Rolling bearing digital twinning dynamic evolution method and system based on continuous learning

Abstract

The invention provides a rolling bearing digital twinning dynamic evolution method and system based on continuous learning, and belongs to the technical field of bearing life prediction. The method comprises the following steps: processing a bearing monitoring signal through short-time Fourier transform to generate standardized time-frequency data; constructing health indexes based on the index degeneration function and dividing health levels; training a life prediction model by using the extended LSTM network and taking the time-frequency data as input; real-time data is collected through a fixed time window to predict the life, and when the root-mean-square error of a predicted value and an actual value exceeds the limit, the edge device is triggered to upload new data; evaluating parameter importance in combination with a Fisher information matrix, and dynamically adjusting a regularization intensity updating model; and monitoring the data standard deviation in real time, triggering shutdown when the data standard deviation exceeds the limit, otherwise, predicting the remaining life by updating the model, and stopping when the remaining life reaches the threshold value. According to the method, dynamic evolution of the digital twin model is realized through continuous learning, and the industrial equipment state monitoring and predictive maintenance capability is effectively improved.

Description

Technical Field

[0001] The present invention relates to a rolling bearing digital twin dynamic evolution method and system based on continuous learning, belonging to the technical field of bearing life prediction. Background Art

[0002] In today's era of rapid industrial development, life monitoring of key components in mechanical equipment is becoming increasingly important. With the rapid development of edge computing, 5G, and big data, the amount of data generated by modern industrial facilities far exceeds that of traditional manufacturing environments. Modern industry urgently needs to seek innovative predictive maintenance solutions to achieve predictive maintenance and health management in dynamic environments.

[0003] Rolling bearings are essential components of industrial equipment such as high-speed trains, aircraft, and marine engines. However, due to harsh operating conditions, they are prone to failure. In some cases, the operating status of rotating machinery is crucially dependent on the health of the bearings. During actual operation, bearings are susceptible to environmental influences, exhibiting varying behavior patterns under different operating conditions. This can lead to wear and tear on internal components, significantly reducing their lifespan.

[0004] Digital twinning is a technology that digitizes physical entities, simulating actual physical processes to enable simulation, data analysis, and design optimization. Therefore, to reduce the gap and uncertainty between models and reality, digital twins need to dynamically evolve to improve the accuracy of rolling bearing life prediction.

[0005] However, traditional static models are typically trained based on offline datasets. Once trained, they require retraining to incorporate new data into the model. Furthermore, predictions using deep learning models are prone to catastrophic forgetting. In real-world applications, datasets often change dynamically, so the performance and accuracy of traditional machine learning methods degrade over time, making them unable to reliably reflect and predict the actual state of the device. Continuous learning is a machine learning technique that incrementally updates a model by continuously introducing new data samples. In continuous learning, a model can improve its predictive capabilities by learning from new data samples without retraining, while preserving existing knowledge and model structure. Summary of the Invention

[0006] The purpose of the present invention is to provide a method and system for the dynamic evolution of rolling bearing digital twins based on continuous learning. Based on the continuous learning method, the dynamic evolution of the digital twin model is realized, aiming to improve the condition monitoring and predictive maintenance capabilities of industrial equipment.

[0007] To achieve the above-mentioned purpose, the present invention is implemented through the following technical solutions: A rolling bearing digital twin dynamic evolution method based on continuous learning includes the following steps: Collect the condition monitoring signal of the rolling bearing and perform short-time Fourier transform to generate time-frequency domain data and standardize it; Based on the time-frequency domain data, a health index is constructed using an exponential degradation function to divide the health status into health levels; An extended long short-term memory network model is used to train a bearing life prediction model with standardized time-frequency domain data as input and health indicators as labels. Set a fixed time window, collect real-time data, and input it into the trained bearing life prediction model for prediction. Calculate the root mean square error between the predicted value and the actual value. If the error exceeds the preset threshold, trigger an edge-side command to send the terminal device to upload new data. A continuous learning method with adaptive elastic weight consolidation is adopted. The importance of model parameters is evaluated based on the Fisher information matrix. The regularization strength is dynamically adjusted through task embedding and cosine similarity to update the bearing life prediction model. Monitor the standard deviation of real-time data. If it is greater than the preset threshold, a shutdown command is triggered. If it is less than or equal to the preset threshold, the remaining life is predicted based on the updated bearing life prediction model. If the predicted value is less than the set life threshold, a shutdown command is triggered.

[0008] Preferably, the state monitoring signal includes a vibration signal, an acoustic emission signal, a temperature signal, a current signal, a rotation speed signal, and a force signal.

[0009] Preferably, the preset threshold is 2 to 3 times the standard deviation of the historical stable operation stage; the set life threshold is 0.2.

[0010] Preferably, the method for constructing the health index includes: Generate comprehensive health indicators by fusing physical health indicators and virtual health indicators; The physical health index is extracted based on the root mean square value of the vibration signal, and the virtual health index is generated by multi-sensor signal fusion; The exponential degradation function is used to describe the nonlinear degradation process, and its formula is: , , in, is the convergence speed hyperparameter, represents the initial value constant of the health index, represents the time scale adjustment parameter, represents the current time point or time step, is the base of natural logarithms, Represents the health index, Indicates the time point at the beginning of the bearing operation. Indicates the time point when the bearing is approaching failure.

[0011] Preferably, the working condition adaptive elastic weight consolidation method specifically includes: Using an autoencoder to extract operating condition features, which are characteristic information that characterizes the equipment's operating environment or working state, including speed, load, temperature, vibration amplitude, and operating condition category labels, to generate a low-dimensional task embedding vector. Based on the cosine similarity between task embedding vectors, the regularization strength is dynamically allocated. The regularization strength is reduced in similar working conditions to allow parameter updates, while the regularization strength is increased in different working conditions to protect important parameters.

[0012] Preferably, the loss function of the bearing life prediction model update process is as follows: , , in, represents the mean square error, represents the final regularization strength, Indicates the total number of parameters, represents the indicator function, represents the network parameters for training new detection tasks when updating digital twins, represents the optimal parameters obtained for the old detection task before the digital twin is updated, Indicates the A detection task, is the actual result, is the predicted result, Indicates the total number of samples involved in the MSE calculation.

[0013] Preferably, the final regularization strength The calculation is as follows: , in, represents the base regularization strength, Represents the similarity between the new and old task embedding vectors.

[0014] Preferably, the basic regularization strength It is determined by the average Fisher value of the bearing life prediction model parameters. The specific formula is as follows: , in, Representation parameters The diagonal Fisher value of is calculated as follows: , in, Indicates the size of the training dataset, represents the maximum likelihood estimate, is the expected function, represents the log-likelihood function with respect to the parameter The derivative of .

[0015] A rolling bearing digital twin dynamic evolution system based on continuous learning, including: Terminal devices and sensor modules: deployed in key locations of rolling bearings to collect status monitoring signals in real time and upload them to edge servers; Edge server module: updates the bearing life prediction model based on the rolling bearing digital twin dynamic evolution method based on continuous learning and generates control instructions; Digital twin decision module: Generates control instructions based on the model prediction results and feeds them back to the terminal device through the edge server to execute operations.

[0016] The advantages of the present invention are: This paper designs a rolling bearing digital twin dynamic evolution system based on continuous learning. This system collects real-time data from rolling bearings through terminal devices and sensors, and performs data preprocessing and real-time model monitoring on edge servers. This system enables real-time monitoring and predictive maintenance of bearings throughout their entire lifecycle, improving the accuracy of equipment maintenance and significantly reducing the likelihood of accidents.

[0017] When the system detects an abnormal signal or determines that the bearing has reached its end of life based on prediction results, it can quickly provide feedback to the terminal equipment and execute shutdown or equipment replacement operations, thereby ensuring stable operation and production efficiency of the equipment.

[0018] This paper uses a regularized continuous learning update strategy to fine-tune the digital twin model by calculating the Fisher information of the model parameters. This method significantly improves the accuracy of the model, enabling it to more accurately reflect and predict the actual state of the device.

[0019] The application of regularization strategies also helps reduce the information transmission, processing, and storage pressure on edge devices. This is because the model can continuously improve its prediction ability by learning from new data samples without retraining, thus reducing the burden on edge devices.

[0020] This invention adopts a condition-adaptive hyperparameter selection strategy. By pre-classifying historical tasks by condition and optimizing hyperparameters, new tasks only need to determine the condition type and directly select appropriate hyperparameters from the corresponding set. This approach significantly saves time and computing resources, making task processing more efficient.

[0021] Especially when faced with a large number of bearing monitoring tasks with high real-time requirements, this strategy can quickly provide configuration solutions to ensure that subsequent analysis and decision-making are not delayed, thereby significantly improving the system's response speed and practicality. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0023] Figure 1 Schematic diagram of the process of the present invention.

[0024] Figure 2 Flowchart for decision-making on building and updating a bearing digital twin.

[0025] Figure 3 Schematic diagram of the dynamic evolution method of rolling bearing digital twins based on continuous learning and the edge-end collaborative system structure. DETAILED DESCRIPTION

[0026] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0027] Example 1 like Figure 1 and Figure 2 The paper shows a method for the dynamic evolution of rolling bearing digital twins based on continuous learning. This method uses continuous learning to construct and dynamically evolve digital twins of rolling bearings. This architecture builds a real-time updating digital twin architecture on the edge computing side, utilizing the digital twin for full lifecycle monitoring to improve real-time processing capabilities. When abnormal signals or predictions are detected, timely feedback is provided to the terminal equipment for shutdown and replacement. This continuous learning approach enables the dynamic evolution of the digital twin model, meeting the real-time requirements of actual operating conditions.

[0028] Specifically, it includes three stages: Phase 1: The construction of a bearing digital twin utilizes historical data cached on edge servers and employs a novel recursive neural network architecture, xLSTM (Extended Long Short-Term Memory). This architecture aims to address some of the limitations of traditional LSTM networks and improve their performance in tasks such as time series forecasting. The model is used to construct a bearing digital twin for bearing life prediction, accurately capturing the dynamic behavior and operating status of the bearing, providing a foundation for subsequent dynamic evolution. This includes: S1: Collect the condition monitoring signal of the rolling bearing and perform short-time Fourier transform to generate time-frequency domain data and standardize it.

[0029] As a refinement of the above embodiment, data processing reveals that in the frequency domain, many bearing faults (such as rolling element defects, inner race defects, and outer race defects) exhibit specific frequency characteristics. These characteristics may not be apparent in time domain signals, but Fourier transforms can more clearly identify them. Considering that data is uploaded to edge devices at long intervals, a short-time Fourier transform (STFT) is used to obtain time-frequency domain signals, which are then standardized. The obtained samples are scaled to fit within the [0–1] range and used as input for the xLSTM model.

[0030] S2: Based on the time-frequency domain data, a health index is constructed using an exponential degradation function to divide the health status into health levels.

[0031] As a refinement of the above embodiment, the construction of bearing health indicators can be divided into two categories: physical health indicators and virtual health indicators. Physical health indicators are related to physical failures and are typically extracted from monitoring signals using statistical methods or signal processing methods (such as the root mean square of vibration signals). Virtual health indicators are typically constructed by fusing multiple physical health indicators or multi-sensor signals. They lose their physical meaning and provide a virtual description of the degradation trend of the machine.

[0032] An exponential degradation function is used to construct a bearing health index. The degradation process of bearings is typically nonlinear, with degradation accelerating particularly towards the end of their lifecycle. The exponential degradation function can better capture this nonlinearity and provide a more accurate health assessment. The health index is divided into five health levels, with a range of 0.2. A health index of [0.8, 1] indicates intact equipment; [0.6, 0.8] indicates degraded equipment functionality; [0.4, 0.6] indicates impaired equipment functionality; [0.2, 0.4] indicates severe functional impairment; and [0, 0.2] indicates complete equipment failure.

[0033] Its exponential decay function is: , , in, is the convergence speed hyperparameter, represents the initial value constant of the health index (usually taken as 1 or slightly greater than 1 for normalization), Represents the time scale adjustment parameter, which is used to control the decay speed of the exponential function. represents the current time point or time step, is the base of natural logarithms, Health Indicator, used to characterize the current health status of the bearing. Indicates the time point at the beginning of the bearing's operation, which corresponds to the best health condition. Indicates the time point when the bearing is close to failure, corresponding to the worst health state. If HI=1, it means the system is operating normally, and if HI=0, it means the system is in complete failure mode.

[0034] S3: Using the extended long short-term memory network model, the standardized time-frequency domain data is used as input and the health indicators are used as labels to train the bearing life prediction model.

[0035] As a refinement of the above embodiment, deep learning model training uses the health index value obtained at each time t as a label. An LSTM model is trained to predict the health status of the bearing, enabling continuous data processing and providing more accurate prediction results. The number of input, hidden, and output layers of the xLSTM is determined, and the Adam optimizer is used for iterative training to fit the health index curve. The loss function is: , in, is the actual value and is the prediction result.

[0036] In this phase, a new recurrent neural network (xLSTM) model is used to construct a digital twin model of the bearing using historical data cached on edge servers to predict bearing life. Because the extent of mechanical bearing damage, such as crack length and wear area, is generally not directly observable, various condition monitoring signals, such as vibration and acoustic emission signals, are typically acquired from operating equipment to estimate the machine's health in real time. These monitoring signals contain a significant amount of health information and measurement noise, necessitating the construction of health indicators that describe historical and ongoing degradation processes. A data-driven approach to bearing life prediction is employed, disregarding the failure mechanisms within rolling bearing components. Instead, implicit bearing health information and its evolution patterns are directly extracted from bearing performance test data and condition monitoring data (such as acceleration and temperature). Deep learning methods are then used to predict the remaining bearing life.

[0037] Phase 2: The dynamic evolution of the bearing digital twin determines whether dynamic evolution of the digital twin is necessary based on a time-triggered scheme. If necessary, the edge side sends instructions to the terminal device to upload collected data. Continuous learning and updating of the model are then used to perform real-time bearing life prediction and dynamic evolution of the digital twin. Data is transmitted and control signals are updated only when pre-designed trigger conditions are met. This reduces unnecessary signal transmission, conserves communication and computing resources, and reduces energy consumption. This includes: S4: Set a fixed time window, collect real-time data and input it into the trained bearing life prediction model for prediction. Calculate the root mean square error between the predicted value and the actual value. If the error exceeds the preset threshold, trigger the edge-side command to send the terminal device to upload new data.

[0038] As a refinement of the above embodiment, it specifically includes: (1) Real-time data monitoring uses a periodic event triggering scheme to determine whether it is necessary to continuously upload data to update the digital twin.

[0039] Specifically, set the time window : Indicates the duration of a single time window or the number of data samples.

[0040] Prediction: The input is real data collected during a window of time. , the predicted value is the interval status.

[0041] Judgment update: comparison is interval Predicted value and actual value that has been collected , using actual data stored in the edge buffer and the relevant simulated output data predicted by the digital twin The error is judged, and the error exceeds the threshold Upload the next real-time data and update the digital twin. Using the updated model, As input, prediction , in this order. Trigger function Using RMSE mean square error function: , in, Indicates the size of the data obtained from the edge side and stored in the buffer. Indicates the actual value, Indicates the prediction result.

[0042] (2) The edge sends instructions and the terminal uploads data. When the error exceeds the specified threshold, the edge sends instructions to the terminal to upload real-time data. The size of the uploaded data is ,in The interval between the last data upload, each transmission time is: , in, is the transmission rate.

[0043] S5: A condition-adaptive elastic weight consolidation continuous learning method is adopted to evaluate the importance of model parameters based on the Fisher information matrix. The regularization strength is dynamically adjusted through task embedding and cosine similarity to update the bearing life prediction model.

[0044] As a refinement of the above embodiment, it specifically includes: (1) Use autoencoder to compress the extracted features of each working condition into a low-dimensional embedding vector Used to quantify task similarity.

[0045] (2) Continuously learn and calculate the Fisher matrix, input the real-time data uploaded in the previous time window into the trained digital twin model, and calculate the Fisher information of the model: is about the parameters described in Eq. The diagonal Fisher information of : , in, Indicates the size of the training dataset, is the maximum likelihood estimate, is the expected function, represents the log-likelihood function with respect to the parameter The derivative of . Save the Fisher information matrix for subsequent parameter grouping and regularization weight assignment.

[0046] (3) Determine according to parameter importance and finally . Determined by the average Fisher value of the task model parameters, The bigger, The bigger , Adaptive Grouping :each group Independently adjust based on task similarity: , in, is the cosine similarity, is the basic regularization strength. Large value): Low, allowing more parameter updates to adapt to similar working conditions; Different tasks ( Small value): High, strengthen the protection of important parameters.

[0047] (4) Model update and bearing life prediction. Based on the calculated model Fisher information, the model is trained with the goal of minimizing loss to obtain an updated model for subsequent bearing life prediction.

[0048] The loss function for iterative model training is: , in, represents the final regularization strength, Indicates the total number of parameters, represents the indicator function, represents the network parameters for training new detection tasks when updating digital twins, represents the optimal parameters obtained for the old detection task before the digital twin is updated, Indicates the detection tasks (related to the task index), is the actual result, is the predicted result, Indicates the total number of samples involved in the MSE calculation.

[0049] (5) The optimization objective based on the update of the bearing digital twin is to minimize the loss function: , Where, Represents minimizing the loss function and regularizing to alleviate the catastrophic forgetting phenomenon of the data-driven digital twin model. It is the optimal parameter of the digital twin after dynamic evolution.

[0050] In this stage, after the prediction is completed within the set fixed time window, the time trigger scheme is used to determine whether dynamic evolution of the digital twin is required. If necessary, the edge side sends instructions to the terminal device to upload the collected data. Then, based on the trained digital twin model, continuous learning is used to update the model for real-time bearing life prediction and dynamic evolution of the digital twin. The continuous learning method uses condition-adaptive elastic weight consolidation (A-EWC) to achieve updates by calculating the Fisher information of the model parameters, realizing the dynamic evolution of the digital twin. The operation is relatively simple without dynamically changing the network architecture, and does not require a large amount of edge memory space. For each data update, the Fisher matrix is ​​used to evaluate the model parameters. Through task embedding (TaskEmbedding) and parameter importance stratification, the regularization strength under different working conditions is dynamically adjusted to achieve parameter sharing within the same working condition and independent working condition adaptation of parameters between different working conditions.

[0051] Phase 3 The bearing digital twin makes decisions based on the dynamic evolution of the digital twin model. When abnormal signals are detected or the predicted lifespan has been reached, timely feedback is provided to the terminal equipment, prompting shutdown or equipment replacement. By analyzing and predicting the operating status of bearings, a basis for maintenance and optimization is provided, thereby improving system reliability and efficiency.

[0052] S6: Monitor the standard deviation of real-time data. If it is greater than a preset threshold (such as 2 to 3 times the standard deviation of the historical stable operation stage), a shutdown command is triggered. If it is less than or equal to the preset threshold, the remaining life is predicted based on the updated bearing life prediction model. If the predicted value is less than the set life threshold (the predicted value is the remaining service life, here HI=0.2 is set as the life threshold), a shutdown command is triggered.

[0053] As a refinement of the above embodiment, when an abnormal signal is detected, the signal fluctuates abnormally within a certain period of time, the standard deviation is large, or the predicted life span is reached, timely feedback is sent to the terminal device, and shutdown or equipment replacement operations are executed. The specific steps are: (1) The device uploaded to the edge server performs standard deviation judgment. If it is greater than a given threshold, it is directly fed back to the terminal bearing device for shutdown inspection. If it does not exceed the threshold, step (2) is executed. The bearing is The number of samples in time is The standard deviation of is: , (2) Determine the bearing life value. If the predicted value If the value is greater than a given threshold, it will be directly fed back to the terminal bearing equipment for shutdown inspection, otherwise it will continue to stage two.

[0054] Example 2 like Figure 3 As shown in FIG, a rolling bearing digital twin dynamic evolution system based on continuous learning includes: Terminal equipment and sensor modules: Deployed at key locations on rolling bearings, they collect status monitoring signals in real time and upload them to edge servers; specifically, these signals include horizontal and vertical vibration signals and temperature signals.

[0055] Edge server module: updates the bearing life prediction model based on the rolling bearing digital twin dynamic evolution method based on continuous learning and generates control instructions; Specifically, the bearing data collected by the sensor and cached at the edge is preprocessed, and deep learning technology is used to generate a digital twin model of the bearing, which can monitor the equipment status in real time and provide timely feedback to the terminal device when abnormal signals are detected or based on prediction results.

[0056] Digital twin decision module: Generates control instructions based on the model prediction results and feeds them back to the terminal device through the edge server to execute operations.

[0057] The system realizes real-time monitoring and predictive maintenance of rolling bearings throughout their life cycle. By utilizing edge computing and continuous learning technologies, it not only improves the system's real-time processing capabilities, but also enables rapid response to abnormal situations, ensuring stable operation and production efficiency of the equipment.

[0058] The disclosed embodiments also provide a rolling bearing digital twin dynamic evolution device based on continuous learning, comprising a processor and memory. Optionally, the device may also include a communication interface and a bus. The processor, communication interface, and memory may communicate with each other via the bus. The communication interface may be used for information transmission. The processor may invoke logic instructions in the memory to execute the rolling bearing digital twin dynamic evolution method based on continuous learning described in the aforementioned embodiment.

[0059] In addition, the logic instructions in the above-mentioned memory can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product.

[0060] Memory, as a computer-readable storage medium, can be used to store software programs and computer-executable programs, such as the program instructions / modules corresponding to the methods in the embodiments of the present disclosure. The processor executes the program instructions / modules stored in the memory to perform functional applications and data processing, thereby implementing the rolling bearing digital twin dynamic evolution method based on continuous learning in the above-mentioned embodiments.

[0061] The memory may include a program storage area and a data storage area. The program storage area may store an operating system and at least one application required for a function; the data storage area may store data generated based on the use of the terminal device. Furthermore, the memory may include high-speed random access memory and non-volatile memory.

[0062] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art will be able to modify the technical solutions described in the aforementioned embodiments or substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A dynamic evolution method of rolling bearing digital twin based on continuous learning, characterized by: The following steps are involved: Collect the condition monitoring signal of the rolling bearing and perform short-time Fourier transform to generate time-frequency domain data and standardize it; Based on the time-frequency domain data, a health index is constructed using an exponential degradation function to divide the health status into health levels; An extended long short-term memory network model is used to train a bearing life prediction model using standardized time-frequency domain data as input and health indicators as labels. Set a fixed time window, collect real-time data, and input it into the trained bearing life prediction model for prediction. Calculate the root mean square error between the predicted value and the actual value. If the error exceeds the preset threshold, trigger an edge-side command to send the terminal device to upload new data. A continuous learning method with adaptive elastic weight consolidation is adopted. The importance of model parameters is evaluated based on the Fisher information matrix. The regularization strength is dynamically adjusted through task embedding and cosine similarity to update the bearing life prediction model. Monitor the standard deviation of real-time data. If it is greater than the preset threshold, a shutdown command is triggered. If it is less than or equal to the preset threshold, the remaining life is predicted based on the updated bearing life prediction model. If the predicted value is less than the set life threshold, a shutdown command is triggered.

2. The rolling bearing digital twin dynamic evolution method based on continuous learning according to claim 1 is characterized in that: The state monitoring signals include vibration signals, acoustic emission signals, temperature signals, current signals, rotation speed signals, and force signals.

3. The rolling bearing digital twin dynamic evolution method based on continuous learning according to claim 1 is characterized in that: The preset threshold is 2 to 3 times the standard deviation of the historical stable operation stage; the set life threshold is 0.

2.

4. The rolling bearing digital twin dynamic evolution method based on continuous learning according to claim 1 is characterized in that: The method for constructing the health indicator includes: Generate comprehensive health indicators by fusing physical health indicators and virtual health indicators; The physical health index is extracted based on the root mean square value of the vibration signal, and the virtual health index is generated by multi-sensor signal fusion; The exponential degradation function is used to describe the nonlinear degradation process, and its formula is: ， ， in, is the convergence speed hyperparameter, represents the initial value constant of the health index, represents the time scale adjustment parameter, represents the current time point or time step, is the base of natural logarithms, Represents the health index, Indicates the time point at the beginning of the bearing operation. Indicates the time point when the bearing is approaching failure.

5. The rolling bearing digital twin dynamic evolution method based on continuous learning according to claim 1 is characterized in that: The working condition adaptive elastic weight consolidation method specifically includes: Using an autoencoder to extract operating condition features, which are characteristic information that characterizes the equipment's operating environment or working state, including speed, load, temperature, vibration amplitude, and operating condition category labels, to generate a low-dimensional task embedding vector. Based on the cosine similarity between task embedding vectors, the regularization strength is dynamically allocated. The regularization strength is reduced in similar working conditions to allow parameter updates, while the regularization strength is increased in different working conditions to protect important parameters.

6. The rolling bearing digital twin dynamic evolution method based on continuous learning according to claim 5 is characterized in that: The loss function of the bearing life prediction model update process is as follows: ， ， in, represents the mean square error, represents the final regularization strength, Indicates the total number of parameters, represents the indicator function, represents the network parameters for training new detection tasks when updating digital twins, represents the optimal parameters obtained for the old detection task before the digital twin is updated, Indicates the A detection task, is the actual result, is the predicted result, Indicates the total number of samples involved in the MSE calculation.

7. The rolling bearing digital twin dynamic evolution method based on continuous learning according to claim 6 is characterized in that: The final regularization strength The calculation is as follows: ， in, represents the base regularization strength, Represents the similarity between the new and old task embedding vectors.

8. The rolling bearing digital twin dynamic evolution method based on continuous learning according to claim 7 is characterized in that: The base regularization strength It is determined by the average Fisher value of the bearing life prediction model parameters. The specific formula is as follows: ， in, Representation parameters The diagonal Fisher value of is calculated as follows: ， in, Indicates the size of the training dataset, represents the maximum likelihood estimate, is the expected function, represents the log-likelihood function with respect to the parameter The derivative of .

9. A rolling bearing digital twin dynamic evolution system based on continuous learning, characterized by: include: Terminal devices and sensor modules: deployed in key locations of rolling bearings to collect status monitoring signals in real time and upload them to edge servers; Edge server module: updates the bearing life prediction model based on the rolling bearing digital twin dynamic evolution method based on continuous learning as described in any one of claims 1 to 8, and generates control instructions; Digital twin decision module: Generates control instructions based on the model prediction results and feeds them back to the terminal device through the edge server to execute operations.

10. A rolling bearing digital twin dynamic evolution device based on continuous learning, comprising a processor and a memory storing program instructions, characterized in that: The processor is configured to execute the rolling bearing digital twin dynamic evolution method based on continuous learning as described in any one of claims 1 to 8 when running the program instructions.

Images