Perception model training method and device, equipment, storage medium and product

Abstract

The invention relates to a perceptual model training method and device, computer equipment, a computer readable storage medium and a computer program product. The method comprises the following steps: receiving local model parameters sent by each sensing node through a central node; determining a first weight of each sensing node based on the index value of each sensing node under the preset dimension; the preset dimension comprises at least one of performance information of a local model corresponding to the sensing node, difference information between local data distribution and global data distribution of local sensing data of the sensing node, signal quality of an area where the sensing node is located and data quality of the local sensing data of the sensing node; and according to the first weight of each sensing node, performing weighted aggregation on the local model parameter corresponding to each sensing node to obtain a global model, so that the global model can integrate the characteristics of each sensing node, thereby improving the precision and generalization ability of the model.

Description

Technical Field

[0001] The present application relates to the field of wireless sensing technology, and in particular to a method, apparatus, computer equipment, computer-readable storage medium and computer program product for training a perception model. Background Art

[0002] Wireless intelligent sensing technology has received extensive attention from academia and industry. Wireless signals are refracted, reflected, and scattered during propagation. The received signals contain a wealth of information about dynamic and static targets in the environment. By analyzing this information, it is possible to detect, identify, locate, and estimate targets.

[0003] At present, in the practical application of wireless intelligent sensing technology, the existing technology usually adopts a relatively unified method to upload data to the server for centralized processing when performing perception tasks, and then builds a global model to realize behavior recognition.

[0004] However, since the existing technology treats the data of each perception node equally when processing data, although it may show certain effects in some simple or single-domain scenarios, it is easy to lead to the problem of reduced generalization ability and stability of the model in multi-domain scenarios. Summary of the invention

[0005] Based on this, it is necessary to provide a training method, device, computer equipment, computer-readable storage medium and computer program product for a perception model that takes into account the differences between perception nodes to improve the generalization ability of the model in order to address the above-mentioned technical problems.

[0006] In a first aspect, the present application provides a method for training a perception model, which is applied to a cross-domain wireless intelligent perception system, wherein the system includes a central node and multiple perception nodes, each perception node being distributed in different areas, and the method includes:

[0007] Receiving local model parameters sent by each sensing node through the central node; the local model parameters are obtained by training each sensing node according to its own local sensing data;

[0008] Determine a first weight of each of the sensing nodes based on an indicator value of each of the sensing nodes under a preset dimension; the preset dimension includes at least one of performance information of a local model corresponding to the sensing node, difference information between a local data distribution and a global data distribution of local sensing data of the sensing node, signal quality of an area where the sensing node is located, and data quality of the local sensing data of the sensing node;

[0009] According to the first weights of the respective sensing nodes, the local model parameters corresponding to the respective sensing nodes are weightedly aggregated to obtain a global model.

[0010] In one embodiment, determining the first weight of each of the sensing nodes based on the index value of each of the sensing nodes under a preset dimension includes:

[0011] Obtain the second weight of each preset dimension;

[0012] For each sensing node, according to the second weight of each preset dimension, weighted aggregation is performed on the indicator value of the sensing node under each preset dimension to obtain a contribution score of the sensing node;

[0013] Based on the contribution score of each sensor node, a first weight of each sensor node is determined.

[0014] In one embodiment, the method further comprises:

[0015] When the perception domain of any perception node changes, obtaining original perception data of the perception node before the perception domain changes and target perception data after the perception domain changes; the perception domain includes at least one of the position of the perception node, the perception environment, and the perception object;

[0016] The features of the target perception data and the original perception data are aligned with each other through a pre-trained feature alignment network to obtain aligned feature data; the aligned feature data has a common feature vector of the target perception data and the original perception data in a common feature space;

[0017] The local model parameters of the perception node after the perception domain changes are updated based on the alignment feature data.

[0018] In one embodiment, the aligning the features of the target perception data and the original perception data through a pre-trained feature alignment network to obtain the aligned feature data includes:

[0019] Inputting the original perception data and the target perception data into a pre-trained feature alignment network;

[0020] Performing feature extraction on the original perception data and the target perception data respectively through the feature extraction module of the feature alignment network to obtain an original feature vector of the original perception data and a target feature vector of the target perception data;

[0021] The common feature vector of the original feature vector and the target feature vector in a common feature space is determined by a generation module of the feature alignment network, and alignment feature data is obtained based on the common feature vector.

[0022] In one of the embodiments, the feature alignment network includes a feature extraction module, a generation module and a discrimination module;

[0023] The feature alignment network is trained by the following method:

[0024] Inputting the first perception data into the feature extraction module for feature extraction to obtain a corresponding first feature vector, and inputting the second perception data into the feature extraction module for feature extraction to obtain a corresponding second feature vector; the first perception data and the second perception data belong to different perception domains;

[0025] Inputting the first feature vector and the second feature vector into the discrimination module to identify the perception domains of the first feature vector and the second feature vector, and performing adversarial training on the discrimination module using a first loss function of the discrimination module until the discrimination module identifies the first feature vector and the second feature vector as the same perception domain;

[0026] The value of the first loss function is passed back to the generation module, and the generation module is adversarially trained using the second loss function of the generation module until the joint loss function converges to obtain a trained feature alignment network; the joint loss function is constructed based on the first loss function and the second loss function according to preset weights.

[0027] In one of the embodiments, after the local model parameters of the perception node after the perception domain changes are updated based on the alignment feature data, the method further includes:

[0028] determining a third weight for each new local model;

[0029] According to the third weights of the new local models, the local model parameters corresponding to the perception nodes are weighted and aggregated to obtain a new global model.

[0030] In one embodiment, the method further comprises:

[0031] When the perception performance of any local model does not meet the preset performance condition, obtaining the local environment characteristics of the local model;

[0032] Inputting the local environment features of the local models into a pre-trained first reinforcement learning model for processing to obtain a local update strategy corresponding to each of the local models;

[0033] The local model is updated based on each of the local update strategies.

[0034] In one embodiment, the first reinforcement learning model is trained by the following method:

[0035] Acquire a first environmental state of the local model; the first environmental state is determined based on the local environmental characteristics;

[0036] Inputting the first environmental state into a pre-built first reinforcement learning model, and determining a first response action of a local model to the first environmental state through the first reinforcement learning model;

[0037] Executing the first response action in the first environmental state; the first response action includes adjusting at least one of a learning rate, a network structure, a migration strategy, and a model parameter of the local model;

[0038] Observe a first result after the first response action is executed;

[0039] The learning parameters of the first reinforcement learning model are updated according to the first result until an end condition is reached to obtain a trained first reinforcement learning model.

[0040] In one embodiment, the method further comprises:

[0041] When the perception domain of any perception node changes or the perception performance of any local model does not meet the preset performance conditions, the state data of the global model and the feedback data received from each perception node are obtained;

[0042] Inputting the state data and feedback data into a pre-trained second reinforcement learning model for processing to obtain a global update strategy for the global model;

[0043] The global model is updated based on the global update strategy.

[0044] In one embodiment, the second reinforcement learning model is trained by the following method:

[0045] Acquire a second environmental state of the global model; the second environmental state is determined based on the state data and the feedback data;

[0046] Inputting the second environmental state into a pre-built second reinforcement learning model, and generating a second response action of the global model to the second environmental state through the second reinforcement learning model;

[0047] Performing the second response action in the second environmental state to update the global model, and determining a reward value based on the model performance of the updated global model and the feedback data of the perception node;

[0048] The state data of the updated global model and the feedback data sent by the perception node based on the updated global model are determined as a new second environment state for iteration until the reward value reaches a preset condition, thereby obtaining a trained second reinforcement learning model.

[0049] In one embodiment, the method further comprises:

[0050] Obtaining perception sample data corresponding to the same perception node under different signal sources;

[0051] Based on the contribution of each signal source to the perception of the perception node, the perception sample data is integrated to obtain local perception data of the perception node;

[0052] Each local model is trained based on the local perception data to obtain a trained local model and local model parameters.

[0053] In one embodiment, the fusing of the perception sample data based on the contribution of each signal source to the perception node in behavior perception includes:

[0054] Extracting data features of each sensing sample data, and determining a fourth weight of each data feature according to the contribution of each signal source to the sensing node during sensing;

[0055] Based on the fourth weight, each data feature is fused to obtain a fused feature vector;

[0056] Based on the attention mechanism, for each data feature in the fused feature vector, determining the attention weight between the data feature and other data features;

[0057] Based on the attention weight, a feature representation of each data feature is determined, and each feature representation is fused to obtain a fused feature representation, and the fused feature representation is determined as the local perception data of the perception node.

[0058] In a second aspect, the present application further provides a training device for a perception model, which is configured with a cross-domain wireless intelligent perception system, the system comprising a central node and a plurality of perception nodes, each perception node being distributed in different areas, the device comprising:

[0059] A parameter receiving module, used to receive local model parameters sent by each sensing node through the central node; the local model parameters are obtained by training each sensing node according to its own local sensing data;

[0060] A weight determination module, configured to determine a first weight of each of the sensing nodes based on an indicator value of each of the sensing nodes under a preset dimension; the preset dimension includes at least one of performance information of a local model corresponding to the sensing node, difference information between a local data distribution and a global data distribution of local sensing data of the sensing node, signal quality of an area where the sensing node is located, and data quality of the local sensing data of the sensing node;

[0061] The model aggregation module is used to perform weighted aggregation on the local model parameters corresponding to each of the perception nodes according to the first weight of each of the perception nodes to obtain a global model.

[0062] In a third aspect, the present application further provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above-mentioned method when executing the computer program.

[0063] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of the method described above are implemented.

[0064] In a fifth aspect, the present application also provides a computer program product, including a computer program, which implements the steps of the method described above when executed by a processor.

[0065] The training method, apparatus, computer equipment, computer-readable storage medium and computer program product of the above-mentioned perception model receive local model parameters sent by each perception node and obtained locally according to their respective local perception data through a central node; determine the first weight of each perception node based on the indicator value of each perception node under a preset dimension; and perform weighted aggregation on the local model parameters corresponding to each perception node according to the first weight of each perception node to obtain a global model. In the process of model fusion, at least one of the performance information of the local model corresponding to each perception node, the difference information between the local data distribution of the local perception data of the perception node and the global data distribution, the signal quality of the area where the perception node is located and the data quality of the local perception data of the perception node is comprehensively considered to make full use of the difference information of each perception node, and perform model training on the local model of each perception node based on its own local perception data, so that the local model can capture the local features and patterns in the local data. By weighted aggregation of the trained local model parameters, the global model can integrate the characteristics of each perception node, thereby improving the accuracy and generalization ability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the drawings required for use in the embodiments of the present application or related technical descriptions will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0067] Figure 1 A diagram of an application environment of a method for training a perception model in one embodiment;

[0068] Figure 2 A flowchart of a method for training a perception model in one embodiment;

[0069] Figure 3 A schematic diagram of a process of updating a local module in one embodiment;

[0070] Figure 4 A schematic diagram of a process for performing data alignment in one embodiment;

[0071] Figure 5 A schematic diagram of a process of training a feature alignment network in one embodiment;

[0072] Figure 6 is a schematic diagram of a process of training a first reinforcement learning model in one embodiment;

[0073] Figure 7 is a schematic diagram of a process of training a second reinforcement learning model in one embodiment;

[0074] Figure 8 is a schematic diagram of a process of fusing perception sample data in one embodiment;

[0075] Fig. 9 is a structural block diagram of a training device for a perception model in one embodiment;

[0076] Fig.10 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0077] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0078] The training method of the perception model provided in the embodiment of the present application can be applied to Figure 1In the application environment shown. Among them, the cross-domain wireless intelligent perception system 102 communicates with the server 104 through the network. The data storage system can store the data that the server 104 needs to process. The data storage system can be integrated on the server 104, or it can be placed on the cloud or other network servers. Among them, the cross-domain wireless intelligent perception system 102 includes a central node and multiple perception nodes, and each perception node may include but is not limited to cameras, infrared sensors, radio frequency devices, positioning devices, etc. distributed in different areas. The server 104 can be an independent physical server, or it can be a server cluster or distributed system composed of multiple physical servers, or it can be a cloud server that provides cloud computing services.

[0079] The cross-domain wireless intelligent perception system 102 refers to a system that can use wireless signals to perceive phenomena and behaviors, so as to realize intelligent perception functions in different perception domains (such as different geographical areas, different signal propagation conditions, different perception objects, etc.). The cross-domain wireless intelligent perception system is configured with a central node and multiple perception nodes, and the multiple perception nodes are configured in different perception domains to perceive wireless signals across different geographical or logical areas.

[0080] Among them, the central node refers to the node in the entire cross-domain wireless intelligent perception system 102 that is responsible for receiving data uploaded from each perception node (in this embodiment, it mainly refers to the local model parameters, feedback data, etc. of the perception node) to process, analyze and make decisions on the received data.

[0081] Among them, perception nodes refer to basic units deployed in different areas and used to collect local perception data; each perception node has the ability to independently collect data and train models based on local data.

[0082] In one embodiment, a method for training a perception model is provided, and the method is applied to Figure 1 The cross-domain wireless intelligent sensing system 102 in FIG. 1 is used as an example to illustrate. It is understandable that the method can also be applied to Figure 1 The server 104 in the embodiment can also be applied to a system including the cross-domain wireless intelligent perception system 102 and the server 104 through a network, and can be realized by the interaction between the cross-domain wireless intelligent perception system 102 and the server 104 through a network.

[0083] The perception model refers to a model that identifies, classifies or predicts specific phenomena or behaviors by learning and analyzing perception data. The perception model is configured to include multiple local models and a global model. The local models are configured in the perception nodes one by one, and the global model is configured in the central node.

[0084] The local model refers to a model that extracts features from the local perception data collected by the perception node to identify characteristic phenomena or behaviors. The local model can be constructed using a feature extraction algorithm, such as at least one of the SIFT algorithm (Scale-Invariant Feature Transform), convolutional neural network, principal component analysis method, edge detection algorithm, etc. In specific implementation, the local models deployed in different perception nodes can be constructed using different feature extraction algorithms according to the perception domain of the perception node.

[0085] Among them, the global model refers to the model jointly constructed according to the local model parameters of each local model. The global model can integrate the information of each perception node to improve the overall performance and robustness of the system.

[0086] like Figure 2 As shown, the method of this embodiment includes the following steps 201 to 203. Among them:

[0087] Step 201: receiving local model parameters sent by each sensing node through the central node.

[0088] Among them, local model parameters refer to the model parameters obtained by each perception node after model training based on the local perception data collected by itself. The local model parameters of different perception nodes vary due to the differences in their local perception data; for example, in a neural network model, local model parameters may include weight matrices and bias vectors of each layer, and in an edge detection model, local model parameters may include gradient threshold parameters, etc.

[0089] Among them, local perception data refers to the data obtained by each perception node through sensing the area where it is located. The local perception data is obtained by the perception node collecting wireless signals locally. The wireless signals include but are not limited to Wi-Fi, ultra-wideband, millimeter wave, terahertz, Bluetooth, FMCW (Frequency Modulated Continuous Wave), visible light, etc.

[0090] Exemplarily, after the local model deployed on each perception node is trained, the trained local model parameters are sent to the central node, and the local model parameters sent by each perception node are received through the central node.

[0091] Step 202: Determine a first weight of each sensing node based on the index value of each sensing node under a preset dimension.

[0092] The preset dimension refers to a measurement dimension used to determine the importance of the perception node in the global model aggregation, and the preset dimension can reflect the characteristics of the perception node and its data from different aspects. The preset dimension includes at least one of the performance information of the local model corresponding to the perception node, the difference information between the local data distribution of the local perception data of the perception node and the global data distribution, the signal quality of the area where the perception node is located, and the data quality of the local perception data of the perception node.

[0093] Among them, the performance information of the local model corresponding to the sensing node refers to the information used to measure the sensing performance of the local model obtained by local training of each sensing node in the wireless signal in the area where it is located; the better the performance of the local model, the more important it is for the construction of the global model. The performance information of the local model includes but is not limited to accuracy, loss value, recall rate, mean square error, etc.; for example, the accuracy is used to measure the prediction correctness of the local model on the local sensing data in the area where the sensing node to which it belongs is located, and can be determined by calculating the classification accuracy of the local model on the corresponding local sensing data. The higher the accuracy, the better the performance of the local model; for another example, the loss value can be calculated using a loss function. The lower the loss value, the better the performance of the local model.

[0094] Among them, the difference information between the local data distribution of the local perception data of the perception node and the global data distribution refers to the information used to measure the degree of difference between the data distribution of the local perception data of a certain perception node and the global data distribution composed of the local perception data of all perception nodes; the local perception data with a smaller difference between the local data distribution and the global data distribution is more likely to represent the global data situation, and therefore, is more valuable to the construction of the global model. The difference information between the local data distribution and the global data distribution can be determined by including but not limited to the maximum mean difference, data balance, etc.; for example, when the maximum mean difference is used, the maximum mean difference value between the local data distribution of a certain perception node and the global data distribution is calculated. If the maximum mean difference value is smaller, it means that the local data distribution of the perception node is closer to the global data distribution, the corresponding local model has better generalization performance, and the contribution of the perception node to the global model is also better, otherwise, the contribution is lower.

[0095] Among them, the signal quality of the area where the sensing node is located refers to the quality of the wireless signal in the area where the sensing node is located. The better the signal quality, the more reliable the local sensing data collected by the sensing node, and correspondingly, the higher the contribution to the global model. The signal quality of the area where the sensing node is located can be determined based on factors including but not limited to the signal-to-noise ratio, signal volatility, signal strength, etc.; for example, the higher the signal-to-noise ratio of the area where a sensing node is located, the better the quality of the local sensing data collected by the sensing node; for another example, by calculating the signal volatility (such as variance, smoothness, etc.) of the area where the sensing node is located, the stability of the data collected by the sensing node in a dynamic environment can be evaluated. The higher the stability of the sensing node, the better the quality of the local sensing data collected.

[0096] Among them, the data quality of the local perception data of the perception node refers to the quality of the local perception data. High-quality local perception data is conducive to improving the accuracy of local model training, thereby having a positive impact on the construction of the global model.

[0097] The index value refers to the specific numerical value or quantitative performance corresponding to each perception node under each preset dimension. For example, in the performance information of the local model corresponding to the perception node, the index value is the accuracy, loss value, recall rate, mean square error, etc. of the model.

[0098] Among them, the first weight refers to the weight value determined for each perception node according to the indicator value under the preset dimension. The first weight is used to reflect the relative importance of each perception node in the global model construction process. The larger the first weight of a perception node, the more important the perception node is in the global model construction process, so as to facilitate the subsequent weighted aggregation of local model parameters, highlight the role of perception nodes that contribute more to the global model, and appropriately reduce the influence of nodes with smaller contributions, so as to comprehensively consider the differences between each perception node in the process of building the global model and improve the generalization performance of the global model.

[0099] Exemplarily, the first weight of each perception node is determined by the central node based on the indicator value of each perception node under a preset dimension.

[0100] Step 203: weighted aggregation is performed on the local model parameters corresponding to each sensing node according to the first weight of each sensing node to obtain a global model.

[0101] Among them, weighted aggregation refers to the process of fusing the local model parameters of multiple perception nodes according to the first weight of the perception node. During the fusion process, the corresponding local model parameters are weighted calculated according to the first weight of each perception node, and then the weighted parameters are summarized. The summarization method includes but is not limited to screening, splicing, merging, etc., so as to obtain a global model to make full use of the information of each perception node and improve the generalization ability and stability of the global model.

[0102] Exemplarily, the central node performs weighted aggregation on the local model parameters corresponding to each perception node according to the first weights corresponding to the perception nodes to obtain a global model.

[0103] Among them, the global model parameters of the global model are expressed as:

[0104] (1)

[0105] in, are global model parameters, is the local model parameter of the i-th sensor node, , is the number of sensing nodes, is the first weight of the i-th sensor node.

[0106] In the above-mentioned training method of the perception model, the local model parameters sent by each perception node and obtained by local training according to their respective local perception data are received by the central node; the first weight of each perception node is determined based on the indicator value of each perception node under the preset dimension; the local model parameters corresponding to each perception node are weighted and aggregated according to the first weight of each perception node to obtain a global model. In the process of model fusion, at least one of the performance information of the local model corresponding to each perception node, the difference information between the local data distribution of the local perception data of the perception node and the global data distribution, the signal quality of the area where the perception node is located, and the data quality of the local perception data of the perception node are comprehensively considered to make full use of the difference information of each perception node, and the local model of each perception node is trained based on its own local perception data, so that the local model can capture the local features and patterns in the local data. By weighted aggregation of the trained local model parameters, the global model can integrate the characteristics of each perception node, thereby improving the accuracy and generalization ability of the model.

[0107] In an exemplary embodiment, determining the first weight of each sensor node includes:

[0108] Obtain the second weight of each preset dimension; for each perception node, perform weighted aggregation on the indicator value of the perception node under each preset dimension according to the second weight of each preset dimension to obtain the contribution score of the perception node; determine the first weight of each perception node based on the contribution score of each perception node.

[0109] The second weight refers to the weight value assigned to each preset dimension, which is used to measure the relative importance of each preset dimension in evaluating the contribution of the sensing node. The second weight can be determined based on experience or multiple experiments during the model training process.

[0110] Among them, weighted aggregation of the indicator values ​​according to the second weight refers to the process of multiplying the indicator values ​​of each sensing node under each preset dimension by the second weight corresponding to the preset dimension, and then accumulating the product results. By weighted aggregation of the indicator values, different preset dimensions and their importance can be comprehensively considered to obtain a value that can reflect the comprehensive performance of the sensing node under all preset dimensions, thereby obtaining a contribution score.

[0111] Among them, the contribution score refers to the score obtained by weighted aggregation of the indicator values ​​under each preset dimension. The contribution score is used to indicate the contribution of the perception node to the global model, so as to measure the performance of the perception node under comprehensive evaluation. If a perception node performs well in all dimensions, then the contribution score of the perception node will be higher, which means that the perception node is more important to the construction of the global model under the comprehensive evaluation of various preset dimensions; conversely, if a perception node performs poorly in some dimensions, then the contribution score of the perception node will be lower, which means that the perception node is less important to the construction of the global model under the comprehensive evaluation of various preset dimensions.

[0112] Exemplarily, the second weight of each preset dimension is obtained through the central node; according to the second weight of each preset dimension, the indicator value of each perception node under each preset dimension is weighted and aggregated to obtain the contribution score of each perception node.

[0113] Among them, the contribution score of the perception node is expressed as:

[0114] (2)

[0115] in, Score the contribution of the i-th sensor node, is the accuracy score of the i-th sensor node, is the loss value of the i-th perception node, is the maximum mean difference between the local data distribution and the global data distribution of the i-th sensing node, is the signal-to-noise ratio of the area where the i-th sensing node is located, is the signal volatility of the i-th sensing node, The accuracy score is , loss value , maximum mean difference , signal-to-noise ratio and signal volatility The corresponding second weight.

[0116] It is understandable that in some other embodiments, when determining the contribution score, it is not limited to using the above-mentioned indicator value and its corresponding second weight for determination, and the contribution score can also be determined by increasing or decreasing the indicator value according to actual needs.

[0117] Subsequently, the system normalizes the contribution scores based on the contribution scores of each perception node to obtain a normalized contribution score.

[0118] Among them, the normalized contribution score is expressed as:

[0119] (3)

[0120] in, is the normalized contribution score of the i-th sensor node, , is the number of sensing nodes.

[0121] Finally, for sensing nodes with poor signal quality, their weights can be further adjusted according to the signal-to-noise ratio and signal volatility to determine the first weight of each sensing node. For example, if the signal quality of a sensing node is unstable, its first weight can be appropriately reduced; conversely, the first weight of the sensing node with good signal quality can be increased. At the same time, based on the feedback of the local environment and the global environment, the weight of each sensing node can also be dynamically adjusted. For example, through the reinforcement learning algorithm, the first weight of the sensing node can be adaptively adjusted according to environmental changes and real-time data feedback. The first weight of each sensing node in the next round of iteration will be adjusted according to the contribution score and model performance feedback of the previous round.

[0122] Among them, the first weight of the perception node is expressed as:

[0123] (4)

[0124] in, is the first weight of the i-th sensor node, is a weighted function, which represents the normalized contribution score of the i-th sensor node , signal-to-noise ratio and signal volatility Perform weighted calculation to obtain the first weight of the i-th sensor node.

[0125] In this embodiment, the second weight of each preset dimension is obtained; for each perception node, the indicator value of the perception node under each preset dimension is weighted and aggregated according to the second weight of each preset dimension to obtain a contribution score of the perception node; based on the contribution score of each perception node, the first weight of each perception node is determined, and the second weight of each preset dimension is comprehensively considered to perform weighted aggregation on the indicator value of the perception node. This can more accurately evaluate the contribution of the perception node, and determine the first weight of each perception node accordingly, which is conducive to further considering the performance of each perception node, thereby improving the accuracy and effectiveness of the evaluation.

[0126] In one embodiment, Figure 3 As shown, the above method also includes:

[0127] Step 301: when the perception domain of any perception node changes, original perception data of the perception node before the perception domain changes and target perception data after the perception domain changes are obtained.

[0128] The perception domain refers to the range and conditions perceived by the perception node, including but not limited to at least one of the location of the perception node, the perception environment in which the perception node is located, and the perception object targeted by the perception node. When the perception domain of a perception node changes, it may cause changes in factors such as the received signal strength, propagation path, and surrounding environmental interference, which in turn affects the characteristics of the perception data. For example, if a perception node is moved from indoors to outdoors, or moved between different spaces indoors, the data obtained by the perception node will vary depending on the location. Therefore, it is necessary to perform adaptive training on the local model of the perception node after the perception domain changes to ensure that the local model can continue to perceive accurately after the perception changes.

[0129] Among them, the sensing node location refers to the installation location of the sensing node. The sensing environment refers to the environmental conditions around the sensing node, including but not limited to channel conditions (such as whether there is multipath fading, noise level, etc.), environmental materials (such as different materials such as metal and wood have different reflection and refraction characteristics of signals) and environmental layout (such as space size, obstacle distribution, etc.). The sensing object refers to the target object or object monitored or identified by the sensing node. For example, in the human behavior perception scenario, the sensing object may be different human movements; for another example, in the target detection scenario, the sensing object may be different types of objects.

[0130] The original perception data refers to the local perception data collected by the perception node in the original perception domain before the perception domain changes. The target perception data refers to the local perception data collected by the perception node in the target perception domain after the perception domain changes. For example, when the perception node moves from the living room to the corridor, the position of the perception node changes. At this time, the original perception data is the local perception data perceived by the perception node in the living room, and the target perception data is the local perception data perceived by the perception node in the corridor.

[0131] Exemplarily, when the system monitors that the perception domain of any perception node has changed, the system obtains the original perception data of the perception node before the perception domain has changed and the target perception data after the perception domain has changed.

[0132] Step 302: align the features of the target perception data and the original perception data through a pre-trained feature alignment network to obtain aligned feature data.

[0133] The pre-trained feature alignment network refers to a feature alignment network that has been trained in advance, which is used to map the features of the original perception data and the target perception data to a common feature space so that the two have similar feature representations in the common feature, thereby achieving feature alignment. In specific implementation, the feature alignment network can be implemented using one or more of, including but not limited to, convolutional neural networks, adversarial training networks, nearest neighbor interpolation methods, etc.

[0134] Among them, aligning the features of target perception data and original perception data refers to the process of using a pre-trained feature alignment network to process the target perception data and the original perception data so that the two tend to be consistent at the feature level, so as to extract the common feature vectors of the target perception data and the original perception data in the common feature space, thereby reducing the impact of perception domain changes on data features and providing a more consistent feature basis for subsequent model parameter updates. Usually, since the perception node has less target perception data in the target perception domain, by aligning the target perception data with the original perception data at the feature level, it is possible to use a small amount of target perception data to perform migration training on the local model to update the local model and realize cross-domain migration perception of the local model.

[0135] Among them, the aligned feature data refers to the aligned perception data obtained based on the common feature vector, which is the result representation after aligning the original perception data and the target perception data at the feature level. The aligned feature data has the common feature vector of the target perception data and the original perception data in the common feature space, so as to eliminate or reduce the difference between the perception data perceived in different perception domains, thereby providing a more stable and universal feature basis for the subsequent update of the local model parameters of the perception node, which helps the model adapt to the changes in the perception domain.

[0136] Exemplarily, the system inputs the target perception data and the original perception data into a pre-trained feature alignment network, extracts features of the target perception data and the original perception data through the pre-trained feature alignment network, and aligns the features of the two sets of data extracted to obtain aligned feature data.

[0137] Step 303: Update the local model parameters of the perception node after the perception domain changes based on the alignment feature data.

[0138] Among them, updating the local model parameters of the perception node after the perception domain changes refers to the process of adjusting or optimizing the local model parameters of the perception node after the perception domain changes based on the alignment feature data. By updating the local model parameters, the performance of the perception node after the perception domain changes can be improved, so that it can better adapt to the new perception environment and perception objects.

[0139] Exemplarily, after the target perception data is aligned with the original perception data, the local model parameters of the perception node after the perception domain is changed are updated based on the obtained alignment feature data.

[0140] In this embodiment, when the perception domain of any perception node changes, the original perception data of the perception node before the perception domain changes and the target perception data after the perception domain changes are obtained, and the features of the target perception data and the original perception data are aligned through a pre-trained feature alignment network to obtain aligned feature data. Finally, the local model parameters of the perception node after the perception domain changes are updated based on the aligned feature data to ensure that the local model can adapt to the changes in the perception domain and improve the adaptability and accuracy of the local model.

[0141] In an exemplary embodiment, Figure 4 As shown, the features of the target perception data and the original perception data are aligned through the pre-trained feature alignment network to obtain aligned feature data, including:

[0142] Step 401: input the original perception data and the target perception data into a pre-trained feature alignment network.

[0143] Among them, the feature alignment network includes a feature extraction module, a generation module and a discrimination module.

[0144] Among them, the feature extraction module is used to extract representative features from the input original perception data and target perception data, and convert them into vector representations to obtain the original feature vectors corresponding to the original perception data and the target feature vectors corresponding to the target perception data. The feature extraction module can adopt one or more implementations including but not limited to convolutional neural networks and recurrent neural networks to convert data into high-dimensional vector representations to obtain corresponding feature vectors. In specific implementations, for different types of feature extraction, the feature extraction module can use different methods to extract features. For example, for time-frequency domain features, such as signal strength, arrival time and other information, a convolutional neural network can be used to extract local patterns and spatial features of the signal; for time series features, a recurrent neural network can be used to extract time series dependencies in a multi-path propagation or dynamically changing environment.

[0145] Among them, the generation module is used to map feature vectors from different data sources to a common feature space and generate a common feature vector, that is, to map the original feature vector of the original perception data and the target feature vector of the target perception data to the common feature space to obtain the common feature vector of the two in the common feature space. The generation module can find features that can represent the commonality between the original perception data and the target perception data by learning the relationship between different original feature vectors and target feature vectors, thereby achieving feature alignment.

[0146] Among them, the discriminant module is used to determine which perceptual domain the input feature vector belongs to, that is, to distinguish whether the common feature vector is the original feature vector or the target feature vector. During the training process, the discriminant module can receive feature vectors from different perceptual domains processed by the feature extraction module, and try to accurately identify the source perceptual domain of the feature vector. Through adversarial training with the generation module, the generation module is prompted to generate feature vectors that are more difficult to distinguish perceptual domains, thereby aligning the features of the target perceptual data with the original perceptual data.

[0147] Exemplarily, the system inputs the original perception data perceived by the perception node in the original perception domain and the target perception data perceived in the target perception domain into a pre-trained feature alignment network.

[0148] Step 402 , extracting features from the original perception data and the target perception data respectively through a feature extraction module of a feature alignment network to obtain an original feature vector of the original perception data and a target feature vector of the target perception data.

[0149] Among them, the original feature vector refers to the vector representation obtained after the feature extraction module processes the original perception data; the original feature vector contains the key information of the original perception data, which is used in the subsequent data alignment task. Similarly, the target feature vector refers to the vector generated by the feature extraction module after processing the target perception data; the target feature vector contains the key information of the target perception data, which is used in the subsequent data alignment task.

[0150] Exemplarily, the system obtains an original feature vector corresponding to the original perception data by inputting the original perception data into the feature extraction module of the feature alignment network for feature extraction, and obtains a target feature vector corresponding to the target perception data by inputting the target perception data into the feature extraction module of the feature alignment network for feature extraction.

[0151] Step 403: determine the common feature vector of the original feature vector and the target feature vector in the common feature space through the generation module of the feature alignment network, and obtain alignment feature data based on the common feature vector.

[0152] The common feature space refers to a unified feature representation space, in which different perception sources or different perception data can be mapped to the common feature space and compared, analyzed and processed in the common feature space.

[0153] Among them, the common feature vector refers to a vector that can represent the common features of the original perception data and the target perception data in the common feature space, and its feature representation is obtained by mapping the original feature vector and the target feature vector to the common feature space through the generation module.

[0154] Exemplarily, the system inputs the original feature vector and target feature vector extracted by the feature extraction module into the generation module of the feature alignment network, and maps the original feature vector and target feature vector to the common feature space through the generation module to obtain the common feature vector of the original feature vector and the target feature vector in the common feature space. Subsequently, the system obtains the alignment feature data based on the common feature vector.

[0155] In this embodiment, the original perception data and the target perception data are input into a pre-trained feature alignment network, and the feature extraction module of the feature alignment network is used to extract features from the original perception data and the target perception data respectively, so as to obtain the original feature vector of the original perception data and the target feature vector of the target perception data. Subsequently, the common feature vector of the original feature vector and the target feature vector in the common feature space is determined through the generation module of the feature alignment network, and the aligned feature data is obtained based on the common feature vector. This is conducive to ensuring the consistency and comparability of the perception data in different perception domains at the feature level, and thus providing a reliable and effective feature basis for subsequent data analysis, pattern recognition or decision support applications.

[0156] In an exemplary embodiment, Figure 5 As shown in Figure 2, the feature alignment network is trained by the following method:

[0157] Step 501, input the first perception data into the feature extraction module for feature extraction to obtain the corresponding first feature vector, and input the second perception data into the feature extraction module for feature extraction to obtain the corresponding second feature vector.

[0158] The first perception data and the second perception data belong to different perception domains and are used as inputs of a feature alignment network to perform adversarial training on the feature alignment network.

[0159] Among them, the first feature vector refers to the vector representation obtained after the feature extraction module processes the first perception data; the first feature vector contains the key information of the first perception data, which is used in the subsequent data alignment task. Similarly, the second feature vector refers to the vector generated by the feature extraction module after processing the second perception data; the second feature vector contains the key information of the second perception data, which is used in the subsequent data alignment task.

[0160] Exemplarily, the system inputs the first perception data into the feature extraction module for feature extraction to obtain a first feature vector corresponding to the first perception data, and then inputs the second perception data into the feature extraction module for feature extraction to obtain a second feature vector corresponding to the second perception data.

[0161] Step 502: Input the first eigenvector and the second eigenvector into a discriminant module to identify the perception domains of the first eigenvector and the second eigenvector, and perform adversarial training on the discriminant module using a first loss function of the discriminant module until the discriminant module identifies the first eigenvector and the second eigenvector as the same perception domain.

[0162] The first loss function refers to a function used to measure the performance of the discriminant module in distinguishing the first feature vector and the second feature vector from different perceptual domains, so as to reflect the accuracy of the discriminant module in identifying the perceptual domains of the first feature vector and the second feature vector. During the training process, the goal of the discriminant module is to minimize the first loss function to improve the ability to recognize feature vectors in different perceptual domains.

[0163] Exemplarily, the system inputs the first eigenvector and the second eigenvector into a discrimination module to identify the perceptual domains of the first eigenvector and the second eigenvector to distinguish which perceptual domain the first eigenvector and the second eigenvector come from, and calculates the value of the first loss function through the discrimination module's judgment result of the perceptual domain of the first eigenvector and the judgment result of the perceptual domain of the second eigenvector, until the discrimination module identifies the first eigenvector and the second eigenvector as the same perceptual domain, that is, the value of the first loss function tends to converge.

[0164] Among them, the first loss function is defined as:

[0165] (5)

[0166] in, represents the first loss function, is the judgment result of the discriminant module on the perception domain (original perception domain) of the first feature vector, is the first eigenvector, is the judgment result of the discriminant module on the perception domain (target perception domain) of the second feature vector, is the second eigenvector, is the expected value.

[0167] Step 503, the value of the first loss function is reversely transferred to the generation module, and the generation module is adversarially trained using the second loss function of the generation module until the joint loss function converges, thereby obtaining a trained feature alignment network.

[0168] The second loss function refers to a function used to measure the performance of the generation module when aligning the first feature vector and the second feature vector from different perceptual domains. During the training process, the generation module will be updated and optimized according to the value of the second loss function and the feedback from the discrimination module. By minimizing the second loss function, the feature vector generated by the generation module can not only reflect the commonality of data from different perceptual domains, but also confuse the discrimination module, making it difficult for the discrimination module to distinguish its source perceptual domain.

[0169] Among them, the joint loss function is constructed based on the first loss function and the second loss function according to the preset weights. The joint loss function comprehensively considers the training objectives of the discriminant module and the generative module. By adjusting the preset weights, the importance of the two modules in the training process can be balanced. During the training process, the goal of the entire feature alignment network is to minimize the joint loss function. When the joint loss function converges, it means that the feature alignment network has reached a relatively stable and effective state to improve its ability to align cross-domain features.

[0170] Exemplarily, the system transmits the value of the first loss function back to the generation module, so that the generation module is adjusted according to the value of the first loss function and generates a new common feature vector of the first feature vector and the second feature vector. Subsequently, the new common feature vector is sent to the judgment module for judgment, and this iteration is performed until the joint loss function converges to obtain a trained feature alignment network.

[0171] Among them, the second loss function is implemented using the maximum mean difference loss function, and the second loss function is expressed as:

[0172] (6)

[0173] in, is the first loss function, is the high-dimensional feature map of the first eigenvector, is the high-dimensional feature map of the second eigenvector.

[0174] Among them, the joint loss function is expressed as:

[0175] (7)

[0176] in, is the joint loss function, It is a hyperparameter used to balance the impact of the first loss function and the second loss function of adversarial training on the final training results.

[0177] In this embodiment, by performing feature extraction, discriminant adversarial and generative adversarial training on the perception data, it is beneficial to achieve feature alignment of the first perception data and the second perception data, so that the two are identified as the same perception domain at the feature level. The entire process continuously optimizes the feature extraction, discrimination and generation modules until the joint loss function converges, and finally obtains the trained feature alignment network, which can effectively improve the consistency and comparability of data across perception domains, and provide a more accurate and reliable foundation for subsequent data processing or analysis.

[0178] In an exemplary embodiment, after the local model parameters of the perception node after the perception domain changes are updated based on the alignment feature data in step S403, the method further includes:

[0179] Determine the third weight of each new local model; and perform weighted aggregation on the local model parameters corresponding to each perception node according to the third weight of each new local model to obtain a new global model.

[0180] The new local model refers to a model obtained by updating the local model after the perception domain of the local model changes.

[0181] Among them, the third weight is similar to the first weight, and is used to reflect the relative importance of each perception node in the process of constructing a new global model. For details, please refer to the records in the above embodiment, and this embodiment will not be repeated.

[0182] Among them, the new global model refers to a new model generated by weighted aggregation of new local model parameters corresponding to each perception node, so as to better cope with the differences in perception node data in different regions and achieve more stable and accurate target recognition or environmental monitoring and other functions.

[0183] Exemplarily, after the local model is updated, the system re-determines the third weight of each new local model based on the updated local model, and performs weighted aggregation on the local model parameters corresponding to each perception node based on the third weight of each new local model to obtain a new global model.

[0184] In this embodiment, by determining the weight of each new local model and performing weighted aggregation on the local model parameters of each perception node, the model information distributed on different perception nodes can be efficiently integrated to obtain a more comprehensive and accurate new global model to enhance the cross-domain perception capability of the model.

[0185] In an exemplary embodiment, the method further includes:

[0186] When the perception performance of any local model does not meet the preset performance conditions, the local environment characteristics of the local model are obtained; the local environment characteristics of the local model are input into the pre-trained first reinforcement learning model for processing to obtain the local update strategy corresponding to each local model; and the local model is updated based on each local update strategy.

[0187] The perceptual performance of the local model refers to the performance or ability of the local model when processing and analyzing its own local perceptual data. The perceptual performance of the local model includes but is not limited to one or more of the accuracy, perceptual speed, and robustness of the model. The preset performance condition refers to the perceptual performance standard or threshold set for the local model, which is used to determine whether the perceptual performance of the local model has decreased.

[0188] Among them, local environmental characteristics refer to various characteristics of the environment in which the perception node is located that are related to the local model. Local environmental characteristics include but are not limited to the geographical location of the perception node, the physical properties of the surrounding environment (such as temperature, humidity, material, etc.), wireless signal-related characteristics (such as signal strength, signal-to-noise ratio, channel fading, etc.), and the characteristics of the perceived object (such as the number of objects, moving speed, category, etc.).

[0189] Among them, the pre-trained first reinforcement learning model refers to a reinforcement learning model that is trained in advance and is used to process the local environment features of the local model and generate the corresponding local update strategy. The first reinforcement learning model learns how to make the best decision by interacting with the environment. The first reinforcement learning model receives the environment state as input, selects and executes actions, and then receives rewards or penalties as feedback. Through continuous trial and error and learning, the first reinforcement learning model gradually learns how to maximize long-term rewards so that it can generate an update strategy suitable for the local model for a given local environment state, thereby improving the performance of the local model in the corresponding environment. In specific implementation, the first reinforcement learning model can be implemented using one or more algorithms including but not limited to Q-Learning (i.e., Quaternary Learning Algorithm or Q Learning Algorithm) or DQN (Deep Q-Network, i.e., Deep Q Network) algorithms.

[0190] The local update strategy refers to a strategy generated by the pre-trained first reinforcement learning model according to the local environment characteristics and used to guide the local model to update. The local update strategy may include but is not limited to adjusting the parameters of the model (such as weights and biases in the neural network), changing the structure of the model (such as increasing or decreasing the number of layers and nodes of the neural network), adjusting the learning rate and other operations to update the local model in a targeted manner, so that the local model can better adapt to the changed local environment, thereby improving the perception performance and meeting the preset performance conditions.

[0191] Among them, updating the local model based on the local update strategy means adjusting and optimizing the local model accordingly according to the local update strategy output by the pre-trained first reinforcement learning model. For example, if the local update strategy recommends increasing the number of nodes in a certain layer of the neural network, then the number of nodes in the corresponding neural network layer of the local model is increased; if the strategy requires adjusting the learning rate, the learning rate parameter during model training is modified in a specified manner, so that the local model can show better perception performance in the current local environment.

[0192] Exemplarily, when the system detects that the perception performance of any local model does not meet the preset performance conditions, the system obtains the local environment characteristics of the local model through the perception node; and inputs the local environment characteristics of the local model into the pre-trained first reinforcement learning model, generates a corresponding local update strategy according to the local environment characteristics through the first reinforcement learning model, and updates each local model under the guidance of the local update strategy to obtain an updated local model.

[0193] In this embodiment, when the perceptual performance of the local model does not meet the standard, by extracting its local environment features and inputting them into the pre-trained first reinforcement learning model, a local update strategy for each local model can be intelligently generated. The local model is updated based on the local update strategy, which can effectively improve the perceptual performance of the model and make it more in line with the preset conditions. It is beneficial to enhance the pertinence and flexibility of the local model update, improve the ability of the local model to adapt to different environments, and at the same time utilize the advantages of reinforcement learning to achieve continuous optimization of model performance.

[0194] In an exemplary embodiment, Figure 6 As shown, the first reinforcement learning model is trained using the following method:

[0195] Step 601, obtaining a first environment state of a local model.

[0196] The first environment state refers to the environment condition or context information of the local model, and the first environment state is an abstract representation of the environment condition of the local model based on the local environment characteristics. In specific implementation, the first environment state includes but is not limited to the channel state (such as signal-to-noise ratio, channel fading, etc.), wireless signal characteristic distribution (such as signal strength, arrival time, multipath components, etc.), user behavior pattern (such as user movement speed, direction, residence time, etc.) and interference level (such as background noise of wireless signals, interference of neighboring signals, etc.).

[0197] Step 602: Input the first environment state into a pre-built first reinforcement learning model, and determine a first response action of the local model to the first environment state through the first reinforcement learning model.

[0198] The first response action refers to the action selected by the first reinforcement learning model for the local model based on the first environment state in the local environment state, which is used to adjust the local model parameters. The first response action includes but is not limited to adjusting one or more of the learning rate, network structure, migration strategy, and model parameters of the local model. By executing the first response action, it is expected that the local model can better adapt to the environment in which it is located to optimize the perception performance of the local model.

[0199] Exemplarily, the system inputs the first environmental state into a pre-built first reinforcement learning model, and determines a first response action of the local model to the first environmental state through the first reinforcement learning model.

[0200] Step 603: Execute a first response action in a first environmental state.

[0201] Step 604: observe a first result after the first response action is executed.

[0202] Among them, the first result refers to the result of the local model observed after the first response action is executed, which is used to evaluate the impact of the first response action on the performance of the local model. The first result can be measured by a variety of indicators, such as the changes in performance indicators such as perception accuracy, recall rate, mean square error, etc. of the local model after the first response action is executed, or the stability and efficiency of the local model when processing new perception data. The first result can be used as a feedback signal of the first reinforcement learning model to feedback the pros and cons of the current first response action to the first reinforcement learning model so that the model can adjust its own parameters.

[0203] Exemplarily, after executing the first response action, the first result after executing the first response action is obtained by observing changes in the performance of the local model and other aspects.

[0204] Step 605: Update the learning parameters of the first reinforcement learning model according to the first result until the end condition is met to obtain a trained first reinforcement learning model.

[0205] Among them, the learning parameters refer to the parameters of the first reinforcement learning model itself. The learning parameters can be adjusted during the training process by observing the first results and updating rules, and are used to instruct the first reinforcement learning model how to select the next first response action and update its own state.

[0206] Exemplarily, the system observes whether the first result meets expectations. If so, the training is terminated. Otherwise, the learning parameters of the first reinforcement learning model are updated according to the first result until the first result meets expectations, thereby obtaining a trained first reinforcement learning model.

[0207] In this embodiment, by obtaining the first environmental state of the local model and inputting it into a pre-built first reinforcement learning model, the first response action of the local model for the first environmental state can be determined through the first reinforcement learning model, which helps to improve the adaptability and performance of the model; after executing the first response action, by observing the first result and feeding it back to the first reinforcement learning model to update the learning parameters, dynamic optimization and adaptive adjustment of the local model performance can be achieved, which is conducive to further improving the generalization ability and practical application effect of the model.

[0208] In an exemplary embodiment, the method further includes:

[0209] When the perception domain of any perception node changes or the perception performance of any local model does not meet the preset performance conditions, the state data of the global model and the feedback data received from each perception node are obtained; the state data and feedback data are input into the pre-trained second reinforcement learning model for processing to obtain the global update strategy of the global model; and the global model is updated based on the global update strategy.

[0210] Among them, state data refers to data that describes the current state of the global model, which is used to reflect the current operating state and performance of the global model.

[0211] Among them, feedback data refers to data from perception nodes, including but not limited to environmental changes and perception results of perception nodes, which are used to guide the update process of the global model.

[0212] Among them, the pre-trained second reinforcement learning model refers to a reinforcement learning model that is trained in advance and is used to process the state data of the global model and the feedback data received from each sensor node, and generate a corresponding global update strategy. The second reinforcement learning model learns how to make the best decision by interacting with the environment. The second reinforcement learning model receives the state data and feedback data as input, selects and executes actions, and then receives rewards or penalties as feedback. Through continuous trial and error and learning, the second reinforcement learning model learns the relationship between the state data of the global model, the feedback data of each sensor node and the appropriate global model update strategy, so as to be able to generate an update strategy suitable for the global model for the given state data and feedback data, thereby improving the performance of the global model in the corresponding environment. In specific implementation, the second reinforcement learning model can be implemented using one or more of the following algorithms, including but not limited to the DDPG (Deep Deterministic Policy Gradient) algorithm and the PPO (Proximal Policy Optimization) algorithm.

[0213] The global update strategy refers to a strategy generated by the pre-trained second reinforcement learning model based on the state data of the global model and the feedback data of each perception node to guide the update of the global model. The global update strategy may include but is not limited to adjusting the parameters of the global model (such as reallocating the weights of the local model parameters of each perception node when aggregating), optimizing the model structure (such as adding or reducing certain layers or modules), changing the model training method (such as adjusting the learning rate, selecting different optimization algorithms), and other operations, so that the global model can better adapt to changes in the perception domain of the perception node or changes in the performance of the local model, and improve the overall perception effect and performance.

[0214] Among them, updating the global model based on the global update strategy means adjusting and optimizing the global model accordingly according to the global update strategy output by the pre-trained second reinforcement learning model. Corresponding to the update of the local model, if the global update strategy recommends increasing the number of nodes in a certain neural network layer, the number of nodes in the corresponding neural network layer of the global model is increased; if the strategy requires adjusting the weights of the local model parameters of the aggregated perception nodes, the weights are modified in a specified manner so that the global model can better adapt to the dynamic changes of the entire perception system and maintain or improve its performance under different conditions.

[0215] In this embodiment, by monitoring the changes in the perception domain of the perception nodes and the perception performance of the local model in real time, the state data of the global model and the feedback data of each perception node are obtained in a timely manner, and the pre-trained second reinforcement learning model is used to process the state data and feedback data. A global update strategy suitable for the global model can be generated in a targeted manner, and the global model can be updated accordingly, which is conducive to improving the real-time and accuracy of the global model update, thereby improving the overall performance and adaptability of the system.

[0216] In an exemplary embodiment, Figure 7 As shown, the second reinforcement learning model is trained using the following method:

[0217] Step 701, obtaining a second environment state of the global model.

[0218] The second environmental state is determined based on the state data and the feedback data.

[0219] The second environmental state refers to the environmental state of the global model. The second environmental state is an abstract representation of the environmental condition of the global model based on the state data and the feedback data.

[0220] Step 702: input the second environment state into a pre-built second reinforcement learning model, and generate a second response action of the global model to the second environment state through the second reinforcement learning model.

[0221] The second response action refers to the action selected by the second reinforcement learning model for the global model based on the state data and feedback data, which is used to adjust the global model parameters. The second response action includes but is not limited to adjusting the global model parameters (such as recalculating and updating the weights of the neural network), changing the model structure (such as adding or deleting certain layers), and optimizing the settings of the model training process (such as adjusting the learning rate and changing the optimization algorithm). By executing the second response action, it is expected that the global model can better adapt to the environment in which it is located to improve the overall performance.

[0222] Exemplarily, the system inputs the second environment state into a pre-built second reinforcement learning model, and generates a second response action of the global model to the second environment state based on the state data and feedback data through the second reinforcement learning model.

[0223] Step 703, performing a second response action in a second environmental state to update the global model, and determining a reward value based on the model performance of the updated global model and the feedback data of the perception node.

[0224] Herein, updating the global model refers to updating the global model parameters according to the second result obtained after executing the second response action to obtain an updated global model.

[0225] The model performance of the updated global model refers to a series of indicators used to measure the performance of the updated global model in the perception task, including but not limited to one or more of the model's accuracy, recall rate, and mean square error.

[0226] The reward value refers to a numerical value determined based on the model performance of the updated global model and the feedback data of the perception node, and the reward value is used to evaluate the effect of executing the second response action under the current second environmental state.

[0227] Exemplarily, the system control center node performs a second response action in the second environmental state to adjust the global model parameters, obtain an updated global model, and determine the reward value based on the model performance of the updated global model and the feedback data of the perception node. If the performance of the updated global model is improved, or the perception node feedback indicates that the overall performance of the system has improved, the reward value is higher; conversely, if the performance of the updated global model is reduced or the perception node feedback is poor, the reward value is lower.

[0228] Step 704, the state data of the updated global model and the feedback data sent by the perception node based on the updated global model are determined as a new second environment state for iteration until the reward value reaches a preset condition, thereby obtaining a trained second reinforcement learning model.

[0229] The new second environment state refers to the new environment state representation determined by integrating the state data of the updated global model and the feedback data sent by the sensing node based on the updated global model. The new second environment state can reflect the new state of the system after executing the second response action once, and serves as the input of the next iteration, so that the second reinforcement learning model generates a new second response action based on the new second environment state, and repeats in this way, gradually optimizing the update strategy of the global model.

[0230] Among them, the preset condition refers to the standard or condition used to determine whether the training of the second reinforcement learning model is completed, that is, whether the reward value reaches a certain threshold. When the preset condition is met, the second reinforcement learning model is considered to have been trained and can be used to generate an effective update strategy for the actual global model. It is understandable that in some other embodiments, the number of iterations can also be used as a condition for the completion of training, that is, when the number of iterations reaches the set number, the second reinforcement learning model is considered to have been trained, and a trained second reinforcement learning model is obtained.

[0231] Exemplarily, the system determines the state data of the updated global model and the feedback data sent by the perception node based on the updated global model as a new second environmental state, and generates a new second response action of the global model for the new second environmental state through the second reinforcement learning model, and repeats the above steps until the reward value reaches the preset conditions to obtain a trained second reinforcement learning model.

[0232] In this embodiment, by constructing a second environmental state and inputting it into a pre-constructed second reinforcement learning model, a second response action of the global model for the second environmental state is generated, and the global model is updated according to the second result of the second response action. In the subsequent iterative training process, the global model is continuously adjusted until the reward value meets the preset conditions, thereby obtaining a trained second reinforcement learning model, which is beneficial to improving the adaptability and performance of the global model. Through continuous learning and optimization, it is ensured that the global model can respond in a timely, rapid and accurate manner when facing different environmental states.

[0233] In an exemplary embodiment, the method further includes:

[0234] Acquire the perception sample data corresponding to the same perception node under different signal sources; fuse the perception sample data based on the contribution of each signal source to the perception of the perception node to obtain the local perception data of the perception node; train each local model based on the local perception data to obtain the trained local model and local model parameters.

[0235] The signal source refers to the source of the sensory data collected by the sensory node. In applications such as the Internet of Things and intelligent monitoring, different signal sources may provide sensory data of different types or formats. In a complex sensory environment, there may be multiple different signal sources, and the signals emitted by each signal source may be received by the sensory node.

[0236] Among them, perception sample data refers to the raw data collected by the perception node from the signal source for subsequent processing and analysis. The perception sample data is used to reflect the perception information obtained by the perception node under the action of a specific signal source. For example, in human behavior perception, the perception sample data under different signal sources may contain information such as changes in wireless signal strength and phase caused by human movements. The perception sample data corresponding to each signal source reflects the characteristics of the perception scene from a certain perspective.

[0237] Among them, the degree of contribution refers to the importance or influence of different signal sources when the sensing node is sensing, which is used to reflect the relative importance of signals from different signal sources for accurate perception. The degree of contribution can be determined based on one or more factors including but not limited to the stability of the signal source, signal strength, and correlation with the target behavior. For example, in a scenario where Wi-Fi signals and Bluetooth signals are used simultaneously for indoor human behavior perception, if the Wi-Fi signal shows more stable and significant feature changes in identifying the direction of human movement, then the contribution of the Wi-Fi signal source in the behavior perception task may be higher than that of the Bluetooth signal source; for example, in an indoor positioning scenario with multiple signal sources, if a certain signal source is less affected by the environment and can accurately reflect the changes in the target position, then its contribution to positioning behavior perception is higher.

[0238] Among them, the fusion of perception sample data refers to the process of aggregating the perception sample data of the same perception node under different signal sources according to the contribution of each signal source. Through fusion, the advantageous information from different signal sources can be integrated, and redundant or interfering information can be removed, so that the obtained data can more comprehensively and accurately reflect the actual situation of the perception scene. In specific implementation, fusion can be achieved by one or more methods including but not limited to weighted fusion, splicing, and merging.

[0239] Among them, local perception data refers to the data obtained by fusing the perception sample data of the same perception node under different signal sources.

[0240] Exemplarily, the system obtains perception sample data collected for the same perception node under different signal sources; determines the contribution of each signal source to the perception node in behavioral perception, and based on the corresponding contribution of each signal source, performs weighted fusion on the perception sample data to obtain local perception data of the perception node; then, the system trains each local model based on the local perception data to obtain the trained local model and local model parameters.

[0241] In this embodiment, by acquiring the perception sample data of the same perception node under different signal sources, the information of multiple signal sources can be integrated to more comprehensively reflect the status and behavior of the perception node and improve the accuracy and reliability of the data. At the same time, data fusion is performed according to the contribution of each signal source to the perception of the perception node, which can ensure that important signal sources are given greater weight, thereby improving the efficiency and accuracy of data fusion and making the generated local perception data more accurate.

[0242] In an exemplary embodiment, Figure 8 As shown, based on the contribution of each signal source to the perception node in behavior perception, the perception sample data is fused, including:

[0243] Step 801: extract data features of each perception sample data, and determine the fourth weight of each data feature according to the contribution of each signal source to the perception of the perception node.

[0244] Among them, data features refer to information extracted from perception sample data that can reflect the characteristics of perception sample data.

[0245] Among them, the fourth weight refers to the weight value determined according to the contribution of each signal source to the perception node during perception, which is used to indicate the relative importance of the corresponding data feature in perception. For example, if the contribution of signal source 1 in perception is greater than that of signal source 2, then the fourth weight of the data feature extracted from the perception sample data of signal source 1 will be higher than the weight of the data feature extracted from signal source 2. By assigning appropriate fourth weights to different data features, it is helpful to highlight the role of important features in the subsequent fusion process.

[0246] Exemplarily, the system extracts data features of each perception sample data, respectively, and determines the contribution of each signal source to the perception of the perception node, and then determines the fourth weight of each data feature based on the contribution.

[0247] Step 802: fuse the data features based on the fourth weight to obtain a fused feature vector.

[0248] The fused feature vector refers to a feature vector obtained by fusing each data feature according to the fourth weight.

[0249] The fusion of the data features refers to the process of aggregating the data features according to the fourth weight of the data features. In specific implementation, the fusion can be implemented in one or more ways including but not limited to weighted fusion, splicing, and merging.

[0250] Exemplarily, the system performs weighted fusion on each data feature based on the fourth weight to obtain a fused feature vector.

[0251] Step 803, based on the attention mechanism, for each data feature in the fused feature vector, determine the attention weight between the data feature and other data features.

[0252] Among them, the attention mechanism is used to analyze the degree of correlation between each data feature and other data features in the fusion feature vector, and assign an attention weight to each data feature to highlight the data features that are closely related to other features and more critical to the overall information expression.

[0253] Among them, attention weight refers to the weight value calculated by the attention mechanism, which is used to measure the importance of each data feature in the feature representation. It is used to characterize the relative prominence of a data feature in the entire fusion feature vector. The higher the attention weight value, the closer the correlation between the data feature and other features, and the more important it is for accurately describing the perceived scene and perception.

[0254] Exemplarily, the system determines the attention weight between each data feature and other data features in the fused feature vector based on the attention mechanism.

[0255] Step 804, determining a feature representation of each data feature based on the attention weight, and fusing each feature representation to obtain a fused feature representation, and determining the fused feature representation as the local perception data of the perception node.

[0256] Among them, feature representation refers to the representation obtained by processing each data feature based on the attention weight, which reflects its role in the overall perceptual information.

[0257] The fused feature representation refers to the result obtained by fusing the feature representations obtained based on the attention weights again, which is used for subsequent tasks such as local model training, in the hope that the model can more accurately learn the patterns and rules in the perceived scene and achieve more precise perception.

[0258] Exemplarily, the system determines the feature representation of each data feature based on the attention weight, and fuses the feature representations according to the attention weight to obtain a fused feature representation. Subsequently, the system determines the fused feature representation as the local perception data of the perception node.

[0259] In this embodiment, by fusing data features based on weights to generate a fused feature vector, the information of multiple features can be combined in the fusion process. At the same time, the attention weight between each data feature and other features is determined by the attention mechanism, which can capture the intrinsic connection and mutual influence between the features, thereby generating a richer and more accurate feature representation.

[0260] It should be understood that, although the various steps in the flowcharts involved in the above-mentioned embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-mentioned embodiments can include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0261] Based on the same inventive concept, the embodiment of the present application also provides a perceptual model training device for implementing the perceptual model training method involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in the embodiments of the training device for one or more perceptual models provided below can refer to the limitations of the perceptual model training method above, and will not be repeated here.

[0262] In an exemplary embodiment, Fig. 9 As shown, a training device for a perception model is provided, comprising: a parameter receiving module 901, a weight determination module 902 and a model aggregation module 903, wherein:

[0263] The parameter receiving module 901 is used to receive local model parameters sent by each sensing node through the central node; the local model parameters are obtained by each sensing node through training based on its own local sensing data.

[0264] The weight determination module 902 is used to determine the first weight of each perception node based on the indicator value of each perception node under a preset dimension; the preset dimension includes at least one of the performance information of the local model corresponding to the perception node, the difference information between the local data distribution and the global data distribution of the local perception data of the perception node, the signal quality of the area where the perception node is located, and the data quality of the local perception data of the perception node.

[0265] The model aggregation module 903 is used to perform weighted aggregation on the local model parameters corresponding to each perception node according to the first weight of each perception node to obtain a global model.

[0266] In an optional embodiment, the weight determination module 902 is also used to obtain the second weight of each preset dimension; for each perception node, according to the second weight of each preset dimension, the indicator value of the perception node under each preset dimension is weighted and aggregated to obtain the contribution score of the perception node; based on the contribution score of each perception node, the first weight of each perception node is determined.

[0267] In an optional embodiment, the training of the perception model further includes:

[0268] The first updating module is used to obtain the original perception data of the perception node before the perception domain changes and the target perception data after the perception domain changes when the perception domain of any perception node changes; the perception domain includes at least one of the position of the perception node, the perception environment and the perception object; through a pre-trained feature alignment network, the features of the target perception data and the original perception data are aligned to obtain aligned feature data; the aligned feature data has a common feature vector of the target perception data and the original perception data in a common feature space; based on the aligned feature data, the local model parameters of the perception node after the perception domain changes are updated.

[0269] In an optional embodiment, the first update module is also used to input the original perception data and the target perception data into a pre-trained feature alignment network; the feature extraction module of the feature alignment network performs feature extraction on the original perception data and the target perception data respectively to obtain the original feature vector of the original perception data and the target feature vector of the target perception data; the generation module of the feature alignment network determines the common feature vector of the original feature vector and the target feature vector in the common feature space, and obtains the alignment feature data based on the common feature vector.

[0270] In an optional embodiment, the first update module is also used to input the first perception data into the feature extraction module for feature extraction to obtain the corresponding first feature vector, and input the second perception data into the feature extraction module for feature extraction to obtain the corresponding second feature vector; the first perception data and the second perception data belong to different perception domains; the first feature vector and the second feature vector are input into the discrimination module to identify the perception domains of the first feature vector and the second feature vector, and the discrimination module is adversarially trained using the first loss function of the discrimination module until the discrimination module identifies the first feature vector and the second feature vector as the same perception domain; the value of the first loss function is reversely transferred to the generation module, and the generation module is adversarially trained using the second loss function of the generation module until the joint loss function converges to obtain a trained feature alignment network; the joint loss function is constructed based on the first loss function and the second loss function according to preset weights.

[0271] In an optional embodiment, the first updating module is also used to determine the third weight of each new local model; according to the third weight of each new local model, the local model parameters corresponding to each perception node are weightedly aggregated to obtain a new global model.

[0272] In an optional embodiment, the training of the perception model further includes:

[0273] The second update module is used to obtain the local environment characteristics of the local model when the perception performance of any local model does not meet the preset performance conditions; input the local environment characteristics of the local model into the pre-trained first reinforcement learning model for processing to obtain the local update strategy corresponding to each local model; and update the local model based on each local update strategy.

[0274] In an optional embodiment, the second update module is also used to obtain a first environmental state of the local model; the first environmental state is determined based on local environmental characteristics; the first environmental state is input into a pre-constructed first reinforcement learning model, and a first response action of the local model to the first environmental state is determined through the first reinforcement learning model; the first response action is performed under the first environmental state; the first response action includes adjusting at least one of the learning rate, network structure, migration strategy, and model parameters of the local model; observing a first result after the first response action is executed; and the learning parameters of the first reinforcement learning model are updated according to the first result until the termination condition is reached to obtain a trained first reinforcement learning model.

[0275] In an optional embodiment, the training of the perception model further includes:

[0276] The third updating module is used to obtain the state data of the global model and the feedback data received from each sensing node when the sensing domain of any sensing node changes or the sensing performance of any local model does not meet the preset performance conditions. The state data and feedback data are input into the pre-trained second reinforcement learning model for processing to obtain the global update strategy of the global model; and the global model is updated based on the global update strategy.

[0277] In an optional embodiment, the third update module is also used to obtain a second environmental state of the global model; the second environmental state is determined based on state data and feedback data; the second environmental state is input into a pre-constructed second reinforcement learning model, and a second response action of the global model to the second environmental state is generated through the second reinforcement learning model; the second response action is performed under the second environmental state to update the global model, and a reward value is determined based on the model performance of the updated global model and the feedback data of the perception node; the state data of the updated global model and the feedback data sent by the perception node based on the updated global model are determined as the new second environmental state for iteration until the reward value reaches a preset condition to obtain a trained second reinforcement learning model.

[0278] In an optional embodiment, the training of the perception model further includes:

[0279] The fusion training module is used to obtain the perception sample data corresponding to the same perception node under different signal sources; based on the contribution of each signal source to the perception of the perception node, the perception sample data is fused to obtain the local perception data of the perception node; based on the local perception data, each local model is trained to obtain the trained local model and local model parameters.

[0280] In an optional embodiment, the fusion training module is also used to extract data features of each perception sample data, and determine the fourth weight of each data feature based on the contribution of each signal source to the perception of the perception node; fuse each data feature based on the fourth weight to obtain a fused feature vector; based on the attention mechanism, for each data feature in the fused feature vector, determine the attention weight between the data feature and other data features; determine the feature representation of each data feature based on the attention weight, and fuse the feature representations to obtain a fused feature representation, and determine the fused feature representation as the local perception data of the perception node.

[0281] Each module in the training device of the above-mentioned perception model can be implemented in whole or in part by software, hardware and their combination. Each of the above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to each of the above modules.

[0282] In an exemplary embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as shown in FIG. Fig.10As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. Among them, the processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store data such as central nodes, perception nodes, local model parameters, global model parameters, various weights, perception data, feature vectors, etc. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a training method for a perception model is implemented.

[0283] Those skilled in the art will understand that Fig.10 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0284] In an exemplary embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the training method of the perception model in the above embodiment is implemented.

[0285] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the training method of the perception model in the above embodiment is implemented.

[0286] In one embodiment, a computer program product is provided, including a computer program, which implements the training method of the perception model in the above embodiment when executed by a processor.

[0287] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0288] A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment method can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, an artificial intelligence (AI) processor, etc., but are not limited to this.

[0289] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0290] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A method for training a perception model, characterized in that: Applied to a cross-domain wireless intelligent sensing system, the system includes a central node and multiple sensing nodes, each sensing node is distributed in different areas, and the method includes: Receiving local model parameters sent by each sensing node through the central node; the local model parameters are obtained by training each sensing node according to its own local sensing data; Determine a first weight of each of the sensing nodes based on an indicator value of each of the sensing nodes under a preset dimension; the preset dimension includes at least one of performance information of a local model corresponding to the sensing node, difference information between a local data distribution and a global data distribution of local sensing data of the sensing node, signal quality of an area where the sensing node is located, and data quality of the local sensing data of the sensing node; According to the first weights of the respective sensing nodes, the local model parameters corresponding to the respective sensing nodes are weightedly aggregated to obtain a global model.

2. The method according to claim 1, characterized in that The determining, based on the index value of each of the sensing nodes under a preset dimension, a first weight of each of the sensing nodes comprises: Obtain the second weight of each preset dimension; For each sensing node, according to the second weight of each preset dimension, weighted aggregation is performed on the indicator value of the sensing node under each preset dimension to obtain a contribution score of the sensing node; Based on the contribution score of each sensor node, a first weight of each sensor node is determined.

3. The method according to claim 1, characterized in that The method further comprises: When the perception domain of any perception node changes, obtaining original perception data of the perception node before the perception domain changes and target perception data after the perception domain changes; the perception domain includes at least one of the position of the perception node, the perception environment, and the perception object; The features of the target perception data and the original perception data are aligned with each other through a pre-trained feature alignment network to obtain aligned feature data; the aligned feature data has a common feature vector of the target perception data and the original perception data in a common feature space; The local model parameters of the perception node after the perception domain changes are updated based on the alignment feature data.

4. The method according to claim 3, characterized in that The process of aligning the features of the target perception data and the original perception data through a pre-trained feature alignment network to obtain aligned feature data includes: Inputting the original perception data and the target perception data into a pre-trained feature alignment network; Performing feature extraction on the original perception data and the target perception data respectively through the feature extraction module of the feature alignment network to obtain an original feature vector of the original perception data and a target feature vector of the target perception data; The common feature vector of the original feature vector and the target feature vector in a common feature space is determined by a generation module of the feature alignment network, and alignment feature data is obtained based on the common feature vector.

5. The method according to claim 3, characterized in that: The feature alignment network includes a feature extraction module, a generation module and a discrimination module; The feature alignment network is trained by the following method: Inputting the first perception data into the feature extraction module for feature extraction to obtain a corresponding first feature vector, and inputting the second perception data into the feature extraction module for feature extraction to obtain a corresponding second feature vector; the first perception data and the second perception data belong to different perception domains; Inputting the first feature vector and the second feature vector into the discrimination module to identify the perception domains of the first feature vector and the second feature vector, and performing adversarial training on the discrimination module using a first loss function of the discrimination module until the discrimination module identifies the first feature vector and the second feature vector as the same perception domain; The value of the first loss function is passed back to the generation module, and the generation module is adversarially trained using the second loss function of the generation module until the joint loss function converges to obtain a trained feature alignment network; the joint loss function is constructed based on the first loss function and the second loss function according to preset weights.

6. The method according to claim 3, characterized in that After the local model parameters of the perception node after the perception domain changes are updated based on the alignment feature data, the method further includes: determining a third weight for each new local model; According to the third weights of the new local models, the local model parameters corresponding to the perception nodes are weighted and aggregated to obtain a new global model.

7. The method according to claim 1, characterized in that The method further comprises: When the perception performance of any local model does not meet the preset performance condition, obtaining the local environment characteristics of the local model; Inputting the local environment features of the local models into a pre-trained first reinforcement learning model for processing to obtain a local update strategy corresponding to each of the local models; The local model is updated based on each of the local update strategies.

8. The method according to claim 7, characterized in that The first reinforcement learning model is trained by the following method: Acquire a first environmental state of the local model; the first environmental state is determined based on the local environmental characteristics; Inputting the first environmental state into a pre-built first reinforcement learning model, and determining a first response action of a local model to the first environmental state through the first reinforcement learning model; Executing the first response action in the first environmental state; the first response action includes adjusting at least one of a learning rate, a network structure, a migration strategy, and a model parameter of the local model; Observe a first result after the first response action is executed; The learning parameters of the first reinforcement learning model are updated according to the first result until an end condition is reached to obtain a trained first reinforcement learning model.

9. The method according to claim 1, characterized in that: The method further comprises: When the perception domain of any perception node changes or the perception performance of any local model does not meet the preset performance conditions, the state data of the global model and the feedback data received from each perception node are obtained; Inputting the state data and feedback data into a pre-trained second reinforcement learning model for processing to obtain a global update strategy for the global model; The global model is updated based on the global update strategy.

10. The method according to claim 9, characterized in that The second reinforcement learning model is trained by the following method: Acquire a second environmental state of the global model; the second environmental state is determined based on the state data and the feedback data; Inputting the second environmental state into a pre-built second reinforcement learning model, and generating a second response action of the global model to the second environmental state through the second reinforcement learning model; Performing the second response action in the second environmental state to update the global model, and determining a reward value based on the model performance of the updated global model and the feedback data of the perception node; The state data of the updated global model and the feedback data sent by the perception node based on the updated global model are determined as a new second environment state for iteration until the reward value reaches a preset condition, thereby obtaining a trained second reinforcement learning model.

11. The method according to claim 1, characterized in that: The method further comprises: Obtaining perception sample data corresponding to the same perception node under different signal sources; Based on the contribution of each signal source to the perception of the perception node, the perception sample data is integrated to obtain local perception data of the perception node; Each local model is trained based on the local perception data to obtain a trained local model and local model parameters.

12. The method according to claim 11, characterized in that The fusing the perception sample data based on the contribution of each signal source to the perception node in behavior perception includes: Extracting data features of each sensing sample data, and determining a fourth weight of each data feature according to the contribution of each signal source to the sensing node during sensing; Based on the fourth weight, each data feature is fused to obtain a fused feature vector; Based on the attention mechanism, for each data feature in the fused feature vector, determining the attention weight between the data feature and other data features; Based on the attention weight, a feature representation of each data feature is determined, and each feature representation is fused to obtain a fused feature representation, and the fused feature representation is determined as the local perception data of the perception node.

13. A training device for a perception model, characterized in that: A cross-domain wireless intelligent sensing system is configured, the system includes a central node and multiple sensing nodes, each sensing node is distributed in different areas, and the device includes: A parameter receiving module, used to receive local model parameters sent by each sensing node through the central node; the local model parameters are obtained by training each sensing node according to its own local sensing data; A weight determination module, configured to determine a first weight of each of the sensing nodes based on an indicator value of each of the sensing nodes under a preset dimension; the preset dimension includes at least one of performance information of a local model corresponding to the sensing node, difference information between a local data distribution and a global data distribution of local sensing data of the sensing node, signal quality of an area where the sensing node is located, and data quality of the local sensing data of the sensing node; The model aggregation module is used to perform weighted aggregation on the local model parameters corresponding to each of the perception nodes according to the first weight of each of the perception nodes to obtain a global model.

14. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 12 are implemented.

15. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 12 are implemented.

16. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 12 are implemented.

Images