Model management method, communication apparatus, storage medium, and program product

Abstract

A model management method, a communication apparatus, a storage medium, and a program product. The method is applied to a first communication device, and comprises: when a currently running initial model does not match a current environment, updating the initial model to obtain a target model. Switching to and using the target model.

Description

A model management method, communication device, storage medium, and program product.

[0001] This disclosure claims priority to Chinese patent application No. 202510012822.X, filed on January 2, 2025, the entire contents of which are incorporated herein by reference. Technical Field

[0002] This disclosure relates to the field of communication technology, and in particular to a model management method, communication device, storage medium, and program product. Background Technology

[0003] With the development and maturation of artificial intelligence technology, the functions of models involved in artificial intelligence are becoming increasingly diverse. For example, models used in communication equipment (such as base stations and user equipment) to improve link performance and optimize network deployment (such as neural network models). Summary of the Invention

[0004] On the one hand, a model management method is provided, which is applied to a first communication device, including: updating the initial model to obtain a target model when the currently running initial model does not match the current environment; and switching to use the target model.

[0005] On the other hand, a model management method is provided, applied to a second communication device, comprising: receiving request information sent by a first communication device, the request information being used to request a model matching the current environment of the first communication device, wherein the second communication device has more second computing resources used for training the model than the first communication device has first computing resources used for training the model; and sending a third model matching the current environment to the first communication device.

[0006] On the other hand, a model management device is provided for use in a first communication device, the device comprising: a processing module.

[0007] The processing module is used to update the initial model to obtain the target model when the currently running initial model does not match the current environment. The processing module is also used to switch the target model.

[0008] On the other hand, a model management device is provided for use in a second communication device, the device comprising a receiving module and a transmitting module.

[0009] The receiving module receives request information sent by the first communication device, which requests a model matching the current environment of the first communication device. The second communication device has more computing resources used for training the model than the first communication device has for training the model. The sending module sends a third model matching the current environment to the first communication device.

[0010] On another front, a model management device is provided, comprising: a memory and a processor. The memory and the processor are coupled. The memory is used to store computer programs. When the processor executes the computer programs, it implements the model management method described above.

[0011] On another front, a computer-readable storage medium is provided, on which computer program instructions are stored, which, when executed by a processor, implement the aforementioned model management method.

[0012] On the other hand, a computer program product is provided, which includes computer program instructions that, when executed, implement the model management method described above. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in this disclosure, the accompanying drawings used in some embodiments of this disclosure will be briefly described below. Obviously, the drawings described below are merely drawings of some embodiments of this disclosure, and those skilled in the art can obtain other drawings based on these drawings.

[0014] Figure 1 is a schematic diagram of a communication system according to some embodiments.

[0015] Figure 2 is a schematic diagram of another communication system according to some embodiments.

[0016] Figure 3 is a flowchart of a model management method according to some embodiments.

[0017] Figure 4 is a schematic diagram of a storage structure according to some embodiments.

[0018] Figure 5 is a flowchart of another model management method according to some embodiments.

[0019] Figure 6 is a flowchart of another model management method according to some embodiments.

[0020] Figure 7 is an example diagram of a communication network according to some embodiments.

[0021] Figure 8 is an example diagram of another communication network according to some embodiments.

[0022] Figure 9 is an example diagram of another communication network according to some embodiments.

[0023] Figure 10 is a block diagram of a model management device according to some embodiments.

[0024] Figure 11 is a block diagram of a model management device according to some embodiments.

[0025] Figure 12 is a block diagram three of a model management device according to some embodiments. Detailed Implementation

[0026] The technical solutions of this disclosure will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0027] It should be noted that, in this disclosure, the words "exemplarily" or "for example" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as "exemplarily" or "for example" in this disclosure should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the words "exemplarily" or "for example" is intended to present the relevant concepts in a specific manner.

[0028] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature.

[0029] In the description of this disclosure, unless otherwise stated, " / " means "or," for example, A / B can mean A or B. "And / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. Furthermore, "at least one" means one or more, and "more than one" means two or more.

[0030] With the development and maturation of artificial intelligence technology, more and more fields have begun to notice its value and are gradually applying it to solve problems that are difficult to solve using traditional methods.

[0031] In the field of wireless communication, artificial intelligence (AI) technology has also achieved great success. During the 5G era, equipment manufacturers and operators have already experimented with AI to improve link performance and optimize network deployment, with promising results. In the early stages of discussions on 6G, numerous experts and scholars from operators and equipment manufacturers quickly reached a consensus that AI would play a crucial role in the 6G standard, and decided to work together to promote its application.

[0032] While there is a consensus from academia and industry that artificial intelligence is a disruptive technology that will revolutionize future communication systems, there are also concerns about its generalization ability due to the black-box nature of neural networks. When the real-world environment differs from the environment in which the model was trained, the model's performance may significantly degrade.

[0033] In summary, how to synchronously update the models in communication devices based on changes in the communication environment has become a pressing technical problem that needs to be solved.

[0034] To address the aforementioned technical problems, this disclosure provides a model management method applicable to model management scenarios involving adjusting network parameters in communication devices. When a model running locally in the communication device is incompatible with the current communication environment, the model can be updated and switched to a new model that matches the environment's requirements. This ensures the normal operation of the communication device and enables synchronous updates of the model within the communication device based on changes in the communication environment.

[0035] The network architecture of the mobile communication network (including but not limited to 2G, 3G, 4G, 5G, and future mobile communication networks such as 5G-A and 6G) ​​may include at least a first communication device and a second communication device. It should be understood that in this example, in the uplink, the first communication device may be a terminal-side device (e.g., including but not limited to a terminal), and the second communication device may be a network-side device (e.g., including but not limited to a base station). Of course, in the downlink, the first communication device may also be a network-side device, and the second communication device may also be a terminal-side device. In device-to-device communication between the two communication devices, both the first and second communication devices can be base stations or terminals. The first and second communication devices may be referred to as the first device and the second device, respectively.

[0036] For example, as shown in FIG1, a schematic diagram of a communication system provided in an embodiment of the present disclosure is provided. The communication system may include: a first communication device 101 and a second communication device 102, and the first communication device 101 is equipped with a model (such as a neural network model) for adjusting network parameters.

[0037] Here, while running the model, the first communication device 101 can detect the matching between the communication environment and the model in real time, and if the communication environment and the model do not match, it can update and switch the model to a new model so that the new model can match the communication environment.

[0038] It should be noted that during the process of the first communication device 101 updating and switching the model to a new model, the first communication device 101 can utilize its local computing resources to retrain (or fine-tune) the model based on the communication environment to obtain the new model. Alternatively, the first communication device 101 can utilize the computing resources shared by other communication devices (such as the second communication device 102) to obtain the new model.

[0039] During the process of the second communication device 102 providing a new model for the first communication device 101, the second communication device 102 can utilize local computing resources to retrain (or fine-tune based on the historical model) the first communication device 101 based on the current communication environment of the first communication device 101 to provide a new model for the first communication device 101.

[0040] Here, fine-tuning is a specialized term in the field of machine learning, distinct from the concept of ordinary neural network training. Its key features include, but are not limited to, using a smaller learning rate, updating some parameters of the network, or performing fewer training iterations.

[0041] In this embodiment of the disclosure, the second computing resources used for training the model in the second communication device are greater than the first computing resources used for training the model in the first communication device.

[0042] In other words, we define a first communication device and a second communication device. Here, the second communication device is equipped with sufficient computing resources (graphics processing unit (GPU), central processing unit (CPU), etc.) and memory resources, and has the ability to train a neural network model from scratch (training from scratch means that the parameters of the model are randomly initialized according to certain rules when training begins), and can store a large number of trained models.

[0043] The first communication device is usually a low-cost device with limited computing power. It can perform inference or fine-tuning on pre-trained neural network models, but it basically does not have the ability to train neural network models from scratch. In addition, the storage space of the first communication device is much smaller than that of the second communication device, and it can only store a limited number of neural network models.

[0044] For example, taking the example of the first communication device 101 using the shared computing resources of the second communication device 102 to obtain a new model, Figure 2 shows a schematic diagram of another communication system. Here, the second communication device 102 includes an AI computing server (including GPU-1, GPU-2, GPU-3, ..., GPU-N) and an AI model memory (including model 1, model 2, model 3, ..., model M), and the first communication device 101 includes the currently used model and the AI ​​model memory (including model 1, model 2, model 3, ..., model K), where K < M. The first communication device 101 can report model training requests to the second communication device 102 and download models from the second communication device 102.

[0045] It should be noted that the embodiments disclosed herein do not limit the first communication device and the second communication device, and the first communication device and the second communication device can satisfy at least one of the following:

[0046] 1. The first communication device is a user equipment, and the second communication device is a base station;

[0047] 2. The first communication device is a user equipment, and the second communication device is an internet device;

[0048] 3. The first communication device is a user equipment, and the second communication device is a core network element;

[0049] 4. The first communication device is a base station, and the second communication device is a core network element.

[0050] User equipment (UE) can be a device with wireless transceiver capabilities. UE can be a mobile phone, tablet, computer with wireless transceiver capabilities, virtual reality (VR) terminal, augmented reality (AR) terminal, wireless terminal in industrial control, wireless terminal in self-driving, wireless terminal in remote medical care, wireless terminal in smart grid, wireless terminal in transportation safety, wireless terminal in smart city, wireless terminal in smart home, etc. The embodiments of this disclosure do not limit the application scenarios. A terminal may also be referred to as a user, terminal, artificial intelligence of things (A-IoT) device, access terminal, UE unit, UE station, mobile station, mobile station, remote station, transmitter, remote terminal, mobile device, UE terminal, wireless communication device, UE agent, or UE device, etc., and the embodiments of this disclosure do not limit this.

[0051] A base station (BS) can be a base station in LTE, Long Term Evolution Advanced (LTEA) or an evolved Node B (eNB or eNodeB), a base station device (gNB) in a 5G network, or a base station in a future communication system. Base stations can include various macro base stations, micro base stations, femtocell base stations, wireless remote extensions, reconfigurable intelligent surfaces (RISS), routers, relay stations, transmission and reception points (TRPs), receivers, access points, wireless fidelity (Wi-Fi) devices, and other network-side equipment. A base station can sometimes also be referred to as a reader / writer used for communication with terminals.

[0052] Core network elements can include various network functions, such as access and mobility management function (AMF), session management function (SMF), user plane function (UPF), policy control function (PCF), unified data management (UDM), location management function (LMF), etc.

[0053] It should be noted that Figure 1 is only an exemplary framework diagram. The number of devices included in Figure 1 and the names of each device are not limited. In addition to the devices shown in Figure 1, the communication system may also include other devices, such as core network devices.

[0054] The application scenarios of the embodiments disclosed herein are not limited. The system architecture and business scenarios described in the embodiments of this disclosure are for the purpose of more clearly illustrating the technical solutions of the embodiments of this disclosure, and do not constitute a limitation on the technical solutions provided by the embodiments of this disclosure. As those skilled in the art will know, with the evolution of network architecture and the emergence of new business scenarios, the technical solutions provided by the embodiments of this disclosure are also applicable to similar technical problems.

[0055] Figure 3 shows a flowchart of a model management method. As shown in Figure 3, the model management method is applied to a first communication device, including S301 and S302.

[0056] In S301, if the currently running initial model does not match the current environment, the initial model is updated to obtain the target model.

[0057] In this embodiment of the disclosure, the first communication device can collect relevant information about the current environment in real time during the running of the initial model, and determine whether the initial model matches the current environment based on the relevant information about the current environment.

[0058] Here, the relevant information about the current environment may include: environmental characteristics of the current environment and model requirements that match the current environment.

[0059] It should be noted that environmental characteristic information refers to objective factors that can affect system performance. Taking the receiver model of a mobile communication system as an example, environmental characteristic information includes, but is not limited to, the signal-to-noise ratio (SNR), Doppler frequency shift, line of sight (LOS) / non-line of sight (NLOS), channel delay spread, etc., and may even include the characteristic information of radio frequency devices.

[0060] The model requirements information that match the current environment include, but are not limited to, the number of model parameters, the computational cost of the model, the number of layers in the model, the type of layers, the structure of the model, and the inference time.

[0061] For example, for a receiver neural network model in a communication system, the environmental feature information on which the matching with the environment is based may include the bit error rate (BER), and the model requirement information may include the model's input layer, hidden layer, and output layer.

[0062] In some embodiments, during the process of determining the matching status between the initial model and the current environment, the first communication device may determine the matching degree between the initial model and the current environment based on the environmental feature information of the current environment and the model requirement information for matching the current environment, and determine the matching status between the initial model and the current environment by comparing it with a pre-stored first preset matching threshold.

[0063] Here, if the matching degree between the initial model and the current environment is greater than the first preset matching threshold, the initial model is matched with the current environment (i.e., it is in a matching state).

[0064] In other words, the first preset matching threshold is used to provide a reference for judging the matching between the model and the environment.

[0065] Similarly, when the matching degree between the initial model and the current environment is less than or equal to the first preset matching threshold, the initial model does not match the current environment.

[0066] It should be noted that the non - matching can include a completely non - matching state and a state between matching and non - matching (i.e., applicable and inapplicable).

[0067] For example, the first communication device may also pre - store a second preset matching threshold, and the first preset matching threshold is greater than the second preset matching threshold. When the matching degree between the initial model and the current environment is greater than the second preset matching threshold and less than or equal to the first preset matching threshold, the initial model and the current environment are in a state between matching and non - matching.

[0068] Or, when the matching degree between the initial model and the current environment is greater than the first preset matching threshold, the initial model and the current environment are in a completely non - matching state.

[0069] It should be noted that the above - mentioned second preset matching threshold and first preset matching threshold are both taken as being positively correlated with the matching situation (i.e., the degree of matching) (that is, the greater the matching degree, the more matching). Similarly, the second preset matching threshold and the first preset matching threshold can also be taken as being negatively correlated with the matching situation (i.e., the degree of matching) (that is, the smaller the matching degree, the more matching).

[0070] That is to say, the result of the matching degree evaluation is described by a coefficient α. Suppose the threshold for model matching is t_match, and the threshold for model non - matching is t_unmatch. If α > t_match (or α < t_match), it means the model and the current environment match; if α < t_unmatch (or α > t_unmatch), it means the model and the current environment do not match; if t_unmatch < α < t_match (or t_unmatch > α > t_match), it means the model and the current environment are in a situation between matching and non - matching.

[0071] As another possible implementation, when the first communication device determines the matching situation between the initial model and the current environment, the first communication device can determine the second preset matching threshold (and / or the first preset matching threshold) corresponding to the current environment based on the environmental feature information of the current environment and the model requirement information for matching the current environment, and compare the status information of the initial model with the second preset matching threshold (and / or the first preset matching threshold) to determine the matching situation between the initial model and the current environment.

[0072] It should be noted that for the introduction of the second preset matching threshold and the first preset matching threshold, reference can be made to the description in the above - mentioned embodiments, and details are not elaborated here.

[0073] For example, taking the receiver neural network model (i.e., the initial model) in a communication system as an example, the matching coefficient α can be the BER information (i.e., the state information of the initial model) calculated for the current set of received data. When the BER is less than the matching threshold t_match, it means that the model and the current environment are matched; when the BER is greater than the non-matching threshold t_unmatch, it means that the model and the current environment are not matched.

[0074] In some embodiments, if the first communication device determines that the initial model matches the current environment, it continues to run the initial model.

[0075] In other words, if the initial model is compatible with changes in the current environment, the initial model can continue to operate normally without needing to synchronize with the changes in the environment.

[0076] In other embodiments, if the first communication device determines that the initial model does not match the current environment, it updates the initial model to obtain the target model.

[0077] Here, the first communication device can determine an update strategy based on the matching status of multiple first historical models stored in the first communication device with the current environment, and update the initial model to obtain the target model based on the update strategy. The first historical models are other models that the first communication device switched to and updated due to changes in the environment before the current environment.

[0078] In other words, if the model running in the communication device does not match the current environment, the existing model stored locally can be used to determine a model that matches the current environment.

[0079] It should be noted that, in the process of determining the update strategy, the first communication device can calculate the matching degree between each first historical model and the current environment, and compare it with the pre-stored first preset matching threshold. Based on the relationship between the matching degree of each first historical model and the first preset matching threshold, the update strategy is determined, and then an appropriate amount of computing power is used to update the initial model to obtain the target model.

[0080] Here, the description of the matching degree between the first historical model and the current environment can be found in the above embodiment, which describes the matching degree between the initial model and the current environment. It will not be repeated here.

[0081] In some embodiments, if there is at least one first model among multiple first historical models that has a matching degree greater than a first preset matching threshold with the current environment, the first communication device may determine an update strategy including: taking the first model with the highest matching degree with the current environment among the at least one first model as the target model.

[0082] In other words, if the model running in the communication device does not match the current environment, but a suitable model exists in the local storage space, the locally stored model can be switched to match the current environment to achieve local processing.

[0083] As another possible implementation, if there is at least one second model among multiple first historical models whose matching degree with the current environment is less than or equal to a first preset matching threshold and greater than a second preset matching threshold, the first communication device may determine that the update strategy includes: fine-tuning the second model with the highest matching degree with the current environment among at least one second model, and using the fine-tuned second model as the target model.

[0084] In other words, if the model running in the communication device does not match the current environment, and there is a flawed model in the local storage space that is between applicable and inapplicable, a small amount of computing power can be used to fine-tune the flawed model to obtain a model that matches the current environment, without having to redefine a new model with a large amount of computing power.

[0085] Similarly, when the matching degree between multiple first historical models and the current environment is less than or equal to the second preset matching threshold, and the matching degree between the initial model and the current environment is greater than the second preset matching threshold, the first communication device can determine that the update strategy includes: fine-tuning the initial model and using the fine-tuned initial model as the target model.

[0086] In other words, when the model running in the communication device is somewhere between matching and not matching the current environment, a small amount of computing power can be used to fine-tune the running model to obtain a model that matches the current environment, without having to switch to another model.

[0087] In summary, communication devices achieve local model matching by analyzing the similarity between the current dataset and the dataset corresponding to the existing model, thereby reducing the need for model training.

[0088] For example, if the matching degree between multiple first historical models and the current environment is less than or equal to the second preset matching threshold, and the matching degree between the initial model and the current environment is less than or equal to the second preset matching threshold, the first communication device may determine that the update strategy includes: using local first computing power resources to train the target model from scratch.

[0089] In other words, if the model running in the communication device is not compatible with the current environment, and the model in the local storage space is also not compatible with the current environment, the local computing resources can be used to retrain a model that is compatible with the current environment. That is, continuous training is carried out during the use of the model to keep it compatible with the current environment.

[0090] In S302, switch to using the target model.

[0091] Understandably, if the model running locally in the communication device does not match the current communication environment, the model can be updated and managed to switch to a new model that matches the environment requirements, so as to ensure the normal operation of the communication device and realize the synchronous update of the model in the communication device based on changes in the communication environment.

[0092] It's important to note that neural networks are now widely used in various systems to replace traditional processing modules, achieving good results. The use of neural networks involves two distinct phases: training and inference. The training phase requires dedicated hardware resources (such as GPUs), consuming significant amounts of electricity and time (ranging from several hours to days or even months), resulting in high costs. In contrast, inference is much simpler, achieving timeframes in the seconds or even milliseconds. Even large language models like GPT can provide inference results within an acceptable timeframe. For low-cost devices, using neural network models for inference is feasible, but training these models is extremely difficult.

[0093] As is well known, neural networks are a computationally intensive technology, requiring expensive dedicated hardware and consuming significant amounts of power and time, especially during the model training phase. Communication devices are cost- and time-sensitive; while deploying neural network inference capabilities on communication devices is feasible, training them is impractical. Even deploying training capabilities on base station equipment would impose enormous power consumption and cost pressures.

[0094] In other words, if the model running in the communication device is incompatible with the current environment, and the model in the local storage space is also incompatible with the current environment, the communication device will need to retrain a new model, which will require a lot of computing resources. However, due to the limited computing resources of the device itself, it may not be able to handle the model training process locally.

[0095] In some embodiments, when the matching degree between multiple first historical models and the current environment is less than or equal to a second preset matching threshold, and the matching degree between the initial model and the current environment is less than or equal to the second preset matching threshold, the first communication device may determine the update strategy by requesting the second communication device to determine the target model.

[0096] Here, the second computing resources used for training the model in the second communication device are greater than the first computing resources used for training the model in the first communication device.

[0097] Understandably, this involves implementing model training and inference on different entities—specifically, deploying only sufficient computing power for model inference on the local device, while migrating model training to the cloud. When the local device needs model training, it sends the training data to the cloud device, which then completes the model training.

[0098] In other words, if a model running on a communication device with limited computing resources is not compatible with the current environment, a communication device with more computing resources can be requested to generate a new model that is compatible with the current environment.

[0099] It should be noted that the embodiments disclosed herein do not limit the first communication device and the second communication device, and the first communication device and the second communication device can satisfy at least one of the following (a)-(d):

[0100] (a) The first communication device is a user equipment, and the second communication device is a base station;

[0101] (b) The first communication device is a user equipment, and the second communication device is an internet device;

[0102] (c) The first communication device is a user equipment, and the second communication device is a core network element;

[0103] (d) The first communication device is a base station, and the second communication device is a core network element.

[0104] In other words, by combining the device identities of the first communication device and the second communication device, the application scenarios of the embodiments of this disclosure can be expanded.

[0105] In some embodiments, in conjunction with the process described above whereby the first communication device updates the initial model based on an update strategy to obtain the target model, the first communication device may send a request message to the second communication device and receive a third model from the second communication device in response to the request message, and then update the initial model based on the third model to obtain the target model.

[0106] Here, the request information is used to request a model that matches the current environment of the first communication device.

[0107] In other words, communication devices with limited computing power can request communication devices with more computing power to generate a new model that matches the current environment through signaling interaction.

[0108] In this embodiment of the disclosure, the request information may include at least one of the following (1)-(4):

[0109] (1) Dataset used for model training;

[0110] (2) Data augmentation algorithms for expanding datasets;

[0111] (3) Environmental characteristics information of the current environment;

[0112] (4) Match the model requirements information of the current environment.

[0113] Here, augmentation algorithms refer to methods that generate more data that can be used for model training based on the training dataset. With the assistance of augmentation algorithms, the first communication device can send only a small amount of training data to the second communication device.

[0114] In other words, when requesting communication devices with abundant computing resources to generate a new model that matches the current environment, corresponding reference information is also provided.

[0115] It should be noted that the second communication device also stores multiple second historical models. The third model can be a model determined by the second communication device from the multiple stored second historical models based on environmental feature information and model feature information. The second historical models are models that the second communication device is currently using / has previously used / generated for other devices.

[0116] In other words, communication devices with abundant computing resources respond to requests by combining existing models stored locally with environmental and model requirements to determine a model that matches the current environment. Furthermore, model sharing is achieved through a cloud-based shared model resource pool, avoiding redundant model training, reducing computing power overhead, and realizing green, energy-saving, and environmentally friendly communication.

[0117] Here, the third model is the model among multiple second historical models whose matching degree with the current environment is greater than the first preset threshold.

[0118] In other words, when a communication device with abundant computing resources responds to a request and determines a new model that matches the current environment, it can select the model that matches the current environment from the multiple models that have been generated by the agent and stored.

[0119] Alternatively, the third model is a model fine-tuned by the second communication device based on the dataset and data augmentation algorithms from the fourth model among multiple second historical models; here, the matching degree of the fourth model with the current environment is less than or equal to the first preset threshold and greater than the second preset threshold.

[0120] In other words, if there are flawed models among the stored models generated by the agent that are between matching and not matching the current environment, the training dataset can be referenced, and data augmentation algorithms and a small amount of computing power can be used to fine-tune the flawed models to obtain a model that matches the current environment, without having to generate a new model from scratch.

[0121] Alternatively, the third model could be a model trained from scratch by the second communication device based on the dataset and data augmentation algorithms.

[0122] In other words, based on the reference training dataset, and using data augmentation algorithms and a large amount of computing resources, a new model that matches the current environment is generated from scratch.

[0123] It should be noted that after a communication device switches to a new model, it can store the new model for future use in similar environments to adapt to the application.

[0124] In some embodiments, the first communication device may store the target model.

[0125] In other words, after obtaining a new model, the communication device can store the new model so that it can directly switch the stored model in future scenarios that repeat the current scenario without having to determine a new model.

[0126] It should be noted that the order of the target model storage steps in this embodiment is not limited to S301 and S302 described above. For example, the first communication device may store the target model after S301 and before S302. Alternatively, the first communication device may store the target model after S302.

[0127] In this embodiment of the disclosure, during the process of storing the target model, the first communication device can also establish a correspondence between the target model and the current environment.

[0128] In other words, by establishing a correspondence with the current environment, it is possible to facilitate subsequent matching and filtering with the stored models.

[0129] It should be noted that, based on the above process of evaluating the matching degree between the model and the environment, the first communication device can analyze the similarity between the data generated by the current environment and the data corresponding to the model. The data corresponding to the model refers to the data used during the model training or data with the same distribution characteristics as the data used during model training.

[0130] For example, as shown in Figure 4, a schematic diagram of the storage structure of a neural network model in the model memory is illustrated. Here, the models (such as model 1, model 2, model 3, ..., model K) and the data corresponding to the models are always stored in pairs (i.e., correspondences) in the model memory.

[0131] Here, when the first communication device searches in the model memory, it uses the data generated by the current environment and the data corresponding to each model in the memory to calculate the similarity. If the similarity reaches the matching threshold, it means that the model matches the current environment. If the similarity reaches the non-matching threshold, it means that the model does not match the current environment. Otherwise, the model and the environment are in a state between matching and non-matching.

[0132] For example, during the storage of the target model, the first communication device can also determine whether its available storage space for the model is less than a preset capacity threshold, and if the available storage space for the model is less than the preset capacity threshold, delete the model with the lowest usage frequency and / or the shortest usage time among multiple first historical models.

[0133] Understandably, the first communication device can manage the models in its model library according to usage frequency or usage duration. If a new model needs to be stored and the current model's storage space is insufficient, the model with the lowest usage frequency or shortest usage duration in the model library will be deleted. In other words, based on usage frequency and usage duration, historically stored models are deleted to free up storage space for models.

[0134] It should be noted that after the first communication device switches to the target model (i.e., S302 above), the first communication device can also conduct a trial run of the target model to verify its effectiveness.

[0135] In some embodiments, after the first communication device switches to use the target model (i.e., S302 above), the first communication device may check the operation of the target model within a preset time period, and if the target model does not match the environment within the preset time period, switch the target model to the initial model or the preset model (i.e., the default model).

[0136] Understandably, after the first communication device replaces its current model with the new one, it will monitor the updated model for a period of time. If the performance does not meet expectations, it will roll back the model and reload either the previous model or the default model. The default model refers to the initial model of the first communication device, which typically has good generalization capabilities and is suitable for various working environments. In other words, after switching to the new model, it is necessary to check whether its operation over a period of time truly matches the environmental requirements. If it does not match, it needs to roll back to the original model or the default initial model to ensure the normal operation of the device.

[0137] This disclosure also provides a model management method applied to a second communication device, as shown in FIG5. The model management method may include S501 and S502.

[0138] In S501, a request message sent by the first communication device is received.

[0139] Here, the request information is used to request a model that matches the current environment of the first communication device.

[0140] It should be noted that the description of the request information can be found in the above embodiments, and will not be repeated here.

[0141] In this embodiment of the disclosure, the second computing resources used for training the model in the second communication device are greater than the first computing resources used for training the model in the first communication device.

[0142] In S502, a third model matching the current environment is sent to the first communication device.

[0143] It should be noted that the description of the third model can be found in the above embodiments, and will not be repeated here.

[0144] In some embodiments, after receiving a model training request (i.e., request information) from the first communication device, the second communication device searches its own model memory for a suitable model based on the model's requirements (i.e., model requirement information) and environmental feature information. The requirements include two aspects: first, the model's compatibility with the environment; and second, the model's own requirements, such as the number of parameters and computational complexity. Depending on the search results, the following processing methods may be used:

[0145] (1) If there are one or more neural network models that meet the requirements, select the model with the highest matching degree (i.e., the third model) and send it to the first communication device;

[0146] (2) If no model that meets the requirements is found, but one or more models that meet the requirements are between matching and not matching, then select the model with the highest matching degree, train the model using the dataset or the data augmented dataset, and send the model (i.e. the third model) to the first communication device after training.

[0147] (3) If none of the models meet the requirements, a neural network model that meets the requirements is constructed according to the model requirement information. The model is trained using environmental feature information, dataset or data augmentation dataset. After training, the model (i.e. the third model) is sent to the first communication device.

[0148] In some embodiments, after the second communication device sends the third model to the first communication device, the second communication device may also store the third model.

[0149] In other words, after the new model is obtained through processing (i.e. fine-tuning or training from scratch), the second communication device can also store the new model so that the stored model can be matched with future scenes that repeat the current scene, without having to determine a new model.

[0150] Similarly, in conjunction with the above embodiments, during the process of storing the third model in the second communication device, the second communication device can establish a correspondence between the third model and the current environment.

[0151] It should be noted that when the second communication device searches in the model memory, it uses the data generated by the current environment and the data corresponding to each model in the memory to calculate the similarity. If the similarity reaches the matching threshold, it means that the model matches the current environment. If the similarity reaches the non-matching threshold, it means that the model does not match the current environment. Otherwise, the model and the environment are in a state between matching and non-matching.

[0152] For example, during the process of storing the third model in the second communication device, the second communication device may delete the models with the lowest usage frequency and / or shortest usage time among multiple second historical models if the available storage space for the model in the second communication device is less than a preset capacity threshold.

[0153] The following describes the model management method provided in the above embodiment, taking the interaction between the first communication device and the second communication device as an example, as shown in Figure 6, including: S601-S607.

[0154] In S601, if the initial model currently in operation does not match the current environment, the first communication device determines an update strategy based on the matching status of multiple first historical models stored in the first communication device with the current environment.

[0155] Here, the update strategy includes: requesting a second communication device to determine the target model.

[0156] In S602, the first communication device sends a request message to the second communication device.

[0157] In S603, the second communication device receives the request information sent by the first communication device.

[0158] In S604, the second communication device sends a third model that matches the current environment to the first communication device.

[0159] In S605, the first communication device receives a third model in response to the second communication device sending a request message.

[0160] In S606, the first communication device updates the initial model based on the third model to obtain the target model.

[0161] In S607, the first communication device switches to the target model.

[0162] The model management method provided in this disclosure embodiment will be described below with reference to specific examples, including steps one, two, three and four.

[0163] Step 1: The first communication device continuously evaluates the matching degree between the current neural network model and the current environment.

[0164] Step 2: The first communication device takes corresponding processing measures based on the matching degree evaluation results.

[0165] If the evaluation results show that the model matches the environment, it means that the current model is suitable for the current environment and no processing is required.

[0166] If the evaluation results show that the model does not match the environment, it means that the current model is not suitable for the current environment and needs to be updated.

[0167] Step 3: The first communication device first searches its own model memory. If a model that matches the current environment exists, it selects one of these models to replace the currently used model. If no model is found, it selects a model for fine-tuning or sends a model request to the second communication device.

[0168] Step 4: After receiving the request, the second communication device first searches its own model memory. If it finds a model that matches the current environment, it selects one of these models and sends it to the first communication device. If it does not find one, it selects a model for fine-tuning or calls on computing resources to train a new model from scratch as required. After completion, it sends the model to the first communication device.

[0169] In some embodiments, in step two above, if the model and the current environment are between a match and a mismatch, the first communication device can select one of the following processing methods:

[0170] (a) Use the computing resources on the first communication device to fine-tune the current model.

[0171] (b) Search the model memory of the first communication device for a model that matches the current environment.

[0172] In some embodiments, in step three above, if the first communication device does not find a model matching the current environment in its own model memory, the first communication device does not rush to send a model training request to the second communication device, but first searches in its local model memory for a model that is somewhere between a match and a non-match:

[0173] (a) If it exists, select the model with the highest matching degree here for fine-tuning.

[0174] (b) If it does not exist, send a request for model training to the second communication device.

[0175] Example 1, taking a deployment configuration where the first communication device is the UE and the second communication device is the base station, as shown in Figure 7, illustrates an example of a communication network. Here, there are one or more first communication devices, all of which establish communication links with the second communication device via the air interface. The second communication device has sufficient computing and storage resources. The first communication device sends a model training request to the second communication device, and the second communication device trains the model according to the request. As the system continues to run, the models in the model memories of the first and second communication devices become increasingly rich. The models in the model memory of the second communication device are shared by UEs within the same cell. The channels between UEs and base stations within the same cell have high similarity, so most model requests can be responded to by looking up the model memory. The actual need for model training decreases, thus achieving the goal of energy conservation and environmental protection.

[0176] In Example 1, a possible scenario is as follows: First communication devices #1, #2, and #3 initially connect to the second communication device. Their model memories contain only one baseline receiver model, adaptable to various channel environments, but with mediocre performance. Assume first communication device #1 is located inside an office building, where its channel characteristics are a low-SNR static channel; first communication device #2 is located on a highway, where its channel characteristics are a channel with rapidly changing SNR and phase; and first communication device #3 is located in an outdoor stadium, where its channel characteristics are a channel with slowly changing SNR and phase (LOS). They each send model training requests to the base station. After a period of time, the second communication device trains three models that meet the requirements and sends them to the three first communication devices, simultaneously adding these three models to its own model memory. After some time, the channel environment of the three first communication devices changes. For example, the channel of first communication device #1 becomes a channel with rapidly changing SNR and phase. At this point, when it sends a model training request to the second communication device, this request contains the current channel characteristic information of first communication device #1. The second communication device checks its own model memory and finds that a model that meets the requirements exists. It then directly sends the model that meets the requirements to the first communication device #1, thereby avoiding the need to train a new model.

[0177] Example 2, taking a deployment method where the first communication device is a UE and the second communication device is a smart computing center (such as a core network element) located on the Internet as an example, is illustrated in Figure 8, which shows an example diagram of another communication network. Here, the first communication device and the second communication device are connected through a base station, and there are one or more first communication devices and one or more second communication devices. One second communication device can serve at least one first communication device (i.e., different first communication devices can access the same second communication device).

[0178] In Example 2, one possible scenario is that the UE chip manufacturer builds its own intelligent computing center to provide the UE with model training capabilities. The UE can establish a connection with the intelligent computing center via Wi-Fi or via a base station.

[0179] Different chip manufacturers can have their own intelligent computing centers, and UEs using chips from the same manufacturer can share models through that manufacturer's intelligent computing center.

[0180] Example 3, taking a deployment method where the first communication device is a base station or UE, and the second communication device is a smart computing center (such as a core network element) located in the core network, as shown in Figure 9, illustrates an example of another communication network. Here, there are one or more first communication devices.

[0181] In Example 3, one possible scenario is that the operator deploys a smart computing center in the core network to provide the ability to train models for base stations and UEs, and the air interface connection between the UE and the base station provides a pipeline for the UE to request and download models.

[0182] In summary, by performing matching analysis on models in the existing model library to reduce the need for model training, the training of neural network models can be migrated to the cloud, and only the inference of the model can be performed on local devices, so that neural networks can also be deployed on inexpensive devices.

[0183] It is understood that, in order to achieve the above-mentioned functions, the model management device includes hardware structures and / or software modules corresponding to the execution of each function. Those skilled in the art should readily recognize that, based on the algorithmic steps of the examples described in conjunction with the embodiments of this disclosure, this disclosure can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed in hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this disclosure.

[0184] This disclosure embodiment can divide the model management device into functional modules according to the above method embodiment. For example, each function can be divided into a separate functional module, or two or more functions can be integrated into one functional module. The integrated module can be implemented in hardware or software. It should be noted that the module division in this disclosure embodiment is illustrative and only represents one logical functional division. In actual implementation, there may be other division methods. The following description uses the example of dividing each functional module according to each function.

[0185] Figure 10 is a block diagram of a model management device according to some embodiments. The model management device can be applied to a first communication device and execute the model management method shown in Figure 3 above, as well as the embodiment on the first communication device side in Figure 6. As shown in Figure 10, the model management device 1000 includes: a processing module 1001.

[0186] The processing module 1001 is used to update the initial model to obtain the target model when the currently running initial model does not match the current environment; the processing module 1001 is also used to switch to the target model.

[0187] In some embodiments, the processing module 1001 is specifically used to determine an update strategy based on the matching status of multiple first historical models stored in the first communication device with the current environment; the processing module 1001 is also used to update the initial model to obtain the target model based on the update strategy.

[0188] In some embodiments, if there is at least one first model among multiple first historical models that has a matching degree greater than a first preset matching threshold with the current environment, the update strategy includes: taking the first model with the highest matching degree with the current environment among the at least one first model as the target model.

[0189] In some embodiments, if there is at least one second model among multiple first historical models whose matching degree with the current environment is less than or equal to a first preset matching threshold and greater than a second preset matching threshold, the update strategy includes: fine-tuning the second model among at least one second model whose matching degree with the current environment is the highest, and using the fine-tuned second model as the target model.

[0190] In some embodiments, when the matching degree between multiple first historical models and the current environment is less than or equal to a second preset matching threshold, and the matching degree between the initial model and the current environment is greater than the second preset matching threshold, the update strategy includes: fine-tuning the initial model and using the fine-tuned initial model as the target model.

[0191] In some embodiments, when the matching degree between multiple first historical models and the current environment is less than or equal to a second preset matching threshold, and the matching degree between the initial model and the current environment is less than or equal to the second preset matching threshold, the update strategy includes: requesting a second communication device to determine the target model, wherein the second computing power resources used for training the model in the second communication device are more than the first computing power resources used for training the model in the first communication device.

[0192] In some embodiments, the update strategy includes: requesting a second communication device to determine a target model; the model management device 1000 further includes: a sending module 1002 and a receiving module 1003. The sending module is used to send request information to the second communication device, the request information being used to request a model that matches the current environment; the receiving module 1003 is used to receive a third model from the second communication device in response to the request information; the processing module 1001 is further used to update the initial model based on the third model to obtain the target model.

[0193] In some embodiments, the request information includes at least one of the following:

[0194] Dataset used for model training;

[0195] Data augmentation algorithms used to expand datasets;

[0196] Environmental characteristics information of the current environment;

[0197] Match the model requirements information to the current environment.

[0198] In some embodiments, the third model is a model determined by the second communication device from a plurality of stored second historical models based on environmental feature information and model feature information.

[0199] In some embodiments, the third model is the model among multiple second historical models whose matching degree with the current environment is greater than a first preset threshold.

[0200] In some embodiments, the third model is a model fine-tuned by the second communication device based on the dataset and data augmentation algorithm on the fourth model among multiple second historical models; here, the matching degree of the fourth model with the current environment is less than or equal to the first preset threshold and greater than the second preset threshold.

[0201] In some embodiments, the third model is a model trained from scratch by the second communication device based on a dataset and a data augmentation algorithm.

[0202] In some embodiments, the processing module 1001 is further configured to store the target model.

[0203] In some embodiments, the processing module 1001 is further configured to establish a correspondence between the target model and the current environment.

[0204] In some embodiments, the processing module 1001 is further configured to delete the models with the lowest usage frequency and / or shortest usage time among a plurality of first historical models when the available storage space of the first communication device for the model is less than a preset capacity threshold.

[0205] In some embodiments, the processing module 1001 is further configured to switch the target model to the initial model or the preset model if the target model does not match the environment within a preset time period.

[0206] In some embodiments, the first communication device and the second communication device satisfy at least one of the following:

[0207] The first communication device is a user equipment, and the second communication device is a base station;

[0208] The first communication device is a user equipment, and the second communication device is an internet device;

[0209] The first communication device is a user equipment, and the second communication device is a core network element;

[0210] The first communication device is a base station, and the second communication device is a core network element.

[0211] Figure 11 is a block diagram of a model management device according to some embodiments. The model management device can be applied to a second communication device and execute the model management method shown in Figure 5 above, as well as the embodiment on the second communication device side in Figure 6. As shown in Figure 11, the model management device 1100 includes a receiving module 1101 and a transmitting module 1102.

[0212] The receiving module 1101 is used to receive request information sent by the first communication device. The request information is used to request a model that matches the current environment of the first communication device. The second computing resources used for training the model in the second communication device are greater than the first computing resources used for training the model in the first communication device. The sending module 1102 is used to send a third model that matches the current environment to the first communication device.

[0213] In some embodiments, the request information includes at least one of the following:

[0214] Dataset used for model training;

[0215] Data augmentation algorithms used to expand datasets;

[0216] Environmental characteristics information of the current environment;

[0217] Match the model requirements information to the current environment.

[0218] In some embodiments, the third model is a model determined by the second communication device from a plurality of stored second historical models based on environmental feature information and model feature information.

[0219] In some embodiments, the third model is the model among multiple second historical models whose matching degree with the current environment is greater than a first preset threshold.

[0220] In some embodiments, the third model is a model fine-tuned by the second communication device based on the dataset and data augmentation algorithm on the fourth model among multiple second historical models; here, the matching degree of the fourth model with the current environment is less than or equal to the first preset threshold and greater than the second preset threshold.

[0221] In some embodiments, the third model is a model trained from scratch by the second communication device based on a dataset and a data augmentation algorithm.

[0222] In some embodiments, the model management device 1100 further includes a processing module 1103. The processing module 1103 is used to store a third model.

[0223] In some embodiments, the processing module 1103 is further configured to establish a correspondence between the third model and the current environment.

[0224] In some embodiments, the processing module 1103 is further configured to delete the models with the lowest usage frequency and / or shortest usage time among a plurality of second historical models when the available storage space of the second communication device for the model is less than a preset capacity threshold.

[0225] In some embodiments, the first communication device and the second communication device satisfy at least one of the following:

[0226] The first communication device is a user equipment, and the second communication device is a base station;

[0227] The first communication device is a user equipment, and the second communication device is an internet device;

[0228] The first communication device is a user equipment, and the second communication device is a core network element;

[0229] The first communication device is a base station, and the second communication device is a core network element.

[0230] In implementing the functions of the integrated modules described above in hardware, this disclosure provides another block diagram three of the model management device involved in the above embodiments. As shown in FIG12, the model management device 1200 includes a processor 1202 and a bus 1204. In some embodiments, the model management device may further include a memory 1201. In some embodiments, the model management device may further include a communication interface 1203.

[0231] Processor 1202 may implement or execute various exemplary logic blocks, modules, and circuits described in conjunction with embodiments of this disclosure. Processor 1202 may be a central processing unit, a general-purpose processor, a digital signal processor, an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It may implement or execute various exemplary logic blocks, modules, and circuits described in conjunction with embodiments of this disclosure. Processor 1202 may also be a combination that implements computing functions, for example, including one or more microprocessor combinations, a combination of a digital signal processor (DSP) and a microprocessor, etc.

[0232] The communication interface 1203 is used to connect with other devices via a communication network. This communication network can be Ethernet, wireless access network, wireless local area network (WLAN), etc.

[0233] The memory 1201 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), disk storage medium or other magnetic storage device, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but is not limited thereto.

[0234] In some embodiments, the memory 1201 may exist independently of the processor 1202. The memory 1201 may be connected to the processor 1202 via a bus 1204 and may be used to store instructions or program code. When the processor 1202 calls and executes the instructions or program code stored in the memory 1201, it can implement the model management method provided in the embodiments of this disclosure.

[0235] In other embodiments, the memory 1201 may also be integrated with the processor 1202.

[0236] Bus 1204 can be an extended industry standard architecture (EISA) bus, etc. Bus 1204 can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in Figure 12, but this does not mean that there is only one bus or one type of bus.

[0237] Some embodiments of this disclosure provide a computer-readable storage medium (e.g., a non-transitory computer-readable storage medium) storing computer program instructions that, when executed on a computer, cause the computer to perform the model management method as described in any of the above embodiments.

[0238] Exemplary examples show that the aforementioned computer-readable storage media may include, but are not limited to: magnetic storage devices (e.g., hard disks, floppy disks, or magnetic tapes), optical discs (e.g., compact disks (CDs), digital versatile disks (DVDs), etc.), smart cards, and flash memory devices (e.g., erasable programmable read-only memory (EPROMs), cards, sticks, or key drives, etc.). The various computer-readable storage media described in this disclosure may represent one or more devices and / or other machine-readable storage media for storing information. The term "machine-readable storage medium" may include, but is not limited to, wireless channels and various other media capable of storing, containing, and / or carrying instructions and / or data.

[0239] This disclosure provides a computer program product containing instructions that, when run on a computer, cause the computer to execute the model management method described in any of the above embodiments.

[0240] The above description is merely a specific embodiment of this disclosure, but the scope of protection of this disclosure is not limited thereto. Any changes or substitutions within the technical scope disclosed in this disclosure should be included within the scope of protection of this disclosure. Therefore, the scope of protection of this disclosure should be determined by the scope of the claims.

Claims

1. A model management method, wherein, The application is applied to a first communication device, comprising: In the case that the currently running initial model does not match the current environment, updating the initial model to obtain a target model; Switching to use the target model.

2. The method of claim 1, wherein, Updating the initial model to obtain the target model comprises: Determining an update strategy based on the matching of a plurality of first historical models stored by the first communication device and the current environment; Updating the initial model to obtain a target model based on the update strategy.

3. The method of claim 2, wherein, In the case that there is at least one first model in the plurality of first historical models that matches the current environment to a degree greater than a first preset matching threshold, the update strategy comprises: taking the first model that matches the current environment to the greatest degree as the target model.

4. The method of claim 2, wherein, In the case that there is at least one second model in the plurality of first historical models that matches the current environment to a degree less than or equal to a first preset matching threshold and greater than a second preset matching threshold, the update strategy comprises: fine-tuning the second model that matches the current environment to the greatest degree among the at least one second model, and taking the fine-tuned second model as the target model.

5. The method of claim 2, wherein, In the case that the matching degrees of the plurality of first historical models and the current environment are all less than or equal to a second preset matching threshold, and the matching degree of the initial model and the current environment is greater than the second preset matching threshold, the update strategy comprises: fine-tuning the initial model, and taking the fine-tuned initial model as the target model.

6. The method of claim 2, wherein, In the case that the matching degrees of the plurality of first historical models and the current environment are all less than or equal to a second preset matching threshold, and the matching degree of the initial model and the current environment is less than or equal to the second preset matching threshold, the update strategy comprises: requesting a second communication device to determine the target model, the second communication device having more second computing resources for training a model than the first communication device.

7. The method of claim 6, wherein, The update strategy comprises: requesting the second communication device to determine the target model; and updating the initial model to obtain a target model based on the update strategy comprises: Sending request information to the second communication device, the request information being used to request a model that matches the current environment; Receiving a third model sent by the second communication device in response to the request information; Updating the initial model based on the third model to obtain the target model.

8. The method of claim 7, wherein, The request information comprises at least one of: A data set used for model training; A data enhancement algorithm used for expanding the data set; Environment feature information of the current environment; Model demand information that matches the current environment.

9. The method of claim 8, wherein, The third model is a model determined by the second communication device from a plurality of second historical models stored by the second communication device based on the environment feature information and the model feature information.

10. The method of claim 9, wherein, The third model is a model in the plurality of second historical models that matches the current environment to a degree greater than a first preset threshold.

11. The method of claim 9, wherein, The third model is a model that is fine-tuned by the second communication device on a fourth model in the plurality of second historical models based on the data set and the data enhancement algorithm. Here, the matching degree of the fourth model with the current environment is less than or equal to a first preset threshold and greater than the second preset threshold.

12. The method of claim 8, wherein, The third model is a model that is trained from scratch by the second communication device based on the data set and the data enhancement algorithm.

13. The method of claim 2, wherein, The method further comprises: storing the target model.

14. The method of claim 13, wherein, The method further comprises: establishing a corresponding relationship between the target model and the current environment.

15. The method of claim 13, wherein, The method further comprises: in a case where the available storage space for models in the first communication device is less than a preset capacity threshold, deleting a model with the lowest frequency of use and / or the shortest length of use in the plurality of first historical models.

16. The method of claim 1, wherein, After the target model is switched to be used, the method further comprises: in a case where the target model does not match the environment within a preset time length, switching the target model to the initial model or a preset model.

17. The method of claim 6, wherein, The first communication device and the second communication device satisfy at least one of the following: The first communication device is a user equipment, and the second communication device is a base station. The first communication device is a user equipment, and the second communication device is an Internet device. The first communication device is a user equipment, and the second communication device is a core network element. The first communication device is a base station, and the second communication device is a core network element.

18. A model management method, wherein, Applied to a second communication device, comprising: receiving request information sent by a first communication device, the request information being used to request a model matching a current environment of the first communication device, second algorithmic resources for training a model in the second communication device being more than first algorithmic resources for training a model in the first communication device; sending a third model matching the current environment to the first communication device.

19. The method of claim 18, wherein, The request information comprises at least one of the following: a data set used for model training; a data enhancement algorithm used for expanding the data set; environment feature information of the current environment; model demand information matching the current environment.

20. The method of claim 19, wherein, The third model is a model determined by the second communication device from a plurality of second historical models based on the environment feature information and the model feature information.

21. The method of claim 20, wherein, The third model is a model in the plurality of second historical models with a matching degree greater than a first preset threshold with the current environment.

22. The method of claim 20, wherein, The third model is a model that is fine-tuned by the second communication device on a fourth model in the plurality of second historical models based on the data set and the data enhancement algorithm. Here, the matching degree of the fourth model with the current environment is less than or equal to a first preset threshold and greater than the second preset threshold.

23. The method of claim 19, wherein, The third model is a model that is trained from scratch by the second communication device based on the data set and the data enhancement algorithm.

24. The method of claim 18, wherein, The method further comprises: storing the third model.

25. The method of claim 24, wherein, The method further comprises: establishing a corresponding relationship between the third model and the current environment.

26. The method of claim 24, wherein, The method further comprises: In a case where available storage space of the second communication device for the model is less than a preset capacity threshold, deleting a model with a lowest frequency of use and / or a shortest length of use from the plurality of second historical models.

27. The method of claim 18, wherein, The first communication device and the second communication device satisfy at least one of the following: The first communication device is a user equipment, and the second communication device is a base station. The first communication device is a user equipment, and the second communication device is an Internet device. The first communication device is a user equipment, and the second communication device is a core network element. The first communication device is a base station, and the second communication device is a core network element.

28. A model management apparatus, wherein, Comprise: a memory and a processor; The memory and the processor are coupled; The memory is configured to store instructions executable by the processor; The processor executes the instructions to perform the method of any one of claims 1-27.

29. A computer readable storage medium, wherein, The computer readable storage medium stores computer instructions, when the computer instructions run on a computer, make the computer execute the method of any one of claims 1-27.

30. A computer program product, wherein, The computer program product comprises computer program instructions, when the computer program instructions are executed, realize the method of any one of claims 1-27.

Images