Method and system for predicting equipment health state based on large model AGENT capability
Through the large-scale AGENT fusion of multi-source data and self-supervised learning, the adaptability and accuracy of equipment health status prediction is solved, the accurate assessment and risk quantification of equipment health status is achieved, maintenance decisions are optimized, and production safety and efficiency are improved.
Patent Information
- Application Number
- CN202510768738.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-10
- Publication Date
- 2025-07-08
- Estimated Expiration
- 2045-06-10
AI Technical Summary
The prior art lacks adaptability and accuracy in predicting equipment health status, making it difficult to adapt to heterogeneous devices, has limited self-learning ability, and cannot effectively quantify the risk of health deterioration, resulting in insufficient prediction accuracy and untimely failure prediction.
Using a method based on big model AGENT, virtual data is generated through data acquisition, preprocessing and digital twin simulation, combined with self-supervised learning and adaptive update mechanisms, a pre-trained large-scale deep neural network is used to fuse multi-source data to calculate health index and risk indicators, and to achieve accurate assessment of device health status and risk quantification.
It realizes accurate health status assessment and risk quantification of different types of equipment, can warning of potential failures in advance, optimize maintenance decisions, reduce unexpected downtime and maintenance costs, and improve production continuity and reliability.
Smart Images

Figure CN120277371A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of industrial equipment health monitoring and predictive maintenance, and specifically relates to a method and system for predicting the health status of equipment based on the capabilities of a large model AGENT. Background Art
[0002] With the improvement of the intelligence level of industrial equipment, monitoring the operation status of key equipment and predicting faults have become important means to ensure production safety and efficiency. Traditional methods usually rely on expert experience and fixed algorithms to analyze equipment data, making it difficult to process a large amount of multi-source data of different types of equipment, and the models lack adaptability.
[0003] For example, Chinese Patent CN119668195A discloses a multi-region monitoring system and method based on an AI vision model, which monitors the multi-region status of equipment by collecting images through a camera. However, this existing technology mainly relies on visual data, and has limited access to the internal performance parameters of the equipment and in-depth health status assessment. In addition, due to the use of a specific AI model, it is difficult to adapt to heterogeneous equipment, lacks the ability of self-evolution learning, and cannot dynamically adjust model parameters according to equipment operation data. Therefore, in practical applications, there are still problems such as insufficient prediction accuracy, poor generalization ability, and inability to quantify risks in advance, making it difficult to detect potential equipment faults in a timely manner and take preventive measures.
[0004] In summary, the existing technology has deficiencies in adaptability and accuracy in predicting the health status of equipment. It is necessary to propose a new technical solution to integrate multi-source data, utilize the capabilities of a more advanced large model Agent, and achieve accurate prediction and risk quantification of the health status of different types of equipment. Summary of the Invention
[0005] Technical Objectives: Aiming at the problems in the prior art such as poor generalization ability of equipment health status prediction means, difficulty in adapting to heterogeneous equipment, insufficient self-learning ability, and lack of quantitative assessment of health degradation risks, the present invention discloses a method and system for predicting the health status of equipment based on the capabilities of a large model AGENT. By using a pre-trained large model intelligent Agent, it effectively integrates multi-source data of the equipment and digital twin simulation, realizes accurate assessment of the operating health status of different types of equipment, and gives risk quantification indicators brought by dynamic changes in the health status, so as to early warn of potential faults and optimize maintenance decisions.
[0006] Technical Solutions: To achieve the above technical objectives, the present invention adopts the following technical solutions: A method for predicting the health status of equipment based on the capabilities of a large model AGENT specifically includes the following steps: Obtain the operation data of the equipment to be monitored through a data acquisition device; Preprocess the operation data to generate a standard time series data vector; Use the digital twin simulation model to generate virtual operation data to expand the abnormal condition samples, and fuse the virtual data with the actual data to form comprehensive input data; Send the comprehensive input data into the pre-trained large model AGENT for analysis to obtain the health index H(t) representing the health state of the device; Calculate the risk index R(t) of the device according to the health index H(t) and its change rate over time; Output the prediction result of the device health state based on the health index H(t) and the risk index R(t).
[0007] Preferably, the risk index R(t) is calculated using the following formula: , where H(t) is the health index of the device at time t, is the change rate of the health index with respect to time, and are weight coefficients that adjust the proportion of the impact of the health index and the change rate on the risk, satisfying , is the sensitivity parameter of the health index impact factor, is the sensitivity parameter of the change rate of the health index with respect to time impact factor.
[0008] Preferably, after outputting the prediction result, the internal parameter weights of the large model AGENT are adaptively updated to improve the subsequent prediction accuracy. The weight update is based on the self-supervised learning rule. When the prediction errors of multiple features deviate from the average level, the corresponding weights are adjusted according to the following formula : , where, is the average value of all feature errors, is the standard deviation of the feature errors, is the learning rate, is a preset minimum constant.
[0009] Preferably, the large model AGENT is composed of a deep neural network, which obtains the general feature extraction ability through pre-training with a large amount of multi-source device data, and is continuously fine-tuned using the device's own operation data through the adaptive update step, so as to adapt to the health state prediction of different devices and working conditions.
[0010] Preferably, it further includes generating virtual operation data of the device by using a digital twin simulation model, and inputting the virtual data and the actually collected data into the large model AGENT for analysis and training to make up for the shortage of real data under abnormal working conditions and improve the model's recognition ability of potential fault modes of the device.
[0011] Preferably, the device operation data includes one or more sensor data of vibration, temperature, pressure, current, and voltage, and the health index H(t) is a dimensionless value from 0 to 1, where 1 indicates that the device is completely healthy and 0 indicates that the device is in a failure state.
[0012] A system for predicting the health state of a device based on the capabilities of a large model AGENT, comprising: A data acquisition module for acquiring the operation data of the device; A data processing module for synchronizing and preprocessing the acquired operation data; A large model AGENT analysis module, connected to the data acquisition module, for receiving the operation data and calculating the device health index H(t); A risk assessment module, connected to the AGENT analysis module, for calculating the risk index R(t) according to the health index H(t) and its change rate; An output module for outputting the prediction results of the health index and the risk index; The large model AGENT analysis module includes a pre-trained deep learning model and an adaptive update unit, configured to execute a method for predicting the health state of a device based on the capabilities of a large model AGENT as described above.
[0013] Preferably, it further includes a digital twin simulation module, which interacts with the large model AGENT analysis module, for generating simulation data of the device under different working conditions and providing it to the large model AGENT analysis module to enrich the model training data source and improve the prediction ability of abnormal states.
[0014] Preferably, the large model AGENT analysis module includes: A feature extraction unit for performing multi-modal feature extraction and fusion on the operation data; A health assessment unit, connected to the feature extraction unit, for calculating the device health index H(t) according to the extracted features; A model training unit for online updating the model parameters of AGENT by using a self-supervised learning algorithm.
[0015] Beneficial effects: The method and system for predicting the health state of a device based on the capabilities of a large model AGENT provided by the present invention have the following beneficial effects: 1. The large model AGENT of the present invention is trained based on a large amount of multi-domain data, has a wide range of knowledge representation capabilities, and is continuously fine-tuned through an adaptive weight update mechanism. It can be applied to devices of different types and working conditions, making up for the deficiencies of traditional models with strong pertinence and poor generality. Whether it is a numerical control machine tool, an engine, or a power equipment, the present invention can effectively predict their health status. At the same time, a self-supervised training strategy is introduced to automatically adjust the model parameters, enabling AGENT to learn from unlabeled data. The model continuously evolves during actual operation, becoming more "intelligent" over time, effectively coping with the slow drift and sudden changes of equipment status, and achieving long-term accurate prediction.
[0016] 2. The present invention generates virtual fault and working condition data through digital twin technology, enriching the training sample space. AGENT can learn the behavior of equipment in extreme situations in a safe and controllable virtual environment, improving the prediction ability for rare faults. This method of simulation integration overcomes the limitation of insufficient actual fault data and enhances the robustness of the model.
[0017] 3. The present invention simultaneously outputs an easily understandable health index H(t) and a quantified risk index R(t). The health index intuitively reflects the current state of the equipment, while the risk index quantifies in advance the possibility and urgency of a fault occurring within a certain period of time in the future. By integrating the state and trend through the risk index, it can provide richer information than a single health score, assisting maintenance personnel in making more scientific decisions on the timing of maintenance and avoiding premature or late maintenance.
[0018] By comprehensively applying the above technical solutions, the present invention can maximize the normal operation time of equipment on the premise of ensuring equipment safety. Compared with traditional empirical judgment or periodic maintenance strategies, using the prediction results of the present invention can transform maintenance from passive to active, intervening before a fault occurs. This will reduce the number of unexpected shutdowns, lower maintenance costs, and improve the continuity and reliability of industrial production. BRIEF DESCRIPTION OF THE DRAWINGS
[0019] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art.
[0020] Figure 1 It is a flowchart of the method for predicting the health status of equipment based on the capabilities of the large model AGENT of the present invention; Figure 2 It is an overall architecture diagram of the system for predicting the health status of equipment based on the capabilities of the large model AGENT of the present invention; Figure 3 It is a relationship diagram of the health index and the risk index of the present invention. DETAILED DESCRIPTION OF THE INVENTION
[0021] The present invention will be more clearly and completely described below by way of a preferred embodiment in conjunction with the accompanying drawings, but the present invention is not limited to the scope of the described embodiments thereby.
[0022] As Figure 1 shown, a method for predicting the health status of a device based on the capabilities of a large model AGENT specifically includes the following steps: S1. Obtain the operation data of the device to be monitored through a data acquisition device; Collect multimodal data during the operation of the target device, including parameters such as vibration, temperature, pressure, current, etc. measured by sensors, as well as optional visual images, control logs, etc.
[0023] S2. Preprocess the operation data to generate a standard time series data vector; Perform preprocessing such as cleaning, filtering, and normalization on the collected data to form a unified time series data input format.
[0024] S3. Use a digital twin simulation model to generate virtual operation data to expand the abnormal condition samples, and fuse the virtual data with the actual data to form comprehensive input data; Build a digital twin model for the target device to simulate the behavior of the device under different working conditions and fault modes. Use the simulation data generated by the digital twin to enhance the training set of the large model AGENT to make up for the lack of actual fault data, thereby improving the model's recognition ability for abnormal conditions. The simulation data can be periodically input into the large model AGENT for self-supervised training, and the AGENT parameters are continuously updated to approach the true state of the device.
[0025] S4. Send the comprehensive input data into a pre-trained large model AGENT for analysis to obtain a health index H(t) representing the health status of the device; Input the preprocessed device data into a pre-trained large model AGENT. This AGENT is composed of a large-scale deep neural network and has general feature extraction and reasoning capabilities obtained through training with a large amount of data, and adapts to specific devices through self-supervised learning. The AGENT performs multi-level feature analysis on the input data, extracts key health feature indicators of the device operation state, and comprehensively outputs the current health status index H(t) of the device. The health index H(t) is a dimensionless value in the range of 0 to 1, where 1 represents that the device is completely healthy and 0 represents that the device fails or is in the worst health state.
[0026] S5. Calculate the risk index R(t) of the device according to the health index H(t) and its change rate over time; Based on the health index H(t) and its change rate, calculate the risk index R(t) of the device health status. The present invention proposes the following dynamic risk quantification formula: , In the formula, and are weight coefficients, respectively adjusting the contribution weights of the current health level and the health change rate of the device to the risk; and are positive sensitivity coefficients, controlling the growth rate of the risk as the health deteriorates and the health change intensifies; H(t) is the health index at time t; is the rate of change of the health index with respect to time. This risk index R(t) increases as the degree of device health deterioration (1 - H(t)) and the health decline rate increase, and is used to quantify the relative risk of future failures. Through the above formula, the present invention can reflect the current state and dynamic deterioration trend of the device in the risk value, realizing the quantitative assessment of potential failure risks.
[0027] S6. Output the prediction result of the device health state based on the health index H(t) and the risk index R(t).
[0028] Output the health index H(t) and the risk index R(t) as the prediction result of the device health state. If the risk index R(t) exceeds the preset threshold, an alarm or a maintenance plan suggestion can be triggered in advance, and the result is provided to the operation and maintenance personnel for decision-making support.
[0029] During the continuous operation of the device, the present invention uses a self-supervised learning mechanism to dynamically update the parameter weights of the large model AGENT. The following weight adaptive update formula is proposed: , In the formula, represents the weight value of the i-th parameter inside the Agent at time t; represents the prediction error or evaluation index calculated based on the i-th data feature of the device at time t; is the average value of all feature errors at this moment; is the standard deviation of the error at this moment; is the learning rate; is a preset minimum constant to avoid division by zero. Through the above update rule, when the error of a certain feature is higher than the average level, its corresponding weight will be positively adjusted (increased), otherwise it will be decreased, so that the large model AGENT pays more attention to abnormal feature signals and corrects its own model. This adaptive adjustment mechanism does not require manual annotation and belongs to the category of self-supervised training. It can make the model continuously optimize with the accumulation of data, enhance the adaptability to the device degradation mode, and achieve the adaptive generalization of heterogeneous devices.
[0030] In a specific embodiment, the method of the present invention specifically operates according to the following process: Initialization: After the system is powered on or started, the pre-trained model parameters of the large model AGENT are loaded, the basic device information and historical data records are read, parameters such as the risk threshold are set, and after confirming that each module is running normally, it enters the monitoring loop.
[0031] Data acquisition and preprocessing: The data acquisition module acquires the data of each sensor of the device at a fixed time interval (such as every second or every minute), and the data processing module synchronizes and filters the acquired multi-source data, etc., and packs the results into the feature vector at the current moment t , which contains the normalized values such as temperature T(t), vibration a(t), pressure P(t), etc.
[0032] AGNET health assessment: The large model AGENT analysis module receives the current feature vector , and uses its internal deep model to calculate and output the current health index H(t). For example, AGNET may calculate that H(t)=0.85, indicating that the device is at 85% health level.
[0033] Risk assessment: The risk assessment module reads the current H(t) and the previous moment H(t - △t), and calculates the health change rate . Substituting into the risk index calculation formula, the value of R(t) can be obtained. For example, if H(t)=0.85 and it has been slowly decreasing in the past period of time, then R(t) may be about 0.2, belonging to the low risk level. The assessment module accordingly obtains the risk level "low", does not trigger an alarm, and only records the status.
[0034] Result output: The current H(t), R(t) and the risk level are displayed through the output module. If R(t) exceeds the threshold, for example, it is calculated that R(t)=0.78 exceeds the warning threshold of 0.7, the output module issues a "medium risk" alarm and recommends arranging maintenance. The operation and maintenance personnel can accordingly plan maintenance in advance to avoid failures.
[0035] Model update: The model update module summarizes the prediction errors of the current cycle. For the unsupervised case, the reconstruction error or prediction residual of the AGNET model for the acquired data can be used as . Suppose that the change of some sensor data patterns in this cycle leads to an increase in the prediction residual, then the of the corresponding feature is higher than the average , and according to the weight adaptive update formula, the weights of these features are increased . Conversely, for features with errors lower than the average, their weights are appropriately reduced. The updated model parameters will be used for the prediction of the next cycle. Through continuous iteration, the AGENT model is gradually optimized for this specific device.
[0036] Loop execution: The system returns to the data acquisition step and enters the next monitoring cycle. This process repeats continuously to achieve continuous monitoring and prediction during the equipment operation. When the equipment stops running or the manual monitoring is stopped, the process ends.
[0037] Formula parameter selection and adjustment: In specific implementations, the parameters of the risk index calculation formula and the weight adaptive update formula can be set and adjusted according to the equipment type and historical data. For example, for key equipment with serious failure consequences, larger values can be selected to increase the sensitivity to sudden changes in health; for slow-changing wear faults, the weight can be increased to enhance the contribution of the declining health level to the risk. The parameters and can be optimized through simulation tests to ensure that the response of the risk R(t) to the health index and its change rate conforms to engineering experience. For example, the value corresponding to the sharp increase in risk when the health index is below 0.4 can be determined by fitting historical fault data. The learning rate generally takes a small value (such as 0.01) to ensure the stable convergence of weight updates; usually takes a magnitude of 10-6 to prevent division by zero. Through the reasonable configuration of the above parameters, the method of the present invention can be customized for different application scenarios to achieve the best prediction performance.
[0038] As Figure 2 shown, a system for predicting the health status of equipment based on the capabilities of the large model AGENT includes: A data acquisition module for obtaining the operation data of the equipment; A data processing module for synchronizing and preprocessing the acquired operation data; A large model AGENT analysis module connected to the data acquisition module for receiving the operation data and calculating the equipment health index H(t); The large model AGENT analysis module includes: A feature extraction unit for performing multi-modal feature extraction and fusion on the operation data; A health assessment unit connected to the feature extraction unit for calculating the equipment health index H(t) based on the extracted features; A model training unit for online updating the model parameters of the AGENT using self-supervised learning algorithms; A risk assessment module connected to the AGENT analysis module for calculating the risk index R(t) based on the health index H(t) and its change rate; An output module for outputting the prediction results of the health index and the risk index; The large model AGENT analysis module includes a pre-trained deep learning model and an adaptive update unit, configured to execute a method for predicting the health status of a device based on the capabilities of the large model AGENT as described above.
[0039] The system of the present invention further includes a digital twin simulation module, which interacts with the large model AGENT analysis module to generate simulation data of the device under different working conditions and provide it to the large model AGENT analysis module, so as to enrich the data source for model training and improve the prediction ability of abnormal states.
[0040] In a specific embodiment, the system of the present invention includes the following components: Data acquisition module: used to obtain operation data from the device to be monitored. The data includes multiple sensor signals, such as vibration acceleration, temperature, pressure, current, voltage, and device operation control parameters, etc. For scenarios with image monitoring requirements, the data acquisition module also includes an industrial camera or a vision sensor to collect image / video information of the device appearance or working conditions. All kinds of raw data are sent to the data processing module after analog-to-digital conversion and preliminary caching.
[0041] Data processing module: synchronizes and preprocesses the acquired multi-source data, including operations such as denoising filtering, outlier removal, time alignment, and normalization, to generate time series data in a unified format. The processed data is used as a feature sequence for subsequent analysis. Through this module, the quality and comparability of different sensor data are ensured, and interference factors are reduced.
[0042] Large model AGNET analysis module: This is the core intelligent analysis unit of the present invention, including a pre-trained large-scale deep learning model and an inference engine. Preferably, the AGNET is based on advanced architectures such as Transformer and Graph Neural Network, and integrates device physical prior knowledge and big data training experience. The feature sequence output by the data processing module is fed into the large model AGNET as input, and after multiple layers of non-linear transformation and feature extraction, the health index H(t) of the device is calculated. Specifically, a representation vector of the device health status is established inside the AGENT, and the attention mechanism is used to fuse the features of each sensor, and the health index H(t) in the range of 0 to 1 is output. The higher the value, the healthier the device. To enhance generalization, the large model AGENT analysis module has been pre-trained with public large industrial datasets and historical operation data at the initial deployment, and can continuously update parameters on-site using the self-supervised mechanism of the present invention to adapt to the specific feature patterns of specific devices.
[0043] Digital Twin Simulation Module: This module is used to build a digital twin model of the monitored device and generate simulation data. This module can be autonomously selected according to the actual usage. The digital twin model can adopt multi-body dynamics simulation, finite element analysis, or a simulation model trained based on historical data to realistically reproduce the behavior of the device under various working conditions. When the device operates within the safe range, the digital twin simulation module can simulate the reaction of the device under abnormal working conditions (such as overloading, component wear and failure, etc.), generating corresponding virtual sensor data and the evolution trajectory of the health state. These simulation data are sent to the large model AGENT analysis module to expand the training set to cover extreme or rare situations, helping the AGENT learn to identify and predict health changes in such situations. Through the combination of real and virtual data, the present invention significantly improves the prediction reliability of the model when facing unknown situations.
[0044] Risk Assessment Module: It is used to calculate the risk index R(t) of the device according to the health index H(t) and its change trend output by the large model AGNET analysis module, such as Figure 3 shown in the figure shows the relationship between the device health index H(t) and the risk R(t). As shown in the formula of the risk index mentioned above, the risk assessment module first calculates the degree of device health deterioration (1 - H(t)) based on the current health index H(t), and combines the difference of the health index at adjacent moments to obtain the health change rate. Subsequently, the preset weight coefficients , and the sensitivity parameters , are substituted into the formula to solve the risk index R(t) in real time. The value range of the risk index R(t) is from 0 to approximately 2. The larger the value, the closer the device is to failure or the higher the probability of failure. The risk assessment module can generate a warning signal according to the comparison result of the calculated R(t) with the threshold. For example, when R(t) exceeds a certain threshold Rth, it is determined that the device enters a high-risk state and needs maintenance.
[0045] Model Update Module: This module implements an adaptive model parameter update mechanism. Specifically, when implemented, the system will, after each prediction cycle ends, compare the health index and risk index obtained from the risk assessment module with the actual device state (such as whether a failure occurs during subsequent operation), and calculate the prediction error of each feature . The error can be obtained by comparing the actual health index of the device at subsequent times (if it can be obtained by manual evaluation or shutdown for maintenance) with the predicted value, or by using unsupervised metrics (such as reconstruction error, residual, etc.) instead. The model update module calculates the average and standard deviation of all feature errors, and then adjusts the weights of the corresponding features inside the large model AGENT analysis module This adjustment is reflected in the software by modifying the weight parameters of some connections of the neural network, so that when predicting the next cycle, AGNET pays more attention to or makes appropriate corrections to features with large errors. This self-learning process does not require human participation. When the operating environment or state statistical characteristics of the equipment change slowly, the model can automatically adapt to maintain stable prediction performance.
[0046] Output module: used to output the equipment health status prediction results, including health index H(t), risk index R(t), risk level and recommended operations. The output can be in the form of real-time display of numerical values and trend graphs on the human-machine interface dashboard, or sending data to a host computer or cloud monitoring system. If the risk index exceeds the threshold or meets the warning conditions, the alarm module triggers an audible and visual alarm or sends an alarm notification to the operation and maintenance personnel, and can also record event logs for future analysis. This module can also archive the prediction data as needed to provide a basis for further offline analysis and model improvement.
[0047] Taking the CNC machine tool spindle system as an example, after the system of the present invention is deployed, the sensor continuously collects signals such as spindle vibration acceleration, temperature and current. The large model AGENT combines the historical data of the machine tool and the wear and failure data simulated by the digital twin for analysis, and outputs the spindle health index H(t). In the early stage of normal operation, H(t) is maintained above 0.9, which is highly healthy. As the usage time accumulates, H(t) slowly decreases to about 0.6, and the vibration signal increases abnormally, resulting in The absolute value of increases. According to the calculation formula of the risk index, the risk index R(t) increases from the initial 0.1 to 0.5, approaching the warning threshold. The system alarm reminds that the spindle bearing needs to be repaired. After the maintenance personnel replaced the bearings, the vibration level dropped, H(t) recovered to 0.85, R(t) dropped below 0.2, and the system cleared the alarm. This process shows that the present invention can effectively track changes in equipment health and give a quantitative risk assessment, providing a scientific basis for maintenance decisions. Compared with manual experience, the method of the present invention is more objective and accurate, and can timely avoid the occurrence of major spindle failures.
[0048] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.
Claims
1. A method for predicting the health status of a device based on the capabilities of a large model AGENT, characterized in that, Specifically, it includes the following steps: Obtain the operation data of the device to be monitored through a data acquisition device; Preprocess the operation data to generate a standard time series data vector; Use the digital twin simulation model to generate virtual operation data to expand the abnormal condition samples, and fuse the virtual data with the actual data to form comprehensive input data; Send the comprehensive input data into the pre-trained large model AGENT for analysis to obtain the health index H(t) representing the health state of the device; Calculate the risk index R(t) of the device according to the health index H(t) and its change rate over time; Output the prediction result of the health state of the device based on the health index H(t) and the risk index R(t).
2. The method for predicting the health status of a device based on the capabilities of a large model AGENT according to claim 1, wherein The calculation of the risk index R(t) adopts the following formula: , Among them, \(H(t)\) is the health index of the device at time \(t\), is the change rate of the health index with respect to time, and are weight coefficients that adjust the proportion of the impact of the health index and the change rate on the risk, satisfying , is the sensitivity parameter of the health index impact factor, is the sensitivity parameter of the impact factor of the change rate of the health index with respect to time.
3. A method for predicting the health state of a device based on the capabilities of a large model AGENT according to claim 1, characterized in that After outputting the prediction results, the internal parameter weights of the large model AGENT are adaptively updated to improve subsequent prediction accuracy. The weight update is based on self-supervised learning rules. When the prediction errors of multiple features are detected deviate from the average level, the corresponding weights are adjusted according to the following formula : , Among them, is the average value of all feature errors, is the standard deviation of feature errors, is the learning rate, is a preset minimum constant.
4. A method for predicting the health state of a device based on the capabilities of a large model AGENT according to claim 1, characterized in that The large model AGENT is composed of a deep neural network, which obtains the general feature extraction ability through pre-training with a large amount of multi-source device data, and continuously fine-tunes and trains by using the device's own operation data through an adaptive update step, so as to adapt to the health state prediction of different devices and working conditions.
5. A method for predicting the health status of a device based on the capabilities of a large model AGENT according to claim 1, characterized in that, It also includes using the digital twin simulation model to generate the virtual operation data of the device, and inputting the virtual data and the actually collected data into the large model AGENT for analysis and training to make up for the deficiency of real data under abnormal conditions and improve the model's ability to identify potential failure modes of the device.
6. The method for predicting the health state of a device based on the capabilities of a large model AGENT according to claim 1, wherein The device operation data includes one or more sensor data of vibration, temperature, pressure, current, and voltage, and the health index H(t) is a dimensionless value from 0 to 1, where 1 represents that the device is completely healthy and 0 represents that the device is in a failure state.
7. A system for predicting the health status of a device based on the capabilities of a large model AGENT, characterized in that, It includes: A data acquisition module for obtaining the operation data of the device; A data processing module for synchronizing and preprocessing the collected operation data; A large model AGENT analysis module connected to the data acquisition module for receiving the operation data and calculating the device health index H(t); A risk assessment module connected to the AGENT analysis module for calculating the risk index R(t) according to the health index H(t) and its change rate; An output module for outputting the prediction results of the health index and the risk index; The large model AGENT analysis module includes a pre-trained deep learning model and an adaptive update unit, configured to execute a method for predicting the health state of a device based on the ability of the large model AGENT according to any one of claims 1 to 6.
8. A system for predicting the health status of a device based on the capabilities of a large model AGENT according to claim 7, characterized in that, It also includes a digital twin simulation module, which interacts with the large model AGENT analysis module for generating simulation data of the device under different working conditions and providing it to the large model AGENT analysis device to enrich the model training data source and improve the prediction ability of abnormal states.
9. The system for predicting the health state of a device based on the capabilities of a large model AGENT according to claim 7, wherein, The large model AGENT analysis module includes: A feature extraction unit for performing multi-modal feature extraction and fusion on the operation data; A health assessment unit connected to the feature extraction unit for calculating the device health index H(t) according to the extracted features; A model training unit for online updating the model parameters of AGENT by using a self-supervised learning algorithm.
Citation Information
Patent Citations
Power plant equipment fault diagnosis method and system based on artificial intelligence and automation
CN119004303A
Deep learning-based multi-energy complementary power distribution system data analysis method and system
CN119416101A
Urban community embedded service facility operation management method and system based on deep learning
CN119417230A
Method and system for predicting equipment health state based on large model AGENT capability
CN119668195A
Integrated power generation equipment health state supervision system
CN120067837A
Cited By
Electromechanical system health monitoring and optimizing method based on artificial intelligence
CN120597642A
Informatization intelligent diagnosis and fault early warning method for operation state of port environmental protection equipment
CN121094785A
Heating ventilation and air conditioning system operation risk assessment method and system based on neural network time sequence model
CN121169119A
Intelligent self-healing method and device for dual-redundancy industrial personal computer, terminal and storage medium
CN121277738A
Method for diagnosing health state of wastewater collection pipeline of electrochemical energy storage prefabricated cabin
CN121412737A