Data prediction method, device and equipment

Abstract

The invention provides a data prediction method, device and equipment, and the method comprises the steps: obtaining a first data sequence of a target optical module in a historical time interval, and the first data sequence comprises the real operation data of a plurality of time points in the historical time interval; the real operation data comprises the voltage of the target optical module and / or the number of port wrong packets, and the real operation data further comprises at least one of the temperature, the bias current, the receiving power and the sending power corresponding to each channel of the target optical module; determining the type of a target optical module based on the number of target channels of the target optical module, and selecting a target prediction model corresponding to the type of the target optical module from all prediction models; and inputting the first data sequence into the target prediction model to obtain an initial prediction result, and determining a target prediction result based on the initial prediction result. According to the technical scheme, the method has the characteristics of high prediction efficiency, high prediction accuracy and the like.

Description

Technical Field

[0001] This application relates to the field of communication technologies, and in particular, to a data prediction method, apparatus, and device. Background Art

[0002] An optical module is composed of optoelectronic devices, functional circuits, optical interfaces, etc. The optical module includes two parts: transmitting and receiving. The function of the optical module is that the optical module at the transmitting end can convert an electrical signal into an optical signal for transmission through an optical fiber, and the optical module at the receiving end can receive the optical signal through the optical fiber and convert the optical signal into an electrical signal. In summary, data transceiver on the network interface needs to be carried out through the optical module.

[0003] When the optical module runs for a long time, it will cause performance attenuation of the optical module, resulting in unstable links. However, this unstable "sub-healthy" state has neither fault alarms nor affects the integrity of data transceiver, thus leading to a decline in the quality of service provided by the network to the service, making the network in a critical state of intermittent interruption between "usable" and "unusable", which greatly affects the perception of service quality. Summary of the Invention

[0004] This application provides a data prediction method, and the method includes: Obtain a first data sequence of a target optical module of a network device in a historical time interval, where the target optical module is any one of all optical modules of the network device, and the first data sequence includes true operation data of multiple time points in the historical time interval; wherein, the true operation data includes the voltage and / or the number of port mispackets of the target optical module, and the true operation data further includes at least one of temperature, bias current, received power, and transmitted power corresponding to each channel of the target optical module; Determine the type of the target optical module based on the number of target channels of the target optical module, and select a target prediction model corresponding to the type of the target optical module from all prediction models supported by the network device; Input the first data sequence into the target prediction model to obtain an initial prediction result, and determine a target prediction result based on the initial prediction result; wherein, the target prediction result includes a second data sequence in a future time interval, and the second data sequence includes predicted operation data of multiple time points in the future time interval, and the predicted operation data is used to determine the health state of the target optical module.

[0005] This application provides a data prediction device, which includes: an acquisition module for acquiring a first data sequence of a target optical module of a network device in a historical time interval, where the target optical module is any one of all optical modules of the network device, and the first data sequence includes true operation data at multiple time points in the historical time interval; wherein the true operation data includes the voltage and / or the number of port mispackets of the target optical module, and the true operation data further includes at least one of the temperature, bias current, received power, and transmitted power corresponding to each channel of the target optical module; a determination module for determining a target optical module type based on the number of target channels of the target optical module, and selecting a target prediction model corresponding to the target optical module type from all prediction models supported by the network device; a processing module for inputting the first data sequence into the target prediction model to obtain an initial prediction result, and determining a target prediction result based on the initial prediction result; wherein the target prediction result includes a second data sequence in a future time interval, and the second data sequence includes predicted operation data at multiple time points in the future time interval, and the predicted operation data is used to determine the health state of the target optical module.

[0006] This application provides an electronic device, which includes: a processor and a machine-readable storage medium, where the machine-readable storage medium stores machine-executable instructions that can be executed by the processor; the processor is used to execute the machine-executable instructions to implement the data prediction method in the above example of this application.

[0007] This application provides a computer program product, which includes a computer program that implements the data prediction method in the above example of this application when executed by a processor.

[0008] This application provides a machine-readable storage medium, which stores machine-executable instructions that can be executed by a processor; wherein the processor is used to execute the machine-executable instructions, and when the machine-executable instructions are executed, the data prediction method in the above example of this application is implemented.

[0009] As can be seen from the above technical solutions, in the embodiments of this application, a first data sequence of a target optical module in a historical time interval can be acquired, the first data sequence is input into a target prediction model to obtain an initial prediction result, and a target prediction result is determined based on the initial prediction result. The target prediction result includes predicted operation data in a future time interval, and the predicted operation data is used to determine the health state of the target optical module. In this way, it is possible to predict the target optical module, obtain the predicted operation data for a period of time in the future, thereby predicting the state change trend of the target optical module and monitoring the health state of the target optical module. Data prediction through the target prediction model has the characteristics of high prediction efficiency, high prediction accuracy, and good long-term prediction effect.

[0010] The first data sequence includes the voltage of the target optical module, the number of mispackets at the port, the temperature, bias current, received power, and transmitted power corresponding to each channel. In this way, by combining the actual operating data of the target optical module and the actual operating data of each channel of the target optical module, the target optical module is predicted. Then, the channel fusion method is adopted to fuse the data of all channels for prediction, further improving the prediction accuracy.

[0011] Determine the type of the target optical module according to the number of target channels of the target optical module, select the target prediction model corresponding to the type of the target optical module from all prediction models, and predict the target optical module through the target prediction model. In this way, the target prediction model most suitable for the target optical module can be selected for prediction, further improving the prediction accuracy, improving the prediction speed, and improving the performance of the optical module prediction algorithm. Description of the Drawings

[0012] Figure 1 is a schematic flowchart of the data prediction method in an embodiment of the present application; Figure 2 is a schematic flowchart of the data prediction method in an embodiment of the present application; Figure 3 is a schematic structural diagram of the target prediction model in an embodiment of the present application; Figure 4 is a schematic structural diagram of the target prediction model in an embodiment of the present application; Figure 5 is a schematic structural diagram of the data prediction device in an embodiment of the present application; Figure 6 is a hardware structure diagram of the electronic device in an embodiment of the present application. Detailed Embodiments

[0013] In an embodiment of the present application, a data prediction method is proposed. This method can be applied to an electronic device. The electronic device can be a network device (such as a router, a switch, etc.), a server, or a cloud device. The type of this electronic device is not limited as long as it can implement the data prediction method. Refer to Figure 1 As shown, it is a schematic flowchart of the data prediction method. The method may include: Step 101: Obtain a first data sequence of a target optical module of a network device in a historical time interval. The target optical module can be any one of all the optical modules of the network device. The first data sequence can include the actual operation data at multiple time points in the historical time interval. Among them, the actual operation data can include the voltage of the target optical module and / or the number of port mispackets; the actual operation data also includes at least one of the temperature, bias current, received power, and transmitted power corresponding to each channel of the target optical module.

[0014] For example, if this method is applied to a network device, the network device itself obtains the first data sequence and performs data prediction based on the first data sequence. Or, if this method is applied to other devices outside the network device, such as a server, the network device itself obtains the first data sequence and sends the first data sequence to the server, and the server performs data prediction based on the first data sequence.

[0015] Step 102: Determine the type of the target optical module based on the number of target channels of the target optical module, and select a target prediction model corresponding to the type of the target optical module from all the prediction models supported by the network device.

[0016] Step 103: Input the first data sequence into the target prediction model to obtain an initial prediction result, and determine the target prediction result based on the initial prediction result. Among them, the target prediction result can include a second data sequence in a future time interval, and the second data sequence can include the predicted operation data at multiple time points in the future time interval. The predicted operation data is used to determine the health status of the target optical module.

[0017] In an example, if the network device includes optical modules of K types of optical module types, the network device supports K prediction models, and the K prediction models correspond one-to-one with the K types of optical module types, where K is a positive integer. Based on this, when selecting a target prediction model corresponding to the type of the target optical module from all the prediction models supported by the network device, the target prediction model can be the prediction model corresponding to the type of the target optical module.

[0018] In an example, determining the type of the target optical module based on the number of target channels of the target optical module can include, but is not limited to: based on the number of target channels of the target optical module, the optical module type corresponding to this number of target channels can be determined as the type of the target optical module. Among them, all the optical modules of the network device support K types of channel numbers in total, and the K types of channel numbers correspond one-to-one with the K types of optical module types. Or, Based on the target channel number and the target FEC type of the target optical module, the optical module type corresponding to the target channel number and the target FEC type can be determined as the target optical module type. Among them, all optical modules of the network device support a total of K parameter pairs, and the K parameter pairs correspond one-to-one with K types of optical module types. For example, for each parameter pair, the parameter pair can include the channel number and the FEC type. Among them, the FEC type includes supporting FEC data and not supporting FEC data; if the target FEC type is supporting FEC data, the actual operation data can also include the FEC data of the target optical module. Or, Based on the target channel number, the target FEC type, and the configured target scenario parameters of the target optical module, the optical module type corresponding to the target channel number, the target FEC type, and the target scenario parameters can be determined as the target optical module type. Among them, all optical modules of the network device support a total of A parameter pairs, and B scenario parameters can be pre-configured. In this way, the K parameter sets composed of the A parameter pairs and the B scenario parameters correspond one-to-one with K types of optical module types. Among them, the scenario parameters can include the total time length of the historical time interval, the time length between two adjacent time points within the historical time interval, the total time length of the future time interval, and the time length between two adjacent time points within the future time interval.

[0019] In an example, determining the target prediction result based on the initial prediction result can include, but is not limited to: determining the initial prediction result as the target prediction result. Or, based on the target model number corresponding to the target optical module, select the target online network layer corresponding to the target model number from all the online network layers supported by the network device; all optical modules of the network device support a total of M model numbers, and the network device supports M online network layers; the M model numbers correspond one-to-one with the M online network layers, and M is a positive integer; input the initial prediction result into the target online network layer, and optimize the initial prediction result through the target online network layer to obtain the target prediction result.

[0020] In an example, for each online network layer supported by the network device, the training process of the online network layer can include, but is not limited to: obtaining the initial network layer to be trained, and obtaining the sample data generated by the sample optical module. The sample optical module can be any one of all the optical modules of the network device; training the initial network layer based on the sample data to obtain the online network layer corresponding to the model number of the sample optical module; among them, the sample data can include the first sample sequence in the first time interval and the second sample sequence in the second time interval. The total time length of the first time interval corresponds to the total time length of the historical time interval, and the total time length of the second time interval corresponds to the total time length of the future time interval.

[0021] Exemplarily, during the training process, the initial network layer is trained based on sample data to obtain the online network layer corresponding to the model of the sample optical module, which may include but is not limited to: inputting the first sample sequence into the prediction model corresponding to the optical module type of the sample optical module to obtain an initial prediction result, inputting the initial prediction result into the initial network layer to obtain a target prediction result, and determining a loss value based on the target prediction result and the second sample sequence; freezing the network parameters of the prediction model, and adjusting the network parameters of the initial network layer based on the loss value to obtain the online network layer corresponding to the model of the sample optical module.

[0022] In one example, the target prediction model includes but is not limited to the iTransformer network, and the iTransformer network may include an embedding network layer, a multi-head attention network layer, a first normalization network layer, a feed-forward network layer, a second normalization network layer, and a prediction network layer. Based on this, inputting the first data sequence into the target prediction model to obtain an initial prediction result may include but is not limited to: inputting the first data sequence into the embedding network layer to obtain a first feature vector; obtaining a Q vector, a K vector, and a V vector based on the first feature vector, and inputting the Q vector, the K vector, and the V vector into the multi-head attention network layer to obtain a second feature vector; generating a third feature vector based on the first feature vector and the second feature vector; inputting the third feature vector into the first normalization network layer to obtain a fourth feature vector; inputting the fourth feature vector into the feed-forward network layer to obtain a fifth feature vector; generating a sixth feature vector based on the fourth feature vector and the fifth feature vector; inputting the sixth feature vector into the second normalization network layer to obtain a seventh feature vector; and inputting the seventh feature vector into the prediction network layer to obtain the initial prediction result.

[0023] As can be seen from the above technical solutions, in the embodiments of the present application, a first data sequence of the target optical module in a historical time interval can be obtained, the first data sequence is input into the target prediction model to obtain an initial prediction result, and a target prediction result is determined based on the initial prediction result. The target prediction result includes predicted operation data in a future time interval, and the predicted operation data is used to determine the health state of the target optical module. In this way, the target optical module can be predicted to obtain the predicted operation data for a future period of time, so as to predict the state change trend of the target optical module and monitor the health state of the target optical module. Data prediction by the target prediction model has the characteristics of high prediction efficiency, high prediction accuracy, and good long-term prediction effect.

[0024] The first data sequence includes the voltage of the target optical module, the number of mispackaged packets at the port, the temperature, bias current, received power, and transmitted power corresponding to each channel. In this way, by combining the actual operation data of the target optical module and the actual operation data of each channel of the target optical module, the target optical module is predicted. Then, the channel fusion method is adopted to fuse the data of all channels for prediction, further improving the prediction accuracy.

[0025] The type of the target optical module is determined by the number of target channels of the target optical module, and the target prediction model corresponding to the type of the target optical module is selected from all prediction models. The target optical module is predicted through the target prediction model. In this way, the target prediction model most suitable for the target optical module can be selected for prediction, further improving the prediction accuracy, improving the prediction speed, and improving the performance of the optical module prediction algorithm.

[0026] The above technical solutions of the embodiments of the present application will be described below in combination with specific application scenarios.

[0027] When the optical module runs for a long time, it will cause the performance degradation of the optical module, resulting in unstable links. To monitor the health status of the optical module, the data of the optical module are collected regularly for analysis, such as analyzing the current health status of the optical module and analyzing the health status of the optical module in a future period of time. To analyze the health status of the optical module in a future period of time, it is necessary to predict based on the data of the optical module to obtain the optical module data in a future period of time, and analyze based on these optical module data to predict the trend of the state change of the optical module.

[0028] When predicting based on the data of the optical module, time series prediction algorithms can be used. Time series prediction algorithms can include ARIMA (Autoregressive Integrated Moving Average Model), Exponential Smoothing, LSTM (Long Short-Term Memory), time series decomposition, DeepAR (Deep Autoregressive Recurrent), etc. However, when using these time series prediction algorithms to predict the data of the optical module, there are problems such as low prediction efficiency, poor prediction effect, and low prediction accuracy.

[0029] In view of the above findings, in an embodiment of the present application, a data prediction method is proposed. Considering that an optical module includes multiple index data (such as the voltage of the target optical module, the number of port mispackets, and FEC (Forward Error Correction) data, the temperature, bias current, received power, and transmitted power corresponding to each channel), there is a certain correlation between these index data. In this embodiment, the prediction model can process multiple index data, so as to make full use of the correlation between multiple index data for data prediction, and has the characteristics of high prediction efficiency, high prediction accuracy, and good long-term prediction effect.

[0030] In one example, the data prediction method can be applied to network devices, such as routers or switches, etc. There is no limitation on the type of such network devices, as long as the network device has an optical module. The network device may include multiple optical modules, such as 10G optical modules, 40G optical modules, 100G optical modules, 200G optical modules, 400G optical modules, etc. There is no limitation on the optical modules of the network device. Alternatively, the data prediction method can be applied to servers or cloud devices, etc., as long as the server or cloud device can collect data of the network device and perform analysis and prediction on the network device based on these data. For the convenience of description, in the following process, the data prediction method is described by taking its application to network devices as an example.

[0031] In one example, assume that all optical modules of the network device support K types of optical module types, that is, the network device may include optical modules of K types of optical module types. Then, the network device can support K prediction models, and the K prediction models correspond one-to-one to the K types of optical module types. K can be a positive integer, such as K>1. Among them, the network device supporting K prediction models means that the network device deploys these K prediction models.

[0032] For example, assume that the network device includes optical modules of optical module type a1 (such as optical module 1 and optical module 2), optical modules of optical module type a2 (such as optical module 3, optical module 4, and optical module 5), and optical modules of optical module type a3 (such as optical module 6), that is, there are a total of 6 optical modules in the network device. Then, the network device can support the prediction model b1 corresponding to optical module type a1, the prediction model b2 corresponding to optical module type a2, and the prediction model b3 corresponding to optical module type a3, that is, a total of 3 prediction models are supported.

[0033] For example, if the server learns that the network device includes an optical module of optical module type a1, the server may send prediction model b1 to the network device. If the server learns that the network device includes an optical module of optical module type a2, the server may send prediction model b2 to the network device. If the server learns that the network device includes an optical module of optical module type a3, the server may send prediction model b3 to the network device.

[0034] In one example, the following method may be used to determine K types of optical module types: Method 1: Determine K types of optical module types based on the number of channels. For example, all optical modules of the network device support K numbers of channels, and the K numbers of channels correspond one-to-one with the K types of optical module types.

[0035] In one example, to achieve high-speed performance, an optical module may include several channels. A channel may also be referred to as a Lane. These channels can receive and transmit data in parallel. For example, the optical module receives data simultaneously through 4 channels and transmits data simultaneously through 4 channels, etc., so as to achieve high-speed performance. On this basis, the number of channels of the optical module can be determined, and the number of channels is an attribute of the optical module. For example, the number of channels of a 10G optical module is 1, the number of channels of a 40G optical module is 4, the number of channels of a 100G optical module is 4, the number of channels of a 200G optical module is 4 or 8, and the number of channels of a 400G optical module is 4 or 8.

[0036] Assume that all optical modules of the network device jointly support K numbers of channels, then the K numbers of channels correspond one-to-one with the K types of optical module types. Taking the K numbers of channels as 3 numbers of channels as an example, that is, channel number 1, channel number 4, and channel number 8, then channel number 1 corresponds to optical module type a1, channel number 4 corresponds to optical module type a2, and channel number 8 corresponds to optical module type a3. On this basis, if optical module 1 and optical module 2 include 1 channel, then optical module 1 and optical module 2 correspond to optical module type a1. If optical module 3, optical module 4, and optical module 5 include 4 channels, then optical module 3, optical module 4, and optical module 5 correspond to optical module type a2. If optical module 6 includes 8 channels, then optical module 6 corresponds to optical module type a3.

[0037] In summary, if all optical modules of the network device support 3 numbers of channels (channel number 1, channel number 4, and channel number 8), then there are 3 types of optical modules in the network device, such as optical module type a1, optical module type a2, and optical module type a3. In this way, the network device can support 3 prediction models, and the 3 prediction models correspond one-to-one with the 3 types of optical module types, such as prediction model b1 corresponding to optical module type a1, prediction model b2 corresponding to optical module type a2, and prediction model b3 corresponding to optical module type a3.

[0038] Method 2: Determine K types of optical module types based on the number of parameter pairs. For example, all optical modules of a network device jointly support K parameter pairs, and the K parameter pairs correspond one-to-one with the K types of optical module types. For example, for each parameter pair, the parameter pair can include the number of channels and the FEC type. The FEC type includes supporting FEC data (i.e., the optical module supports the transmission and processing of FEC data) and not supporting FEC data.

[0039] In an example, FEC data is a channel coding data that recovers lost data packets by adding redundant data, and can be divided into correctable FEC data and uncorrectable FEC data. An optical module may support the transmission and processing of FEC data, or may not support the transmission and processing of FEC data. Based on this, the FEC type of the optical module can be determined, and the FEC type is an attribute of the optical module. For example, the FEC type of a 10G optical module indicates that it does not support FEC data, the FEC type of a 40G optical module indicates that it does not support FEC data, the FEC type of a 100G optical module indicates that it supports FEC data, the FEC type of a 200G optical module indicates that it supports FEC data, and the FEC type of a 400G optical module indicates that it supports FEC data.

[0040] Assume that all optical modules of a network device support K parameter pairs, then the K parameter pairs correspond one-to-one with the K types of optical module types. For example, if all optical modules support 3 numbers of channels (i.e., number of channels 1, number of channels 4, and number of channels 8), then at most 6 parameter pairs are supported (number of channels 1 + support FEC data, number of channels 1 + do not support FEC data, number of channels 4 + support FEC data, number of channels 4 + do not support FEC data, number of channels 8 + support FEC data, number of channels 8 + do not support FEC data).

[0041] Regarding how many parameter pairs all optical modules of a network device support, it is related to the actual situation of the optical modules. For example, optical module 1 and optical module 2 include 1 channel, optical module 3, optical module 4, and optical module 5 include 4 channels, optical module 6 includes 8 channels, optical module 1 supports FEC data, optical module 2 does not support FEC data, optical module 3 supports FEC data, optical module 4 and optical module 5 do not support FEC data, and optical module 6 supports FEC data. Then, all optical modules of the network device support 5 parameter pairs.

[0042] The optical module 1 corresponds to the parameter pair c1 (number of channels 1 + supporting FEC data), the parameter pair c1 corresponds to the optical module type a1, the optical module 2 corresponds to the parameter pair c2 (number of channels 1 + not supporting FEC data), the parameter pair c2 corresponds to the optical module type a2, the optical module 3 corresponds to the parameter pair c3 (number of channels 4 + supporting FEC data), the parameter pair c3 corresponds to the optical module type a3, the optical modules 4 and 5 correspond to the parameter pair c4 (number of channels 4 + not supporting FEC data), the parameter pair c4 corresponds to the optical module type a4, and the optical module 6 corresponds to the parameter pair c5 (number of channels 8 + supporting FEC data), the parameter pair c5 corresponds to the optical module type a5.

[0043] In summary, if all the optical modules of a network device support 5 parameter pairs, then there are 5 types of optical modules in the network device, such as optical module type a1, optical module type a2, optical module type a3, optical module type a4, and optical module type a5. In this way, the network device can support 5 prediction models, and these 5 prediction models correspond one-to-one with the 5 types of optical modules, such as the prediction model b1 corresponding to the optical module type a1, the prediction model b2 corresponding to the optical module type a2, the prediction model b3 corresponding to the optical module type a3, the prediction model b4 corresponding to the optical module type a4, and the prediction model b5 corresponding to the optical module type a5.

[0044] In an example, considering that the number of channels of a 10G optical module is 1 and the FEC type indicates not supporting FEC data, the number of channels of a 40G optical module is 4 and the FEC type indicates not supporting FEC data, the number of channels of a 100G optical module is 4 and the FEC type indicates supporting FEC data, the number of channels of a 200G optical module is 4 or 8 and the FEC type indicates supporting FEC data, the number of channels of a 400G optical module is 4 or 8 and the FEC type indicates supporting FEC data. On this basis, there are a total of 4 parameter pairs, namely: parameter pair c1 (number of channels 1 + not supporting FEC data), parameter pair c2 (number of channels 4 + not supporting FEC data), parameter pair c3 (number of channels 4 + supporting FEC data), parameter pair c4 (number of channels 8 + supporting FEC data).

[0045] As can be seen from the above, the parameter pair c1 can be applied to a 10G optical module, the parameter pair c2 can be applied to a 40G optical module, the parameter pair c3 can be applied to a 100G optical module, a 4-channel 200G optical module, and a 400G optical module, and the parameter pair c4 can be applied to an 8-channel 200G optical module and a 400G optical module.

[0046] Method 3: Determine K types of optical module types based on the number of parameter pairs and the number of scenario parameters. For example, all optical modules of a network device jointly support A parameter pairs, and B scenario parameters are pre-configured. The K parameter sets (i.e., A * B = K) composed of the A parameter pairs and the B scenario parameters correspond one-to-one with the K types of optical module types. For example, for each parameter pair, the parameter pair includes the number of channels and the FEC type, and the FEC type includes supporting FEC data and not supporting FEC data. The scenario parameters can include the total time length of the historical time interval, the time length between two adjacent time points within the historical time interval (i.e., the sampling frequency), the total time length of the future time interval, and the time length between two adjacent time points within the future time interval.

[0047] In an example, assume that all optical modules of a network device support A parameter pairs (similar to the K parameter pairs in Method 2, just replace K with A), and B scenario parameters are pre-configured. Then there are a total of A * B parameter sets, and the A * B parameter sets can be denoted as K parameter sets. For example, assume that all optical modules of a network device support 5 parameter pairs, and 2 scenario parameters are predicted and configured. Then there are a total of 10 parameter sets, and there are 10 types of optical module types in the network device, such as optical module type a1 - optical module type a10. The network device can support 10 prediction models corresponding one-to-one with the 10 types of optical module types, such as prediction model b1 - prediction model b10. Assume that all optical modules of a network device support 4 parameter pairs, and 3 scenario parameters are predicted and configured. Then there are a total of 12 parameter sets, and there are 12 types of optical module types and 12 prediction models in the network device. And so on, the optical module types and the parameter sets correspond one-to-one.

[0048] For example, if the A parameter pairs are parameter pair c1, parameter pair c2, parameter pair c3, and parameter pair c4 in sequence, and the B scenario parameters are scenario parameter d1, scenario parameter d2, and scenario parameter d3 in sequence, then parameter set e1 is parameter pair c1 and scenario parameter d1, and parameter set e1 corresponds to optical module type a1, parameter set e2 is parameter pair c1 and scenario parameter d2, and parameter set e2 corresponds to optical module type a2, parameter set e3 is parameter pair c1 and scenario parameter d3, and parameter set e3 corresponds to optical module type a3, parameter set e4 is parameter pair c2 and scenario parameter d1, and parameter set e4 corresponds to optical module type a4, and so on.

[0049] In an example, the scenario parameters can be configured according to actual business requirements. The scenario parameters can include the total time length of the historical time interval, the time length between two adjacent time points within the historical time interval, the total time length of the future time interval, and the time length between two adjacent time points within the future time interval.

[0050] For example, according to the actual business requirements, based on a 30 - minute frequency, using 7 - day (7 * 48 = 336) data to predict the data for the next 3 days (3 * 48 = 144), that is, the historical data length is 336 and the prediction length is 144. Then, the scenario parameter d1 includes the total time length of the historical time interval (such as 7 days), the time length between two adjacent time points within the historical time interval (such as 30 minutes), the total time length of the future time interval (such as 3 days), and the time length between two adjacent time points within the future time interval (such as 30 minutes).

[0051] For example, according to the actual business requirements, based on a 1 - hour frequency, using 10 - day (10 * 24 = 240) data to predict the data for the next 5 days (5 * 24 = 120), that is, the historical data length is 240 and the prediction length is 120. Then, the scenario parameter d2 includes the total time length of the historical time interval (such as 10 days), the time length between two adjacent time points within the historical time interval (such as 60 minutes), the total time length of the future time interval (such as 5 days), and the time length between two adjacent time points within the future time interval (such as 60 minutes).

[0052] Of course, the above are just two examples of scenario parameters, and there are no restrictions on these scenario parameters. The scenario parameters can be configured according to the actual business requirements, as long as the scenario parameters include the above 4 parameter values.

[0053] In summary, if all optical modules of the network device support A parameter pairs and B scenario parameters are pre - configured, then there are a total of A * B (i.e., K) types of optical modules in the network device. The network device can support A * B prediction models, and the A * B prediction models correspond one - to - one with the A * B types of optical modules.

[0054] In one example, taking Method 3 as an example, the training process of the prediction model corresponding to the optical module type (such as the prediction model b1 corresponding to the optical module type a1) is introduced. Since the training process of each prediction model is the same, the training process of the prediction model b1 is taken as an example here. The prediction model b1 can be trained on the server and sent by the server to the network device, or the prediction model b1 can also be trained on the network device.

[0055] First, obtain the sample data of the optical module type a1. Considering that the optical module type a1 corresponds to the parameter pair c1 (channel number 1 + does not support FEC data) and the scenario parameter d1, therefore, it is necessary to obtain the sample data corresponding to the parameter pair c1 and the scenario parameter d1. For example, a large amount of real - running data generated by a specified optical module (i.e., an optical module with a channel number of 1 and does not support FEC data) can be collected (see the subsequent embodiments for the content of the real - running data). Based on these real - running data, multiple sample data are constructed.

[0056] For example, sample data 1 includes the actual operation data at multiple time points in a historical time interval (with a total time length of 7 days, such as from the 1st day to the 7th day), where the time length between adjacent time points is 30 minutes, and the actual operation data at multiple time points in a future time interval (with a total time length of 3 days, such as from the 8th day to the 10th day, i.e., the 3 days after the historical time interval), where the time length between adjacent time points is 30 minutes. Sample data 2 includes the actual operation data at multiple time points from the 2nd day to the 8th day and the actual operation data at multiple time points from the 9th day to the 11th day. Sample data 3 includes the actual operation data at multiple time points from the 3rd day to the 9th day and the actual operation data at multiple time points from the 10th day to the 12th day. By analogy, a large number of sample data can be obtained. All these sample data include the actual operation data generated by the specified optical module and are constructed according to the requirements of scenario parameter d1.

[0057] Then, based on multiple sample data, train the prediction model to be trained (i.e., the pre-configured prediction model to be trained) to obtain prediction model b1. There is no restriction on the training process of this prediction model b1. During the training process, the actual operation data at multiple time points in the historical time interval is used as input data. These actual operation data need to be input into the prediction model to be trained for prediction to obtain prediction results. The actual operation data at multiple time points in the future time interval is used as label data. Based on the label data and the prediction results, calculate the loss value, and adjust the network parameters of the prediction model to be trained through the loss value.

[0058] In summary, the prediction model b1 can be trained, and then the network device supports the prediction model b1.

[0059] In one example, to train the prediction models corresponding to all optical module types, a large amount of actual operation data needs to be collected. For example, collect a large amount of actual operation data from the experimental network and classify these actual operation data according to the parameter pairs. For example, all the actual operation data can be divided into the actual operation data of parameter pair c1 (channel number 1 + non-FEC-supported data) (i.e., generated by the optical module corresponding to parameter pair c1), the actual operation data of parameter pair c2 (channel number 4 + non-FEC-supported data), the actual operation data of parameter pair c3 (channel number 4 + FEC-supported data), and the actual operation data of parameter pair c4 (channel number 8 + FEC-supported data). In addition, for the actual operation data of parameter pair c3, after removing the FEC data, it can also be used as the actual operation data of parameter pair c2.

[0060] Taking the scenario parameters d1 and d2 as examples, based on the actual operation data of the parameter pair c1, a training data set corresponding to the optical module type a1 (parameter pair c1 + scenario parameter d1) can be generated. The training data set can include multiple sample data, and a training data set corresponding to the optical module type a2 (parameter pair c1 + scenario parameter d2) can be generated. Based on the actual operation data of the parameter pair c2, a training data set corresponding to the optical module type a3 (parameter pair c2 + scenario parameter d1) can be generated, and a training data set corresponding to the optical module type a4 (parameter pair c2 + scenario parameter d2) can be generated. And so on, training data sets corresponding to 8 optical module types can be obtained. On this basis, based on the training data set corresponding to the optical module type a1, a prediction model b1 corresponding to the optical module type a1 can be trained. Based on the training data set corresponding to the optical module type a2, a prediction model b2 corresponding to the optical module type a2 can be trained. And so on, 8 prediction models corresponding to 8 optical module types can be trained, and the 8 prediction models and the 8 optical module types are in one-to-one correspondence.

[0061] Under the above application scenario, in an embodiment of the present application, a data prediction method is proposed, which can be applied to a network device. Refer to Figure 2 As shown, it is a schematic flowchart of the data prediction method. The method may include: Step 201, obtain a first data sequence of the target optical module in the historical time interval.

[0062] In one example, the target optical module may be any one of all the optical modules of the network device. For example, for each optical module of the network device, data prediction can be performed on the optical module. Subsequently, taking the data prediction process of one optical module as an example, this optical module can be referred to as the target optical module.

[0063] In one example, the configured target scenario parameters can be obtained. The target scenario parameters may include the total time length of the historical time interval, the time length between two adjacent time points in the historical time interval (i.e., the sampling frequency), the total time length of the future time interval, and the time length between two adjacent time points in the future time interval. For example, the target scenario parameters may be scenario parameter d1 or scenario parameter d2.

[0064] Taking the target scenario parameter as scenario parameter d1 as an example, then, the total time length of the historical time interval is 7 days, the time length between two adjacent time points in the historical time interval is 30 minutes, the total time length of the future time interval is 3 days, and the time length between two adjacent time points in the future time interval is 30 minutes.

[0065] In one example, the first data sequence may include the actual operation data of multiple time points in a historical time interval. For example, the last time point in the historical time interval may be the current time point. Starting from the current time point, traversing 336 (7 * 48 = 336) time points (including the current time point) forward at a granularity of 30 minutes, these time points are the multiple time points in the historical time interval.

[0066] To sum up, the total time length of the historical time interval is 7 days, the time length between two adjacent time points in the historical time interval is 30 minutes, and the historical time interval includes 336 time points.

[0067] In one example, for each time point (taking one time point as an example), the actual operation data of this time point may include the voltage of the target optical module and / or the number of port error packets. If the target FEC type of the target optical module indicates support for FEC data, the actual operation data may further include the FEC data of the target optical module. If the target FEC type indicates no support for FEC data, the actual operation data does not include FEC data.

[0068] Regarding the voltage of the target optical module (Voltage), the network device can periodically collect the voltage of the target optical module, and there is no restriction on this process. In this way, the voltage of the target optical module can be obtained.

[0069] Regarding the number of port error packets (in_error_packets) of the target optical module, when the network device receives a data packet through the target optical module, it can know whether the data packet is a correct data packet or an error data packet. If the data packet is an error data packet, the number of port error packets can be incremented by 1. In this way, the number of port error packets of the target optical module in a specified time period can be counted, and the number of port error packets is used as the actual operation data.

[0070] Regarding the FEC data of the target optical module, the FEC data can be classified into correctable FEC data (FEC_Correctable) and uncorrectable FEC data (FEC_UnCorrectable). When the network device receives a data packet through the target optical module, it can also receive the FEC data of multiple data segments of the data packet received through the target optical module. If an error occurs in a certain data segment and the FEC data of this data segment can correct the error, the correctable FEC data is incremented by 1. If the FEC data of this data segment cannot correct the error, the uncorrectable FEC data is incremented by 1. In this way, the FEC data (correctable FEC data and uncorrectable FEC data) of the target optical module in a specified time period can be counted.

[0071] In one example, to achieve high-speed performance, the target optical module includes at least one channel (which can also be referred to as a Lane), and these channels can transmit and receive data in parallel. For example, the target optical module includes 1 channel, the target optical module includes 4 channels, or the target optical module includes 8 channels. On this basis, for each time point (taking one time point as an example), the actual operating data at this time point may include at least one of the temperature, bias current, received power, and transmitted power corresponding to each channel.

[0072] Taking the example where the target optical module includes 4 channels, there are 4-channel circuits inside the target optical module. For example, circuit 1 is used to implement the transceiver function of channel 1, circuit 2 is used to implement the transceiver function of channel 2, circuit 3 is used to implement the transceiver function of channel 3, and circuit 4 is used to implement the transceiver function of channel 4.

[0073] Regarding the temperature of each channel, the network device can periodically collect the temperature of each channel (i.e., the temperature of the circuit of this channel), and there is no limitation on this process. In this way, the temperature of each channel of the target optical module can be obtained. Regarding the bias current of each channel, the network device can periodically collect the bias current of each channel (i.e., the bias current of the circuit of this channel), and in this way, the bias current of each channel of the target optical module can be obtained. Regarding the received power of each channel, the network device can periodically collect the received power of each channel (i.e., the received power of the circuit of this channel), and in this way, the received power of each channel of the target optical module can be obtained. Regarding the transmitted power of each channel, the network device can periodically collect the transmitted power of each channel (i.e., the transmitted power of the circuit of this channel), and in this way, the transmitted power of each channel of the target optical module can be obtained.

[0074] In summary, for each time point, the actual operating data at this time point may include the voltage of the target optical module, the number of port mispackets of the target optical module, the FEC data of the target optical module, the temperature of each channel of the target optical module, the bias current of each channel of the target optical module, the received power of each channel of the target optical module, and the transmitted power of each channel of the target optical module. In this way, using the channel fusion method, all channel data of the target optical module are fused, and combined with the voltage, port mispacket number, and FEC data of the target optical module, they jointly form the actual operating data to participate in subsequent prediction. Since the above data has a certain correlation, fusing these data into a multi-variable time series data is beneficial to improving the prediction accuracy.

[0075] If the target optical module includes 4 channels and supports FEC data, the actual operating data can be expressed as: Voltage + channel1 (Temperature, Current Bias, Rx Power, Tx power) + channel2 (Temperature, Current Bias, Rx Power, Tx power) + channel3 (Temperature, Current Bias, Rx Power, Tx power) + channel4 (Temperature, Current Bias, Rx Power, Tx power) + in_error_packets + FEC_Correctable + FEC_UnCorrectable. If it does not support FEC data, the actual operating data is expressed as: Voltage + channel1 (Temperature, Current Bias, Rx Power, Tx power) + channel2 (Temperature, Current Bias, Rx Power, Tx power) + channel3 (Temperature, Current Bias, Rx Power, Tx power) + channel4 (Temperature, Current Bias, Rx Power, Tx power) + in_error_packets.

[0076] Step 202: Determine the type of the target optical module based on the number of target channels of the target optical module, and select the target prediction model corresponding to the type of the target optical module from all the prediction models supported by the network device.

[0077] In one example, considering that the number of channels of different optical modules may be different, and the FEC types (supporting FEC data or not supporting FEC data) of different optical modules may also be different, according to the hardware differences of the optical modules, all optical modules can be divided into multiple optical module types, and prediction models are supported for each optical module type respectively. On this basis, determine the type of the target optical module based on the number of target channels of the target optical module, and select the target prediction model corresponding to the type of the target optical module from all the prediction models supported by the network device.

[0078] In one example, if method 1 is used to determine K types of optical module types, then the target channel number of the target optical module can be determined. Based on the target channel number of the target optical module, the optical module type corresponding to this target channel number can be determined as the target optical module type. For example, if the target channel number is channel number 1, then the optical module type a1 corresponding to channel number 1 is used as the target optical module type, and the prediction model b1 corresponding to the optical module type a1 is used as the target prediction model. If the target channel number is channel number 4, then the optical module type a2 corresponding to channel number 4 is used as the target optical module type, and the prediction model b2 corresponding to the optical module type a2 is used as the target prediction model, and so on.

[0079] In one example, if method 2 is used to determine K types of optical module types, then the target channel number of the target optical module and the target FEC type of the target optical module can be determined. Based on the target channel number and the target FEC type of the target optical module, the optical module type corresponding to the target channel number and the target FEC type can be determined as the target optical module type. For example, the target channel number and the target FEC type can correspond to a target parameter pair, and the optical module type corresponding to the target parameter pair is determined as the target optical module type.

[0080] For example, taking the parameter pairs c1 (channel number 1 + does not support FEC data), c2 (channel number 4 + does not support FEC data), c3 (channel number 4 + supports FEC data), and c4 (channel number 8 + supports FEC data) as examples, if the target channel number is channel number 1 and the target FEC type is does not support FEC data, then the target parameter pair is parameter pair c1, and the optical module type a1 corresponding to parameter pair c1 is used as the target optical module type, and the prediction model b1 corresponding to the optical module type a1 is used as the target prediction model. If the target channel number is channel number 4 and the target FEC type is does not support FEC data, then the target parameter pair is parameter pair c2, and the optical module type a2 corresponding to parameter pair c2 is used as the target optical module type, and the prediction model b2 corresponding to the optical module type a2 is used as the target prediction model, and so on.

[0081] In one example, if method 3 is used to determine K types of optical module types, then the target channel number of the target optical module, the target FEC type of the target optical module, and the configured target scenario parameters are determined. Based on the target channel number, the target FEC type, and the target scenario parameters of the target optical module, the optical module type corresponding to the target channel number, the target FEC type, and the target scenario parameters is determined as the target optical module type. For example, the target channel number and the target FEC type can correspond to a target parameter pair, and the optical module type corresponding to the parameter set composed of the target parameter pair and the target scenario parameters is determined as the target optical module type.

[0082] For example, if the number of target channels is Channel Number 1 and the target FEC type is not supporting FEC data, the target parameter pair is Parameter Pair c1. Assuming the target scenario parameter is Scenario Parameter d1, the parameter set composed of Parameter Pair c1 and Scenario Parameter d1 corresponds to optical module type a1. That is, optical module type a1 is used as the target optical module type, and the prediction model b1 corresponding to optical module type a1 is used as the target prediction model. Assuming the target scenario parameter is Scenario Parameter d2, the parameter set composed of Parameter Pair c1 and Scenario Parameter d2 corresponds to optical module type a2, and the prediction model b2 corresponding to optical module type a2 is used as the target prediction model.

[0083] If the number of target channels is Channel Number 4 and the target FEC type is not supporting FEC data, the target parameter pair is Parameter Pair c2. Assuming the target scenario parameter is Scenario Parameter d1, the parameter set composed of Parameter Pair c2 and Scenario Parameter d1 corresponds to optical module type a3. That is, optical module type a3 is used as the target optical module type, and the prediction model b3 corresponding to optical module type a3 is used as the target prediction model, and so on.

[0084] In summary, for different optical modules, the optical module types can be distinguished, a suitable target prediction model can be selected for the target optical module based on the optical module type, and the target optical module can be predicted based on the target prediction model.

[0085] In one example, the parameter pair c1 (number of channels 1 + FEC data not supported) can be applicable to a 10G optical module. The actual operating data of this type of optical module can be expressed as: Voltage + (Temperature, Current Bias, RxPower, Tx power) + in_error_packets. The parameter pair c2 (number of channels 4 + FEC data not supported) can be applicable to a 40G optical module. The actual operating data of this type of optical module can be expressed as: Voltage + 4 * (Temperature, Current Bias, Rx Power, Tx power) + in_error_packets. The parameter pair c3 (number of channels 4 + FEC data supported) can be applicable to 100G optical modules, 4-channel 200G optical modules, and 400G optical modules. The actual operating data of this type of optical module can be expressed as: Voltage + 4 * (Temperature, Current Bias, Rx Power, Tx power) + in_error_packets + FEC_Correctable + FEC_UnCorrectable. The parameter pair c4 (number of channels 8 + FEC data supported) can be applicable to 8-channel 200G optical modules and 400G optical modules. The actual operating data of this type of optical module can be expressed as: Voltage + 8 * (Temperature, Current Bias, Rx Power, Tx power) + in_error_packets + FEC_Correctable + FEC_UnCorrectable.

[0086] Step 203: Input the first data sequence into the target prediction model to obtain an initial prediction result.

[0087] In one example, for each prediction model (taking the target prediction model as an example for illustration), the target prediction model can be any type of time series prediction model, such as an ARIMA prediction model, an exponential smoothing prediction model, an LSTM prediction model, a time series decomposition prediction model, a DeepAR prediction model, etc. Considering that the prediction effects of these time series prediction models are not good, a Transformer network or an iTransformer network can also be used as the target prediction model. Taking the iTransformer network as the target prediction model as an example.

[0088] The iTransformer network is a multivariate time-series data prediction network. After inputting the actual operating data (the voltage, port error packet count, and FEC data of the target optical module, and the temperature, bias current, received power, and transmitted power of each channel of the target optical module) into the iTransformer network, the iTransformer network can make predictions based on the multivariate time-series data, can consider the correlations between multiple variables (such as data from multiple channels), can better handle long-term dependencies, can better describe the correlations between various data, extract multi-layer representations of the sequence, effectively mine the correlations between multiple variables (sequences), has good performance, and has a better operation rate and prediction accuracy, meeting the requirements for optical module fault prediction.

[0089] In one example, the target prediction model includes, but is not limited to, the iTransformer network and the Transformer network. The type of this target prediction model is not restricted. Taking the iTransformer network as an example for illustration, of course, when using other types of target prediction models, their implementation methods are similar and will not be elaborated here.

[0090] In one example, the first data sequence can be input into the target prediction model (the iTransformer network) to obtain an initial prediction result. Taking the target scenario parameter as scenario parameter d1, then, the first data sequence includes the actual operating data at 336 time points. For each time point, the actual operating data at this time point can include the voltage, port error packet count, and FEC data of the target optical module, and the temperature, bias current, received power, and transmitted power of each channel of the target optical module. The target prediction model (the iTransformer network) makes predictions based on the actual operating data at 336 time points to obtain an initial prediction result.

[0091] The initial prediction result can include a third data sequence in the future time interval, and this third data sequence can include the predicted operating data at multiple time points in the future time interval. For example, under scenario parameter d1, the total time length of the future time interval is 3 days, and the time length between two adjacent time points in the future time interval is 30 minutes. Based on this, the first time point in the future time interval can be the current time point + a fixed duration (such as 30 minutes). Starting from the first time point, with a granularity of 30 minutes as the interval, traverse 144 (3 * 48 = 144) time points backward. These time points are the multiple time points in the future time interval.

[0092] In summary, the total time length of the future time interval is 3 days, the time length between two adjacent time points in the future time interval is 30 minutes, and the future time interval includes 144 time points.

[0093] The third data sequence includes predicted operation data at 144 time points. For each time point, the predicted operation data at that time point may include the voltage of the target optical module, the number of mispacketed ports, and FEC data, as well as the temperature, bias current, received power, and transmitted power of each channel of the target optical module. Obviously, the data types corresponding to the predicted operation data are the same as those corresponding to the real operation data, except that the predicted operation data is predicted by the target prediction model, while the real operation data is actually collected from the target optical module.

[0094] In one example, refer to Figure 3 As shown, it is a schematic structural diagram of the target prediction model (iTransformer network). The iTransformer network may include an Embedding network layer, a MultivariateAttention network layer, a first LayerNorm network layer, a Feed-forward network layer, a second LayerNorm network layer, and a Projection network layer. Figure 3 This is just an example of the iTransformer network, and the structure of the iTransformer network is not limited thereto.

[0095] In one example, based on Figure 3 the iTransformer network shown, in step 203, the process of inputting the first data sequence (such as real operation data at 336 time points) into the iTransformer network to obtain an initial prediction result (such as predicted operation data at 144 time points) may include: Input the first data sequence into the Embedding network layer to obtain a first feature vector. For example, Embedding is a technique for converting data into vector representations. Embedding can convert words, sentences, or documents into low-dimensional vectors, thereby enabling better understanding and processing of text data. The core idea of Embedding is to map data to a low-dimensional vector space by learning the internal structure and semantic information of the data. In this vector space, similar data points will be mapped to nearby positions, so the similarity between data can be measured by calculating the similarity between vectors. In this example, Embedding can regard the historical data of each variable as a Token, so that the multi-variable sequence (i.e., the first data sequence) corresponds to multiple Tokens. In summary, after inputting the first data sequence into the Embedding network layer, the Embedding network layer can process the first data sequence to obtain a first feature vector, and the processing process is not limited thereto.

[0096] After obtaining the first feature vector, the Q vector, K vector, and V vector can be obtained based on the first feature vector. For example, the input data of the Multivariate Attention network layer is the Q vector, K vector, and V vector. Therefore, the Q vector, K vector, and V vector can be obtained based on the first feature vector, and there is no restriction on the acquisition method, as long as the Q vector, K vector, and V vector can be obtained.

[0097] After obtaining the Q vector, K vector, and V vector, the Q vector, K vector, and V vector are input into the Multivariate Attention network layer to obtain the second feature vector. For example, the Multivariate Attention network layer is a Self-Attention (self-attention layer) used to obtain the correlation between multiple sequences. After inputting the Q vector, K vector, and V vector into the Multivariate Attention network layer, the second feature vector can be obtained based on the processing of the Q vector, K vector, and V vector, and there is no restriction on this processing process.

[0098] After obtaining the second feature vector, the third feature vector is generated based on the first feature vector and the second feature vector. For example, the third feature vector is obtained by performing an addition operation on the first feature vector and the second feature vector.

[0099] After obtaining the third feature vector, the third feature vector can be input into the first LayerNorm (Layer Normalization) network layer to obtain the fourth feature vector. For example, LayerNorm is a regularization technique for neural networks that reduces the problem of internal covariate shift by normalizing the activation values of neurons within a single sample. The core idea of LayerNorm is to normalize the activation values of neurons in each layer. Specifically, for an input vector (x), LayerNorm first calculates the mean and standard deviation of this layer, then uses these statistics to normalize the activation values, and finally scales and translates through learnable parameters. In summary, after inputting the third feature vector into the first LayerNorm network layer, the first LayerNorm network layer can process the third feature vector to obtain the fourth feature vector.

[0100] After obtaining the fourth feature vector, the fourth feature vector is input into the Feed-forward network layer to obtain the fifth feature vector. For example, Feed-forward can also be FFN (Feed-forward Network), and FFN is a fully connected layer used to encode each sequence. After inputting the fourth feature vector into the Feed-forward network layer, the Feed-forward network layer can process the fourth feature vector to obtain the fifth feature vector.

[0101] After obtaining the fifth eigenvector, generate a sixth eigenvector based on the fourth eigenvector and the fifth eigenvector. For example, perform an addition operation on the fourth eigenvector and the fifth eigenvector to obtain the sixth eigenvector.

[0102] After obtaining the sixth eigenvector, input the sixth eigenvector into the second layer normalization network layer (LayerNorm) to obtain a seventh eigenvector. For example, after inputting the sixth eigenvector into the second layer normalization network layer, the seventh eigenvector can be obtained based on the processing of the sixth eigenvector.

[0103] After obtaining the seventh eigenvector, input the seventh eigenvector into the prediction network layer (Projection) to obtain the initial prediction result and output the initial prediction result. For example, the prediction network layer can be an MPL (Multiple Layer Perceptron). Based on this, the seventh eigenvector can be processed by the MPL to finally obtain the initial prediction result, and this process is not limited.

[0104] In summary, after inputting the first data sequence into the iTransformer network, the iTransformer network can process the first data sequence and finally obtain the initial prediction result.

[0105] Step 204: Based on the target model number of the target optical module, select the target online network layer corresponding to the target model number from all the online network layers (Online layer) supported by the network device.

[0106] In one example, the network device may include optical modules of multiple models (styles). The model number of the optical module can indicate which product the optical module belongs to. The model numbers of optical modules from different manufacturers are different, and the model numbers of optical modules from the same manufacturer may be the same (i.e., optical modules of the same batch), or the model numbers of optical modules from the same manufacturer may also be different (i.e., optical modules of different batches). For optical modules of different model numbers, when using the same prediction model to predict these optical modules, there may be prediction errors, that is, there will be differences in the predictions of optical modules of different model numbers. Based on this, in this embodiment, an online network layer (Online layer) can also be added on the basis of the prediction model, that is, an online network layer is added for each model number.

[0107] When adding an online network layer, instead of deploying the online network layer inside the prediction model, an additional online network layer is added. For example, assume there are 10 prediction models. If an online network layer is added inside the prediction model, such as adding 3 online network layers for 3 models, then 30 prediction models need to be deployed. If 3 additional online network layers are added for 3 models, then 10 prediction models and 3 online network layers need to be deployed. In this way, the number of models can be reduced, avoiding the deployment of a large number of prediction models.

[0108] In one example, assume that all optical modules of a network device support M models. Then the network device only needs to support M online network layers, and the M models correspond one-to-one with the M online network layers. M is a positive integer. The network device supporting M online network layers can mean that the network device deploys M online network layers. For example, if all optical modules of the network device support model f1 and model f2, then the network device can support the online network layer g1 corresponding to model f1 and the online network layer g2 corresponding to model f2.

[0109] In one example, for the training process of the online network layer, since the training process of each online network layer is the same, here the training process of the online network layer g1 is taken as an example. The online network layer g1 is trained on the network device, that is, the online network layer g1 is trained online based on the data generated locally by the network device.

[0110] First, an initial network layer to be trained can be obtained. That is, all online network layers can share the same initial network layer. That is, the online network layer g1 is trained based on the initial network layer, and the online network layer g2 is trained based on the initial network layer. Here, the training of the online network layer g1 is taken as an example.

[0111] Then, multiple sample data generated by a sample optical module are obtained. The sample optical module can be any one of all the optical modules of the network device. For example, when training the online network layer g1, the optical module corresponding to model f1 can be used as the sample optical module. The number of sample optical modules can be at least one. For example, if the models of optical module 1 and optical module 2 are model g1, then optical module 1 and optical module 2 can be used as the sample optical modules to obtain the sample data generated by optical module 1 and the sample data generated by optical module 2.

[0112] For each sample data, the sample data can include a first sample sequence in a first time interval and a second sample sequence in a second time interval. The total time length of the first time interval corresponds to the total time length of the historical time interval, and the total time length of the second time interval corresponds to the total time length of the future time interval.

[0113] For example, a large amount of real operating data generated by the optical module 1 (optical module 2) can be collected, and multiple sample data can be constructed based on this real operating data. The sample data 1 includes the real operating data at multiple time points (the time length between two adjacent time points is 30 minutes) in the first time interval (corresponding to the historical time interval, such as the total time length is 7 days, such as from the 1st day to the 7th day), and the real operating data at multiple time points (the time length between two adjacent time points is 30 minutes) in the second time interval (corresponding to the future time interval, such as the total time length is 3 days, such as from the 8th day to the 10th day, that is, 3 days after the first time interval). The sample data 2 includes the real operating data at multiple time points from the 2nd day to the 8th day, the real operating data at multiple time points from the 9th day to the 11th day, and so on.

[0114] Then, the sample data can include a first sample sequence (the content of the first sample sequence is similar to the content of the first data sequence, such as the real operating data at 336 time points) and a second sample sequence (the content of the second sample sequence is similar to the content of the first data sequence, such as the real operating data at 144 time points). The first sample sequence can be input into the prediction model corresponding to the optical module type of the sample optical module to obtain an initial prediction result.

[0115] For example, for the sample data of the optical module 1, the parameter pair of the optical module 1 can be determined, and a prediction model can be selected from all the prediction models in combination with the scenario parameters. The selection method refers to the above method 3, and this prediction model is used as the prediction model corresponding to the optical module 1. In this way, the first sample sequence can be input into this prediction model to obtain an initial prediction result, and the processing process of this prediction model refers to step 203.

[0116] Then, the initial prediction result is input into the initial network layer to obtain a target prediction result. For example, the initial network layer is used to optimize the initial prediction result to obtain the target prediction result. The initial network layer does not modify the dimension of the initial prediction result, but only optimizes and adjusts the values in the initial prediction result.

[0117] For example, referring to step 203, the initial prediction result includes the predicted operating data at 144 time points. Therefore, the target prediction result includes the predicted operating data at 144 time points. The predicted operating data in the initial prediction result includes the voltage of the optical module, the number of port mispackets, and FEC data, and the temperature, bias current, received power, and transmitted power of each channel. The predicted operating data in the target prediction result includes the voltage of the optical module, the number of port mispackets, and FEC data, and the temperature, bias current, received power, and transmitted power of each channel. The initial network layer may optimize and adjust the voltage, may optimize and adjust the number of port mispackets, may optimize and adjust the bias current, and so on. There is no limit to this adjustment process.

[0118] Then, a loss value can be determined based on the target prediction result and the second sample sequence. For example, the target prediction result includes predicted operation data at 144 time points, and the second sample sequence includes actual operation data at 144 time points. The second sample sequence can be used as the label data for the target prediction result, and a loss value can be determined based on the target prediction result and the second sample sequence. For example, the MSE (Mean Square Error) algorithm can be used to determine the loss value, or the MAE (Mean Absolute Error) algorithm can be used to determine the loss value. There is no limitation on the way to determine this loss value.

[0119] After obtaining the loss value, the network parameters of the initial network layer can be adjusted based on the loss value to obtain the online network layer corresponding to the model of the sample optical module, so as to obtain the online network layer corresponding to the model of the sample optical module. There is no limitation on this network parameter adjustment process. For example, the gradient descent method can be used to adjust the network parameters. During the training process, the online network layer can be finally trained through multiple iterations based on a large amount of sample data.

[0120] When adjusting the network parameters of the initial network layer, it is necessary to freeze the network parameters of the prediction model and not adjust the network parameters of the prediction model, but only adjust the network parameters of the initial network layer.

[0121] In summary, the online network layer g1 corresponding to model f1 is trained. Similarly, the online network layer g2 corresponding to model f2 can be trained, that is, the network device supports online network layers of all models.

[0122] In the above process, the online network layer is trained online according to the actual operation data of the optical module, which greatly improves the adaptability of the online network layer and further improves the prediction accuracy.

[0123] In step 204, based on the target model corresponding to the target optical module, the target online network layer corresponding to the target model is selected from all online network layers. If the target model corresponding to the target optical module is model f1, then the online network layer g1 corresponding to model f1 is used as the target online network layer. If the target model corresponding to the target optical module is model f2, then the online network layer g2 corresponding to model f2 is used as the target online network layer.

[0124] Step 205: Input the initial prediction result into the target online network layer, and optimize the initial prediction result through the target online network layer to obtain the target prediction result. Among them, the target prediction result may include a second data sequence in a future time interval, and the second data sequence may include predicted operation data at multiple time points in the future time interval, and the predicted operation data is used to determine the health status of the target optical module.

[0125] In one example, refer to Figure 4 As shown in the figure, it is a schematic diagram of adding a target online network layer behind the prediction model. In step 203, the first data sequence is input into the target prediction model to obtain the initial prediction result. In this way, the initial prediction result can be input into the target online network layer, and the initial prediction result is optimized through the target online network layer to obtain the target prediction result, and this optimization process is not limited.

[0126] For example, the target online network layer is used to optimize the initial prediction result to obtain the target prediction result. The target online network layer does not modify the dimension of the initial prediction result, but only optimizes and adjusts the values in the initial prediction result. For example, the initial prediction result includes predicted operation data at 144 time points. Therefore, the target prediction result includes predicted operation data at 144 time points (that is, the second data sequence in the future time interval, and the second data sequence includes predicted operation data at multiple time points in the future time interval).

[0127] For example, when the predicted operation data in the initial prediction result includes the voltage of the optical module, the number of port mispackets, and FEC data, and the temperature, bias current, received power, and transmitted power of each channel, the predicted operation data in the target prediction result includes the voltage of the optical module, the number of port mispackets, and FEC data, and the temperature, bias current, received power, and transmitted power of each channel. The target online network layer may optimize and adjust the voltage and may optimize and adjust the number of port mispackets, and this adjustment process is not limited.

[0128] After obtaining the predicted operation data at multiple time points in the future time interval, the health status of the target optical module can be determined based on the predicted operation data, and the analysis process of this health status is not limited.

[0129] As can be seen from the above technical solutions, in this embodiment, by combining the actual operation data (voltage, number of port mispackets, FEC data) of the target optical module and the actual operation data of each channel of the target optical module (temperature, bias current, received power, and transmitted power corresponding to each channel), the target optical module is predicted. Thus, the channel fusion method is adopted to fuse the data of all channels for prediction, improving the prediction accuracy. It has the characteristics of high prediction efficiency, high prediction accuracy, and good long-term prediction effect. By selecting the target prediction model corresponding to the type of the target optical module for prediction, the most suitable target prediction model can be selected for prediction, further improving the prediction accuracy, enhancing the prediction speed, and improving the performance of the optical module prediction algorithm. An online network layer (online layer) is added for different models, and the online network layer is trained online according to the actual scenario data of the optical module of this model, thereby improving the adaptability of the online network layer.

[0130] Based on the same application concept as the above method, in an embodiment of this application, a data prediction device is proposed, which is applied to a network device. Refer to Figure 5 As shown, it is a schematic structural diagram of the device. The device includes: An acquisition module 51, configured to acquire a first data sequence of a target optical module of a network device in a historical time interval. The target optical module is any one of all optical modules of the network device. The first data sequence includes the actual operation data at multiple time points in the historical time interval. Among them, the actual operation data includes the voltage and / or the number of port mispackets of the target optical module, and the actual operation data further includes at least one of the temperature, bias current, received power, and transmitted power corresponding to each channel of the target optical module; A determination module 52, configured to determine the type of the target optical module based on the number of target channels of the target optical module, and select a target prediction model corresponding to the type of the target optical module from all prediction models supported by the network device; A processing module 53, configured to input the first data sequence into the target prediction model to obtain an initial prediction result, and determine a target prediction result based on the initial prediction result. Among them, the target prediction result includes a second data sequence in a future time interval. The second data sequence includes the predicted operation data at multiple time points in the future time interval. The predicted operation data is used to determine the health status of the target optical module.

[0131] In one example, if the network device includes optical modules of K types of optical module types, the network device supports K prediction models, and the K prediction models correspond one-to-one to the K types of optical module types, where K is a positive integer; among them, when the determining module 52 selects a target prediction model corresponding to the target optical module type from all the prediction models supported by the network device, the target prediction model is the prediction model corresponding to the target optical module type.

[0132] In one example, when the determining module 52 determines the target optical module type based on the target channel number of the target optical module, it is specifically configured to: based on the target channel number of the target optical module, determine the optical module type corresponding to the target channel number as the target optical module type; among them, all the optical modules of the network device support K types of channel numbers, and the K types of channel numbers correspond one-to-one to the K types of optical module types; Or, based on the target channel number and target FEC type of the target optical module, determine the optical module type corresponding to the target channel number and the target FEC type as the target optical module type; among them, all the optical modules of the network device support K parameter pairs, and the K parameter pairs correspond one-to-one to the K types of optical module types, and the parameter pair includes a channel number and an FEC type; among them, the FEC type includes supporting FEC data and not supporting FEC data; if the target FEC type is supporting FEC data, the actual operation data further includes the FEC data of the target optical module; Or, based on the target channel number, target FEC type and configured target scenario parameters of the target optical module, determine the optical module type corresponding to the target channel number, the target FEC type and the target scenario parameters as the target optical module type; among them, all the optical modules of the network device support A parameter pairs, and B scenario parameters are pre-configured, and the K parameter sets composed of the A parameter pairs and the B scenario parameters correspond one-to-one to the K types of optical module types; the scenario parameters include the total time length of the historical time interval, the time length between two adjacent time points in the historical time interval, the total time length of the future time interval, and the time length between two adjacent time points in the future time interval.

[0133] In one example, when determining the target prediction result based on the initial prediction result, the processing module 53 is specifically configured to: determine the initial prediction result as the target prediction result; or, based on the target model corresponding to the target optical module, select a target online network layer corresponding to the target model from all the online network layers supported by the network device; where all the optical modules of the network device support M models in total, the network device supports M online network layers, the M models and the M online network layers are in one-to-one correspondence, and M is a positive integer; input the initial prediction result into the target online network layer, and optimize the initial prediction result through the target online network layer to obtain the target prediction result.

[0134] In one example, the target prediction model includes an iTransformer network; where the iTransformer network includes an embedding network layer, a multi-head attention network layer, a first normalization network layer, a feed-forward network layer, a second normalization network layer, and a prediction network layer; where when the processing module 53 inputs the first data sequence into the target prediction model to obtain an initial prediction result, it is specifically configured to: Input the first data sequence into the embedding network layer to obtain a first feature vector; Based on the first feature vector, obtain a Q vector, a K vector, and a V vector, and input the Q vector, the K vector, and the V vector into the multi-head attention network layer to obtain a second feature vector; Generate a third feature vector based on the first feature vector and the second feature vector; Input the third feature vector into the first normalization network layer to obtain a fourth feature vector; Input the fourth feature vector into the feed-forward network layer to obtain a fifth feature vector; Generate a sixth feature vector based on the fourth feature vector and the fifth feature vector; Input the sixth feature vector into the second normalization network layer to obtain a seventh feature vector; Input the seventh feature vector into the prediction network layer to obtain the initial prediction result.

[0135] Based on the same inventive concept as the above method, an electronic device is proposed in an embodiment of the present application. Refer to Figure 6 As shown, the electronic device includes: a processor 61 and a machine-readable storage medium 62, and the machine-readable storage medium 62 stores machine-executable instructions that can be executed by the processor 61; the processor 61 is configured to execute the machine-executable instructions to implement the data prediction method disclosed in the above examples of the present application.

[0136] Based on the same application concept as the above method, an embodiment of the present application further provides a machine-readable storage medium, on which a number of computer instructions are stored. When the computer instructions are executed by a processor, the data prediction method disclosed in the above examples of the present application can be implemented.

[0137] Among them, the above machine-readable storage medium can be any electronic, magnetic, optical or other physical storage device, which can contain or store information, such as executable instructions, data, etc. For example, the machine-readable storage medium can be: RAM (Random Access Memory), volatile memory, non-volatile memory, flash memory, storage drive (such as a hard disk drive), solid-state drive, any type of storage disk (such as an optical disc, DVD, etc.), or a similar storage medium, or a combination thereof.

[0138] Based on the same application concept as the above method, an embodiment of the present application further provides a computer program product, which may include a computer program. Among them, when the computer program is executed by a processor, the data prediction method disclosed in the above examples of the present application can be implemented.

[0139] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the embodiments of the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0140] The above are only the embodiments of the present application and are not used to limit the present application. For those skilled in the art, various changes and modifications can be made to the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included within the scope of the claims of the present application.

Claims

1. A data prediction method, characterized in that, The method includes: Obtaining a first data sequence of a target optical module of a network device in a historical time interval, where the target optical module is any one of all optical modules of the network device, and the first data sequence includes actual operation data at multiple time points in the historical time interval; wherein, the actual operation data includes the voltage and / or the number of port mispackets of the target optical module, and the actual operation data further includes at least one of the temperature, bias current, received power, and transmitted power corresponding to each channel of the target optical module; Determining a target optical module type based on the number of target channels of the target optical module, and selecting a target prediction model corresponding to the target optical module type from all prediction models supported by the network device; Inputting the first data sequence into the target prediction model to obtain an initial prediction result, and determining a target prediction result based on the initial prediction result; wherein, the target prediction result includes a second data sequence in a future time interval, and the second data sequence includes predicted operation data at multiple time points in the future time interval, and the predicted operation data is used to determine the health status of the target optical module.

2. The method according to claim 1, characterized in that, If the network device includes optical modules of K types of optical module types, the network device supports K prediction models, and the K prediction models correspond one-to-one to the K types of optical module types, where K is a positive integer; wherein, when selecting a target prediction model corresponding to the target optical module type from all prediction models supported by the network device, the target prediction model is the prediction model corresponding to the target optical module type.

3. The method according to claim 1 or 2, characterized in that The determining the target optical module type based on the number of target channels of the target optical module includes: Based on the number of target channels of the target optical module, determining the optical module type corresponding to the number of target channels as the target optical module type; wherein, all optical modules of the network device support a total of K types of channel numbers, and the K types of channel numbers correspond one-to-one to the K types of optical module types; Alternatively, based on the number of target channels and the target FEC type of the target optical module, determining the optical module type corresponding to the number of target channels and the target FEC type as the target optical module type; wherein, all optical modules of the network device support a total of K parameter pairs, and the K parameter pairs correspond one-to-one to the K types of optical module types, and the parameter pair includes the number of channels and the FEC type; wherein, the FEC type includes supporting FEC data and not supporting FEC data; if the target FEC type is supporting FEC data, the actual operation data further includes the FEC data of the target optical module; Alternatively, based on the number of target channels, the target FEC type, and the configured target scenario parameters of the target optical module, determine the optical module type corresponding to the number of target channels, the target FEC type, and the target scenario parameters as the target optical module type; wherein, all optical modules of the network device support a total of A parameter pairs, and B scenario parameters are pre-configured, and the K parameter sets composed of the A parameter pairs and the B scenario parameters correspond one-to-one with the K optical module types; the scenario parameters include the total time length of the historical time interval, the time length between two adjacent time points within the historical time interval, the total time length of the future time interval, and the time length between two adjacent time points within the future time interval.

4. The method according to claim 1, wherein The determining the target prediction result based on the initial prediction result includes: Determining the initial prediction result as the target prediction result; or, Based on the target model number corresponding to the target optical module, select the target online network layer corresponding to the target model number from all the online network layers supported by the network device; wherein, all optical modules of the network device support a total of M model numbers, the network device supports M online network layers, and the M model numbers correspond one-to-one with the M online network layers, and M is a positive integer; input the initial prediction result into the target online network layer, and optimize the initial prediction result through the target online network layer to obtain the target prediction result.

5. The method according to claim 4, characterized in that, For each online network layer supported by the network device, the training process of this online network layer includes: Obtain the initial network layer to be trained, and obtain the sample data generated by the sample optical module, where the sample optical module is any optical module among all optical modules of the network device; train the initial network layer based on the sample data to obtain the online network layer corresponding to the model number of the sample optical module; wherein, the sample data includes the first sample sequence in the first time interval and the second sample sequence in the second time interval, the total time length of the first time interval corresponds to the total time length of the historical time interval, and the total time length of the second time interval corresponds to the total time length of the future time interval; The training the initial network layer based on the sample data to obtain the online network layer corresponding to the model number of the sample optical module specifically includes: input the first sample sequence into the prediction model corresponding to the optical module type of the sample optical module to obtain an initial prediction result, input the initial prediction result into the initial network layer to obtain a target prediction result, and determine the loss value based on the target prediction result and the second sample sequence; freeze the network parameters of the prediction model, and adjust the network parameters of the initial network layer based on the loss value to obtain the online network layer corresponding to the model number of the sample optical module.

6. The method according to claim 1, wherein The target prediction model includes an iTransformer network, and the iTransformer network includes an embedding network layer, a multi-head attention network layer, a first normalization network layer, a feed-forward network layer, a second normalization network layer, and a prediction network layer; Said inputting the first data sequence into the target prediction model to obtain an initial prediction result includes: Inputting the first data sequence into the embedding network layer to obtain a first feature vector; Obtaining a Q vector, a K vector, and a V vector based on the first feature vector, and inputting the Q vector, the K vector, and the V vector into the multi-attention network layer to obtain a second feature vector; Generating a third feature vector based on the first feature vector and the second feature vector; Inputting the third feature vector into the first layer normalization network layer to obtain a fourth feature vector; Inputting the fourth feature vector into the feed-forward network layer to obtain a fifth feature vector; Generating a sixth feature vector based on the fourth feature vector and the fifth feature vector; Inputting the sixth feature vector into the second layer normalization network layer to obtain a seventh feature vector; Inputting the seventh feature vector into the prediction network layer to obtain the initial prediction result.

7. A data prediction device, characterized in that, The device includes: An acquisition module, configured to acquire a first data sequence of a target optical module of a network device in a historical time interval, where the target optical module is any one of all optical modules of the network device, and the first data sequence includes actual operation data of multiple time points in the historical time interval; wherein, the actual operation data includes the voltage of the target optical module and / or the number of port mispackets, and the actual operation data further includes at least one of temperature, bias current, received power, and transmitted power corresponding to each channel of the target optical module; A determination module, configured to determine a target optical module type based on the number of target channels of the target optical module, and select a target prediction model corresponding to the target optical module type from all prediction models supported by the network device; A processing module, configured to input the first data sequence into the target prediction model to obtain an initial prediction result, and determine a target prediction result based on the initial prediction result; wherein, the target prediction result includes a second data sequence in a future time interval, and the second data sequence includes predicted operation data of multiple time points in the future time interval, and the predicted operation data is used to determine the health status of the target optical module.

8. The device according to claim 7, characterized in that, If the network device includes optical modules of K types of optical module types, the network device supports K prediction models, and the K prediction models correspond to the K types of optical module types one by one, where K is a positive integer; wherein, when the determination module selects a target prediction model corresponding to the target optical module type from all prediction models supported by the network device, the target prediction model is the prediction model corresponding to the target optical module type.

9. The device according to claim 7 or 8, characterized in that When the determination module determines the target optical module type based on the number of target channels of the target optical module, it specifically is used for: Based on the number of target channels of the target optical module, determining the optical module type corresponding to the number of target channels as the target optical module type; wherein, all optical modules of the network device support K types of channel numbers in total, and the K types of channel numbers correspond to the K types of optical module types one by one; Alternatively, based on the target channel number and target FEC type of the target optical module, determine the optical module type corresponding to the target channel number and the target FEC type as the target optical module type; wherein, all optical modules of the network device support a total of K parameter pairs, and the K parameter pairs correspond one-to-one with the K types of optical module types, and the parameter pair includes a channel number and an FEC type; wherein, the FEC type includes supporting FEC data and not supporting FEC data; if the target FEC type is supporting FEC data, the actual operation data further includes the FEC data of the target optical module; Alternatively, based on the target channel number, target FEC type, and configured target scenario parameters of the target optical module, determine the optical module type corresponding to the target channel number, the target FEC type, and the target scenario parameters as the target optical module type; wherein, all optical modules of the network device support a total of A parameter pairs, and B scenario parameters are pre-configured, and the K parameter sets composed of the A parameter pairs and the B scenario parameters correspond one-to-one with the K types of optical module types; the scenario parameters include the total time length of the historical time interval, the time length between two adjacent time points within the historical time interval, the total time length of the future time interval, and the time length between two adjacent time points within the future time interval.

10. The device according to claim 7, characterized in that, When the processing module determines the target prediction result based on the initial prediction result, it specifically is used for: Determine the initial prediction result as the target prediction result; or, Based on the target model number corresponding to the target optical module, select the target online network layer corresponding to the target model number from all the online network layers supported by the network device; wherein, all optical modules of the network device support a total of M model numbers, the network device supports M online network layers, and the M model numbers correspond one-to-one with the M online network layers, and M is a positive integer; input the initial prediction result into the target online network layer, and optimize the initial prediction result through the target online network layer to obtain the target prediction result.

11. The device according to claim 7, characterized in that, The target prediction model includes an iTransformer network; wherein, the iTransformer network includes an embedding network layer, a multi-attention network layer, a first normalization network layer, a feed-forward network layer, a second normalization network layer, and a prediction network layer; wherein, when the processing module inputs the first data sequence into the target prediction model to obtain an initial prediction result, it specifically is used for: Input the first data sequence into the embedding network layer to obtain a first feature vector; Based on the first feature vector, obtain a Q vector, a K vector, and a V vector, and input the Q vector, the K vector, and the V vector into the multi-attention network layer to obtain a second feature vector; Generate a third feature vector based on the first feature vector and the second feature vector; Input the third feature vector into the first normalization network layer to obtain a fourth feature vector; Input the fourth feature vector into the feed-forward network layer to obtain a fifth feature vector; Generate a sixth eigenvector based on the fourth eigenvector and the fifth eigenvector; Input the sixth eigenvector into the second layer normalization network layer to obtain a seventh eigenvector; Input the seventh eigenvector into the prediction network layer to obtain the initial prediction result.

12. An electronic device, characterized in that, Comprising: A processor and a machine-readable storage medium, the machine-readable storage medium storing machine-executable instructions that can be executed by the processor; The processor is configured to execute the machine-executable instructions to implement the method according to any one of claims 1-6.

Images