Automatic high-voltage aging method and system based on local large model

Abstract

The invention relates to an automatic high-voltage aging method and system based on a local large model, and relates to the technical field of high-voltage aging control. The method comprises the following steps: deploying a large model on a local server; in the high-voltage aging process, multi-source sensor data of a high-voltage aging system are synchronously collected, multiple aging record data are generated, and each aging record data comprises a timestamp, multi-source sensor data, an ignition mark and a corresponding next-step voltage set value; performing fine tuning and training on the large model in combination with locally collected historical aging record data; docking the large model with PLC equipment control logic; and predicting a next voltage set value through the large model, and executing voltage adjustment by the PLC equipment. The method has the effect of meeting the high requirements of high-voltage aging on real-time performance and safety as much as possible.

Description

Technical Field

[0001] This application relates to the technical field of high-voltage aging control, and in particular to an automated high-voltage aging method and system based on a local large model. Background Technology

[0002] High-voltage aging is a crucial step in the manufacturing and maintenance of high-voltage equipment. Its purpose is to gradually increase voltage and monitor performance to detect potential insulation defects early, thereby improving equipment reliability. Traditional high-voltage aging processes rely on manual experience; operators must manually adjust the pressurization strategy based on real-time data such as voltage, current, and vacuum level. This process is not only time-consuming and labor-intensive but also highly subjective, prone to misjudgments due to lack of experience or fatigue, thus affecting equipment safety and aging efficiency.

[0003] With the development of artificial intelligence and automation technologies, using large deep learning models to assist decision-making has become a trend. However, traditional large deep learning models usually rely on cloud training and inference, which has problems such as data privacy, response latency, and insufficient generalization ability of large models, making it difficult to meet the high requirements of real-time performance and security in high-pressure training. Summary of the Invention

[0004] In order to meet the high requirements of real-time performance and safety in high-voltage aging as much as possible, this application provides an automated high-voltage aging method and system based on a local large model.

[0005] On the one hand, the automated high-pressure aging method based on a local large model provided in this application adopts the following technical solution: An automated high-pressure aging method based on a local large model includes the following steps: Deploy large models on local servers; During the high-voltage aging process, multi-source sensor data is collected synchronously and multiple aging record data are generated. Each aging record data includes a timestamp, multi-source sensor data, ignition mark and corresponding next voltage setting value. The large model is fine-tuned and trained by combining locally collected historical data. Connect the large model and the PLC device control logic; The PLC device performs voltage adjustment based on the prediction of the next voltage setpoint using the large model.

[0006] By adopting the above technical solutions, large models are deployed locally to avoid data transmission delays and leakage risks in the cloud, thus meeting the real-time performance and data security requirements of high-voltage aging. Multi-dimensional data is collected simultaneously to provide the model with comprehensive operating condition characteristics, improving prediction accuracy. The prediction results of the large model are directly connected to the control logic to achieve automatic control from data acquisition to voltage adjustment. Preferably, the process also includes manually screening high-quality aging records; the high-quality aging records are data that restores the voltage to the set value after eliminating sparking.

[0007] By adopting the above technical solutions, noise and erroneous data can be eliminated as much as possible, allowing the model to focus on learning effective control strategies and avoiding overfitting.

[0008] Preferably, the large model is fine-tuned and trained, including the following steps; Data set preparation involves dividing the historical mature record data into a training set and a validation set. Set the parameters of the large model; Input the training set and validation set into the large model and begin training; After training is complete, the fine-tuned large model is saved to the local server.

[0009] By adopting the above technical solution, fine-tuning and training of the large model can be achieved, so as to predict the next voltage setting value.

[0010] Preferably, setting the parameters of the large model includes setting the output directory, log recording directory, large model storage strategy, learning rate, weight decay, batch size, number of training rounds, and evaluation strategy.

[0011] By adopting the above technical solutions, setting up a large model output directory facilitates the management of output results, setting up a log recording directory allows for the recording of training process information, setting up a large model saving strategy ensures reasonable model preservation, setting the learning rate controls the model training speed and accuracy, setting weight decay prevents overfitting, setting the batch size optimizes training efficiency, setting the number of training rounds controls the degree of training, and setting an evaluation strategy allows for the evaluation of model performance. This improves the rationality and effectiveness of large model training and enhances the accuracy and reliability of automated high-pressure aging methods based on local large models.

[0012] Preferably, the historical mature records are processed by word segmentation encoding, attention mask generation, and maximum length limitation.

[0013] By adopting the above technical solution, unstructured sensor data is transformed into time-series tensors that can be processed by the model, thus solving the problem of inconsistent lengths of multi-source data.

[0014] Preferably, the control logic interface includes the following: The next voltage setting value predicted by the large model is written into the PLC register via the Modbus TCP protocol; Limit voltage variation, holding time, and current threshold in PLC equipment.

[0015] By adopting the above technical solution, communication with the PLC via the Modbus TCP protocol can be made compatible with existing industrial equipment without the need for hardware modification; the voltage change amplitude and holding time are rigidly limited in the PLC equipment to form a dual protection mechanism, which can prevent equipment damage even if the model predicts abnormalities.

[0016] Preferably, the method also includes periodically collecting new sensor data, which is used as a training set for incremental training.

[0017] By adopting the above technical solution, periodic incremental training enables the model to dynamically adapt to new working conditions such as equipment aging and environmental changes, and incremental training is only performed on new data, without the need for retraining.

[0018] On the other hand, the automated high-pressure aging system based on a local large model provided in this application adopts the following technical solution: An automated high-pressure aging system based on a local large model includes: Local servers are used to deploy large models and support local inference and fine-tuning; The data acquisition unit is connected to the local server via a signal and is used to synchronously acquire multi-source sensor data, generate structured mature records, and send them to the local server. The control unit, connected to the local server, is used to convert the next voltage setpoint predicted by the large model into PLC instructions.

[0019] By adopting the above technical solution, a large model is deployed on the local server and supports local inference and fine-tuning. It can be fine-tuned and trained by combining locally collected historical aging record data, and the large model is used to predict the next voltage setpoint. The data acquisition unit synchronously acquires multi-source sensor data and generates structured aging records, which can provide key data in the high-voltage aging process. The control unit converts the next voltage setpoint predicted by the large model into PLC instructions to realize the automated control of the high-voltage aging system and improve the efficiency and accuracy of high-voltage aging.

[0020] Preferably, the control unit further includes: The conservative control logic module sets the limit range of a single voltage adjustment and sets the holding time after voltage is applied; The anomaly detection module triggers the conservative control logic module when the output of the large model contradicts the current trend.

[0021] By adopting the above technical solution, the anomaly detection module identifies model output anomalies through current trend analysis (such as volatility exceeding the threshold) and triggers a conservative strategy.

[0022] Preferably, the data acquisition unit includes: Multiple types of sensors are used for real-time monitoring of output voltage, current, leakage current, and vacuum level; A time synchronization mechanism ensures that the timestamps of all sensor data are aligned.

[0023] By adopting the above technical solution, the time synchronization mechanism controls the timestamp error of multi-source sensors within a reasonable range, ensures strict alignment of time series data, and avoids misjudgment of the model due to data misalignment.

[0024] In summary, this application includes at least one of the following beneficial technical effects: 1. Large-scale models are deployed locally throughout the entire process, with data remaining within the factory premises, greatly improving data security. Local inference response latency is far superior to cloud-based models, meeting the real-time requirements of high-pressure, mature systems. 2. The large model is logically integrated with the PLC device to achieve closed-loop automatic control, supporting manual intervention and dual verification, improving system safety and reliability, and solving the problems of traditional solutions that are difficult to achieve fully automatic closed-loop control and lack control accuracy and real-time performance; 3. Supports online incremental learning. By continuously collecting new sensor data for incremental training, the model can continuously evolve and adapt to equipment aging and changes in new operating conditions. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the overall process of Embodiment 1 of this application; Figure 2 This is a flowchart illustrating step S1 in Embodiment 1 of this application; Figure 3 This is a flowchart illustrating step S2 in Embodiment 1 of this application.

[0026] Figure 4 This is a schematic diagram of Embodiment 2 of this application. Detailed Implementation

[0027] The following combination Figures 1-4 This application will be described in further detail.

[0028] Example 1: This application discloses an automated high-pressure aging method based on a local large model.

[0029] Reference Figure 1 An automated high-pressure aging method based on a local large model includes the following steps: S0. Deploy a large model on a local server; S1. During the high-voltage aging process, multi-source sensor data is collected synchronously and multiple aging record data are generated. Each aging record data includes a timestamp, multi-source sensor data, ignition mark and corresponding next voltage setting value.

[0030] S2. Fine-tune and train the large model by combining locally collected historical data. S3. Connect the large model and the PLC device control logic; S4. The PLC device performs voltage adjustment after predicting the corresponding next voltage setting value through the large model.

[0031] Specifically, step S0 involves deploying the large model on the local server, including the following: A local server provides a stable operating environment. In this embodiment, a GPU server is used, whose powerful computing capabilities improve the running efficiency of large models. Furthermore, in this embodiment, the local server deploys the DeepSeekR1 distillation version using frameworks such as Ollam. The Ollam framework is an effective deployment tool that enables local inference and fine-tuning capabilities; other frameworks with similar functionalities can also be used. The DeepSeekR1 distillation version model excels at handling time-series data and prediction tasks, making it suitable for voltage setpoint sequence decision prediction in high-voltage mature control.

[0032] Reference Figure 2 In step S1, during the high-voltage aging process, multi-source sensor data is collected synchronously, and multiple aging record data are generated. Each aging record data includes a timestamp, multi-source sensor data, an ignition marker, and the corresponding next voltage setting value. The process also includes the following steps: S11: Acquire real-time data and generate aging records. During the high-voltage aging process, the high-voltage aging equipment collects real-time data on the high-voltage power supply output voltage, feedback current, leakage current of the tested equipment, and vacuum level through multiple types of sensors installed on the equipment. A unified time synchronization mechanism is used to ensure that all sensor data are strictly aligned under timestamps, guaranteeing data accuracy and consistency. After each aging process, the system automatically packages the collected real-time data and process events to generate a structured aging record. The aging record includes a timestamp, saved data, ignition markers, and the corresponding next voltage setting. The next voltage setting is set by the operator. Furthermore, the aging record is automatically saved after each aging process.

[0033] S12: Manually select high-quality, mature records as subsequent datasets.

[0034] High-quality data is mainly defined as follows: when ignition characteristics appear, the ignition state is successfully cleared by adjusting the voltage, and then the voltage is actively increased to the pre-ignition voltage value. After the voltage stabilizes at the current value, the voltage is further increased. All state data of successfully completing this process is recorded and saved as high-quality data for subsequent reinforcement training.

[0035] Operators can use visualization tools such as time series curves and event logs to check the data and confirm the following: whether the arcing event is real, whether the arc suppression operation is effective, and whether the voltage has been successfully restored to the target voltage and stabilized. Only mature records that meet the definition of high quality are retained.

[0036] Manual screening can remove noisy data (such as sensor malfunctions and misoperation records) to prevent large models from learning incorrect patterns.

[0037] S13: Format the filtered aging records into a format suitable for the large model. For example, store the aging records in JSON format. The timestamp, saved data from multiple sensors, and ignition markers are inputs, and the corresponding next-step voltage setting is the output.

[0038] Reference Figure 3 Step S2 combines locally collected historical mature data to fine-tune and train the large model, including the following steps: S21: Dataset Preparation. High-quality, mature records are loaded using the Hugging Face datasets library and divided into an 80% training set and a 20% validation set to ensure consistent data distribution for both large-scale model training and evaluation. The training set is used to train the large model, allowing it to learn patterns and features from the data; the test set is used to evaluate the model's performance, monitoring its mean absolute error and response latency on the test set to ensure accurate and real-time decision-making.

[0039] S22: Preprocessing mature record data. First, the mature record data is segmented and encoded using a tokenizer, converting the input into a numerical sequence composed of vocabulary indices, with a maximum length limit. Excessively long content is truncated, and insufficient content is padded with special padding characters. Simultaneously, corresponding attention masks are generated to distinguish between the real content and the padded parts. This ensures that all mature record data has a uniform tensor dimension, facilitating batch training.

[0040] Finally, the preprocessing function is applied in batches using the `ataset.map()` method. This batch processing significantly improves the efficiency of transforming large datasets. The entire process transforms unstructured text dialogue data into a structured numerical format containing input recognition sequences, attention masks, and label sequences, providing standardized input for subsequent large model training.

[0041] S23: Set large model parameters. Use the TrainingArguments class from the Hugging Face Transformer library to configure the core parameters for training the large model. It defines how the entire fine-tuning process runs and the training strategy.

[0042] First, the basic runtime configuration for the training process was set. This includes the output directory, log directory, and a large model saving strategy, ensuring that all output files during training are systematically saved. Second, learning parameters were also configured. These parameters include the learning rate and weight decay, which control the convergence speed of large models and prevent overfitting.

[0043] In terms of training process control, batch size, number of training rounds, and evaluation strategies are set to determine how the large model will traverse the training data and when validation and evaluation will be performed. Furthermore, by setting log recording step frequency, evaluation execution step frequency, and large model saving step frequency, fine-grained monitoring of the training process is further realized, enabling developers to regularly check training metrics, evaluate the performance of the large model, and save checkpoints, which is beneficial for monitoring and debugging long-term training tasks.

[0044] S24: Input the dataset and begin the training process.

[0045] During training, a validation set is used for large-scale model evaluation to monitor the performance of the large model. For example, MAE (mean absolute error) is used to measure the accuracy of the large model in predicting the next set voltage value.

[0046] Specifically, in high-voltage aging control tasks, when the large model needs to predict continuous variables (such as "next setpoint = 225V"), the large model will automatically calculate the MAE between the predicted voltage and the actual operating voltage, which can intuitively reflect the average deviation level of the large model's control voltage, and measure the accuracy of the large model in predicting the next setpoint voltage value.

[0047] S25: Save the large model.

[0048] After training is complete, the fine-tuned large model and tokenizer are saved to the local server for subsequent inference or deployment.

[0049] Step S3 connects the large model and the PLC device control logic, including the following steps; S31: Control logic interface.

[0050] The experienced system calls the local large model API to obtain the next pressure increase instruction (such as "next set value = 225V"), and converts the instruction into Modbus TCP protocol and writes it into the PLC register to facilitate subsequent voltage regulation.

[0051] S32: Incremental training.

[0052] In practical applications, the system continuously collects new sensor data. This data will be used as new training samples and incorporated into the training set for incremental training to improve the predictive ability of the large model.

[0053] Set up incremental training cycles, with each cycle automatically triggering incremental training. Each time new data is collected, use Hugging Face Trainer's incremental training function to continue training the large model without having to start training from scratch. Incremental training allows the model to adapt to device aging and learn new operating conditions.

[0054] Step S4 uses a large model to predict the corresponding next voltage setpoint, and the PLC device performs voltage adjustment, including the following: The finely tuned large model will be used for voltage setpoint prediction. Each time sensor data is input into the large model, it will output the next voltage setpoint and send it to the high-voltage aging system to perform voltage adjustment.

[0055] Furthermore, a dual verification mechanism is added to the PLC device program: if the output of the large model contradicts the current pressure trend, a conservative control strategy is prioritized to prevent misoperation. An example of adding conservative control logic to the PLC device is as follows: the voltage setting change range does not exceed ±5kV; after each voltage adjustment, a 3-minute hold period is required before increasing the voltage again; voltage reduction is not subject to the hold time limit; and the current threshold does not exceed 50uA. The above conservative control logic can be modified according to actual conditions, and more complex voltage and current trend change judgment logic can be added, but the main purpose is to prevent errors in the large AI model's decision-making, resulting in abnormally set voltage and current values ​​that could endanger the equipment.

[0056] The implementation principle of the automated high-pressure aging method based on a local large model in Embodiment 1 of this application is as follows: This application deploys the entire process locally, with data remaining within the factory area, greatly improving data security; the local inference response latency is far superior to that of a cloud-based large model, meeting the real-time requirements of high-pressure aging; it supports online incremental learning, allowing the large model to continuously evolve, whereas traditional deep learning large models typically require offline retraining and are difficult to adapt quickly to new working conditions; it directly connects to PLC equipment for closed-loop automatic control, whereas traditional solutions are mostly auxiliary decision-making and difficult to achieve fully automatic closed-loop control; it supports manual intervention and dual verification, improving system security and reliability.

[0057] Example 2: This application discloses an automated high-pressure aging system based on a local large model.

[0058] Reference Figure 4An automated high-voltage aging system based on a local large model includes a local server, a data acquisition unit, and a control unit. The local server is used to deploy the large model and support local inference and fine-tuning. The data acquisition unit is connected to the local server and is responsible for synchronously acquiring multi-source sensor data and generating structured aging records to be sent to the local server. The control unit is connected to the local server and is used to convert the next voltage setpoint predicted by the large model into PLC instructions, thereby achieving the effect of automated setting of the high-voltage aging voltage.

[0059] The data acquisition unit includes multiple types of sensors and a time synchronization mechanism. The sensors include voltage sensors, current sensors, leakage current sensors, and vacuum sensors, used to monitor output voltage, current, leakage current, and vacuum level in real time. The time synchronization mechanism can employ the NTP time synchronization protocol to ensure timestamp alignment of all sensor data.

[0060] The control unit includes a conservative control logic module and an anomaly detection module. The conservative control logic module sets the limit range for single voltage adjustment and the hold-up time after voltage application, thus preventing damage to the equipment from excessive voltage adjustment. The anomaly detection module triggers the conservative control logic module when the output of the large model contradicts the current trend, ensuring the safety and stability of the system.

[0061] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. An automated high-pressure aging method based on a local large model, characterized in that: Includes the following steps: Deploy large models on local servers; During the high-voltage aging process, multi-source sensor data is collected synchronously and multiple aging record data are generated. Each aging record data includes a timestamp, multi-source sensor data, ignition mark and corresponding next voltage setting value. The large model is fine-tuned and trained by combining locally collected historical data. Connect the large model and the PLC device control logic; The PLC device performs voltage adjustment based on the prediction of the next voltage setpoint using the large model.

2. The automated high-pressure aging method based on a local large model according to claim 1, characterized in that: It also includes manually screening high-quality aging records; the high-quality aging records are data that restores the voltage to the set value after eliminating sparking.

3. The automated high-pressure aging method based on a local large model according to claim 1, characterized in that: The large model was then fine-tuned and trained. Includes the following steps; Data set preparation involves dividing the historical mature record data into a training set and a validation set. Set the parameters of the large model; Input the training set and validation set into the large model and begin training; After training is complete, the fine-tuned large model is saved to the local server.

4. The automated high-pressure aging method based on a local large model according to claim 3, characterized in that: Setting the parameters of the large model includes setting the output directory, log recording directory, large model storage strategy, learning rate, weight decay, batch size, number of training epochs, and evaluation strategy.

5. The automated high-pressure aging method based on a local large model according to claim 3, characterized in that: The historical records are segmented and encoded, attention masks are generated, and maximum length limits are applied.

6. The automated high-pressure aging method based on a local large model according to claim 1, characterized in that: The control logic interface includes the following: The next voltage setting value predicted by the large model is written into the PLC register via the Modbus TCP protocol; Limit voltage variation, holding time, and current threshold in PLC equipment.

7. The automated high-pressure aging method based on a local large model according to claim 1, characterized in that: It also includes periodically collecting new sensor data, which is used as a training set for incremental training.

8. An automated high-pressure aging system based on a local large model, applicable to the high-pressure aging method described in any one of claims 1-7, characterized in that: include; Local servers are used to deploy large models and support local inference and fine-tuning; The data acquisition unit is connected to the local server via a signal and is used to synchronously acquire multi-source sensor data, generate structured mature records, and send them to the local server. The control unit, connected to the local server, is used to convert the next voltage setpoint predicted by the large model into PLC instructions.

9. The automated high-pressure aging system based on a local large model according to claim 8, characterized in that: The control unit further includes: The conservative control logic module sets the limit range of a single voltage adjustment and sets the holding time after voltage is applied; The anomaly detection module triggers the conservative control logic module when the output of the large model contradicts the current trend.

10. An automated high-pressure aging system based on a local large model according to claim 8, characterized in that: The data acquisition unit includes: Multiple types of sensors are used for real-time monitoring of output voltage, current, leakage current, and vacuum level; A time synchronization mechanism ensures that the timestamps of all sensor data are aligned.

Images