Multimode Crosstalk Analysis Method Based on Distributed Deep Neural Network Mode Field Recognition
By optimizing the distributed deep neural network model and cross-entropy cost function, combined with the dropout method, the problem of pattern crosstalk identification in multimode fiber transmission was solved, achieving fast and accurate pattern crosstalk analysis, improving the system's fault tolerance and reducing communication costs.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NANJING UNIV OF INFORMATION SCI & TECH
- Filing Date
- 2023-05-30
- Publication Date
- 2026-06-30
AI Technical Summary
Existing optical communication systems cannot effectively measure the intermode crosstalk level of discrete mode devices in multimode optical fibers, and traditional deep neural networks suffer from computational delays and high energy consumption, making it difficult to meet the real-time control requirements of radio resources.
By employing a distributed deep neural network model, and through training and cross-entropy cost function optimization, combined with the dropout method, we can achieve fast and accurate identification of crosstalk in multimode fiber transmission modes.
It improves the accuracy of pattern crosstalk identification, reduces communication costs, enhances the system's fault tolerance and data privacy, and is suitable for edge computing environments.
Smart Images
Figure CN116634311B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to optical transmission technology in the field of communication technology, and more particularly to a multimode crosstalk analysis method based on distributed deep neural network mode field identification. Background Technology
[0002] With the continuous increase in network users and the emergence of new network data services, the demand for network capacity is constantly growing. Optical communication systems based on ordinary single-mode optical fibers can no longer cope with the booming demand for network capacity. Multimode optical fibers, as a new dimension of optical multiplexing, have received widespread attention worldwide.
[0003] To implement a mode-division multiplexing (MDD) optical transmission system, in addition to multimode fiber, it is also necessary to use matching discrete mode devices, such as mode converters and mode multiplexers / demultiplexers. These discrete mode devices, together with the multimode fiber, determine the performance of the MMD optical transmission system. Therefore, accurately measuring the inter-mode crosstalk (IMC) level of these discrete mode devices has become a top priority. As a representative of discrete mode devices, the measurement of discrete IMC in mode multiplexers / demultiplexers has been extensively studied.
[0004] Deep neural networks have a multi-layered structure, and their representation learning is also hierarchically distributed. For input vectors, layer-by-layer transmission introduces latency to later layers of the deep neural network. Moreover, as computational parameters accumulate, computational energy consumption increases layer by layer, which is detrimental to the real-time control of radio resources in next-generation mobile networks. Based on this, a distributed deep neural network model is proposed, which has a distributed computing hierarchical structure. Distributed deep neural networks oriented towards edge computing refer to mapping parts of a single deep neural network onto distributed heterogeneous devices, including cloud, edge, and geographically distributed terminal devices.
[0005] Distributed deep neural networks (DNNs) not only enable deep neural network inference in the cloud but also allow for fast, localized inference using shallower parts of the neural network on edge and terminal devices. Supported by a scalable distributed computing hierarchy, DNNs can scale both in terms of network size and geographical reach. Due to their distributed nature, DNNs enhance sensor fusion, system fault tolerance, and data privacy. In implementing DNNs, deep neural network components are mapped onto a distributed computing hierarchy. By jointly training these components, device communication and resource usage are minimized, while maximizing the utilization of features extracted from cloud computing. DNNs can leverage the geographical diversity of sensors to improve target recognition accuracy and reduce communication costs. Summary of the Invention
[0006] Purpose of the invention: The purpose of this invention is to provide a multimode crosstalk analysis method based on distributed deep neural network mode field identification.
[0007] Technical solution: The method of the present invention includes the following steps:
[0008] (1) The distributed deep neural network used is trained, and different crosstalk information is obtained by changing the propagation mode of the optical signal in different optical transmission systems.
[0009] (2) The crosstalk information was generated into an image and input into a distributed deep neural network for training, thus obtaining the required distributed deep neural network pattern crosstalk recognition model.
[0010] (3) Combine the obtained crosstalk identification model with different optical transmission systems to obtain corresponding faster and more accurate pattern crosstalk information.
[0011] Furthermore, in step (1), the distributed deep neural network extracts features from the image layer by layer and learns more complex features step by step.
[0012] Furthermore, in step (2), the distributed deep neural network pattern crosstalk identification model uses the cross-entropy cost function instead of the mean square error cost function.
[0013] Furthermore, step (3) includes multiplexing the optical signals emitted by multiple optical transmitters, and after transmission through multimode optical fiber, dividing the optical signals into two parts, 1% and 99%. The 1% optical signal enters the image sensor to generate an image, and these images are processed by a pre-trained distributed deep neural network model to obtain faster and more accurate pattern crosstalk information.
[0014] Furthermore, the optical signals emitted by the plurality of optical transmitters are multiplexed by a mode multiplexer.
[0015] Furthermore, the optical signal is separated by an optical isolator.
[0016] Furthermore, the optical signal is multiplexed and coupled into a multimode optical fiber.
[0017] Furthermore, in step (2), the method of adding dropout to the distributed deep neural network is described.
[0018] Furthermore, the dropout method calculates the average of multiple models by changing the structure of the neural network model.
[0019] Furthermore, the distributed deep neural network in step (2) is computed layer by layer according to forward propagation.
[0020] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: By using a distributed deep neural network model to identify crosstalk in multimode fiber transmission modes, the present invention can improve the accuracy of target recognition and reduce communication costs. Compared with traditional image recognition technology, the distributed deep neural network improves the accuracy of image recognition, communication and latency requirements, while achieving higher fault tolerance and privacy. Attached Figure Description
[0021] Figure 1 This is a flowchart of the present invention;
[0022] Figure 2 Diagram of a distributed deep neural network pattern crosstalk recognition model;
[0023] Figure 3 This is a diagram of a multimode crosstalk model based on distributed deep neural network mode field recognition. Detailed Implementation
[0024] The technical solution of the present invention will be further described below with reference to the accompanying drawings.
[0025] This invention is a multimode crosstalk analysis method based on distributed deep neural network mode field recognition, the flowchart of which is shown below. Figure 1 As shown, firstly, the distributed deep neural network used is trained to obtain different crosstalk information by changing the propagation mode of the optical signal in different optical transmission systems. Secondly, the crosstalk information is used to generate images and input into the distributed deep neural network for training, thus obtaining the required distributed deep neural network pattern crosstalk recognition model. Finally, the obtained crosstalk recognition model is combined with different optical transmission systems to obtain corresponding faster and more accurate pattern crosstalk information.
[0026] Deep neural networks can extract image features layer by layer and learn more complex features progressively. Therefore, it is natural to think of using distributed deep neural network models for pattern crosstalk recognition. Based on a pattern crosstalk digit database, the recognition model used is, for example... Figure 2 As shown, the deep neural network has 784 neurons in the first layer, three hidden layers (500 units in the first, 300 units in the second, and 300 units in the third), and 10 neurons in the output layer. The activation function for each neuron is the sigmoid function, and the output layer uses softmax. Furthermore, the cost function is changed to the cross-entropy function, and dropout is added to prevent overfitting.
[0027] Traditional neural network pattern recognition methods:
[0028] A neural network is a feedforward network that computes layer by layer using forward propagation. The input to each neuron in a layer is a weighted sum of the outputs of the previous layer, and this process continues until the output layer is reached. The output value is then compared to the expected value. The performance of the network can be measured using the mean squared error cost function, which is defined as follows:
[0029]
[0030] In equation (1), n is the total number of training samples, a is the network output vector, and y is the training label vector. Then, the network undergoes backpropagation. Based on the chain rule of calculus, gradient descent is used to adjust the weights of each layer of the network using the error. The learning algorithm for adjusting the weights is as follows:
[0031] ω j(k+1) =ω j(k) +Δω ij =w j(k) -ηg k (2)
[0032] In equation (2), w ij(k+1 ) represents the updated weight, w ij(k) η is the current weight value, η is the learning rate, and g is the learning rate. k This is the current gradient.
[0033] Improvements to the cost function of deep neural networks:
[0034] To enable neural networks to learn faster, the cross-entropy cost function is used instead of the mean squared error cost function. Analyzing the previous formula (1) for the mean squared error cost function of deep neural networks, assuming there is only one training sample (n=1) and only one neuron, and σ represents the sigmoid function, the cost function becomes:
[0035]
[0036] In equation (3), a = σ(z), z = wx + b. Since the gradient descent method is used, the partial derivatives with respect to w and b are obtained respectively.
[0037]
[0038]
[0039] For the sigmoid function, when the neuron output is close to 1 or close to 0, the curve is very flat, causing the derivative σ′(z) to be close to 0, which makes the partial derivatives with respect to w and b close to 0, thus making the network update very slow.
[0040] Define the cross-entropy cost function (assuming it's a binary classification problem).
[0041]
[0042] From equation (6), we get: when the expected value y is equal to the neural network output, c = 0. Calculate its partial derivative with respect to the weights.
[0043]
[0044] Equation (7) shows that the learning of weights mainly depends on σ(z)-y, which is the error between the output value and the expected value. The advantage of this cost function is that the weights are updated quickly when the deviation is large, and the weights are updated slowly when the deviation is small. Therefore, the cross-entropy cost function is chosen.
[0045] Implementation of methods to prevent overfitting: Deep neural networks often exhibit overfitting problems where good performance on the training set fails to generalize to the test set. To address this, dropout is introduced into deep neural networks. The essence of dropout is to modify the neural network model structure and then calculate the average of multiple models. During dropout, some neurons are temporarily de-outputted with a probability p, while the remaining neurons (with a probability of 1-p) are trained. Afterward, all neurons are restored, and in the next training iteration, some neurons are randomly selected again with probability p to temporarily de-output, repeating this process. Using dropout is equivalent to selecting a different network structure for each training iteration, thus reducing the adaptability and dependence between neurons. Adding dropout allows the network to learn more robust and resilient features; an empirical value of 0.7 is selected here.
[0046] A multimode crosstalk analysis method based on distributed deep neural network mode field identification: In mode-division multiplexing optical transmission systems, mode crosstalk between discrete mode devices is unavoidable. However, there is still no convenient and fast method for analyzing mode crosstalk. In this invention, we use an existing distributed deep neural network, trained and applied to the analysis of mode crosstalk. The multimode crosstalk analysis method based on distributed deep neural network mode field identification is as follows: Figure 3 As shown, optical signals from n optical transmitters are multiplexed using a mode multiplexer and coupled into a multimode fiber. After transmission through the multimode fiber, an optical isolator divides the optical signal into two parts: 1% and 99%. The 1% optical signal is then sent to an image sensor to generate an image. These images, processed by a pre-trained distributed deep neural network model, quickly provide the mode crosstalk information after transmission through the optical transmission system.
Claims
1. A multimode crosstalk analysis method based on distributed deep neural network mode field recognition, characterized in that: The method includes the following steps: (1) The distributed deep neural network used is trained, and different crosstalk information is obtained by changing the propagation mode of the optical signal in different optical transmission systems; (2) The crosstalk information was used to generate an image and input into a distributed deep neural network for training, thus obtaining the required distributed deep neural network pattern crosstalk recognition model. (3) Combine the obtained crosstalk identification model with different optical transmission systems to obtain corresponding faster and more accurate pattern crosstalk information; The method of adding dropout to the distributed deep neural network in step (2) is described above; The dropout method calculates the average of multiple models by changing the structure of the neural network model; In step (1), the distributed deep neural network extracts the features of the image layer by layer and learns more complex features step by step. The distributed deep neural network pattern crosstalk identification model in step (2) uses the cross-entropy cost function instead of the mean square error cost function.
2. The multimode crosstalk analysis method based on distributed deep neural network mode field recognition according to claim 1, characterized in that: Step (3) includes multiplexing the optical signals emitted by multiple optical transmitters. After transmission through multimode optical fiber, the optical signals are divided into two parts: 1% and 99%. The 1% optical signal enters the image sensor to generate images. These images are then processed by a pre-trained distributed deep neural network model to obtain faster and more accurate pattern crosstalk information.
3. The multimode crosstalk analysis method based on distributed deep neural network mode field recognition according to claim 2, characterized in that: The optical signals emitted by the multiple optical transmitters are multiplexed by a mode multiplexer.
4. The multimode crosstalk analysis method based on distributed deep neural network mode field recognition according to claim 2, characterized in that: The optical signal is separated by an optical isolator.
5. The multimode crosstalk analysis method based on distributed deep neural network mode field recognition according to claim 2, characterized in that: The optical signal is multiplexed and coupled into the multimode optical fiber.
6. The multimode crosstalk analysis method based on distributed deep neural network mode field recognition according to claim 1, characterized in that: In step (2), the distributed deep neural network is computed layer by layer according to forward propagation.