Model training method, electronic device, storage medium, and computer program product

Abstract

The present application relates to the technical field of optical access networks, and discloses a model training method, an electronic device, a storage medium, and a computer program product. The model training method comprises: sending a first model parameter of a target model to at least one slave device in an optical access network, wherein the target model is used for detecting anomaly information of the optical access network and / or detecting an environmental change of the optical access network on the basis of sensing information in the optical access network; a master device receiving a second model parameter of the target model sent by the at least one slave device, wherein the second model parameter is a model parameter of the target model obtained after the slave device uses sensing information collected by the slave device to train the target model based on the first model parameter, and the sensing information is used for indicating a network condition of the optical access network and / or an environmental state of the optical access network; and the master device updating model parameters of the target model on the basis of the received second model parameter.

Description

Model training method, electronic device, storage medium and computer program product

[0001] Cross-reference to related applications

[0002] The present application claims priority to the Chinese patent application No. 202411318436.5, filed on September 20, 2024, and entitled "Model training method, electronic device, storage medium and computer program product", the content of which is incorporated herein by reference in its entirety. TECHNICAL FIELD

[0003] Embodiments of the present application relate to the technical field of optical access networks, in particular to a model training method, an electronic device, a storage medium and a computer program product. BACKGROUND

[0004] In an optical access network, there can be many communication terminal devices inside a home. These devices can provide sensing functions in addition to communication functions, realizing the integration of sensing and communication. Using the sensing information provided by these communication terminal devices, an artificial intelligence (AI) model trained can be used to detect abnormal information or changes in the network environment. Generally, an AI model with excellent performance requires a large amount of data support, so each communication terminal device needs to provide a large amount of sensing data for the training of the AI model. However, direct data interaction can lead to security problems of private data. SUMMARY

[0005] Embodiments of the present application provide a model training method, an electronic device, a storage medium and a computer program product, which can ensure the security of private data.

[0006] In a first aspect, a model training method is provided, applied to a master device in an optical access network. The method comprises: sending, to at least one slave device in the optical access network, first model parameters of a target model, wherein the target model is used to detect abnormal information of the optical access network and / or detect changes in the environment of the optical access network based on sensing information in the optical access network; receiving, by the master device, second model parameters of the target model sent by the at least one slave device, wherein the second model parameters are model parameters of the target model obtained after the slave device trains the target model based on the first model parameters using sensing information collected by the slave device, and the sensing information is used to indicate network conditions of the optical access network and / or environmental states of the optical access network; and updating, by the master device, the model parameters of the target model based on the received second model parameters.

[0007] In a second aspect, a model training method is provided, applied to a slave device in an optical access network. The method comprises: receiving, by the slave device, first model parameters of a target model sent by a master device, wherein the target model is used to detect abnormal information of the optical access network and / or detect environmental changes of the optical access network based on sensing information in the optical access network; collecting, by the slave device, sensing information, wherein the sensing information is used to indicate network conditions of the optical access network and / or environmental states of the optical access network; training, by the slave device, the target model based on the first model parameters using the collected sensing information to obtain second model parameters of the target model; and sending, by the slave device, the second model parameters to the master device.

[0008] In a third aspect, an electronic device is provided, which comprises a processor and a memory. The memory stores programs or instructions executable on the processor. When the programs or instructions are executed by the processor, the steps of the method according to the first aspect or the steps of the method according to the second aspect are implemented.

[0009] In a fourth aspect, a readable storage medium is provided, which stores programs or instructions. When the programs or instructions are executed by a processor, the steps of the method according to the first aspect or the steps of the method according to the second aspect are implemented.

[0010] In a fifth aspect, a computer program product is provided, which comprises a computer program stored on a non-transitory computer readable storage medium. The computer program comprises program instructions which, when executed by a computer, cause the computer to perform the steps of the method according to the first aspect or the steps of the method according to the second aspect. BRIEF DESCRIPTION OF DRAWINGS

[0011] The accompanying drawings, which are incorporated herein and form part of the specification, illustrate embodiments consistent with the present application and, together with the description, further serve to explain the principles of the application.

[0012] FIG. 1 shows a flowchart of a model training method according to an example embodiment of the present application;

[0013] FIG. 2 shows a flowchart of a model training method according to another example embodiment of the present application;

[0014] FIG. 3 shows a flowchart of a model training method according to yet another example embodiment of the present application;

[0015] FIG. 4 shows three state diagrams of a master device according to an example embodiment of the present application;

[0016] FIG. 5 shows three state diagrams of a slave device according to an example embodiment of the present application;

[0017] Figure 6 shows a schematic diagram of a model training method in a PON network according to an example embodiment of the present application;

[0018] Figure 7 shows a schematic diagram of a flow executed by an OLT in state one according to an example embodiment of the present application;

[0019] Figure 8 shows a schematic diagram of a flow executed by an OLT in state two according to an example embodiment of the present application;

[0020] Figure 9 shows a schematic diagram of a flow executed by an OLT in state three according to an example embodiment of the present application;

[0021] Figure 10 shows a schematic diagram of a flow executed by an ONU in state one according to an example embodiment of the present application;

[0022] Figure 11 shows a schematic diagram of a flow executed by an ONU in state two according to an example embodiment of the present application;

[0023] Figure 12 shows a schematic diagram of a flow executed by an ONU in state three according to an example embodiment of the present application;

[0024] Figure 13 shows a schematic diagram of a flow executed by an MFU in state one according to an example embodiment of the present application;

[0025] Figure 14 shows a schematic diagram of a flow executed by an MFU in state two according to an example embodiment of the present application;

[0026] Figure 15 shows a schematic diagram of a flow executed by an MFU in state three according to an example embodiment of the present application;

[0027] Figure 16 shows a schematic diagram of a flow executed by an SFU in state one according to an example embodiment of the present application;

[0028] Figure 17 shows a schematic diagram of a flow executed by an SFU in state two according to an example embodiment of the present application;

[0029] Figure 18 shows a schematic diagram of a flow executed by an SFU in state two according to another example embodiment of the present application;

[0030] Figure 19 shows a schematic diagram of a flow executed by an ONU in state three according to an example embodiment of the present application;

[0031] Figure 20 is a block diagram of an electronic device according to an example embodiment. DETAILED DESCRIPTION

[0032] The exemplary embodiments will be described in detail below with reference to the accompanying drawings. The following description is only exemplary and is not intended to represent all embodiments consistent with the present application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the present application as detailed in the appended claims.

[0033] FIG. 1 shows a flowchart of a model training method provided by an exemplary embodiment of the present application, which can be applied to a master device in an optical access network. As shown in FIG. 1, the method mainly includes the following steps.

[0034] S110, the master device sends first model parameters of a target model to at least one slave device in the optical access network.

[0035] In the embodiments of the present application, the target model is used to detect abnormal information of the optical access network and / or detect environmental changes of the optical access network based on sensing information in the optical access network.

[0036] In the embodiments of the present application, the first model parameters are current model parameters of the target model, that is, the first model parameters are model parameters of the target model before starting the current round of training. For example, the master device can pre-train the target model, and send the model parameters of the target model after the pre-training to each slave device. Alternatively, the master device can initiate the next round of training based on the training result after completing a round of training by the slave device, and send the model parameters of the target model after the last round of training to each slave device.

[0037] The target model in the embodiments of the present application can be an AI model trained by unsupervised learning, for example, the target model can be trained by federated learning. Federated learning is a distributed machine learning framework with privacy protection and secure encryption technology, which aims to enable dispersed participants to collaborate on model training for machine learning while meeting the requirement of not disclosing private data to other participants. Federated learning replaces the original remote data transmission with encrypted parameter transmission in the training process, ensuring the security and privacy of data from all parties, while meeting the requirements of existing laws and regulations on data security.

[0038] In the embodiments of the present application, the at least one slave device can be all slave devices in the optical access network, or can be part of the slave devices in the optical access network, for example, slave devices that can perform training of the target model in the optical access network, or slave devices that have the function of collecting sensing information in the optical access network.

[0039] In some embodiments, the master device can broadcast the first model parameter to at least one slave device in the optical access network in a broadcast manner, and the slave device supporting the training of the target model can train the target model based on the received first model parameter, thereby saving network resources.

[0040] In other embodiments, the master device can also send the first model parameter to each slave device in a unicast manner one by one, for example, the master device can unicast the first model parameter to each slave device supporting the training of the target model, thereby avoiding unnecessary devices receiving the first model parameter.

[0041] In S112, the master device receives the second model parameter of the target model sent by the at least one slave device.

[0042] In the embodiments of the present application, the second model parameter is the model parameter of the target model obtained after the slave device trains the target model based on the first model parameter using the collected perception information, and the collected perception information is used to indicate the network condition of the optical access network and / or the environmental state of the optical access network. That is, the second model parameter is the model parameter of the target model after the slave device trains the target model using the collected perception information.

[0043] In S114, the master device updates the model parameter of the target model based on the received second model parameter.

[0044] In the embodiments of the present application, the master device can update the model parameter of the target model based on the second model parameter sent by each slave device, thereby completing the training of the target model in this round.

[0045] In the embodiments of the present application, the model parameter includes but is not limited to gradient information, each slave device can train the target model based on the gradient value corresponding to the first model parameter and the locally collected perception information, obtain a new gradient value, and send the new gradient value to the master device, and the master device performs model aggregation based on the gradient value sent by each slave device to update the model parameter of the target model.

[0046] Through the technical solutions provided in the embodiments of the present application, in each round of training of the target model, the master device of the optical access network can send the current model parameters of the target model, i.e., the first model parameters, to at least one slave device in the optical access network, and each slave device can collect sensing information after receiving the current model parameters of the target model, train the target model by using the collected sensing information, and send the model parameters of the target model after training, i.e., the second model parameters, to the master device, so that the master device updates the model parameters of the target model based on the second model parameters sent by each slave device, thereby completing one round of training of the target model. In the technical solutions provided in the embodiments of the present application, the slave device does not need to send the collected sensing information to the master device, but uses the collected sensing information to train the target model, and transmits the model parameters of the target model between the master device and the slave device, thereby avoiding the security problem of private data caused by directly transmitting the sensing information to the master device, and ensuring the security of the private data.

[0047] In some embodiments, in order to enable the slave device to know in advance that the target model needs to be trained, before S110, the method can further include: the master device sends a first control message to the at least one slave device, the first control message being used to instruct the at least one slave device to prepare to train the target model. The slave device can know that the target model needs to be trained based on the first control message, so as to prepare in advance for the training of the target model.

[0048] In the above embodiments, the first control message can carry information of at least one of the following.

[0049] 1) Model type information of the target model; the model type information can be used to indicate the type of the target model.

[0050] 2) Number of iterations of training of the target model by the slave device; the number of iterations can be used to indicate the number of model iteration updates that the slave device needs to perform in the current round of training.

[0051] 3) Number of times of sending the model parameters of the target model by the master device. Through the number of times, it can be indicated that the master device performs the sending of the target model for the nth time.

[0052] In some embodiments, in order to enable the master device to know whether the slave device successfully receives the first model parameters, after S110, the method can further include: the master device receives an acknowledgement message sent by the at least one slave device, wherein the acknowledgement message is used to indicate that the slave device successfully receives the first model parameters.

[0053] It should be noted that the master device can receive the confirmation message after sending the first model parameter in the case of sending the first control message, or can receive the confirmation message after sending the first model parameter in the case of not sending the first control message.

[0054] In some embodiments, before S112, the method can further include the following steps: step 1, the master device receives a second control message sent by the at least one slave device, wherein the second control message is used to indicate that the slave device has completed the training of the target model; step 2, based on the second control message, the master device allocates uplink transmission resources for the at least one slave device, wherein the uplink transmission resources are used to transmit the second model parameter.

[0055] In an example, the uplink transmission resources include but are not limited to uplink bandwidth or uplink time interval.

[0056] In an example, S112 can include that the master device receives the second model parameter sent by the at least one slave device on the allocated uplink transmission resources.

[0057] Through the above embodiments, the master device can allocate uplink transmission resources for the slave device to transmit the second model parameter after receiving the second control message indicating that the slave device has completed the training of the target model, so as to ensure the timely transmission of the second model parameter and improve the training efficiency of the target model.

[0058] FIG. 2 shows a flowchart of a model training method provided by another exemplary embodiment of the present application, which can be executed by a slave device in an optical access network. As shown in FIG. 2, the model training method mainly includes the following steps.

[0059] S210, the slave device receives the first model parameter of the target model sent by the master device.

[0060] In the embodiments of the present application, the target model is used to detect abnormal information of the optical access network and / or detect environmental changes of the optical access network based on sensing information in the optical access network.

[0061] Wherein, the master device can send the first model parameter by using the method shown in FIG. 1. For example, in some embodiments, the slave device receiving the first model parameter of the target model sent by the master device can include that the slave device receives the first model parameter broadcasted by the master device. In other embodiments, the slave device receiving the first model parameter of the target model sent by the master device can include that the slave device receives the first model parameter sent by the master device in a unicast manner.

[0062] In the embodiment, the slave device can be any slave device in the optical access network, or the slave device can be a specific slave device in the optical access network, for example, a slave device supporting target model training, or a slave device having a function of collecting sensing information.

[0063] S212, the slave device collects sensing information, where the sensing information is used to indicate a network condition of the optical access network and / or an environmental state of the optical access network.

[0064] S214, the slave device trains the target model based on the first model parameter to obtain a second model parameter of the target model.

[0065] In the embodiment, the slave device can train the target model using the collected sensing information to obtain the second model parameter of the target model after the training.

[0066] S216, the slave device sends the second model parameter to the master device.

[0067] Through the above technical solution provided in the embodiment, in each round of training of the target model, the slave device of the optical access network receives the current model parameter of the target model, i.e., the first model parameter, sent by the master device of the optical access network, collects sensing information, trains the target model using the collected sensing information, and sends the model parameter of the target model after the training, i.e., the second model parameter, to the master device, so that the master device can update the model parameter of the target model based on the second model parameters sent by each slave device, thereby completing one round of training of the target model. In the technical solution provided in the embodiment, the slave device does not need to send the collected sensing information to the master device, but trains the target model using the collected sensing information, and transmits the model parameter of the target model between the master device and the slave device, thereby avoiding the security problem of private data caused by directly transmitting the sensing information to the master device, and ensuring the security of the private data.

[0068] In some embodiments, in order to enable the slave device to know in advance that the target model will be trained, before S210, the method can further include: the slave device receives a first control message sent by the master device, where the first control message is used to instruct the slave device to train the target model.

[0069] In an example, the first control message can carry information of at least one of the following: model type information of the target model, iteration number of the slave device training the target model, and number of times of the master device issuing the model parameter of the target model. The slave device can perform iteration training on the target model for a corresponding number of times according to the iteration number carried in the first control message in S214.

[0070] In some embodiments, in order to enable the master device to learn that the slave device successfully receives the first model parameter, after S210, the method can further include: sending, by the slave device, an acknowledgement message to the master device, where the acknowledgement message is used to indicate that the slave device successfully receives the first model parameter.

[0071] In some embodiments, the slave device can be a device with a sensing function, and then the slave device collecting the sensing information can include: the slave device starting its sensing function, and collecting the sensing information through the sensing function.

[0072] For example, for a passive optical network (PON), the slave device is an optical network unit (ONU), each ONU can implement an intrusion detection sensing function in a different home scenario, and the ONU can start a distributed fiber sensing function to collect timing signals, i.e., sensing data.

[0073] In some embodiments, in order to save the power consumption of the slave device, after S214, the method can further include: the slave device closing its sensing function.

[0074] In other embodiments, the slave device can collect sensing information from wireless fidelity (WIFI) nodes connected thereto, and therefore, in these embodiments, the slave device collecting the sensing information can include: the slave device starting the sensing function of the WIFI nodes connected thereto, and receiving the sensing information uploaded by the WIFI nodes.

[0075] For example, for a fiber to the room (FTTR) network, the slave device is a WIFI device subordinate to a sub-FTTR unit (SFU), and the WIFI device can support a person detection function. After receiving the first model parameter, the SFU starts the sensing function of the WIFI device to collect channel state information (CSI). Since the presence of a target to be sensed (such as a human body) can interfere with wireless signals, causing changes in the wireless channel, the target can be sensed through the CSI to meet the application requirements of wireless sensing such as human presence detection, posture recognition, fall detection, and intrusion detection.

[0076] In some embodiments, in order to save the power consumption of WIFI, after S214, the method can further include: turning off the sensing function of the WIFI node connected to the slave device. In this embodiment, after the training of the target model is completed, the slave device does not need to obtain sensing information from the WIFI node for training the target model, and therefore, the slave device can control the WIFI node to turn off the sensing function to save the power consumption of the WIFI node.

[0077] In some embodiments, in order to upload the second model parameter in time, before S216, the method can include the following steps: step 1, the slave device sends a second control message to the master device, wherein the second control message is used to indicate that the slave device has completed the training of the target model; step 2, the slave device obtains the uplink transmission resource allocated by the master device for the slave device, wherein the uplink transmission resource is used to transmit the second model parameter.

[0078] Through the above-mentioned embodiments, after the training of the target model is completed, the slave device can notify the master device, so that the master device allocates the uplink transmission resource for the slave device to transmit the second model parameter.

[0079] In the above-mentioned embodiments, S216 can include: the slave device sends the second model parameter to the master device on the allocated uplink transmission resource. Through this implementation, after the training of the target model is completed, the slave device can notify the master device, so that the master device allocates the uplink transmission resource for the slave device to transmit the second model parameter, and then the slave device can transmit the second model parameter through the allocated uplink transmission resource without waiting for the available uplink transmission resource, thereby improving the transmission efficiency of the second model parameter.

[0080] FIG. 3 shows a flowchart of a model training method according to another example embodiment of the present application. In the flowchart shown in FIG. 3, the master device acts as a server role of federated learning, and initially stores an artificial intelligence model that is preliminarily trained. In the federated learning process, the master device performs transmission and reception of model parameters, aggregation of model parameters, and sending and receiving of control messages. The slave device acts as a worker role of federated learning, and in the federated learning process, the slave device performs transmission and reception of model parameters, collection of local data, training of model parameters, and receiving of control messages from the master device and responding. In the training process, data transmission is performed in the form of model parameters, which avoids possible data leakage risk and protects the privacy and security of users. At the same time, the final model training is jointly performed by data from each slave device, which enhances the model effect.

[0081] In the flow shown in FIG. 3, in the above implementation process, the master device plays a server role, as shown in FIG. 4, and has three states in total: State 1: the master device sends a message to the slave device, broadcasts the local model parameters, and waits to receive an acknowledgement message from the slave device. After receiving the acknowledgement message from the slave device, the master device leaves this state; State 2: the master device waits for an acknowledgement message from the slave device in this state, assigns uplink bandwidth to the slave device for sending the updated model parameters after receiving the message, and sends an acknowledgement message after the receiving is completed. After receiving the model parameters from all slave devices, the master device leaves this state; State 3: the master device processes the received model parameters in this state, updates the local model, and leaves this state after storing the local model.

[0082] In the flow shown in FIG. 3, in the above implementation process, the slave device plays a worker role, as shown in FIG. 5, and has three states in total: State 1: the slave device waits for a control message from the master device in this state, receives the model parameters sent by the master device, and sends an acknowledgement message after the receiving is completed and leaves this state; State 2: the slave device starts the sensing function in this state, collects sensing information, and trains the model to update the model parameters. After the training is completed, the sensing function is closed and the slave device leaves this state; State 3: the slave device sends a control message to the master device in this state, sends the trained model parameters to the master device after receiving the bandwidth allocation from the master device, and leaves this state after receiving the acknowledgement message from the master device.

[0083] As shown in FIG. 3, the model training method provided in this example mainly includes the following steps.

[0084] S301: The master device starts the training process of the target model.

[0085] S302: The master device notifies the slave device to train the target model.

[0086] From this step, the master device enters State 1 of the master device.

[0087] After receiving the notification from the master device, the slave device enters State 1 of the slave device.

[0088] S303: The master device sends the current model parameters of the target model, i.e., the first model parameters.

[0089] The master device can broadcast the above model parameters in the optical access network, or can send the above model parameters to the slave devices in the optical access network in a unicast manner.

[0090] S304: The slave device receives the model parameters sent by the master device.

[0091] S305: The slave device sends an acknowledgement message to the master device, which indicates that the slave device has successfully received the above model parameters through the acknowledgement message.

[0092] After receiving the confirmation message, the master device leaves the state one of the master device and enters the state two of the master device.

[0093] After sending the confirmation message, the slave device leaves the state one of the slave device and enters the state two of the slave device.

[0094] S306, the slave device starts the sensing function and collects sensing information.

[0095] The slave device can start the sensing function of the slave device itself and collect sensing information, or start the sensing function of the WIFI node subordinate to the slave device and collect sensing information through the WIFI node.

[0096] S307, the slave device uses the collected sensing information to train the target model and update the model parameters of the target model.

[0097] S308, the slave device stops the sensing function.

[0098] If the sensing function of the slave device itself is started, the sensing function of the slave device itself is stopped, and if the sensing function of the WIFI node subordinate to the slave device is started, the sensing function of the WIFI node subordinate to the slave device is stopped.

[0099] After the step is performed, the slave device leaves the state two of the slave device.

[0100] S309, the slave device notifies the master device that the training is completed.

[0101] The slave device enters the state three of the slave device.

[0102] S310, after receiving the notification of the slave device, the master device configures uplink transmission resources for the slave device.

[0103] S311, the slave device transmits the trained model parameters through the uplink transmission resources configured by the master device.

[0104] The slave device leaves the state three.

[0105] S312, the master device sends a confirmation message to the slave device to confirm that the model parameters sent by the slave device are received.

[0106] The master device leaves the state two of the slave device and enters the state three.

[0107] S313, the master device updates the model parameters of the target model based on the received trained model parameters, i.e., the second model parameters.

[0108] S314, the master device ends the current training of the target model, and if the next round of training is needed, returns to S302, otherwise, ends the training process of the target model.

[0109] The above technical solutions provided by the embodiments of the present application can realize AI model training of a slave device based on federated learning, meet sensing requirements in different scenarios, and protect user privacy.

[0110] The technical solutions provided by the embodiments of the present application will be described below by taking PON networks and FTTR networks as examples.

[0111] In a PON network, each ONU is deployed inside a home, and the sensing function of each ONU faces different home scenarios to realize the sensing function of intrusion detection. Since the difference and randomness of the environment where the ONU is located cannot be artificially deployed in advance, a federated learning method can be used to train the artificial intelligence model at each ONU online to support the intrusion detection requirements of each ONU.

[0112] In this scenario, the master device of the above method is an optical line termination (OLT), and the slave device is an ONU, wherein the OLT and the ONU additionally collect an artificial intelligence module, which has a computing function and a storage function, and can perform model training and application and model storage. An example of the AI model is a Long Short-Term Memory (LSTM) model, the input of which is waveform data, for example, a time series signal sensed by a distributed optical fiber of the ONU, and the output can be the judgment result of an intrusion event: there is an intrusion / no intrusion.

[0113] In the present embodiment, as shown in FIG. 6, the execution process of the OLT is as follows.

[0114] State one: In this state, the OLT starts the training process and issues local model parameters to the ONU, as shown in FIGS. 6 and 7, which can include the following steps.

[0115] S601, the OLT sends a broadcast or unicast management message 1, for example, a Physical layer Operations, Administration and Maintenance (PLOAM) message 1 or an ONT Management and Control Interface (OMCI) message 1, to notify the ONU to enter the federated learning process and configure the training parameters (including the model type, the number of node iterations, the global iteration number, the maximum number of global iterations, etc.).

[0116] In some embodiments, an example structure of the PLOAM message 1 in S601 can be as shown in Table 1, and an example structure of the OMCI message 1 can be as shown in Table 2.

[0117] Table 1.

[0118] Wherein, the model type, the node iteration number and the global iteration number are new fields. The model type indicates the type of artificial intelligence model; the node iteration number represents the number of model iteration updates required when the ONU side performs model training; and the global iteration number represents the number of local model downlink performed by the OLT.

[0119] Table 2.

[0120] Wherein, the model type, the node iteration number and the global iteration number are new fields. The model type indicates the type of artificial intelligence model; the node iteration number represents the number of model iteration updates required when the ONU side performs model training; and the global iteration number represents the number of local model downlink performed by the OLT.

[0121] S602, the OLT reads the model parameters from the artificial intelligence model, takes the read model parameters as service data, and broadcasts them to all ONUs in the downlink traffic.

[0122] S603, after the ONU receives the completion, the control message (PLOAM / OMCI) in S601 is responded, and an acknowledgement message is sent. After the OLT receives the acknowledgement messages of all ONUs, it enters the next state.

[0123] State two: the OLT allocates uplink bandwidth to the ONU for transmitting the model parameters updated by the ONU training. As shown in FIGS. 6 and 8, this state can include the following steps.

[0124] S604, the OLT receives the control message (PLOAM / OMCI) 2 from the ONU, which indicates that the ONU has completed the model training and is waiting for the upload of the model parameters. The control message can be PLOAM message 2 or OCMI message 2.

[0125] Wherein, an example structure of the PLOAM message 2 can be shown in Table 3, and an example structure of the OCMI message 2 can be shown in Table 4.

[0126] Table 3.

[0127] Wherein, the global iteration number is a new field. The global iteration number represents the number of times of model parameter uplink performed by the ONU, and its value is the same as the global iteration number of the PLOAM message 1 received by the ONU in the cycle.

[0128] Table 4.

[0129] Wherein, the global iteration number is a new field. The global iteration number indicates that the ONU performs the model parameter uplink for the first time, and the value of the global iteration number is the same as the global iteration number of the OMCI message 1 received by the ONU in the loop.

[0130] S605, the OLT sends a bandwidth map (BWmap) to allocate uplink bandwidth for the ONUs, and the uplink bandwidth is used for the ONUs to upload the model parameters.

[0131] S606, the OLT receives the model parameters from the corresponding ONUs in the uplink bandwidth allocated in S605, and stores the received model parameters in the artificial intelligence module. After the OLT receives the model parameters of all the ONUs, the OLT enters the next state.

[0132] State three: the artificial intelligence module in the OLT processes the model parameters obtained in state two and updates the local model. As shown in FIG. 6 and FIG. 9, the state can include the following steps.

[0133] S607, the artificial intelligence module processes the model parameters from all the ONUs, for example, average operation, obtains new model parameters, and stores the new model parameters as the local model of the OLT. The artificial intelligence module informs the OLT that the current round of model training has been completed. If the iteration round is less than the global iteration number in S601, the next round of training is performed; if the iteration round is equal to the global iteration number in S601, the training is ended.

[0134] In the embodiment, the execution process of the OLT is as follows.

[0135] State one: the ONU receives the model parameters from the OLT. As shown in FIG. 10, the state one of the ONU mainly includes the following steps.

[0136] S1001, the ONU receives the management and control message (PLOAM / OMCI) from the OLT, starts the training process, and obtains the related parameters of the training.

[0137] S1002, the ONU receives the local model parameters sent by the OLT.

[0138] S1003, the ONU stores the received local model parameters into the artificial intelligence module on the ONU side.

[0139] S1004, the ONU sends an acknowledgement message (PLOAM / OMCI) to inform the OLT that the model parameter reception has been completed, and enters the next state after the reception is completed.

[0140] State two: the ONU starts the sensing function training model. As shown in FIG. 11, the state two of the ONU mainly includes the following steps.

[0141] S1101, the ONU starts the sensing function of the sensing module.

[0142] S1102, the ONU starts the training process of the artificial intelligence module, the perception module inputs the real-time collected perception data into the artificial intelligence module, and the artificial intelligence module performs model training according to the parameters obtained in S1101.

[0143] S1103, the artificial intelligence module notifies the ONU after completing the training, and the ONU sends a perception function of the perception module to close, and enters the next state after closing.

[0144] State three, ONU uploads model parameters, as shown in FIG. 12, state three of the ONU mainly includes the following steps.

[0145] S1201, the ONU sends a control message (PLOAM / OMCI) to notify the OLT that the current model training has been completed, and waits for the uplink bandwidth to upload the model parameters.

[0146] S1202, the ONU receives the BWmap of the OLT to obtain the uplink bandwidth configured by the OLT.

[0147] S1203, the ONU reads the model parameters after training from the artificial intelligence module, and transmits the model parameters to the OLT using the uplink bandwidth obtained in S1202.

[0148] S1204, the ONU receives a confirmation message (PLOAM / OMCI) from the OLT, confirming that the current round of training has been completed, and enters the next state.

[0149] Through the above process, each ONU completes the training of the AI model through federated learning, and the trained model can be applied to intrusion detection at all ONUs.

[0150] In the FTTR network, each SFU is deployed inside the home, and the SFU itself or the WIFI device subordinate to it can support personnel detection function. Due to the difference and randomness of the environment where the SFU is located, the artificial intelligence model cannot be deployed in advance, and the federated learning method can be used to train the artificial intelligence model at each SFU online to support the personnel detection demand at each SFU.

[0151] In this scenario, the master device of the above method is the main FTTR unit (MFU), and the slave device is the SFU, wherein the MFU and the SFU have an artificial intelligence module, which has a computing function and a storage function, can perform model training and application and model storage, and the AI model used is an example of RCNN, the input data is the collected CSI data, and the output result is whether there is a person in the perception area and the position of the person.

[0152] In this embodiment, the implementation process of the MFU is as follows.

[0153] State one, in this state, MFU starts the training process and sends local model parameters to SFU. As shown in FIG. 13, this state mainly includes the following steps.

[0154] S1301, MFU sends a broadcast management message, for example, a fiber internal network physical layer operation management and maintenance (F-PLOAM) message, a fiber management and control interface (FMCI) message, a wireless management and control interface (WMCI) message, to inform SFU to enter the federated learning process and configure training parameters (including model type, node iteration number, global iteration number, maximum global iteration number, etc.).

[0155] S1302, MFU reads model parameters from the artificial intelligence module as service data and broadcasts them to all SFUs in the downlink traffic.

[0156] S1303, after SFU receives the message, it sends a confirmation message in response to the management message (F-PLOAM / FMCI / WMCI) in S1301. After MFU receives the confirmation messages from all SFUs, it enters the next state.

[0157] State two, in this state, MFU allocates uplink time intervals for SFUs to transmit updated model parameters after training, as shown in FIG. 14, which mainly includes the following steps.

[0158] S1401, MFU receives a management message (F-PLOAM / FMCI / WMCI) from SFU, which indicates that the SFU has completed model training and is waiting to upload model parameters.

[0159] S1402, MFU sends TAmap to allocate uplink time intervals for SFUs to upload model parameters.

[0160] S1403, MFU receives model parameters from the corresponding SFU in the uplink time interval allocated in S1402 and stores the received model parameters in the artificial intelligence module.

[0161] State three, in this state, the artificial intelligence module in MFU processes the model parameters obtained in state two and updates the local model, as shown in FIG. 15, which mainly includes the following steps.

[0162] S1501, the artificial intelligence module processes the model parameters from all SFUs, for example, takes the average, obtains new model parameters, and stores them as the local model of the MFU.

[0163] S1502, the artificial intelligence model informs the MFU that the current round of model training has been completed. If the iteration round is less than the maximum global iteration number, the next round of training is performed; if the iteration round is equal to the maximum global iteration number in S1301, the training is ended.

[0164] In this embodiment, the execution process of the SFU is as follows.

[0165] State one, in this state, the SFU receives the model parameters from the MFU, as shown in FIG. 16, which mainly includes the following steps.

[0166] S1601, the SFU receives the management and control message (F-PLOAM / FMCI / WMCI) from the MFU, starts the training process, and obtains the related parameters of the training.

[0167] S1602, the SFU receives the local model parameters sent by the MFU.

[0168] S1603, the SFU stores the received local model parameters into the artificial intelligence module on the SFU side.

[0169] S1604, the SFU sends a confirmation management and control message (F-PLOAM / FMCI / WMCI) to inform that the SFU has completed the model parameter reception, and enters the next state after sending.

[0170] State two, the SFU starts the training model of the perception function, in an embodiment, as shown in FIG. 17, which mainly includes the following steps.

[0171] S1701, the SFU starts the perception function of the perception module.

[0172] S1702, the SFU starts the training process of the artificial intelligence module, the perception module inputs the real-time collected perception data into the artificial intelligence module, and the artificial intelligence module performs model training according to the parameters obtained in S1701.

[0173] S1703, after the artificial intelligence module completes the training, it informs the SFU, the SFU closes the perception function of the perception module, and after closing, the SFU enters the next state.

[0174] In another embodiment, as shown in FIG. 18, state two can include the following steps.

[0175] S1801, the SFU sends a message (such as a WiFi management and control message) to start the perception function of the WiFi node.

[0176] S1802, the SFU starts the training process of the artificial intelligence module, and the WiFi node sends the perception data collected in real time to the SFU, the SFU inputs the perception data into the artificial intelligence module, and the artificial intelligence module performs model training according to the parameters obtained in step one.

[0177] S1803, after the artificial intelligence module completes the training, the SFU is notified, and the SFU sends a message (such as a WiFi control message) to close the perception function of the WiFi node, and after being closed, the SFU enters the next state.

[0178] State three, SFU uploads model parameters, as shown in FIG. 19, which mainly includes the following steps.

[0179] S1901, the SFU sends a control message (F-PLOAM / FMCI / WMCI) to notify the MFU that the current model training has been completed, and waits for an uplink time interval to upload the model parameters.

[0180] S1902, the SFU receives the TAmap of the MFU and obtains the uplink time interval configured by the MFU.

[0181] S1903, the SFU reads the model parameters after training from the artificial intelligence module, and uses the uplink time interval obtained in S1902 to transmit the model parameters to the MFU.

[0182] S1904, the SFU receives a confirmation control message (F-PLOAM / FMCI / WMCI) from the MFU, confirming that the current round of training has been completed, and entering the next state.

[0183] Through the above technical solutions provided by the embodiments of the present application, through the signaling interaction process in the optical access network, the federated learning-based slave AI model training is realized, the sensing needs in different scenarios are met, and the safety of private data is ensured.

[0184] As shown in FIG. 20, the embodiments of the present application also provide an electronic device 2000, which includes a processor 2001 and a memory 2002, and the memory 2002 stores programs or instructions executable on the processor 2001. For example, when the electronic device 2000 is a master device, the programs or instructions are executed by the processor 2001 to implement each step of the model training method embodiment shown in FIG. 1, and achieve the same technical effects. When the electronic device 2000 is a slave device, the programs or instructions are executed by the processor 2001 to implement each step of the model training method embodiment shown in FIG. 2, and achieve the same technical effects. To avoid repetition, they will not be described here.

[0185] In an exemplary embodiment, a readable storage medium is also provided, in which a program or instruction is stored, the storage program or instruction is executed by a processor to implement all or part of the steps of the model training method, and the same technical effects can be achieved, and here will not be repeated to avoid repetition.

[0186] The processor can be the processor in the master device or the slave device described in the above embodiments. The readable storage medium includes a computer readable storage medium, such as a computer readable only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc. In some examples, the readable storage medium can be a non-transitory readable storage medium.

[0187] In an exemplary embodiment, a computer program product is also provided, the computer program / product is stored in a storage medium, and the computer program / product is executed by at least one processor to implement various processes of the above model training method embodiments, and the same technical effects can be achieved, and here will not be repeated to avoid repetition.

[0188] Other embodiments of the application will be apparent to those skilled in the art from consideration of the specification and practice of the application disclosed herein. It is intended that the specification and examples be considered as exemplary only, with the true scope and spirit of the application being indicated by the following claims.

[0189] It should be understood that the application is not limited to the precise construction that has been described above and shown in the accompanying drawings, and that various modifications and changes can be made by those skilled in the art without departing from the scope of the application. The scope of the application is limited only by the appended claims.

Claims

1. A model training method applied to a master device in an optical access network, the method comprising: sending, by the master device, first model parameters of a target model to at least one slave device in the optical access network, wherein the target model is used to detect abnormal information of the optical access network and / or detect environmental changes of the optical access network based on sensing information in the optical access network; receiving, by the master device, second model parameters of the target model sent by the at least one slave device, wherein the second model parameters are model parameters of the target model obtained after the slave device trains the target model based on the first model parameters using sensing information collected by the slave device, the sensing information being used to indicate network conditions of the optical access network and / or environmental states of the optical access network; updating, by the master device, the model parameters of the target model based on the received second model parameters.

2. The method of claim 1, wherein, Before the master device sends the first model parameters of the target model, the method further comprises: sending, by the master device, a first control message to the at least one slave device, wherein the first control message is used to instruct the at least one slave device to prepare to train the target model.

3. The method of claim 2, wherein, The first control message carries at least one of the following: model type information of the target model, the number of iterations of training the target model by the slave device, and the number of times of sending model parameters of the target model by the master device.

4. The method according to any one of claims 1 to 3, wherein, After the master device sends the first model parameters of the target model, the method further comprises: receiving, by the master device, an acknowledgement message sent by the at least one slave device, wherein the acknowledgement message is used to indicate that the slave device successfully receives the first model parameters.

5. The method according to any one of claims 1 to 3, wherein, Before the master device receives the second model parameters of the target model sent by the at least one slave device, the method further comprises: receiving, by the master device, a second control message sent by the at least one slave device, wherein the second control message is used to indicate that the slave device has completed training of the target model; allocating, by the master device, uplink transmission resources for the at least one slave device based on the second control message, wherein the uplink transmission resources are used to transmit the second model parameters.

6. The method of claim 5, wherein, The master device receiving the second model parameters of the target model sent by the at least one slave device comprises: receiving, by the master device, the second model parameters sent by the at least one slave device on the allocated uplink transmission resources.

7. The method according to any one of claims 1 to 3, wherein, The master device sending the first model parameters of the target model to the at least one slave device in the optical access network comprises: the master device broadcasting the first model parameters to the at least one slave device; or the master device sending the first model parameters to each of the slave devices in a unicast manner. 8.A model training method applied to a slave device in an optical access network, the method comprising: receiving, by the slave device, first model parameters of a target model sent by a master device, wherein the target model is used to detect abnormal information of the optical access network and / or detect environmental changes of the optical access network based on sensing information in the optical access network; The slave device collects sensing information, wherein the sensing information is used to indicate network conditions of the optical access network and / or environmental states of the optical access network. The slave device trains the target model based on the first model parameter using the collected sensing information to obtain second model parameters of the target model. The slave device sends the second model parameters to the master device.

9. The method of claim 8, wherein, Before the slave device receives the first model parameters of the target model sent by the master device, the method further comprises: The slave device receives a first control message sent by the master device, wherein the first control message is used to instruct the slave device to train the target model.

10. The method of claim 9, wherein, The first control message carries at least one of the following: model type information of the target model, the number of iterations of training the target model by the slave device, and the number of times of sending model parameters of the target model by the master device.

11. The method according to any one of claims 8 to 10, wherein, After the slave device receives the first model parameters of the target model sent by the master device, the method further comprises: The slave device sends an acknowledgement message to the master device, wherein the acknowledgement message is used to indicate that the slave device successfully receives the first model parameters.

12. The method of claim 8, wherein, The slave device collects sensing information, comprising: the slave device turning on a sensing function of the slave device, and collecting the sensing information through the sensing function; After the slave device trains the target model based on the first model parameter using the collected sensing information, the method further comprises: the slave device turning off the sensing function of the slave device.

13. The method of claim 8, wherein, The slave device collects sensing information, comprising: the slave device turning on a sensing function of a wireless fidelity (WIFI) node connected to the slave device, and receiving sensing information uploaded by the WIFI node; After the slave device trains the target model based on the first model parameter using the collected sensing information, the method further comprises: the slave device turning off the sensing function of the WIFI node connected to the slave device.

14. The method of any one of claims 8 to 10, wherein, Before the slave device sends the second model parameters to the master device, the method further comprises: The slave device sends a second control message to the master device, wherein the second control message is used to indicate that the slave device has completed training of the target model; The slave device obtains uplink transmission resources allocated by the master device for the slave device, wherein the uplink transmission resources are used to transmit the second model parameters.

15. The method of claim 14, wherein, The slave device sends the second model parameters to the master device, comprising: The slave device sends the second model parameters to the master device on the allocated uplink transmission resources.

16. The method of any one of claims 8 to 10, wherein, The slave device receives the first model parameters of the target model sent by the master device, comprising: The slave device receives the first model parameters broadcasted by the master device; or The slave device receives the first model parameters sent by the master device in a unicast manner.

17. An electronic device comprising a processor and a memory, the memory storing a program or instructions executable on the processor, the program or instructions, when executed by the processor, implement the steps of the model training method according to any one of claims 1 to 16.

18. A readable storage medium having stored thereon a program or instructions, the program or instructions, when executed by a processor, implement the steps of the model training method according to any one of claims 1 to 16.

19. A computer program product comprising a computer program stored on a non-transitory computer-readable storage medium, the computer program comprising program instructions that, when executed by a computer, cause the computer to perform the steps of the model training method according to any one of claims 1 to 16.

Images