Equipment generation type modeling method based on DBSCAN clustering and hybrid expert model

Abstract

The invention relates to the technical field of automatic control and industrial equipment management, and discloses an equipment generation type modeling method based on DBSCAN clustering and a hybrid expert model, which relates to the technical field of automatic control and industrial equipment management, and comprises the steps of establishing uniform identification and data representation of equipment, processing data, calculating Pearson correlation coefficients or mutual information to screen key variables, carrying out dimensionality reduction on operation data, and utilizing a DBSCAN clustering algorithm to carry out modeling on the operation data. Clustering analysis is carried out on the standardized equipment data, a multi-level hybrid expert model is constructed, new equipment feature data are classified into corresponding clusters according to the new equipment feature data, a common model and a personalized model are selected to be combined, and a prediction model of the new equipment is generated through data fine tuning; the method provided by the invention has good expansibility, adapts to modeling requirements of different types of equipment, not only can perform rapid modeling on new equipment based on limited local data, but also can realize knowledge migration by utilizing the learned common characteristics of the same cluster of equipment, and can efficiently generate a model of the new equipment.

Description

Technical Field

[0001] The present invention relates to the technical fields of automatic control and industrial equipment management, and particularly to a device generative modeling method based on DBSCAN clustering and a hybrid expert model. Background Art

[0002] In an industrial control system, the device states and environmental conditions are usually extremely complex, and the types, models, parameters, etc. of devices are different. How to efficiently and accurately establish a device model for prediction and control is the key to improving the operation efficiency of the industrial system.

[0003] Traditional device modeling methods often rely on a large amount of historical data and expert knowledge, and mainly have the following problems:

[0004] 1. Strong data dependence: Traditional model identification methods require a large amount of historical data for training, and the data acquisition cost is high;

[0005] 2. High professional knowledge dependence: Due to the variety of device types and different models, the selection and construction of models seriously depend on the domain knowledge and experience judgment of professionals, with low automation, high labor costs, and a large subjective factor influence;

[0006] 3. Poor model generality: Existing device models are often difficult to be directly applied to new devices, and the cost of model migration is high.

[0007] DBSCAN (Density-Based Spatial Clustering of Applications with Noise) is a density-based clustering algorithm that can discover clusters of any shape and identify noise points. DBSCAN does not require the number of clusters to be specified in advance and is robust to noise and outliers. Therefore, in view of the diversity of devices and complex states, the present invention proposes a device generative modeling method based on DBSCAN clustering and a hybrid expert model to solve the deficiencies in the prior art. Summary of the Invention

[0008] In view of the above problems, the purpose of the present invention is to provide a device generative modeling method based on DBSCAN clustering and a hybrid expert model. The method has good scalability and can meet the modeling requirements of different types of devices. For new devices, not only can a quick model be established based on limited local data, but also the common features learned by devices in the same cluster can be utilized to achieve knowledge migration, and the model of the new device can be efficiently generated, thus significantly reducing the cost of device modeling.

[0009] To achieve the above purpose, the present invention provides the following technical solutions:

[0010] Device generative modeling method based on DBSCAN clustering and mixture of experts model, comprising the following steps:

[0011] Step 1: Establish unified identification and data representation of the device

[0012] Collect the historical operation data of the device, store it in time series, and then perform unified identification and data representation;

[0013] Step 2: Analysis of historical operation data

[0014] Perform data processing on the collected historical operation data, and in the feature selection stage, calculate the Pearson correlation coefficient or mutual information to screen key variables, and apply principal component analysis to reduce the dimension of the operation data of the device after data processing to form standardized device data;

[0015] Step 3: Clustering analysis based on DBSCAN

[0016] Use the DBSCAN clustering algorithm to perform clustering analysis on the standardized device data to identify the similarity between devices;

[0017] Step 4: Construct a multi-level mixture of experts model

[0018] Based on the clustering analysis results, construct a multi-level mixture of experts model, where the mixture of experts model includes a basic shared expert and a personalized expert to adapt to the characteristics of different devices;

[0019] Step 5: Model identification and optimization of new devices

[0020] Classify the new device according to its characteristic data into the corresponding clustering cluster, select a common model and a personalized model for combination, and then perform data fine-tuning to generate a prediction model for the new device.

[0021] The further improvement lies in: when performing unified identification and data representation on the historical operation data of the device in step 1, let the data of each device be represented as , where represents the th feature, and through feature standardization and unified processing, convert the original data of different devices into a unified feature representation , and each standardized feature satisfies:

[0022]

[0023] where and respectively represent the mean and standard deviation of the th feature.

[0024] A further improvement lies in that: before processing the collected historical operation data in the second step, first form a data set from the collected historical operation data:

[0025]

[0026] wherein, is the input feature at the t-th time point, is the corresponding output response.

[0027] A further improvement lies in that: when processing the collected historical operation data in the second step, it includes preprocessing the collected device operation data set, such as cleaning, denoising, and filling missing values, to ensure data quality.

[0028] A further improvement lies in that: in the feature selection stage of the second step, calculate the Pearson correlation coefficient or mutual information to screen key variables, and apply principal component analysis to reduce the dimension of the operation data of the device after data processing to form standardized device data. Specifically:

[0029] Extract the key features that mainly affect the device performance and reduce the data dimension:

[0030]

[0031] wherein, is the feature transformation matrix, is the feature vector after dimension reduction.

[0032] A further improvement lies in that: the specific process of the third step is as follows: DBSCAN divides data clusters through the density threshold and the minimum number of points MinPts, where:

[0033] Density neighborhood definition: For each feature vector after dimension reduction of the device, define its neighborhood as:

[0034]

[0035] When , then is a core point, and then form a clustering cluster by connecting core points and density-reachable points;

[0036] Clustering result: Obtain the class distribution of the device, where each cluster represents a class of devices with similar features.

[0037] A further improvement lies in that: in step four, the basic shared expert module is used to capture the common features of each device within the same clustering cluster. Through the shared expert model, the general behavioral features of different devices under similar operating conditions can be effectively modeled, enabling the devices within the same cluster to share some common feature expressions for the clustering cluster , train the basic commonality expert network :

[0038]

[0039] wherein, are the parameters of the shared expert model.

[0040] A further improvement lies in that: in step four, the personalized expert module is for different devices within the same clustering cluster and is used to learn the unique features of the devices. For the th device in the clustering cluster , train the personalized expert network :

[0041]

[0042] wherein, represents the dataset of the th device in the cluster , and are the parameters of the personalized expert network;

[0043] Fuse the basic shared expert and the personalized expert to form a hybrid expert model, and use the gating function to determine the weights of each model:

[0044]

[0045] The gating function is adaptively adjusted according to the device features or data volume and is set as:

[0046]

[0047] wherein, is the Sigmoid function, and are the parameters of the gating function.

[0048] A further improvement lies in that: when classifying new devices in step five, for the dimensionality-reduced feature vector of the new device, calculate its distance from the center of each clustering cluster:

[0049]

[0050] Assign the new device to the clustering cluster with the minimum distance.

[0051] A further improvement lies in that when identifying the new device in step five, the new device is classified into the clustering cluster with the smallest distance. For the new device with limited local data Then, a preliminary model identification is carried out using the hybrid expert model:

[0052]

[0053] Based on the data of the new device, the personalized expert network and the parameters of the gating function are fine-tuned to minimize the loss function:

[0054]

[0055] The parameters of the shared expert model are fine-tuned or remain unchanged.

[0056] The beneficial effects of the present invention are as follows: By combining DBSCAN clustering analysis and the hybrid expert model, the present invention can identify and generate device models under limited local data, reducing the dependence on a large amount of historical data;

[0057] By using the clustering analysis results to achieve automatic model selection, modeling can be carried out without the in-depth participation of domain experts, which is applicable to diverse device scenarios; <L

[0058] The multi-level hybrid expert model not only has generality but also can provide personalized device feature modeling, which helps to improve the adaptability and accuracy of the model;

[0059] The unified device representation system and the multi-level expert model framework enable the method to have good scalability and adapt to the modeling needs of different types of devices. For new devices, not only can rapid modeling be carried out based on limited local data, but also the common features learned by the devices in the same cluster can be utilized to achieve knowledge transfer, and the model of the new device can be efficiently generated, thus significantly reducing the cost of device modeling. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 It is a schematic diagram of the method framework of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0061] In order to deepen the understanding of the present invention, the following will further elaborate on the present invention in combination with embodiments. These embodiments are only used to explain the present invention and do not constitute a limitation on the protection scope of the present invention.

[0062] Embodiment 1

[0063] According to Figure 1As shown in the figure, this embodiment proposes a device generative modeling method based on DBSCAN clustering and a mixture of experts model, including the following steps:

[0064] Step 1: Establish a unified identification and data representation of the device

[0065] Collect the historical operation data of the device, store it in time series, and then perform unified identification and data representation;

[0066] When performing unified identification and data representation on the historical operation data of the device, assume that the data of each device is represented as , where represents the th feature (such as device type, model, parameters, operating environment, etc.). Through feature standardization and unified processing, the original data of different devices is converted into a unified feature representation , and each standardized feature satisfies:

[0067]

[0068] Among them, and respectively represent the mean and standard deviation of the th feature.

[0069] Step 2: Analysis of historical operation data

[0070] Perform data processing on the collected historical operation data, and in the feature selection stage, calculate the Pearson correlation coefficient or mutual information to screen key variables, and apply principal component analysis to reduce the dimension of the operation data of the device after data processing to form standardized device data;

[0071] In this step, before performing data processing on the collected historical operation data, first form a data set from the collected historical operation data:

[0072]

[0073] Among them, is the input feature at the t-th time point, is the corresponding output response;

[0074] When performing data processing on the collected historical operation data, it includes preprocessing such as cleaning, denoising, and filling missing values on the collected device operation data set to ensure data quality.

[0075] In this implementation, data cleaning removes outliers through the IQR (interquartile range) method, performs smoothing processing using the moving average method, and at the same time performs Z-score or Min-Max normalization on the data to eliminate scale differences.

[0076] In the feature selection stage, the Pearson correlation coefficient or mutual information (MI) is calculated to screen key variables, and principal component analysis (PCA) is applied for dimensionality reduction to improve computational efficiency and ensure the robustness and accuracy of equipment state modeling;

[0077] When extracting the key features that mainly affect equipment performance and reducing the data dimension,

[0078]

[0079] where, is the feature transformation matrix, is the feature vector after dimensionality reduction.

[0080] Step 3: Clustering analysis based on DBSCAN

[0081] The DBSCAN clustering algorithm is used to perform clustering analysis on the standardized equipment data to identify the similarities between equipment;

[0082] The specific process is as follows: DBSCAN divides data clusters through the density threshold and the minimum number of points MinPts, where:

[0083] Definition of density neighborhood: For the feature vector after dimensionality reduction of each equipment, its neighborhood is defined as:

[0084]

[0085] When , then is the core point, and then by connecting the core point and the density-reachable points, the clustering cluster is formed;

[0086] Clustering result: Obtain the class distribution of the equipment, where each cluster represents a class of equipment with similar features.

[0087] Step 4: Construct a multi-level hybrid expert model

[0088] Based on the clustering analysis results, a multi-level hybrid expert model is constructed. Among them, the hybrid expert model includes a basic shared expert and a personalized expert to adapt to the characteristics of different equipment;

[0089] The basic shared expert module is used to capture the common features of each equipment within the same clustering cluster. Through the shared expert model, the common behavioral features of different equipment under similar operating conditions can be effectively modeled, enabling the equipment within the same cluster to share some common feature expressions. For the clustering cluster , train the basic common expert network :

[0090]

[0091] Among them, are the parameters of the shared expert model.

[0092] The personalized expert module is for different devices in the same clustering cluster, and is used to learn the unique features of the devices. For the th device in the clustering cluster, train the personalized expert network :

[0093]

[0094] Among them, represents the dataset of the th device in the cluster, are the parameters of the personalized expert network;

[0095] Fuse the basic shared expert and the personalized expert to form a hybrid expert model, and use the gating function to determine the weights of each model:

[0096]

[0097] The gating function is adaptively adjusted according to the device features or the data volume, and is set as:

[0098]

[0099] Among them, is the Sigmoid function, are the parameters of the gating function.

[0100] Step Five: Model Identification and Optimization of New Devices

[0101] Classify the new device into the corresponding clustering cluster according to the new device feature data, select a combination of the common model and the personalized model, and then perform data fine-tuning to generate a prediction model for the new device;

[0102] When classifying the new device, for the dimensionality-reduced feature vector of the new device, calculate its distance from the center of each clustering cluster:

[0103]

[0104] Assign the new device to the clustering cluster with the minimum distance .

[0105] When identifying a new device, classify the new device into the cluster with the smallest distance. For new devices with limited local data Then use the mixture of experts model for preliminary model identification:

[0106]

[0107] Based on the data of the new device, fine-tune the personalized expert network and the parameters of the gating function to minimize the loss function:

[0108]

[0109] Fine-tune or keep the parameters of the shared expert model unchanged.

[0110] Example 2

[0111] According to Figure 1 As shown, this example proposes a device generative modeling method based on DBSCAN clustering and mixture of experts model (device modeling of the chilled water source group control system in intelligent buildings), including the following steps:

[0112] Step 1: Establish the unified identification and data representation of devices

[0113] Collect the operation data of devices such as chillers, cooling towers, and circulation pumps, including temperature, pressure, flow rate, load rate, and energy consumption, etc., and store them in time series.

[0114] Step 2: Analysis of historical operation data

[0115] Data cleaning, remove outliers through the IQR (Interquartile Range) method, and perform smoothing processing using the moving average method. At the same time, perform Z-score or Min-Max normalization on the data to eliminate scale differences;

[0116] In the feature selection stage, calculate the Pearson correlation coefficient or mutual information (MI) to screen key variables, and apply principal component analysis (PCA) for dimensionality reduction to improve the calculation efficiency and ensure the robustness and accuracy of device state modeling.

[0117] Step 3: Clustering analysis based on DBSCAN

[0118] In the DBSCAN three-dimensional density clustering process, first set the initial search radius and the minimum number of samples in the neighborhood , optimize the parameters by calculating the Silhouette Score, and traverse different and The combination selects the parameters that maximize the silhouette coefficient to ensure the clustering quality. Further, the dynamic search radius method is adopted to adaptively adjust according to the density distribution to improve the adaptability to the operating states of different devices. Based on the optimized parameters, three-dimensional DBSCAN clustering is performed to classify the device states such as chillers, cooling towers, and circulation pumps into categories such as high-load operation, low-load operation, standby, and overload anomalies. For the noise points (isolated points) identified during the clustering process, the Robust Least Squares method is used to correct the data to reduce the impact of abnormal data on the model.

[0119] Step 4: Construct a multi-level hybrid expert model

[0120] In the hybrid expert model, first, a shared-parameter neural network (Shared-Parameter Model) is constructed to train the basic expert model to learn the global common features of the cold source system. This model takes device operation data (temperature, pressure, flow rate, load rate, energy consumption, etc.) as input and extracts the general features of the devices. Then, personalized sub-expert models are trained according to the device categories (chillers, cooling towers, circulation pumps), where:

[0121] The chiller expert fits the relationship between refrigeration capacity and energy consumption;

[0122] The cooling tower expert models in combination with heat dissipation performance and ambient temperature and humidity;

[0123] The circulation pump expert predicts the correlation between flow velocity and power consumption.

[0124] Finally, a gating mechanism is used for model fusion to calculate the expert weights under different device states and perform weighted summation with the output of the expert model to obtain the final prediction result:

[0125]

[0126] where: represents the input features of the device, is the final prediction output (such as future energy consumption, operating state, etc.). This ensures the sharing of global features while providing personalized modeling for different device types, improving the prediction accuracy and generalization ability.

[0127] Step 5: Model identification and optimization of new devices

[0128] When a new device is connected to the system, first extract its feature vector , the vector is composed of the historical and current operation data of the device, including key parameters such as inlet and outlet water temperatures, flow rate, pressure, load rate, energy consumption, control status, etc., and is uniformly normalized in combination with time tags (such as seasons, operation cycles). Subsequently, is projected onto the existing DBSCAN clustering space, and its similarity with each cluster center is measured by Mahalanobis distance or cosine similarity. If the distance is lower than the set threshold , it is classified into the corresponding category; otherwise, it is regarded as a new category, triggering model retraining.

[0129] To improve adaptability, the new device model adopts distillation learning, using the original expert model as the teacher, taking its output probability distribution as the soft label to guide the student model to learn, so as to maintain the global performance unchanged under limited samples. At the same time, combined with the Fine-tuning strategy, only the sub-expert model corresponding to the new device is fine-tuned freezing the shared layer to quickly adapt to its unique operation characteristics. In the online stage, a sliding window is used to dynamically monitor the change of the input distribution, and the weights of the gating network are adjusted in real time ; if it is judged by the LPIPS semantic loss that the feature distribution differs from the existing categories by more than the threshold , it will automatically enter the model update process.

[0130] Chilled water unit prediction and control: Input historical and real-time operation data (temperature, load, flow rate, etc.), use LSTM or Transformer to predict the load in the next 30 minutes, and the optimization goal is to minimize the MSE error. When the predicted load , start the standby equipment in advance. If , reduce the operating power to save energy.

[0131] Example 3

[0132] According to Figure 1 shown, this example proposes a device generative modeling method based on DBSCAN clustering and a mixture of experts model (taking the intelligent status monitoring of industrial equipment as an example), including the following steps:

[0133] Step 1: Establish a unified identification and data representation of the device

[0134] Collect multi-source time series data during the operation of the generator, covering key parameters such as voltage, current, active / reactive power, temperature, vibration, frequency, etc., and the sampling frequency is usually 1Hz to 1kHz.

[0135] Step 2: Analysis of historical operation data

[0136] For outliers and drift noise in the original data, the IQR method or sliding Z-score is used for anomaly detection. The exponential weighted moving average (EWMA) is used to smooth the time series signal, and Z-score or Min-Max normalization is uniformly adopted to eliminate the influence of dimensions. Finally, a standardized multi-dimensional feature vector is constructed , as the input basis for subsequent clustering and model training.

[0137] Step 3: Clustering analysis based on DBSCAN

[0138] Reduce the device operation data to a three-dimensional feature space (using PCA or t-SNE), and perform DBSCAN clustering to identify the generator operation mode. Among them, the clustering radius is initially set empirically according to the amplitude of feature changes (such as the current or vibration change range), and the minimum number of neighborhood samples is usually set to the time length under continuous stable state. Through traversal and optimization of the Silhouette Score and , high-quality clustering labels are obtained , representing categories such as "steady-state operation", "overload", and "vibration anomaly". This clustering result is not only used as the training supervision signal but also for judging the state attribution of new samples.

[0139] Step 4: Build a multi-level hybrid expert model

[0140] Establish a hybrid expert model architecture that includes a shared backbone network and multiple sub-experts. The shared backbone part extracts the general representation of the input feature sequence through multiple layers of LSTM or Transformer encoders , retaining time dynamics and device operation trend information. Each DBSCAN clustering category corresponds to a personalized expert model , whose structure can be LSTM or VAE, and is used for time series prediction or reconstruction modeling respectively. For example, the expert model simulates the steady-state current fluctuation , predicts the temperature rise curve under overload , and fits the vibration spectrum change. The expert output calculates the normalized weight through the gating network to achieve the final prediction:

[0141]

[0142] Among them is adaptively calculated based on the current input features, reflecting the interpretability credibility of each expert for the current state. During model training, the prediction error is minimized, and if it is VAE, the KL divergence term is introduced to form the reconstruction loss.

[0143] Step Five: Model Identification and Optimization of New Equipment

[0144] The hybrid expert structure forms a conditional generation model, that is, given the current operating characteristics , it generates the predicted state or future behavior trajectory of the equipment. The key parameters of the model include: input features (collected from real-time sensor data), clustering labels , expert model outputs (predicted power, power fluctuations, vibration spectra, etc.), gating weights , etc. The generative modeling ability stems from the modeling ability of the expert model for different state distributions and the dynamic control of state transitions by the gating mechanism, thereby achieving high-precision state simulation and prediction of the generator under any state.

[0145] Through the combination of DBSCAN clustering analysis and the hybrid expert model, the present invention can identify and generate equipment models with limited local data, reducing the dependence on a large amount of historical data; by using the clustering analysis results to achieve automatic model selection, modeling can be carried out without the in-depth participation of domain experts, and it is applicable to diverse equipment scenarios; the multi-level hybrid expert model is both general and can provide personalized equipment feature modeling, which helps to improve the adaptability and accuracy of the model; the unified equipment characterization system and multi-level expert model framework make the method have good scalability and can adapt to the modeling needs of different types of equipment. For new equipment, not only can it quickly model based on limited local data, but it can also utilize the common features learned by the equipment in the same cluster to achieve knowledge transfer, efficiently generate the model of the new equipment, and thus significantly reduce the cost of equipment modeling.

[0146] The above has shown and described the basic principles, main features and advantages of the present invention. Those skilled in the art of this industry should understand that the present invention is not limited by the above embodiments. What is described in the above embodiments and the specification only illustrates the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. A device generative modeling method based on DBSCAN clustering and a mixture of experts model, characterized in that It includes the following steps: Step 1: Establish the unified identification and data representation of the device Collect the historical operation data of the device, store it in time series, and then perform unified identification and data representation; Step 2: Analyze the historical operation data Perform data processing on the collected historical operation data. In the feature selection stage, calculate the Pearson correlation coefficient or mutual information to screen key variables, and apply principal component analysis to reduce the dimension of the operation data of the device after data processing to form standardized device data; Step 3: Clustering analysis based on DBSCAN Use the DBSCAN clustering algorithm to perform clustering analysis on the standardized device data to identify the similarity between devices; Step 4: Construct a multi-level hybrid expert model Based on the clustering analysis results, construct a multi-level hybrid expert model, where the hybrid expert model includes a basic shared expert and a personalized expert to adapt to the characteristics of different devices; Step 5: Model identification and optimization of new devices Classify the new device into the corresponding clustering cluster according to the new device feature data, select a common model and a personalized model for combination, and then perform data fine-tuning to generate a prediction model for the new device.

2. The device generative modeling method based on DBSCAN clustering and mixture of experts model according to claim 1, characterized in that: When uniformly identifying and data representing the historical operation data of the device in the first step, let the data of each device be expressed as , where represents the th feature. Through feature standardization and unified processing, the original data of different devices is converted into a unified feature representation , and each standardized feature satisfies: Among them, and respectively represent the mean and standard deviation of the th feature.

3. The device generative modeling method based on DBSCAN clustering and mixture of experts model according to claim 1, characterized in that: Before performing data processing on the collected historical operation data in Step 2, first form a data set from the collected historical operation data: Among them, is the input feature at the t-th time point, is the corresponding output response.

4. The device generative modeling method based on DBSCAN clustering and hybrid expert model according to claim 3, characterized in that: When performing data processing on the collected historical operation data in Step 2, it includes preprocessing such as cleaning, denoising, and filling missing values for the collected device operation data set to ensure data quality.

5. The device generative modeling method based on DBSCAN clustering and mixture of experts model according to claim 4, characterized in that: When calculating the Pearson correlation coefficient or mutual information to screen key variables and applying principal component analysis to reduce the dimension of the operation data of the device after data processing to form standardized device data in the feature selection stage of Step 2, specifically: Extract the key features that mainly affect the device performance and reduce the data dimension: Among them, is the feature transformation matrix, is the feature vector after dimensionality reduction.

6. The device generative modeling method based on DBSCAN clustering and a mixture of experts model according to claim 1, characterized in that: The specific process of the third step is as follows: DBSCAN divides data clusters through a density threshold and the minimum number of points MinPts. Among them, Definition of density neighborhood: For the feature vector after dimensionality reduction of each device , its neighborhood is defined as: When , then is the core point, and then by connecting the core point and the density-reachable points, a clustering cluster is formed ; Clustering result: Obtain the class distribution of the devices , where each cluster represents a class of devices with similar characteristics.

7. The device generative modeling method based on DBSCAN clustering and mixture of experts model according to claim 1, characterized in that: In the fourth step, the basic shared expert module is used to capture the common features of each device within the same clustering cluster. Through the shared expert model, the general behavioral characteristics of different devices under similar operating conditions can be effectively modeled, enabling the devices within the same cluster to share some common feature expressions for the clustering cluster , train the basic commonality expert network : Among them, are the parameters of the shared expert model.

8. The device generative modeling method based on DBSCAN clustering and hybrid expert model according to claim 7, characterized in that: In step 4, the personalized expert module is used to learn the unique features of different devices in the same cluster. For the th device in the cluster, train the personalized expert network as follows: Among them, represents the dataset of the th device in the cluster, which are the parameters of the personalized expert network; Fuse the basic sharing expert and the personalized expert to form a hybrid expert model and use a gating function Determine the weights of each model: The gating function is adaptively adjusted according to the device characteristics or data volume and is set to: Among them, is the Sigmoid function, is the parameter of the gating function.

9. The device generative modeling method based on DBSCAN clustering and mixture expert model according to claim 1, characterized in that: When classifying the new devices in the fifth step, for the feature vectors of the new devices after dimensionality reduction , calculate the distances between them and the centers of each clustering cluster : Assign the new device to the cluster with the minimum distance .

10. The device generative modeling method based on DBSCAN clustering and a mixture of experts model according to claim 9, characterized in that: When identifying new devices in the fifth step, classify the new devices into the clustering cluster with the smallest distance. For new devices with limited local data Then, use the hybrid expert model for preliminary model identification: Based on the data of the new device, for the personalized expert network and the parameters of the gating function perform fine-tuning to minimize the loss function: Share expert model parameters Fine-tune or keep unchanged.

Images