CN-121567208-B - Transmitting power determining method and training method of transmitting power determining model

CN121567208BCN 121567208 BCN121567208 BCN 121567208BCN-121567208-B

Abstract

The application provides a transmitting power determining method and a training method of a transmitting power determining model, which can be applied to the technical fields of optical communication technology and artificial intelligent data centers. The method for determining the transmitting power comprises the steps of obtaining historical operation data sets of a plurality of channels of a multimode optical module, determining candidate transmitting power characteristics of the channels at the current moment according to a historical operation data sequence obtained based on the historical operation data sets of the channels by utilizing a time neural network branch of a transmitting power determination model, extracting characteristics of multi-channel characteristic tensors determined by the historical operation data sets by utilizing a space neural network branch of the transmitting power determination model to obtain multi-channel fusion characteristics, and determining actual transmitting power of the channels at the current moment according to the multi-channel fusion characteristics and the candidate transmitting power characteristics of the channels.

Inventors

LI CHAO
WANG CONG
JIANG JINKUN
CHEN XIANG

Assignees

苏州元脑智能科技有限公司

Dates

Publication Date: 20260505
Application Date: 20260122

Claims (9)

1. A transmit power determination method for a multimode optical module for optical communication of a data center network, the multimode optical module comprising a plurality of parallel channels for data transmission, the method comprising: Acquiring a historical operation data set of each of a plurality of channels of the multimode optical module, wherein the historical operation data set comprises a plurality of historical operation data with historical time stamps; Determining candidate transmission power characteristics of each of a plurality of channels at the current moment according to a historical operation data sequence obtained based on each historical operation data set of the plurality of channels by utilizing a time neural network branch of a transmission power determination model; the method comprises the steps of utilizing a space neural network branch of a transmitting power determining model to perform feature extraction on multi-channel feature tensors determined by a plurality of historical operation data sets to obtain multi-channel fusion features, wherein the multi-channel fusion features characterize single-channel operation states of a plurality of channels and inter-channel interference environments among the channels; The method comprises the steps of respectively fusing multi-channel fusion characteristics with candidate transmission power characteristics of a plurality of channels by utilizing a fusion sub-module of an output module of a transmission power determination model to obtain target fusion characteristics after correction of the candidate transmission power characteristics of the channels; And respectively carrying out regression prediction on the target fusion characteristics of each of the channels by utilizing a prediction submodule of an output module of the transmitting power determination model, and outputting the real transmitting power of each of the channels.
2. The method of claim 1, wherein the historical operating data includes historical bias current, historical drive voltage and historical temperature, wherein using the spatial neural network branch of the transmit power determination model to perform feature extraction on the multi-channel feature tensors determined from the plurality of sets of historical operating data to obtain the multi-channel fusion feature comprises: Aligning and splicing a plurality of historical bias currents, a plurality of historical driving voltages and a plurality of historical temperatures in each historical operation data set of a plurality of channels according to historical time stamps to form a single-channel feature matrix of each channel, and stacking the single-channel feature matrix of each channel according to channel dimensions to obtain a multi-channel feature tensor for each channel; And carrying out feature extraction on the multi-channel feature tensor based on the spatial neural network branch to obtain the multi-channel fusion feature.
3. The method of claim 2, wherein aligning and stitching the plurality of historical bias currents, the plurality of historical drive voltages, and the plurality of historical temperatures in the historical operating dataset for each of the plurality of channels according to the historical timestamps to form a single channel feature matrix for each of the plurality of channels, and stacking the single channel feature matrix for each of the plurality of channels according to the channel dimension to obtain the multi-channel feature tensor for the plurality of channels comprises: Sequentially arranging a plurality of historical bias currents according to the sequence of the historical time stamps to obtain a historical bias current sequence, sequentially arranging a plurality of historical driving voltages according to the sequence of the historical time stamps to obtain a historical driving voltage sequence, and sequentially arranging a plurality of historical temperatures according to the sequence of the historical time stamps to obtain a historical temperature sequence; For each channel, performing time stamp alignment on a historical bias current sequence, a historical driving voltage sequence and a historical temperature sequence according to the sequence of the historical time stamps, then splicing to obtain a two-dimensional feature matrix, taking the two-dimensional feature matrix as a single-channel feature matrix, and arranging rows and columns of the two-dimensional feature matrix according to the historical time stamps and a preset data sequence respectively; stacking the single-channel feature matrixes of the channels according to the channel dimension to obtain a three-dimensional feature tensor, and taking the three-dimensional feature tensor as a multi-channel feature tensor.
4. The method of claim 1, wherein the historical operating data further includes historical transmit power, and wherein determining, for each channel, candidate transmit power characteristics at the current time instance based on the historical operating data sequence derived based on the historical operating data sequence using the temporal neural network branch of the transmit power determination model comprises: sequentially arranging a plurality of historical transmitting powers in the historical operating data set according to the sequence of the historical time stamps to obtain a historical transmitting power sequence; And inputting the historical transmission power sequence into a time neural network branch, and outputting the candidate transmission power characteristic at the current time.
5. The method of claim 1, wherein the prediction submodule includes a fully-connected neural network, wherein the regression prediction is performed on the target fusion features of each of the plurality of channels based on the prediction submodule of the output module, and wherein outputting the true transmit power of each of the plurality of channels includes, for each of the channels: Based on the fully connected neural network, nonlinear transformation is carried out on the target fusion characteristics to obtain hidden characteristics, regression prediction is carried out on the hidden characteristics, and real transmitting power is output.
6. The method of claim 1 or 4, wherein the temporal neural network branches include at least one of temporal neural network branches based on a time-series convolutional network, temporal neural network branches based on a cyclic neural network, temporal neural network branches based on a long-short-term memory network.
7. The method of claim 1 or 2, wherein the spatial neural network branches comprise convolutional neural network-based spatial neural network branches.
8. A training method for a transmit power determination model of a multimode optical module, the method comprising: acquiring a sample operation data set of each of a plurality of sample channels of a sample multimode optical module, wherein the sample operation data set comprises a plurality of sample operation data with sample time stamps; Determining a first transmission power characteristic of each of a plurality of sample channels at the current moment of a sample according to a sample operation data sequence obtained based on each sample operation data set of the plurality of sample channels by utilizing a time neural network branch of a candidate transmission power determination model; The method comprises the steps of utilizing a space neural network branch of a candidate transmitting power determining model to perform feature extraction on sample multi-channel feature tensors determined by a plurality of sample operation data sets to obtain sample multi-channel fusion features, wherein the sample multi-channel fusion features characterize respective single-channel operation states of a plurality of sample channels and inter-channel interference environments among multi-sample channel environments; The method comprises the steps of respectively fusing sample multi-channel fusion characteristics with respective first transmission power characteristics of a plurality of sample channels by utilizing a fusion submodule of an output module of a candidate transmission power determination model to obtain corrected characteristics of the respective first transmission power characteristics of the plurality of sample channels, wherein the corrected characteristics represent transmission power change trend of the sample channels and inter-channel interference environment where the sample channels are positioned; Carrying out regression prediction on the respective corrected characteristics of the plurality of sample channels by using a prediction submodule of an output module of the candidate transmission power determination model, and outputting respective transmission power prediction values of the plurality of channels; And training the candidate transmission power determination model according to the respective transmission power predicted values of the plurality of sample channels and the respective transmission power true values of the plurality of sample channels to obtain a transmission power determination model, wherein the respective transmission power true values of the plurality of sample channels are the optical power received by the opposite-end optical module for carrying out data transmission with the sample multimode optical module through the plurality of sample channels at the current time of the sample.
9. An electronic device, comprising: One or more processors; a memory for storing one or more computer programs, Characterized in that the one or more processors execute the one or more computer programs to implement the steps of the method according to any one of claims 1-8.

Description

Transmitting power determining method and training method of transmitting power determining model Technical Field The application relates to the technical field of optical communication technology and artificial intelligent data centers, in particular to a transmitting power determining method and a training method of a transmitting power determining model. Background With the rapid development of cloud computing, big data, and artificial intelligence technologies, data center networks are evolving towards high bandwidth, low latency, and high reliability. In a data center network system, an optical module is used as a core component for realizing optical communication, and the stability of the working state of the optical module directly determines the transmission quality of the whole data link. Digital diagnostic monitoring (Digital Diagnostic Monitoring, DDM) techniques, for example, may be employed to monitor operating parameters of the optical module (e.g., transmit optical power, receive optical power, etc.) to enable quality monitoring and fault early warning for the optical module. However, a significant error exists between the related optical module transmitting power monitored by adopting the DDM technology and the real transmitting power of the optical module, so that the reliability of the transmitting power in the monitoring result is lower. Disclosure of Invention In view of the above, the present application provides a transmission power determination method and a training method of a transmission power determination model. In addition, the application also provides a transmitting power determining device, a training device of the transmitting power determining model, equipment, a medium and a program product. According to one aspect of the application, a transmitting power determining method for a multimode optical module is provided, and the transmitting power determining method comprises the steps of obtaining a historical operation data set of each of a plurality of channels of the multimode optical module, wherein the historical operation data set comprises a plurality of historical operation data with historical time stamps, determining candidate transmitting power characteristics of each of the plurality of channels at the current moment by utilizing a time neural network branch of a transmitting power determining model according to a historical operation data sequence obtained based on each of the historical operation data sets of the plurality of channels, performing characteristic extraction on multi-channel characteristic tensors determined by the historical operation data sets by utilizing the space neural network branch of the transmitting power determining model to obtain multi-channel fusion characteristics, wherein the multi-channel fusion characteristics represent single-channel operation states of each of the plurality of channels and inter-channel interference environments among the plurality of channels, and determining real transmitting power of each of the plurality of channels at the current moment according to the multi-channel fusion characteristics and the candidate transmitting power characteristics of each of the plurality of channels. According to another aspect of the application, a training method for a transmission power determination model of a multimode optical module is provided, and the training method comprises the steps of obtaining a sample operation data set of each of a plurality of sample channels of the sample multimode optical module, wherein the sample operation data set comprises a plurality of sample operation data with sample time stamps, determining a transmission power prediction value of each of the plurality of sample channels at the current moment of a sample according to a sample operation data sequence obtained based on the sample operation data set of each of the plurality of sample channels by utilizing a candidate transmission power determination model, performing feature extraction on a sample multi-channel feature tensor determined by the sample operation data set to obtain a sample multi-channel fusion feature, characterizing a single-channel operation state of each of the plurality of sample channels and an inter-channel interference environment between the sample multi-channel environment by utilizing the candidate transmission power determination model, determining a transmission power prediction value of each of the plurality of sample channels at the current moment of a sample according to the transmission power prediction value of each of the plurality of sample channels and the sample operation data set, and determining a real power of each of the sample channels by utilizing the candidate transmission power determination model, and the real power determination model is performed when the sample transmission power of each of the sample channels is real, and the real power determination model is performed by the real power o