual connection and layer normalization. Subsequently, the final output of the feedforward sublayer is passed to subsequent layers of the diffusion denoising backbone network to complete all remaining diffusion denoising steps. The final output is a noiseless estimation sample that is completely consistent with the original time series sample in terms of dimension and structure. This noiseless estimation retains the statistical characteristics of the real industrial time series while also meeting the constraints of the target task and condition labels.

[0150] Specifically, after the hybrid expert routing network completes the feature transformation, it returns to the feedforward sublayer to perform residual connections and layer normalization. Then, the final output of the feedforward sublayer is passed to subsequent layers of the diffusion denoising backbone network to complete all remaining diffusion denoising steps. The mathematical formula for recovering noise-free estimated samples from predicted noise is:

[0151]

[0152] in, The noiseless estimated samples output by the model. Let be the cumulative signal retention coefficients from step 1 to step t, where , Let be the signal retention coefficient for the i-th diffusion step, where , Let be the noise intensity at the i-th diffusion step.

[0153] The final output of noiseless estimated samples is completely consistent with the original time series samples in terms of dimension and structure, which not only preserves the statistical characteristics of real industrial time series, but also meets the constraints of the target task and condition label.

[0154] The task-oriented multi-source data generation method for large-scale industrial time-series models provided in this application reduces the complexity of long sequence processing by dividing the original time series into fixed-length segments, ensuring that each segment contains sufficient equipment operating status information. Independent mapping for each channel preserves the differences in channel physical characteristics, eliminates dimensional interference, and prevents weak signal channels from being overwhelmed by strong signals. Concatenating channel representations and adding time position encoding explicitly expresses the temporal sequence relationship, generating standardized input features suitable for diffusion models. Converting to a channel subspace structure tensor preserves independent channel boundaries, providing precise targets for subsequent channel attention operations. Through time and channel-dimensional attention aggregation, temporal dynamics and channel-related features can be extracted, generating vectors reflecting the overall equipment status. Global modulation vectors fuse local and global information, providing comprehensive contextual constraints for subsequent feature transformations. The intermediate output of the feedforward sublayer integrates multi-dimensional equipment operating features, providing sufficient basis for expert selection. Hybrid expert routing combined with task-oriented methods completes feature transformation, ultimately generating high-quality, noise-free estimation samples that meet the constraints.

[0155] Figure 4 A flowchart illustrating the task-oriented multi-source data generation method for large-scale industrial time-series models provided in this application embodiment. Figure 3 ,like Figure 4 As shown, this embodiment, based on the above embodiments, provides a detailed explanation of the specific process of inputting the intermediate output of the feedforward sublayer into each expert subnetwork of the hybrid expert routing network, including:

[0156] S401. Map the task type of the target task to a deterministic task pattern vector.

[0157] For example, the task type is obtained by encoding the fault category label, operating condition label, or equipment health status label through a pre-trained task embedding layer. The same task type is always mapped to a fixed deterministic vector:

[0158]

[0159] in, For task type The corresponding deterministic task mode vector, Embed a function for the task.

[0160] Further, the task mode is mapped to an expert routing bias:

[0161]

[0162] in, For expert routing bias vectors, Let E be the routing bias matrix, and E be the total number of expert subnetworks.

[0163] This deterministic task pattern vector serves as a task condition bias, guiding the subsequent expert selection process to lean towards the current target task and ensuring that the generated features match the task requirements.

[0164] S402. Calculate the routing probability of each expert subnetwork based on the intermediate output of the feedforward sublayer and the deterministic task mode vector.

[0165] For example, the intermediate outputs of the feedforward sub-layers are input together with the deterministic task mode vectors into the routing controller to calculate the routing probability of each expert sub-network:

[0166]

[0167] in, This is the routing probability vector for the nth token at layer l. For normalization function, Let l be the routing weight matrix. This represents the feature representation of the nth token in the l-th layer.

[0168] The routing controller employs a lightweight linear structure to avoid introducing excessive computational overhead, and the sum of the routing probabilities of all experts is 1.

[0169] S403. Based on the routing probability, sort each expert subnetwork from high to low, and select the top k target expert subnetworks to form a target expert subnetwork set.

[0170] In this embodiment, k is a positive integer.

[0171] For example, the value of k is typically set to 2 to 4, controlling computational complexity while ensuring diversity in feature representation. Only the top k expert subnetworks are activated and participate in subsequent feature computation; the parameters of unselected expert subnetworks are not updated in the current iteration. This sparse activation mechanism significantly reduces the inference and training costs of hybrid expert networks.

[0172] Specifically, a deterministic Top-K routing method is used to activate the expert set:

[0173]

[0174] in, For the set of activation experts for the nth token in layer l, The function is used to select the indices corresponding to the top K maximum values, where K is the number of activated experts.

[0175] Only the top K expert subnetworks are activated and participate in subsequent feature computation; the parameters of unselected expert subnetworks are not updated in the current iteration. This sparse activation mechanism significantly reduces the inference and training costs of hybrid expert networks.

[0176] S404. Determine the output of each target expert subnetwork based on the target expert subnetwork set.

[0177] In this embodiment, the output of the target expert subnetwork is a dedicated high-dimensional feature vector adapted to a single equipment operating condition or fault scenario. The outputs of different expert subnetworks correspond to the feature transformation results of different industrial operation modes.

[0178] For example, each target expert subnetwork adopts a two-layer feedforward network structure with independent parameters, and the mathematical expression is:

[0179]

[0180] in, This is the output of the e-th expert subnetwork in layer l. Let e ​​be the weight matrices of the two linear transformation layers for the e-th expert. It is the Sigmoid linear unit activation function. The conditional embedding weight matrix for the e-th expert, A conditional embedding consisting of a diffusion step, conditional labels, and task semantics. This is an element-wise multiplication operation.

[0181] Different expert subnetworks automatically learn the feature transformation rules of different industrial operation modes during training, such as adapting to different scenarios such as normal operation, bearing failure, and gear failure. The output special feature vectors can accurately capture the unique statistical characteristics of the corresponding modes.

[0182] In one possible implementation, after determining the output of each target expert subnetwork based on the target expert subnetwork set, the method further includes:

[0183] Calculate the average route importance and actual load ratio for each expert subnetwork; generate the expert load balancing loss based on the average route importance and actual load ratio; and adjust the parameters of the hybrid expert routing network based on the expert load balancing loss.

[0184] In this embodiment, the average route importance and actual load ratio are used for subsequent processing. Multiple expert subnetworks in the hybrid expert routing network can correspond to different operating conditions, fault modes, or device health status modes. The routing controller outputs the routing probability of each expert based on the input characteristics and, after selecting the target expert subnetwork set, calculates its activation frequency and routing weight.

[0185] For example, the expert load balancing loss can be constructed from the deviation between the average route importance and the actual load ratio; the larger the deviation, the higher the corresponding loss. This loss is applied to the route controller and its associated parameter matrix through backpropagation, updating the route probability distribution towards a more balanced direction and giving different expert subnetworks more similar usage opportunities during training. Parameter tuning can be accomplished through gradient descent, updating the route controller's gating weights, bias parameters, and projection parameters related to expert selection.

[0186] Specifically, after the target expert subnetwork set is output, the system synchronously calculates the average route importance and actual load ratio of each expert, then generates the expert load balancing loss and participates in model training, ensuring that the routing network does not consistently favor a few expert subnetworks in subsequent generation processes. By continuously adjusting the parameters of the hybrid expert routing network, multiple expert subnetworks can form a relatively stable and clearly defined response mechanism, thereby maintaining the consistency of expert utilization during the conditional diffusion generation process.

[0187] With this implementation, routing skew is suppressed, the use of expert subnetworks is more balanced, model training is more stable, and the generated paths can maintain coverage of different industrial operation modes, making the output multi-source time series samples more reliable in terms of condition adaptability and task consistency.

[0188] S405. The outputs of each target expert subnetwork are weighted and summed to obtain the output of the hybrid expert routing network.

[0189] For example, the weights used in the weighted summation are the routing probabilities corresponding to each target expert sub-network calculated in S402. Weighted fusion is performed only on the outputs of activated target expert sub-networks; inactive experts are not included in the calculation. The output dimension after weighted fusion remains consistent with the output dimension of a single expert sub-network, ensuring the continuity and stability of the feature scale. The enhanced fusion formula is:

[0190]

[0191] in, The output of the layer l hybrid expert routing network, Let n be the routing probability of the nth token in layer l to the eth expert. The diffusion step t, conditional label y, and task semantics are used to define the process. Conditional embedding of composition Let be the global modulation vector of the b-th sample.

[0192] S406. Return the output of the hybrid expert routing network to the feedforward sub-layer for residual connection and layer normalization processing to obtain the corresponding feedforward sub-layer output. Then, input the feedforward sub-layer output to the diffusion denoising backbone network to obtain noise-free estimated samples.

[0193] For example, the residual connection operation adds element-wise the output of the hybrid expert routing network to the original global modulation vector input to the feedforward sub-layer, followed by layer normalization to stabilize the feature distribution and prevent gradient vanishing. The final output of the feedforward sub-layer is then passed to subsequent layers of the diffusion denoising backbone network to complete all remaining diffusion denoising steps, ultimately outputting a noiseless estimated sample with dimensions and structure completely identical to the original time series sample. The specific calculation formula is as follows:

[0194]

[0195] in, This represents the feature representation of the nth token in the (l+1)th layer.

[0196] The task-oriented multi-source data generation method for large-scale industrial time-series models provided in this application maps target task types to deterministic task pattern vectors, providing stable task condition biases and guiding the expert selection process towards current task requirements. By combining the intermediate outputs of the feedforward sublayer with the task pattern vectors to calculate routing probabilities, task conditions and input features jointly determine expert activation weights. Selecting top-ranked experts to form an activation set enables a sparse activation mechanism, preserving feature diversity while controlling computational costs. Activating expert outputs are tailored to specific operating conditions or faults, accurately capturing the unique statistical characteristics of different industrial operating modes. Weighted summation of the activated expert outputs according to routing probabilities fuses multi-expert features, ensuring continuity of feature scale and stability of expression. Residual connections and layer normalization processes stabilize feature distribution, ultimately generating high-quality, noise-free estimation samples that meet task requirements.

[0197] Figure 5 A flowchart illustrating the task-oriented multi-source data generation method for large-scale industrial time-series models provided in this application embodiment. Figure 4 ,like Figure 5 As shown, this embodiment, based on the above embodiments, provides a detailed description of the construction process of the frozen task evaluator, including:

[0198] S501. Obtain historical time series samples collected by multiple sensors of industrial equipment as a time series training set, as well as the corresponding task supervision signal.

[0199] In this embodiment, the task supervision signal is the condition label corresponding to the time series training set.

[0200] For example, historical time-series samples of industrial equipment can be constructed by synchronously collecting data from multiple sensor channels, such as temperature, vibration, pressure, and current. The data from each channel is first precisely time-aligned according to the collection timestamp, then short-term missing values ​​are filled in using linear interpolation, and finally, amplitude normalization is achieved through Z-score standardization, forming uniformly formatted training input data. The task supervision signal corresponds one-to-one with the time-series training set, directly using pre-labeled conditional labels such as fault category, operating condition, and remaining service life.

[0201] S502. Based on the target task, determine the corresponding downstream tasks.

[0202] For example, downstream tasks are divided according to the application scenario of the upper-level target task. If the target task is the full life cycle health management of industrial equipment, it is uniformly divided into three core downstream tasks: fault classification, life regression prediction, and equipment anomaly detection. Each type of downstream task corresponds to a specific condition label type, which realizes different dimensions of equipment status assessment capabilities.

[0203] S503. For each type of downstream task, train multiple independent task evaluation models.

[0204] In this embodiment, the downstream tasks include fault classification tasks, lifetime regression prediction tasks, and equipment anomaly detection tasks. The task evaluation model is an evaluation model used to input industrial multi-source time series and output task prediction indicators. The task evaluation model includes any one of classification evaluation model, regression prediction evaluation model, and anomaly detection evaluation model.

[0205] For example, suitable network architectures are matched for different types of downstream tasks: fault classification tasks use a one-dimensional convolutional neural network with fully connected layers, lifetime regression prediction tasks use a long short-term memory network with an attention mechanism, and equipment anomaly detection tasks use a stacked autoencoder structure. This application defines a unified task reward function for each of the three types of tasks:

[0206] Classification task reward:

[0207]

[0208] in, The reward value for categorizing tasks. Let cross-entropy be the loss function. The output of the task evaluation model, For task monitoring signals, All of these are hyperparameters (weighting coefficients). The function for calculating the classification margin. This is the entropy calculation function.

[0209] Regression prediction task reward:

[0210]

[0211] in, The reward value for the regression prediction task. The industrial risk weighting function is related to the direction of error. This is the smoothed mean absolute error (Huber) loss function.

[0212] Anomaly detection task reward:

[0213]

[0214] in, This is the reward value for the anomaly detection task. This is the cross-entropy loss function for binary classification.

[0215] All task evaluation models use independent parameter spaces, and the training process does not interfere with each other.

[0216] S504. Input the time series training set into the corresponding type of task evaluation model, and combine it with the task supervision signal to fit the model parameters to obtain multiple fitted target task evaluation models.

[0217] For example, different loss functions are used for parameter optimization for different tasks: the classification evaluation model uses the cross-entropy loss function, the regression prediction evaluation model uses the mean squared error loss function, and the anomaly detection evaluation model uses the reconstruction error loss function. During training, the training set and the validation set are divided in a 7:3 ratio. An early stopping strategy is used to prevent overfitting. Training is terminated when the performance of the validation set does not improve for 10 consecutive epochs, resulting in a converged evaluation model for the target task.

[0218] S505. Lock the network parameters of each target task evaluation model to obtain multiple locked target task evaluation models.

[0219] For example, a deep learning framework can be used to set all network parameters of the evaluation model for all target tasks to an untrainable state, i.e., freeze the gradient calculation of all weights and biases. These parameters will remain constant throughout the subsequent training of the conditional diffusion generative model and will not change as the parameters of the generated model are updated.

[0220] S506. Integrate the evaluation models of each locked target task to obtain the frozen task evaluator.

[0221] For example, a parallel evaluation unit architecture is used for integration to build a unified input-output interface. The input interface simultaneously distributes noise-free estimation samples to all locked target task evaluation models, and the output interface collects the prediction results of each model, automatically selects the corresponding evaluation signal according to the type of the currently executed target task, or fuses multi-task evaluation signals according to preset weights to form the final output task evaluation signal.

[0222] The task-oriented multi-source data generation method for large-scale industrial time-series models provided in this application ensures training data quality and provides a reliable foundation for model fitting by acquiring preprocessed historical time-series training sets and corresponding supervision signals. Dividing downstream tasks according to the target task clarifies the training objectives of each evaluation model, comprehensively covering core industrial application scenarios. Training independent evaluation models for different downstream tasks matches the characteristics of each task, improving the evaluation accuracy of corresponding dimensions. Combining supervision signals to complete model parameter fitting enables each model to possess accurate task prediction and evaluation capabilities. Locking all evaluation model parameters ensures the stability of evaluation criteria and avoids drift during the generation process. Integrating multiple models to form a unified frozen task evaluator provides reliable and consistent task constraints for the generated models.

[0223] Figure 6 A flowchart illustrating the task-oriented multi-source data generation method for large-scale industrial time-series models provided in this application embodiment. Figure 5 ,like Figure 6 As shown, this embodiment, based on the above embodiments, provides a detailed explanation of the routing probability adjustment process, including:

[0224] S601. Perform batch standardization processing on the task evaluation signal to obtain a standardized evaluation signal.

[0225] For example, firstly, the task evaluation signals corresponding to all noiseless estimation samples within the current training batch are collected, and the mean and variance of the batch signals are calculated. Then, each evaluation signal is converted into a standardized evaluation signal with a mean of 0 and a variance of 1 according to the Z-score standardization formula. This process can eliminate the scale difference of task evaluation signals between different batches, avoid gradient instability caused by excessive fluctuations in evaluation signals, and ensure the consistency of subsequent advantage function calculations.

[0226] S602. Calculate the difference between the standardized evaluation signal and the corresponding exponential moving average baseline of the task type to obtain the corresponding advantage function value.

[0227] In this embodiment, the exponential moving average baseline is a historical performance score baseline that is pre-maintained for each task type.

[0228] For example, the system maintains a separate exponential moving average baseline for each type of downstream task, initially set to 0. After each iteration, the baseline is updated using the average standardized evaluation signal of the corresponding task in the current batch, with the smoothing coefficient typically set to 0.9. The advantage function value represents the degree to which the task performance of the current sample is better or worse than the historical average; a positive value indicates performance better than the historical average, and a negative value indicates performance worse than the historical average. The mathematical expression is:

[0229]

[0230] in, For the task The corresponding advantage function value, The standardized task reward value. For the task The corresponding exponential moving average baseline.

[0231] S603. Generate the route-oriented task feedback loss based on the advantage function value.

[0232] For example, first define the set of expert paths selected by the current sample across all Hybrid Expert Structure (MoE) layers and all tokens, with the log probability as:

[0233]

[0234] in, For expert path set, Let be the logarithmic probability of the path set.

[0235] The feedback loss for route-oriented tasks is expressed as the product of a negative dominance function value and the log probability of the path:

[0236]

[0237] in, Feedback on the loss for the routing task. To stop the gradient operation.

[0238] When the dominance function is positive, the loss value is small, which enhances the activation tendency of the corresponding expert path; when the dominance function is negative, the loss value is large, which weakens the activation tendency of the corresponding expert path. This loss only applies to the routing controller of the hybrid expert routing network and does not affect other parameters of the diffusion denoising backbone network.

[0239] S604. Based on the feedback loss from the routing guidance task, adjust the parameters of the routing controller in the hybrid expert routing network to obtain the adjusted routing probability.

[0240] For example, the route-oriented task feedback loss is passed to the route controller via backpropagation. Only the gradients of the weights and bias parameters within the route controller are calculated, and these parameters are updated using an Adaptive Moment Estimation (Adam) optimizer. After the parameters are updated, the route controller recalculates the routing probabilities of each expert subnetwork based on the new parameters, forming a new expert activation distribution. This makes the subsequent generation process more inclined to select expert paths that can produce higher task evaluation signals.

[0241] In one possible implementation, after obtaining the adjusted routing probabilities, the method further includes:

[0242] Based on the noiseless estimated samples and the original time series samples, the corresponding diffusion denoising loss is determined; the preset weighting weights corresponding to the diffusion denoising loss are obtained; the first weighting weight of the routing task feedback loss is determined using a preset delayed start and cosine growth strategy; the second weighting weight of the expert load balancing loss is determined using a preset gradual descent and lower bound retention strategy; based on the preset weighting weights, the first weighting weight, and the second weighting weight, the diffusion denoising loss, the expert load balancing loss, and the routing task feedback loss are weighted and summed to obtain the overall training loss; based on the overall training loss, the model parameters of the conditional diffusion generation model are updated.

[0243] For example, after outputting noiseless estimated samples, the conditional diffusion generation model calculates the difference between these samples and the original time series samples under the same time-aligned benchmark to obtain the diffusion denoising loss. This loss, along with preset weighted weights, participates in the construction of the overall objective. The route-oriented task feedback loss and the expert load balancing loss correspond to the task adaptability and expert call uniformity of the route controller, respectively. These three losses are dynamically weighted to form the overall training loss, which is used to jointly update the parameters of the diffusion denoising backbone network, the feedforward sublayer, and the hybrid expert routing network. The expert load balancing loss formula is:

[0244]

[0245] in, For expert load balancing losses, The average route importance of the e-th expert at layer l. , This represents the actual load ratio of the e-th expert in layer l. , This is an indicator function, which is 1 when e belongs to the set of activation experts, and 0 otherwise.

[0246] The task feedback weighting adopts a delayed start and cosine growth strategy as follows:

[0247]

[0248] in, The task feedback loss weights are used for the e-th training round, where e is the current training round. The delayed start rounds for task feedback loss For the total number of training rounds, The maximum weight for task feedback loss.

[0249] The load balancing weight adopts a strategy of gradually decreasing while retaining a lower limit:

[0250]

[0251] in, The load balancing loss weights are used for the e-th training round. This represents the minimum weight (lower bound) for load balancing losses. The initial weights for load balancing losses.

[0252] The overall training loss is:

[0253]

[0254] in, This represents the overall training loss.

[0255] In practical applications, the preset weighting weight, the first weighting weight, and the second weighting weight can also be preset and configured according to training stability, the fluctuation range of task evaluation signals, and expert activation distribution. This application embodiment does not limit this.

[0256] This approach enables the diffusion reconstruction objective, task feedback objective, and expert equilibrium objective to collaboratively constrain model training within the same optimization framework. This allows the routing controller's adjustments to simultaneously consider the authenticity of generated samples, task relevance, and the balance of expert calls. Consequently, after the model parameters are updated, the conditional diffusion generation model can more stably generate industrial multi-source time series samples that conform to conditional labels and task types, and enhance the samples' ability to support downstream tasks.

[0257] The task-oriented multi-source data generation method for large-scale industrial time-series models provided in this application eliminates scale differences between batches by batch standardizing the task evaluation signal, avoiding gradient instability and ensuring consistency in subsequent calculations. Calculating the difference between the standardized signal and the exponential moving average baseline of the corresponding task objectively measures the performance of the current sample relative to the historical average. Generating a route-oriented task feedback loss based on the dominance function value enables differentiated rewards and penalties for different expert paths, guiding the route towards higher performance. Adjusting the route controller parameters based on the feedback loss dynamically optimizes the expert activation distribution, making the generation process more aligned with the target task requirements.

[0258] Figure 7 This is a schematic diagram of a task-oriented multi-source data generation device for large-scale industrial time-series models provided in an embodiment of this application. The device in this embodiment can be in software and / or hardware form. Figure 7 As shown in the embodiment of this application, the task-oriented multi-source data generation device 700 for industrial time-series large models includes: an acquisition module 701, a determination module 702, a first processing module 703, a second processing module 704, an adjustment module 705, and a generation module 706.

[0259] The acquisition module 701 is used to acquire the raw time series samples collected by multiple sensors of industrial equipment and the task type of the target task;

[0260] The determination module 702 is used to determine the corresponding condition labels based on the original time series samples. The condition labels include at least one of the following: fault category label, operating condition label, equipment health status label, remaining service life interval label, and continuous remaining service life label.

[0261] The first processing module 703 is used to perform conditional diffusion generation processing based on the original time series samples, task type and condition label using a pre-trained conditional diffusion generation model to obtain the corresponding noiseless estimation samples. The conditional diffusion generation model is a neural network model used to learn the distribution characteristics of real industrial time series and generate new industrial time series. The conditional diffusion generation model includes a diffusion denoising backbone network. The feedforward sublayer of the diffusion denoising backbone network is equipped with a hybrid expert routing network. The hybrid expert routing network includes multiple expert subnetworks and a routing controller. The routing controller is used to calculate the routing probability of each expert subnetwork and select the expert subnetwork to be activated.

[0262] The second processing module 704 is used to input the noiseless estimation sample into the pre-built frozen task evaluator to obtain the corresponding task evaluation signal. The frozen task evaluator is a task evaluation device used to output the performance score corresponding to the target task based on the noiseless estimation sample.

[0263] The adjustment module 705 is used to adjust the routing probability of the routing controller in the conditional diffusion generation model according to the task evaluation signal, so as to obtain the adjusted routing probability.

[0264] The generation module 706 is used to generate target industrial multi-source time series samples that meet the conditional labels and task types by using a conditional diffusion generation model based on the adjusted routing probabilities.

[0265] In one possible implementation, the first processing module 703 is further configured to:

[0266] The original time series sample is divided into multiple time segments of fixed length along the time axis. The original time series sample contains multiple sensor data acquisition channels.

[0267] Each time segment corresponding to a sensor data acquisition channel is independently mapped to obtain a subspace representation for each sensor data acquisition channel.

[0268] The subspace representations of each sensor data acquisition channel within the same time segment are spliced ​​together and time position encoding is added to generate a time segment-level feature sequence.

[0269] The time-segment-level feature sequences, task types, and conditional labels are input into the diffusion denoising backbone network for processing to obtain noise-free estimated samples.

[0270] In one possible implementation, the first processing module 703 is further configured to:

[0271] The time-segment level feature sequence is converted into a tensor containing a channel subspace structure, which is composed of the subspace representation corresponding to each sensor data acquisition channel;

[0272] In the time dimension, attention aggregation is performed on the features of each sensor data acquisition channel to obtain the time aggregated representation of each sensor data acquisition channel.

[0273] Attention weighting is applied to the temporal aggregation representation of each sensor data acquisition channel at the channel dimension to obtain the global running state vector.

[0274] A global modulation vector is generated based on the global operating state vector and the time aggregation representation of each sensor data acquisition channel.

[0275] The global modulation vector is input into the feedforward sub-layer of the diffusion denoising backbone network to obtain the intermediate output of the feedforward sub-layer. The intermediate output of the feedforward sub-layer is a high-dimensional feature vector containing the temporal variation features, inter-channel correlation features and global operating status features of the multi-sensor data of industrial equipment.

[0276] The intermediate output of the feedforward sublayer is input into each expert subnetwork of the hybrid expert routing network to obtain noiseless estimation samples.

[0277] In one possible implementation, the first processing module 703 is further configured to:

[0278] Map the task type of the target task to a deterministic task pattern vector;

[0279] The routing probability of each expert subnetwork is calculated based on the intermediate output of the feedforward sublayer and the deterministic task mode vector.

[0280] Based on the routing probability, each expert subnetwork is sorted from high to low, and the top k target expert subnetworks are selected to form a target expert subnetwork set, where k is a positive integer;

[0281] Based on the set of target expert subnetworks, the output of each target expert subnetwork is determined. The output of the target expert subnetwork is a dedicated high-dimensional feature vector adapted to a single equipment operating condition or fault scenario. The outputs of different expert subnetworks correspond to the feature transformation results of different industrial operation modes.

[0282] The outputs of each target expert subnetwork are weighted and summed to obtain the output of the hybrid expert routing network.

[0283] The output of the hybrid expert routing network is returned to the feedforward sublayer for residual connection and layer normalization processing to obtain the corresponding feedforward sublayer output.

[0284] The output of the feedforward sublayer is input into the diffusion denoising backbone network to obtain noise-free estimated samples.

[0285] In one possible implementation, the first processing module 703 is further configured to:

[0286] Calculate the average route importance and actual load ratio for each expert subnetwork;

[0287] Based on average route importance and actual load ratio, generate expert load balancing loss;

[0288] The parameters of the hybrid expert routing network are adjusted based on the expert load balancing loss.

[0289] In one possible implementation, the second processing module 704 is further configured to freeze the task evaluator, which is constructed through the following steps:

[0290] Historical time series samples collected by multiple sensors of industrial equipment are obtained as the time series training set, along with the corresponding task supervision signals. The task supervision signals are the conditional labels corresponding to the time series training set.

[0291] Based on the target task, identify the corresponding downstream tasks;

[0292] For each type of downstream task, multiple independent task evaluation models are trained. The downstream tasks include fault classification tasks, lifetime regression prediction tasks, and equipment anomaly detection tasks. The task evaluation model is an evaluation model used to input industrial multi-source time series and output task prediction indicators. The task evaluation model can be any one of classification evaluation model, regression prediction evaluation model, or anomaly detection evaluation model.

[0293] Input the time series training set into the corresponding type of task evaluation model, and combine the task supervision signal to fit the model parameters to obtain multiple fitted target task evaluation models.

[0294] The network parameters of each target task evaluation model are locked to obtain multiple locked target task evaluation models.

[0295] The evaluation models of each locked target task are integrated to obtain the frozen task evaluator.

[0296] In one possible implementation, the adjustment module 705 is further configured to:

[0297] Batch standardization processing is performed on the task evaluation signal to obtain a standardized evaluation signal;

[0298] The difference between the standardized evaluation signal and the corresponding exponential moving average baseline of the task type is calculated to obtain the corresponding advantage function value. The exponential moving average baseline is a historical performance score baseline that is pre-maintained for each task type.

[0299] Based on the advantage function value, generate the route-oriented task feedback loss;

[0300] Based on the feedback loss from the routing guidance task, the parameters of the routing controller in the hybrid expert routing network are adjusted to obtain the adjusted routing probability.

[0301] In one possible implementation, the adjustment module 705 is further configured to:

[0302] Based on the noiseless estimation sample and the original time series sample, determine the corresponding diffusion denoising loss;

[0303] Obtain the preset weighting weights corresponding to the diffusion denoising loss;

[0304] A preset delayed start and cosine growth strategy is used to determine the first weighted weight of the feedback loss of the routing task;

[0305] A pre-defined strategy of gradually decreasing while retaining a lower limit is adopted to determine the second weighting weight of the expert load balancing loss;

[0306] Based on the preset weighting weights, the first weighting weight, and the second weighting weight, the diffusion denoising loss, the expert load balancing loss, and the route-oriented task feedback loss are summed to obtain the overall training loss.

[0307] Update the model parameters of the conditional diffusion generative model based on the overall training loss.

[0308] The task-oriented multi-source data generation device for industrial time-series large models provided in this embodiment can execute the methods provided in the above-described method embodiments. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0309] Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 8 As shown, the electronic device 800 provided in this embodiment includes at least one processor 801 and a memory 802. Optionally, the device 800 further includes a communication component 803. The processor 801, memory 802, and communication component 803 are connected via a bus.

[0310] In a specific implementation, at least one processor 801 executes computer execution instructions stored in memory 802, causing at least one processor 801 to perform the above-described method.

[0311] The specific implementation process of processor 801 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0312] 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 (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0313] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0314] 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. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0315] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0316] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0317] The aforementioned 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 storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0318] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from 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 reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0319] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0320] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0321] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0322] If a function is implemented as 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 this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0323] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0324] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A task-oriented multi-source data generation method for large-scale industrial time-series models, characterized in that, include: Acquire raw time-series samples collected by multiple sensors in industrial equipment and the task type of the target task; Based on the original time series sample, the corresponding condition label is determined, wherein the condition label includes at least one of the following: fault category label, operating condition label, equipment health status label, remaining service life interval label, and continuous remaining service life label. A pre-trained conditional diffusion generation model is used to perform conditional diffusion generation processing based on the original time series samples, the task type, and the conditional label to obtain corresponding noise-free estimated samples. The conditional diffusion generation model is a neural network model used to learn the distribution characteristics of real industrial time series and generate new industrial time series. The conditional diffusion generation model includes a diffusion denoising backbone network. The feedforward sublayer of the diffusion denoising backbone network is equipped with a hybrid expert routing network. The hybrid expert routing network includes multiple expert subnetworks and a routing controller. The routing controller is used to calculate the routing probability of each expert subnetwork and select the expert subnetwork to be activated. The noiseless estimation sample is input into a pre-built frozen task evaluator to obtain the corresponding task evaluation signal. The frozen task evaluator is a task evaluation device used to output the performance score corresponding to the target task based on the noiseless estimation sample. Based on the task evaluation signal, the routing probability of the routing controller in the conditional diffusion generation model is adjusted to obtain the adjusted routing probability; Based on the adjusted routing probability, target industrial multi-source time series samples that conform to the conditional label and the task type are generated through the conditional diffusion generation model.

2. The method according to claim 1, characterized in that, The conditional diffusion generation process based on the industrial multi-source time series samples, the task type, and the conditional label to obtain the corresponding noiseless estimation samples includes: The original time series sample is divided into multiple time segments of fixed length along the time axis, wherein the original time series sample contains multiple sensor data acquisition channels; Each time segment corresponding to a sensor data acquisition channel is independently mapped to obtain a subspace representation corresponding to each sensor data acquisition channel. The subspace representations of each sensor data acquisition channel within the same time segment are spliced ​​together and time position encoding is added to generate a time segment-level feature sequence. The time-segment-level feature sequence, the task type, and the conditional label are input into the diffusion denoising backbone network for processing to obtain the noise-free estimation sample.

3. The method according to claim 2, characterized in that, The step of inputting the time-segment-level feature sequence, the task type, and the conditional label into the diffusion denoising backbone network for processing to obtain the noise-free estimated sample includes: The time-segment level feature sequence is converted into a tensor containing a channel subspace structure, wherein the channel subspace structure is composed of the subspace representation corresponding to each of the sensor data acquisition channels; In the time dimension, attention aggregation is performed on the features of each sensor data acquisition channel to obtain the time aggregated representation of each sensor data acquisition channel; Attention weighting is applied to the temporal aggregation representation of each sensor data acquisition channel in the channel dimension to obtain a global running state vector. A global modulation vector is generated based on the global operating state vector and the time aggregation representation of each sensor data acquisition channel; The global modulation vector is input into the feedforward sub-layer of the diffusion denoising backbone network to obtain the intermediate output of the feedforward sub-layer. The intermediate output of the feedforward sub-layer is a high-dimensional feature vector containing the temporal variation features, inter-channel correlation features and global operating status features of the multi-sensor data of industrial equipment. The intermediate output of the feedforward sublayer is input into each expert subnetwork of the hybrid expert routing network to obtain the noiseless estimation sample.

4. The method according to claim 3, characterized in that, The step of inputting the intermediate output of the feedforward sublayer into each expert subnetwork of the hybrid expert routing network to obtain the noise-free estimation samples includes: Map the task type of the target task to a deterministic task pattern vector; The routing probability of each expert subnetwork is calculated based on the intermediate output of the feedforward sublayer and the deterministic task mode vector. Based on the routing probabilities, each expert subnetwork is sorted from high to low, and the top k target expert subnetworks are selected to form a target expert subnetwork set, where k is a positive integer; Based on the set of target expert sub-networks, the output of each target expert sub-network is determined, wherein the output of the target expert sub-network is a dedicated high-dimensional feature vector adapted to a single equipment operating condition or fault scenario, and the outputs of different expert sub-networks correspond to the feature transformation results of different industrial operation modes. The outputs of each target expert subnetwork are weighted and summed to obtain the output of the hybrid expert routing network. The output of the hybrid expert routing network is returned to the feedforward sublayer for residual connection and layer normalization processing to obtain the corresponding feedforward sublayer output. The output of the feedforward sublayer is input into the diffusion denoising backbone network to obtain the noise-free estimation sample.

5. The method according to claim 4, characterized in that, After determining the output of each target expert subnetwork based on the set of target expert subnetworks, the method further includes: Calculate the average route importance and actual load ratio for each of the expert subnetworks; Based on the average route importance and the actual load ratio, an expert load balancing loss is generated. The parameters of the hybrid expert routing network are adjusted based on the expert load balancing loss.

6. The method according to claim 1, characterized in that, The frozen task evaluator is constructed through the following steps: The historical time series samples collected by the multiple sensors of the industrial equipment are obtained as a time series training set, and the corresponding task supervision signal is obtained. The task supervision signal is a condition label corresponding to the time series training set. Based on the target task, determine the corresponding downstream tasks; For each type of downstream task, multiple independent task evaluation models are trained. The downstream tasks include fault classification tasks, lifetime regression prediction tasks, and equipment anomaly detection tasks. The task evaluation model is an evaluation model used to input industrial multi-source time series and output task prediction indicators. The task evaluation model includes any one of classification evaluation model, regression prediction evaluation model, and anomaly detection evaluation model. The time series training set is input into the corresponding type of task evaluation model, and the model parameters are fitted by combining the task supervision signal to obtain multiple fitted target task evaluation models. The network parameters of each target task evaluation model are locked to obtain multiple locked target task evaluation models. The frozen task evaluator is obtained by integrating each of the locked target task evaluation models.

7. The method according to claim 5, characterized in that, The step of adjusting the routing probability of the routing controller in the conditional diffusion generation model based on the task evaluation signal to obtain the adjusted routing probability includes: The task evaluation signal is batch standardized to obtain a standardized evaluation signal; The difference between the standardized evaluation signal and the corresponding exponential moving average baseline of the task type is calculated to obtain the corresponding advantage function value, wherein the exponential moving average baseline is a historical performance score baseline that is pre-maintained for each task type; Based on the aforementioned advantage function value, generate the route-oriented task feedback loss; Based on the feedback loss from the routing task, the parameters of the routing controller in the hybrid expert routing network are adjusted to obtain the adjusted routing probability.

8. The method according to claim 7, characterized in that, After adjusting the parameters of the route controller in the hybrid expert routing network based on the feedback loss from the route guidance task to obtain the adjusted route probability, the method further includes: Based on the noiseless estimation sample and the original time series sample, determine the corresponding diffusion denoising loss; Obtain the preset weighting weight corresponding to the diffusion denoising loss; A preset delayed start and cosine growth strategy is used to determine the first weighted weight of the feedback loss of the routing task; A second weighting weight for the expert load balancing loss is determined by adopting a preset strategy of gradually decreasing while retaining a lower limit. According to the preset weighted weights, the first weighted weights and the second weighted weights, the diffusion denoising loss, the expert load balancing loss and the routing task feedback loss are weighted and summed to obtain the overall training loss; The model parameters of the conditional diffusion generation model are updated based on the overall training loss.

9. A task-oriented multi-source data generation device for large-scale industrial time-series models, characterized in that, include: The acquisition module is used to acquire raw time series samples collected by multiple sensors of industrial equipment and the task type of the target task; The determination module is used to determine the corresponding condition label based on the original time series sample, wherein the condition label includes at least one of the following: fault category label, operating condition label, equipment health status label, remaining service life interval label, and continuous remaining service life label. The first processing module is used to perform conditional diffusion generation processing based on the original time series samples, the task type, and the conditional label using a pre-trained conditional diffusion generation model to obtain corresponding noise-free estimated samples. The conditional diffusion generation model is a neural network model used to learn the distribution characteristics of real industrial time series and generate new industrial time series. The conditional diffusion generation model includes a diffusion denoising backbone network. The feedforward sublayer of the diffusion denoising backbone network is equipped with a hybrid expert routing network. The hybrid expert routing network includes multiple expert subnetworks and a routing controller. The routing controller is used to calculate the routing probability of each expert subnetwork and select the expert subnetwork to be activated. The second processing module is used to input the noiseless estimation sample into a pre-built frozen task evaluator to obtain the corresponding task evaluation signal, wherein the frozen task evaluator is a task evaluation device used to output the performance score corresponding to the target task based on the noiseless estimation sample. An adjustment module is used to adjust the routing probability of the routing controller in the conditional diffusion generation model according to the task evaluation signal, so as to obtain the adjusted routing probability; The generation module is used to generate target industrial multi-source time series samples that conform to the conditional labels and the task type based on the adjusted routing probabilities and the conditional diffusion generation model.

10. An electronic device, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-8.

Images