A cloud-based industrial device multitask integrated operation and maintenance method and system

Abstract

The application provides a cloud-based industrial equipment multi-task integrated operation and maintenance method and system. The method comprises the following steps: obtaining an operation and maintenance task of an industrial equipment, constructing a multi-task integrated operation and maintenance basic model based on the operation and maintenance task; obtaining multi-modal feature data of the industrial equipment, and presetting a cross-modal feature fusion mechanism based on the multi-modal feature data; training the multi-task integrated operation and maintenance basic model by using a multi-modal cloud joint method based on the preset cross-modal feature fusion mechanism; obtaining a multi-task integrated operation and maintenance large model of the industrial equipment based on the trained multi-task integrated operation and maintenance basic model; and performing multi-task integrated operation and maintenance of the industrial equipment, such as anomaly detection, fault diagnosis and life prediction, based on the multi-task integrated operation and maintenance large model. The multi-task integrated operation and maintenance large model formed by the application significantly improves the operation and maintenance efficiency and reliability of the industrial equipment in a complex scene.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and industrial equipment operation and maintenance technology, specifically relating to a cloud-based multi-task integrated operation and maintenance method and system for industrial equipment. Background Technology

[0002] Industrial equipment, with its complex functional requirements, shows broad application prospects in fields such as intelligent manufacturing, service interaction, and special environment operation. However, its complex mechanical structure (usually containing dozens of degrees of freedom and hundreds of precision parts) and the unstructured dynamic scenarios it faces lead to significant challenges in system operation and maintenance: on the one hand, the time-varying loads and complex working conditions generated by the coordinated movement of multiple components can easily induce a variety of component failure modes; on the other hand, the signal characteristics of multiple sensory modalities such as vision, hearing, and touch of industrial equipment, such as robots, are deeply coupled in dynamic tasks, further exacerbating the difficulty of operation and maintenance feature extraction.

[0003] Currently, academia and industry have developed various intelligent operation and maintenance (O&M) models, demonstrating good O&M efficiency in specific standardized scenarios. However, when these models are directly applied to industrial equipment systems, they fail to achieve the expected O&M accuracy. The root cause lies in the inherent defects of traditional deep learning models: First, their feature extraction capabilities are limited by fixed network architectures, making it difficult to adapt to the massive and heterogeneous nature of multimodal data from industrial equipment; second, the models lack the ability to model the deep coupling relationships of multi-sensor signals, leading to the loss of crucial O&M information during feature fusion; third, when faced with high-dimensional, non-stationary time-series data, existing models struggle to effectively capture the feature evolution patterns at different time scales, thus hindering the real-time performance and accuracy of O&M decisions.

[0004] Therefore, how to enable the operation and maintenance model to achieve accurate diagnosis when faced with different forms of data (such as time series data, image data, text data, etc.) has become a core technical challenge that urgently needs to be overcome in the current multi-task integrated operation and maintenance methods for industrial equipment. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides a cloud-based multi-task integrated operation and maintenance method and system for industrial equipment, aiming to solve the problem that the feature learning capabilities of ordinary deep learning models are limited, making it difficult to achieve accurate operation and maintenance of industrial equipment based on different forms of data.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] A cloud-based multi-task integrated operation and maintenance method for industrial equipment, the method comprising:

[0008] S1. Obtain the operation and maintenance tasks of industrial equipment, and construct a multi-task integrated operation and maintenance basic model based on the operation and maintenance tasks;

[0009] S2. Obtain the multimodal feature data of the industrial equipment, and based on the multimodal feature data, preset a cross-modal feature fusion mechanism;

[0010] S3. Based on the preset cross-modal feature fusion mechanism, the multi-modal cloud joint method is used to train the multi-task integrated operation and maintenance basic model;

[0011] S4. Based on the trained multi-task integrated operation and maintenance basic model, obtain the multi-task integrated operation and maintenance large model of the industrial equipment.

[0012] S5. Based on the aforementioned multi-task integrated operation and maintenance model, perform multi-task integrated operation and maintenance on the industrial equipment, including anomaly detection, fault diagnosis, and lifespan prediction.

[0013] Preferably, S1 includes:

[0014] S11. Use the Transformer architecture to build a multi-task integrated operation and maintenance basic model;

[0015] S12. Using a model self-growth strategy based on Mu-Scaling technology, a dynamic expansion strategy for the multi-task integrated operation and maintenance basic model is set according to the complexity and data volume of the operation and maintenance tasks of the industrial equipment.

[0016] S13. By adopting parameter migration and feature distillation techniques, a knowledge inheritance method is set for the multi-task integrated operation and maintenance basic model during the growth process.

[0017] Preferably, S2 includes:

[0018] S21. Acquire the multimodal feature data of the industrial equipment and perform preprocessing;

[0019] S22. Based on the preprocessed multimodal feature data, a cross-modal feature fusion mechanism is preset: adaptively adjusting the multimodal feature data; using a cross-modal attention mechanism to extract features from the multimodal feature data; and using a progressive fusion method to fuse the extracted features.

[0020] Preferably, S3 includes:

[0021] S31. The multi-task integrated operation and maintenance basic model is trained using a supervised training strategy with frozen weight parameters:

[0022] Based on tensor and hybrid parallel architecture, dynamic allocation of cloud resources is achieved;

[0023] Based on the progressive parameter freezing mechanism, fine-tuning of the multi-task integrated operation and maintenance basic model is achieved to adapt it to the domain.

[0024] S32. The multi-task integrated operation and maintenance basic model is trained using online instruction fine-tuning driven by feedback reinforcement learning:

[0025] Based on the dynamic instruction generator, an adversarial sample generation algorithm is used to create training instructions, and the multi-task integrated operation and maintenance basic model is trained based on the training instructions.

[0026] Based on a two-stage strategy of offline reinforcement learning and online real-time feedback, fine-tuning of the multi-task integrated operation and maintenance basic model is achieved.

[0027] The present invention also provides a cloud-based multi-task integrated operation and maintenance system for industrial equipment, used to implement the aforementioned cloud-based multi-task integrated operation and maintenance method for industrial equipment. The system includes: a basic model module, a fusion mechanism module, a training module, a model acquisition module, and an operation and maintenance module.

[0028] The basic model module is used to acquire the operation and maintenance tasks of industrial equipment and, based on the operation and maintenance tasks, construct a multi-task integrated operation and maintenance basic model.

[0029] The fusion mechanism module is used to acquire multimodal feature data of the industrial equipment and, based on the multimodal feature data, preset a cross-modal feature fusion mechanism.

[0030] The training module is used to train the multi-task integrated operation and maintenance basic model based on a preset cross-modal feature fusion mechanism and a multi-modal cloud joint method.

[0031] The model acquisition module is used to obtain a large multi-task integrated operation and maintenance model of the industrial equipment based on the trained multi-task integrated operation and maintenance basic model.

[0032] The operation and maintenance module is used to perform multi-task integrated operation and maintenance of the industrial equipment based on the multi-task integrated operation and maintenance model, including anomaly detection, fault diagnosis, and life prediction.

[0033] Preferably, the basic model module includes: a construction unit, a dynamic expansion unit, and a knowledge inheritance unit;

[0034] The building unit is used to build a multi-task integrated operation and maintenance basic model using the Transformer architecture;

[0035] The dynamic expansion unit is used to set the dynamic expansion strategy of the multi-task integrated operation and maintenance basic model according to the complexity and data volume of the operation and maintenance tasks of the industrial equipment through a model self-growth strategy based on Mu-Scaling technology.

[0036] The knowledge inheritance unit is used to set the knowledge inheritance method of the multi-task integrated operation and maintenance basic model during the growth process by adopting parameter transfer and feature distillation techniques.

[0037] Preferably, the fusion mechanism module includes: a data acquisition unit and a mechanism preset unit;

[0038] The data acquisition unit is used to acquire the multimodal feature data of the industrial equipment and perform preprocessing.

[0039] The mechanism preset unit is used to preset a cross-modal feature fusion mechanism based on the preprocessed multimodal feature data: adaptively adjust the multimodal feature data; extract features from the multimodal feature data using a cross-modal attention mechanism; and fuse the extracted features using a progressive fusion method.

[0040] Preferably, the training module includes: a supervised training unit and an online instruction fine-tuning unit;

[0041] The supervised training unit is used to train the multi-task integrated operation and maintenance basic model using a supervised training strategy with frozen weight parameters.

[0042] Based on tensor and hybrid parallel architecture, dynamic allocation of cloud resources is achieved;

[0043] Based on the progressive parameter freezing mechanism, fine-tuning of the multi-task integrated operation and maintenance basic model is achieved to adapt it to the domain.

[0044] The online instruction fine-tuning unit is used to train the multi-task integrated operation and maintenance basic model using feedback reinforcement learning-driven online instruction fine-tuning.

[0045] Based on the dynamic instruction generator, an adversarial sample generation algorithm is used to create training instructions, and the multi-task integrated operation and maintenance basic model is trained based on the training instructions.

[0046] Based on a two-stage strategy of offline reinforcement learning and online real-time feedback, fine-tuning of the multi-task integrated operation and maintenance basic model is achieved.

[0047] The present invention also provides an electronic device, including at least one processor, at least one network interface, a user interface, a memory, at least one communication bus, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the aforementioned method.

[0048] The present invention also provides a computer-readable storage medium storing a computer program that, when executed, implements the aforementioned method.

[0049] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0050] This invention proposes a cloud-based multi-task integrated operation and maintenance method and system for industrial equipment. It employs an adaptive Transformer architecture and a Mu-Scaling dynamic expansion strategy to construct a basic multi-task integrated operation and maintenance model suitable for different operation and maintenance tasks. By designing a cross-modal fusion mechanism between different modal data features, it achieves progressive fusion of multi-modal features. Using a supervised training strategy with frozen weight parameters, combined with cloud-based joint pre-training and feedback reinforcement learning, it eliminates the "illusion" of the operation and maintenance model, forming a large-scale multi-task integrated operation and maintenance model with a super-large model structure and strong feature capture capabilities, significantly improving the operation and maintenance efficiency and reliability in complex scenarios. Attached Figure Description

[0051] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 This is a flowchart of the cloud-based multi-task integrated operation and maintenance large model construction method for industrial equipment according to an embodiment of the present invention;

[0053] Figure 2 This is an overall framework diagram of the multi-task integrated operation and maintenance large model construction method according to an embodiment of the present invention;

[0054] Figure 3 This is a schematic diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0057] Example 1

[0058] like Figure 1-2 As shown, this invention provides a cloud-based multi-task integrated operation and maintenance method for industrial equipment, including:

[0059] S1. Obtain the operation and maintenance tasks of industrial equipment, and build a multi-task integrated operation and maintenance basic model based on the operation and maintenance tasks;

[0060] S2. Acquire multimodal feature data of industrial equipment, and based on the multimodal feature data, preset a cross-modal feature fusion mechanism;

[0061] S3. Based on the preset cross-modal feature fusion mechanism, a multi-modal cloud joint method is used to train the multi-task integrated operation and maintenance basic model;

[0062] S4. Based on the trained multi-task integrated operation and maintenance basic model, obtain a large multi-task integrated operation and maintenance model for industrial equipment.

[0063] S5. Based on a multi-task integrated operation and maintenance model, it performs multi-task integrated operation and maintenance of industrial equipment, including anomaly detection, fault diagnosis, and life prediction.

[0064] Furthermore, the specific implementation process of the present invention is as follows:

[0065] S1 acquires the operation and maintenance tasks of industrial equipment, and based on these tasks, constructs a multi-task integrated operation and maintenance basic model, including:

[0066] S11. Set up a basic multi-task integrated operation and maintenance model: Use the Transformer architecture to build an initial small-scale multi-task integrated operation and maintenance basic model, and make full use of its advantages in processing sequential data and capturing long-distance dependencies.

[0067] S12. Set a dynamic expansion strategy for the multi-task integrated operation and maintenance basic model. Through the model self-growth strategy based on Mu-Scaling technology, the depth and width of the multi-task integrated operation and maintenance basic model are dynamically adjusted according to the complexity of the operation and maintenance tasks of industrial equipment and the growth of data volume. This enables the adaptive expansion of the multi-task integrated operation and maintenance basic model and ensures that the multi-task integrated operation and maintenance basic model can flexibly cope with operation and maintenance tasks of different scales and difficulties.

[0068] The dynamic expansion strategy process of the multi-task integrated operation and maintenance basic model is as follows:

[0069] S121. Based on the complexity of the operation and maintenance tasks of industrial equipment and the growth of data volume, set the depth and width of the multi-task integrated operation and maintenance basic model; the width refers to the dimension of the input and output of the embedding layer and each Transformer block, and the depth refers to the number of stacked layers in the neural network.

[0070] S122. Determine the range of hyperparameters for hyperparameter search, such as setting the learning rate to 0.0005 to 0.1, the batch size to 32 to 256, and the number of iterations to 50 to 300 rounds; then try different combinations of hyperparameters on the multi-task integrated operation and maintenance basic model, train each combination, and record the optimal hyperparameters.

[0071] S123. Train multi-task integrated operation and maintenance basic models with different widths and depths through optimal hyperparameters, and obtain the loss value by fitting the loss of multi-task integrated operation and maintenance basic models with different widths and depths.

[0072] S124. Based on the loss value, select the multi-task integrated operation and maintenance basic model with the lowest loss under a specific width and depth.

[0073] S13. Define a knowledge inheritance method for the multi-task integrated operation and maintenance basic model during its growth process. By adopting parameter transfer and feature distillation techniques, the weight coefficients and feature representations of the trained multi-task integrated operation and maintenance basic model are inherited to ensure the continuity and stability of the multi-task integrated operation and maintenance basic model during the iteration process, avoid the waste of computing resources caused by retraining, and promote the effective transfer and accumulation of knowledge.

[0074] The multi-task integrated operation and maintenance basic model constructed in this invention is a self-growing structural model. By constructing the multi-task integrated operation and maintenance basic model, a super-large-scale topology structure can be obtained to adapt to the needs of different operation and maintenance tasks.

[0075] S2 acquires multimodal feature data of industrial equipment. Based on this multimodal feature data, a pre-defined cross-modal feature fusion mechanism is established, including:

[0076] S21. Acquire multimodal feature data of industrial equipment and perform preprocessing:

[0077] Multimodal feature data, including time-series signal data, image data, and text data, was collected from multiple sources such as industrial equipment, monitoring equipment, and maintenance records. These data were then subjected to a series of preprocessing operations. Preprocessing operations included: filtering and denoising to eliminate interference and noise in the data; standardization, scaling, and normalization to unify the data scale; pruning to remove irrelevant information; cleaning to correct errors and fill in missing values; and word segmentation and vectorization for subsequent feature extraction and analysis. These preprocessing steps aimed to ensure a reliable dataset, laying the foundation for subsequent feature fusion and training.

[0078] S22. Based on the preprocessed multimodal feature data, a cross-modal feature fusion mechanism is preset: adaptively adjusting the multimodal feature data; using a cross-modal attention mechanism to extract features from the multimodal feature data; and employing a progressive fusion method to fuse the extracted features.

[0079] To achieve adaptive adjustment of multimodal feature data, a cross-modal attention mechanism is used for feature extraction. This mechanism dynamically adjusts the weights of feature extraction, enabling the multi-task integrated operation and maintenance (MOP) foundation model to focus more on features relevant to the task objective, thereby effectively associating and fusing features from different modalities. Simultaneously, the learning rate and the contribution ratio of each modality's feature data are adaptively adjusted based on data complexity and the MOP's performance during training. This step aims to optimize training efficiency, improve the performance of the MOP, and ensure that it fully captures effective information from the multimodal data. Feature fusion is implemented in stages using a progressive framework, starting with the fusion of some modalities and gradually transitioning to deep multimodal fusion, ensuring that the MOP foundation model can robustly improve its robustness and generalization ability.

[0080] The adaptive adjustment of the learning rate and the contribution ratio of each modality can be expressed as:

[0081] 1. The multi-task integrated operation and maintenance basic model has low complexity, low computational requirements, and a high contribution ratio. By choosing a smaller learning rate, we can ensure that the multi-task integrated operation and maintenance basic model learns the optimal feature information.

[0082] 2. The multi-task integrated operation and maintenance basic model is highly complex, has large computational requirements, and contributes a small proportion. Therefore, a larger learning rate is selected to ensure the efficiency of training the multi-task integrated operation and maintenance basic model.

[0083] 3. For multi-task integrated operation and maintenance basic models with excellent training performance, select a high contribution ratio; for multi-task integrated operation and maintenance basic models with poor training performance, select a low contribution ratio.

[0084] The processing flow of the cross-modal attention mechanism is as follows:

[0085] Feature extraction is performed on data of different modalities (such as text, images, and videos) to obtain their respective feature sequences;

[0086] The attention score is obtained by calculating the dot product between the feature sequences of one modality as the query and the feature sequences of the other modality as the key.

[0087] The calculated attention scores are used to weight the Value (the feature sequence of another modality) to generate the fused feature sequence.

[0088] The progressive framework starts with the fusion of simple partial modalities (such as images and text) in stages, helping the multi-task integrated operation and maintenance basic model learn simple relationships. Then, based on this, other modal information is gradually added to learn more complex feature associations, and finally, the deep fusion of multimodal information is achieved.

[0089] S3, based on a pre-defined cross-modal feature fusion mechanism, employs a multi-modal cloud-based joint method to train a multi-task integrated operation and maintenance basic model, including:

[0090] S31. The supervised training strategy of weight parameter freezing is adopted to train the multi-task integrated operation and maintenance basic model. The supervised training strategy of weight parameter freezing includes: tensor and hybrid parallel architecture and progressive parameter freezing mechanism.

[0091] The tensor and hybrid parallel architecture employs a three-dimensional parallel strategy (data parallelism × model parallelism × tensor parallelism). It dynamically allocates computing tasks through a cloud resource scheduler, achieving dynamic allocation of cloud resources. The three-dimensional parallel strategy refers to distributing different parts of the multi-task integrated operation and maintenance basic model to different nodes according to requirements (model parallelism), dividing data into multiple subsets and distributing them to different nodes (data parallelism), and performing tensor partitioning and computation on the subsets (tensor parallelism). By adopting a multi-dimensional parallel architecture, it achieves efficient training of ultra-large-scale multi-task integrated operation and maintenance basic models.

[0092] The progressive parameter freezing mechanism employs a multi-stage model weight freezing approach to achieve fine-tuning of the training multi-task integrated operation and maintenance basic model for domain adaptation. The progressive parameter freezing mechanism is implemented in three stages: the first freezing stage freezes the bottom layer of the multi-task integrated operation and maintenance basic model, fine-tuning only the top layer to adapt to the new domain; the second freezing stage gradually unfreezes the intermediate layers according to importance and introduces knowledge distillation to inject domain knowledge; the third freezing stage is used for global unfreezing, fine-tuning all parameters with a low learning rate to achieve deep knowledge fusion while preventing catastrophic forgetting.

[0093] S32. The multi-task integrated operation and maintenance basic model is trained using online instruction fine-tuning driven by feedback reinforcement learning:

[0094] Operation and maintenance (O&M) models sometimes output content that deviates from user input or factual knowledge; this phenomenon is known as O&M model "illusion." Since erroneous O&M results generated by this illusion can directly harm the quality and safety of industrial equipment O&M operations, feedback reinforcement learning is used to perform online instruction fine-tuning during the training process of the multi-task integrated O&M basic model. This guides the multi-task integrated O&M basic model to overcome its erroneous knowledge perception; this is the feedback reinforcement learning-driven online instruction fine-tuning stage, which includes a dynamic instruction generator and a two-stage strategy optimization.

[0095] The dynamic instruction generator uses adversarial example generation algorithms to dynamically create diverse training instructions to enhance the robustness and generalization ability of the model.

[0096] The two-stage policy optimization employs a two-stage approach, including an offline reinforcement learning stage and an online real-time feedback stage. The offline reinforcement learning stage uses the Proximal Policy Optimization (PPO) algorithm, which sets a multi-objective reward function. The algorithm can be expressed as:

[0097] R = β1·ACC hallucination +β2·Δ confidence

[0098] Where R represents the multi-objective reward function, ACC hallucination The accuracy rate of hallucination detection represents the model's ability to correctly identify and correct hallucination phenomena, Δ. confidence β1 and β2 represent the weighting coefficients, reflecting the degree of fluctuation in the model's output confidence level, indicating decision stability. By optimizing the multi-objective reward function using the PPO algorithm, the multi-task integrated operation and maintenance basic model can learn a decision-making strategy that balances illusion detection accuracy and decision stability in an offline environment, laying the foundation for the subsequent online real-time feedback stage.

[0099] The specific process of the Proximal Policy Optimization (PPO) algorithm in the illusion correction task of the multi-task integrated operation and maintenance basic model is as follows:

[0100] First, the behavior of the multi-task integrated operation and maintenance basic model is evaluated through a predefined multi-objective reward function, where the hallucination detection accuracy measures the factual consistency of the content generated by the multi-task integrated operation and maintenance basic model, and the decision stability is used to avoid policy drift caused by frequent corrections.

[0101] Next, PPO uses a policy network (i.e., a Transformer-based generative model) to iteratively sample the dataset. By clipping the objective function, it limits the policy update magnitude to avoid unstable updates caused by noise in the reward signal. At the same time, it uses the advantage function to reduce variance and improve sample efficiency.

[0102] In each iteration, the multi-task integrated operation and maintenance basic model generates actions based on the old strategy and uses importance sampling to estimate the advantage of the new strategy. If the new strategy does not significantly reduce stability while maintaining high accuracy, it is accepted for updating; otherwise, it reverts to the conservative strategy through a pruning mechanism.

[0103] Ultimately, after multiple iterations, the multi-task integrated operation and maintenance basic model learned a strategy to balance illusion correction capability and decision robustness in the offline phase, and dynamically adjusted the confidence threshold.

[0104] The local fine-tuning mechanism rapidly corrects the multi-task integrated operations and maintenance (O&M) foundation model by freezing some parameters and increasing the learning rate. Freezing some parameters helps maintain the stable decision-making strategies learned by the O&M foundation model during the offline learning phase, avoiding the uncertainty brought about by global adjustments. Increasing the learning rate accelerates the correction process of the O&M foundation model, enabling it to adapt to changes in actual O&M tasks more quickly.

[0105] S4. Based on the trained multi-task integrated operation and maintenance basic model, obtain a large multi-task integrated operation and maintenance model for industrial equipment.

[0106] S5. Based on a multi-task integrated operation and maintenance model, it performs multi-task integrated operation and maintenance of industrial equipment, including anomaly detection, fault diagnosis, and life prediction.

[0107] Through the above steps, a multi-task integrated operation and maintenance model for industrial equipment is finally formed, which is used to process operation and maintenance tasks such as anomaly detection, fault diagnosis and life prediction. This model breaks through the feature learning bottleneck of traditional models and improves operation and maintenance efficiency and system reliability in complex scenarios.

[0108] In summary, this invention proposes a cloud-based multi-task integrated operation and maintenance method for industrial equipment. It employs an adaptive Transformer architecture and a Mu-Scaling dynamic expansion strategy to construct a basic multi-task integrated operation and maintenance model suitable for different operation and maintenance tasks. By designing a cross-modal fusion method between different modal data features, it achieves progressive fusion of multi-modal features. Using a supervised training strategy with frozen weight parameters, combined with cloud-based joint pre-training and feedback reinforcement learning, it eliminates the "illusion" of a large operation and maintenance model, forming a large-scale multi-task integrated operation and maintenance model with a super-large model structure and strong feature capture capabilities, encompassing "anomaly detection, fault diagnosis, and lifespan prediction." This significantly improves the operation and maintenance efficiency and reliability in complex scenarios.

[0109] Example 2

[0110] The present invention also provides a cloud-based industrial equipment multi-task integrated operation and maintenance large model construction system, which is used to implement the cloud-based industrial equipment multi-task integrated operation and maintenance large model construction method of the aforementioned embodiments. The system includes: a basic model module, a fusion mechanism module, a training module, a model acquisition module, and an operation and maintenance module.

[0111] The basic model module is used to acquire the operation and maintenance tasks of industrial equipment and build a multi-task integrated operation and maintenance basic model based on the operation and maintenance tasks.

[0112] The fusion mechanism module is used to acquire multimodal feature data of industrial equipment and, based on the multimodal feature data, preset a cross-modal feature fusion mechanism;

[0113] The training module is used to train the multi-task integrated operation and maintenance basic model based on a preset cross-modal feature fusion mechanism and a multi-modal cloud joint method.

[0114] The model acquisition module is used to obtain a large multi-task integrated operation and maintenance model for industrial equipment based on the trained multi-task integrated operation and maintenance basic model.

[0115] The operation and maintenance module is used for multi-task integrated operation and maintenance of industrial equipment based on a multi-task integrated operation and maintenance model, which includes anomaly detection, fault diagnosis and life prediction.

[0116] Furthermore, the basic model module includes: building blocks, dynamic expansion blocks, and knowledge inheritance blocks;

[0117] Building blocks are used to construct a multi-task integrated operation and maintenance basic model using the Transformer architecture;

[0118] The dynamic expansion unit is used to set a dynamic expansion strategy for the multi-task integrated operation and maintenance basic model based on the complexity and data volume of the operation and maintenance tasks of industrial equipment through a model self-growth strategy based on Mu-Scaling technology.

[0119] The knowledge inheritance unit is used to define the knowledge inheritance method of the multi-task integrated operation and maintenance basic model during the growth process by adopting parameter migration and feature distillation techniques.

[0120] The fusion mechanism module includes: a data acquisition unit and a mechanism preset unit;

[0121] The data acquisition unit is used to acquire multimodal feature data of industrial equipment and perform preprocessing.

[0122] The mechanism preset unit is used to preset a cross-modal feature fusion mechanism based on the preprocessed multimodal feature data: adaptively adjust the multimodal feature data; use a cross-modal attention mechanism to extract features from the multimodal feature data; and use a progressive fusion method to fuse the extracted features.

[0123] The training module includes: a supervised training unit and an online instruction fine-tuning unit;

[0124] Supervised training units are used to train the multi-task integrated operation and maintenance basic model using a supervised training strategy with frozen weight parameters.

[0125] Based on tensor and hybrid parallel architecture, dynamic allocation of cloud resources is achieved;

[0126] Based on the progressive parameter freezing mechanism, fine-tuning of the multi-task integrated operation and maintenance basic model is achieved to adapt to the domain.

[0127] The online instruction fine-tuning unit is used to train the multi-task integrated operation and maintenance basic model using feedback reinforcement learning-driven online instruction fine-tuning.

[0128] Based on the dynamic instruction generator, an adversarial sample generation algorithm is used to create training instructions, and the multi-task integrated operation and maintenance basic model is trained based on the training instructions.

[0129] Based on a two-stage strategy of offline reinforcement learning and online real-time feedback, fine-tuning of the multi-task integrated operation and maintenance basic model is achieved.

[0130] Example 3

[0131] like Figure 3 As shown, the present invention also provides an electronic device 500, including at least one processor 501, at least one network interface 504, a user interface 503, a memory 505, at least one communication bus 502, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method of the foregoing embodiments.

[0132] The communication bus 502 is used to enable communication between these components.

[0133] The processor 501 is used to execute the instructions stored in the memory 505 so that the electronic device 500 performs the above-mentioned industrial equipment status monitoring method.

[0134] The user interface 503 may include a display screen, and optionally, the user interface 503 may also include a standard wired interface or a wireless interface.

[0135] The network interface 504 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).

[0136] Example 4

[0137] The present invention also provides a computer-readable storage medium storing a computer program that, when executed, implements the methods of the foregoing embodiments.

[0138] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A cloud-based multi-task integrated operation and maintenance method for industrial equipment, characterized in that, The method includes: S1. Obtain the operation and maintenance tasks of industrial equipment, and construct a multi-task integrated operation and maintenance basic model based on the operation and maintenance tasks; S2. Obtain the multimodal feature data of the industrial equipment, and based on the multimodal feature data, preset a cross-modal feature fusion mechanism; S3. Based on the preset cross-modal feature fusion mechanism, the multi-modal cloud joint method is used to train the multi-task integrated operation and maintenance basic model; S4. Based on the trained multi-task integrated operation and maintenance basic model, obtain the multi-task integrated operation and maintenance large model of the industrial equipment. S5. Based on the aforementioned multi-task integrated operation and maintenance model, perform multi-task integrated operation and maintenance on the industrial equipment, including anomaly detection, fault diagnosis, and life prediction. S1 includes: S11. Use the Transformer architecture to build a multi-task integrated operation and maintenance basic model; S12. Using a model self-growth strategy based on Mu-Scaling technology, a dynamic expansion strategy for the multi-task integrated operation and maintenance basic model is set according to the complexity and data volume of the operation and maintenance tasks of the industrial equipment. S13. By adopting parameter migration and feature distillation techniques, the knowledge inheritance method of the multi-task integrated operation and maintenance basic model is set in the growth process; S2 includes: S21. Acquire the multimodal feature data of the industrial equipment and perform preprocessing; S22. Based on the preprocessed multimodal feature data, a cross-modal feature fusion mechanism is preset: adaptively adjusting the multimodal feature data; using a cross-modal attention mechanism to extract features from the multimodal feature data; and adopting a progressive fusion method to fuse the extracted features: the progressive framework starts with the fusion of simple partial modalities in stages to help the multi-task integrated operation and maintenance basic model learn simple relationships, and then gradually adds other modal information based on it to gradually learn more complex feature associations, and finally achieves deep fusion of multimodal information. The simple partial modalities include images and text. S3 includes: S31. The multi-task integrated operation and maintenance basic model is trained using a supervised training strategy with frozen weight parameters: Based on tensor and hybrid parallel architecture, dynamic allocation of cloud resources is achieved; Based on the progressive parameter freezing mechanism, fine-tuning of the multi-task integrated operation and maintenance basic model is achieved to adapt it to the domain. S32. The multi-task integrated operation and maintenance basic model is trained using online instruction fine-tuning driven by feedback reinforcement learning: Based on the dynamic instruction generator, an adversarial sample generation algorithm is used to create training instructions, and the multi-task integrated operation and maintenance basic model is trained based on the training instructions. Based on a two-stage strategy of offline reinforcement learning and online real-time feedback, the multi-task integrated operation and maintenance basic model is fine-tuned. The PPO algorithm is optimized by using a proximal strategy in the offline reinforcement learning stage.

2. A cloud-based multi-task integrated operation and maintenance system for industrial equipment, used to implement the method of claim 1, characterized in that, The system includes: a basic model module, a fusion mechanism module, a training module, a model acquisition module, and an operation and maintenance module; The basic model module is used to acquire the operation and maintenance tasks of industrial equipment and, based on the operation and maintenance tasks, construct a multi-task integrated operation and maintenance basic model. The fusion mechanism module is used to acquire multimodal feature data of the industrial equipment and, based on the multimodal feature data, preset a cross-modal feature fusion mechanism. The training module is used to train the multi-task integrated operation and maintenance basic model based on a preset cross-modal feature fusion mechanism and a multi-modal cloud joint method. The model acquisition module is used to obtain a large multi-task integrated operation and maintenance model of the industrial equipment based on the trained multi-task integrated operation and maintenance basic model. The operation and maintenance module is used to perform multi-task integrated operation and maintenance of the industrial equipment based on the multi-task integrated operation and maintenance model, including anomaly detection, fault diagnosis, and life prediction.

3. The system according to claim 2, characterized in that, The basic model module includes: a construction unit, a dynamic expansion unit, and a knowledge inheritance unit; The building unit is used to build a multi-task integrated operation and maintenance basic model using the Transformer architecture; The dynamic expansion unit is used to set the dynamic expansion strategy of the multi-task integrated operation and maintenance basic model according to the complexity and data volume of the operation and maintenance tasks of the industrial equipment through a model self-growth strategy based on Mu-Scaling technology. The knowledge inheritance unit is used to set the knowledge inheritance method of the multi-task integrated operation and maintenance basic model during the growth process by adopting parameter transfer and feature distillation techniques.

4. The system according to claim 2, characterized in that, The fusion mechanism module includes: a data acquisition unit and a mechanism preset unit; The data acquisition unit is used to acquire the multimodal feature data of the industrial equipment and perform preprocessing. The mechanism preset unit is used to preset a cross-modal feature fusion mechanism based on the preprocessed multimodal feature data: adaptively adjust the multimodal feature data; extract features from the multimodal feature data using a cross-modal attention mechanism; and fuse the extracted features using a progressive fusion method.

5. The system according to claim 2, characterized in that, The training module includes: a supervised training unit and an online instruction fine-tuning unit; The supervised training unit is used to train the multi-task integrated operation and maintenance basic model using a supervised training strategy with frozen weight parameters. Based on tensor and hybrid parallel architecture, dynamic allocation of cloud resources is achieved; Based on the progressive parameter freezing mechanism, fine-tuning of the multi-task integrated operation and maintenance basic model is achieved to adapt it to the domain. The online instruction fine-tuning unit is used to train the multi-task integrated operation and maintenance basic model using feedback reinforcement learning-driven online instruction fine-tuning. Based on the dynamic instruction generator, an adversarial sample generation algorithm is used to create training instructions, and the multi-task integrated operation and maintenance basic model is trained based on the training instructions. Based on a two-stage strategy of offline reinforcement learning and online real-time feedback, fine-tuning of the multi-task integrated operation and maintenance basic model is achieved.

6. An electronic device, characterized in that, It includes at least one processor, at least one network interface, a user interface, a memory, at least one communication bus, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the method as described in claim 1.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed, implements the method as described in claim 1.

Images