A time-delay reserve pool calculation system and method for high-speed channel equalization
Through deep structure delayed reserve pool and complementary matrix processing, the problem of mask size limitation in high-speed channel equalization is solved, the processing speed is improved and performance is maintained, and the speed and accuracy balance is achieved.
Patent Information
- Application Number
- CN202411179370.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-27
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2044-08-27
AI Technical Summary
The computing speed of existing delayed reserve pool computing systems in high-speed channel equalization tasks is limited by the input mask size. Reducing the mask size will lead to performance degradation, and existing methods increase hardware cost and integration difficulty.
Using a deep structure delayed reserve pool and data processing method based on complementary matrix, the input mask matrix size is reduced through the utilization of nonlinear characteristics and memory length, the channel equalization processing speed is improved, and performance loss is alleviated through complementary matrix. .
It effectively reduces the mask size, improves the channel equalization speed, maintains system performance, and achieves a balance between speed and accuracy.
Smart Images

Figure CN118713761B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of photon computing technology, and in particular relates to a time-delay reserve pool computing system and method for high-speed channel equalization. Background Art
[0002] In recent years, fiber optic communication technology has played an extremely important role in the information society. Signals will be distorted after passing through the fiber optic channel, so channel equalization is required to restore the original signal transmitted. Neural network technology can effectively perform channel equalization. Reservoir Computing (RC), as a new neural network architecture that simulates the principles of biological neural networks, is characterized by its streamlined architecture and significantly simplified training process. It effectively avoids the problems of slow convergence, easy to fall into local optimality, and complex parameter tuning in neural network methods.
[0003] RC can cleverly utilize existing optical components and system platforms to build physical neural networks at low hardware cost through time division multiplexing. This form of RC is called Time-Delay Reservoir Computing (TDRC). This feature enables TDRC to be better integrated into optical fiber communication systems in hardware.
[0004] However, high-speed signal transmission requires TDRC's high-speed channel equalization, and the operation speed of the TDRC architecture is limited by the input mask. The mask can be regarded as an input matrix, which provides richer dynamic characteristics for the distorted signal. The larger the mask, the larger the amount of input signal data generated, and therefore the longer it takes to complete the channel equalization task. Reducing the mask size can achieve hardware-level acceleration, but it will weaken the dynamic capabilities of the system, resulting in a decrease in TDRC performance. In order to support high-speed and real-time channel equalization tasks, there is an urgent need for systems and methods that enable TDRC to perform high-speed operations while minimizing performance losses.
[0005] A literature search found that Zheng Li et al. from China published a paper titled "Packet header recognition utilizing an all-optical reservoir based on reinforcement-learning-optimized double-ring resonators" in the IEEE Journal of Selected Topics in Quantum Electronics in 2023. The paper proposed using all-optical hardware encoding at the input end to replace the digital mask of the traditional method in the form of hardware, so that the input signal can be input into the system in parallel, thereby improving the information processing speed of TDRC. In addition, by implementing it in an all-optical way, the time required for photoelectric conversion can be reduced, further improving the information processing speed of TDRC. However, this method increases the hardware required at the input end, which increases the overall cost of the system and places higher requirements on future integration processes.
[0006] A literature search revealed that Junfeng Zhang et al. from China published a paper titled "High-speed parallel processing with photonic feedforwardreservoir computing" in the journal Optics Express in 2023. The paper proposed expanding TDRC in the wavelength dimension and improving the information processing capability by increasing the reservoir. In this way, the increase in the reservoir can reduce the performance loss caused by the reduction of the mask. However, this method requires careful control of the synchronization between different wavelengths. Summary of the invention
[0007] In view of the above, the purpose of the present invention is to provide a time-delayed reserve pool calculation system and method for high-speed channel equalization, which utilizes the deep structure of the reserve pool and the data processing acceleration based on complementary matrices, effectively utilizes different degrees of nonlinear characteristics and memory lengths, and thus can reduce the size of the mask matrix at the input end, improve the processing speed of the channel equalization task, and at the same time does not cause a significant impact on the performance.
[0008] To achieve the above-mentioned purpose of the invention, an embodiment provides a delay-type reserve pool calculation system for high-speed channel equalization, comprising:
[0009] A signal preprocessing unit, which is used to perform mask processing on the distorted signal obtained after the original signal is transmitted through the optical fiber channel to obtain the initial input signal of the time-delay type depth reserve pool;
[0010] A time-delay deep reserve pool, comprising an input layer, a state layer, and an output layer, wherein the state layer is composed of a plurality of state calculation sublayers connected in sequence to form a deep structure, each state calculation sublayer comprises a nonlinear node and is combined with a time-delay loop composed of a plurality of virtual nodes, and the nonlinear node of the latter state calculation sublayer of two adjacent state calculation sublayers is connected to the virtual node or nonlinear node of the previous state calculation sublayer, the initial input signal is input through the input layer, and nonlinear mapping and interaction are performed through each state calculation sublayer, and the state of the virtual node is read by the output layer through a readout matrix;
[0011] The signal post-processing unit is used to construct a complementary matrix and multiply the complementary matrix with the readout matrix to achieve channel equalization to obtain an equalized signal.
[0012] Preferably, in the signal preprocessing unit, a mask matrix of size M×K and randomly generated is selected to multiply the distorted signal of size K×T to implement mask processing, and the distorted signal is converted into an initial input signal of dimension M×T associated with the mask matrix.
[0013] Preferably, the initial input signal is expanded into a one-dimensional signal, encoded into the amplitude or phase of the photoelectric domain, and then serially input into the time-delay type deep reserve pool.
[0014] Preferably, each nonlinear node is used to perform nonlinear mapping on the input signal and map it to a high-dimensional space, and the nonlinear mapping is achieved by using a circuit, an optical path, or an optical circuit method, wherein the optical circuit method includes a Mach-Zehnder modulator.
[0015] Preferably, each delay loop is connected by means of an electric circuit, an optical circuit, or an optical circuit, wherein the optical circuit includes an optical fiber delay line.
[0016] Preferably, each virtual node corresponds to a section of the delay loop, and the delay loop is virtually divided into multiple sections according to the requirements of the channel equalization task. Each section corresponds to a virtual node, and the encoded information on the virtual node is the state value of the virtual node, which corresponds to the amplitude or phase of the optical domain.
[0017] Preferably, the nodes of two adjacent state calculation sublayers are connected in the form of circuits, optical paths, or optical circuits, wherein the optical path includes an optical fiber delay line.
[0018] Preferably, the dimension of the complementary matrix is N×T, and N is greater than M. The complementary matrix includes the state values of the virtual nodes in all state calculation sublayers, or also includes at least one of the state values of the virtual nodes of different state calculation sublayers in additional time steps and the initial input signal.
[0019] Preferably, the readout matrix is subjected to parameter optimization before being applied. During the optimization, the parameters of the readout matrix are adjusted by comparing the difference between the equalized signal output by the signal post-processing unit and the original signal.
[0020] To achieve the above-mentioned purpose of the invention, an embodiment of the present invention further provides a method for calculating a delay-type reserve pool for high-speed channel equalization. The method adopts the above-mentioned delay-type reserve pool calculation system and includes the following steps:
[0021] The signal preprocessing unit is used to perform mask processing on the distorted signal obtained after the original signal is transmitted through the optical fiber channel to obtain the initial input signal of the time-delay type depth reserve pool;
[0022] The time-delayed deep reserve pool is used to calculate the state value of each position based on the initial input signal. The state value corresponds to the amplitude or phase of the photoelectric domain, and the state of the virtual node is read through the readout matrix.
[0023] The signal post-processing unit is used to multiply the constructed complementary matrix with the readout matrix to achieve channel equalization and obtain an equalized signal.
[0024] Compared with the prior art, the present invention has the following beneficial effects:
[0025] The state layer of the time-delay deep reserve pool constructed by the present invention is composed of multiple state calculation sublayers connected in sequence, and has a certain depth structure. The depth structure is combined with a complementary matrix, so that the rich dynamic characteristics inside the reserve pool can be fully utilized in the complementary matrix, thereby effectively reducing the size of the mask, reducing the amount of data generated, and improving the channel equalization speed while minimizing performance damage, ultimately achieving a balance between speed and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0026] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.
[0027] Figure 1 It is a schematic diagram of the structure of a time-delayed reserve pool computing system for high-speed computing provided in an embodiment;
[0028] Figure 2is a schematic diagram of mask preprocessing provided by an embodiment;
[0029] Figure 3 is a schematic diagram of the structure of a deep time-delay storage tank provided in an embodiment;
[0030] Figure 4 is a schematic diagram of a complementary matrix provided in an embodiment;
[0031] Figure 5 It is a flow chart of a delay-type reserve pool calculation method for high-speed channel equalization provided in an embodiment. DETAILED DESCRIPTION
[0032] To make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific implementation methods described herein are only used to explain the present invention and do not limit the scope of protection of the present invention.
[0033] The embodiment provides a time delay reserve pool calculation system for high-speed channel equalization to solve the technical problems of existing TDRC in achieving high-speed channel equalization goals, such as Figure 1 As shown, the time-delay reservoir computing system includes a signal preprocessing unit, a time-delay depth reservoir, and a signal post-processing unit.
[0034] In an embodiment, the original signal used for communication will be distorted during long-distance optical fiber signal transmission to form a distorted signal. The distorted signal needs to be masked in a signal preprocessing unit before being input into the delay-type deep reserve pool for high-speed channel equalization. The amount of input generated is reduced by reducing the mask size, thereby accelerating the channel equalization calculation speed of the delay-type deep reserve pool at a given input rate.
[0035] like Figure 2 As shown, a mask matrix W of size M×K is selected and randomly generated in The distorted signal u of size K×T is multiplied to realize mask processing. After passing through the mask matrix, the distorted signal u will be converted into the mask matrix W in The relevant initial input signal x is of dimension M×T in , since the initial input signal x in The state and mask matrix W in It is related to the selection of and can be regarded as a random linear combination of the distorted signal u. Figure 2 The n in is the index of the discrete time step, for example, the signal sequence is sorted in time u(1),u(2),...,u(n).
[0036] The mask size is reduced by reducing the mask matrix W inAfter reducing M, the initial input signal x in The amount of data will be reduced, so the time required to complete the channel equalization task will be shortened, achieving the purpose of accelerating information processing speed without causing a significant impact on performance.
[0037] In an embodiment, the time-delayed depth reserve pool is used to complete the channel equalization calculation task based on the initial input signal. Figure 3 As shown in the figure, it includes an input layer, a state layer, and an output layer, wherein the input layer is used to input the initial input signal. The state layer is used to calculate the state values of the initial input signal at different positions and times. Specifically, the state layer is composed of multiple state calculation sublayers connected in sequence to form a deep structure. Figure 3 An example is given of a deep structure composed of two state calculation sublayers connected in sequence. Each state calculation sublayer contains a nonlinear node NL and is combined with multiple virtual nodes (such as Figure 3 The delay loop is composed of two adjacent state calculation sublayers (as shown by the black dots in the figure), and the two adjacent state calculation sublayers are connected through nodes to realize the transmission of state values. Specifically, the nonlinear node of the latter state calculation sublayer is connected to the virtual node or nonlinear node of the former state calculation sublayer.
[0038] In the time-delay deep reserve pool, each nonlinear node is used to perform nonlinear mapping on the input signal and map it to a high-dimensional space to solve the linear inseparability problem. Specifically, each nonlinear node can use a circuit, an optical path, or an optical circuit to achieve nonlinear mapping, wherein the optical circuit method includes a Mach-Zehnder modulator, that is, the input signal in optical form is nonlinearly mapped to an electrical signal through a Mach-Zehnder modulator through photoelectric conversion.
[0039] In the time-delay deep reserve pool, each delay loop is connected by circuit, optical path, or optical circuit to interact the input signals in time. Each delay loop reads the state after reaching the steady state through the internal cycle. The optical path includes the optical fiber delay line. Since the delay loop is very short, even if it passes through the optical fiber delay line, the distortion effect is negligible. Each virtual node on the delay loop corresponds to a section of the delay loop. Specifically, the delay loop is virtually divided into multiple sections according to the requirements of the channel equalization task. Each section corresponds to a virtual node. The encoded information on the virtual node is the state value of the virtual node, which corresponds to the amplitude or phase of the optical domain.
[0040] In the time-delay deep reserve pool, the connection between the nodes of two adjacent state calculation sublayers can be made by circuit, optical path, or optical circuit, wherein the optical path includes an optical fiber delay line. By connecting multiple layers of state calculation sublayers to one layer in this way, and treating each layer of state calculation sublayer as a depth, a multi-depth deep structure is formed. Such a deep structure can further improve the nonlinear ability of the system and enrich the dynamic characteristics of the system, which is convenient for the subsequent construction of the complementary matrix C.
[0041] When using the time-delay deep reserve pool constructed above to calculate the trust balance task, the initial input signal is expanded into a one-dimensional signal, encoded into the amplitude or phase of the photoelectric domain, and then serially input into the input layer of the time-delay deep reserve pool. Then, nonlinear mapping and interaction are performed through each state calculation sublayer, and the output layer reads the state of the virtual node through the readout matrix.
[0042] In the embodiment, the signal post-processing unit is used to construct a complementary matrix, and multiply the complementary matrix with the readout matrix to achieve channel equalization to obtain an equalized signal. Figure 4 As shown, the constructed complementary matrix C is an expanded state matrix with a dimension of N×T, where N is greater than M. The complementary matrix includes the state values of the virtual nodes in all state calculation sublayers, such as Figure 4 As shown, based on Figure 3 In the state layer structure shown, the complementary matrix includes the state value of the first state calculation sublayer of size M×T and the state value of the second state calculation sublayer, or on the basis of the virtual nodes in all state calculation sublayers, it can also include at least one of the state values of the virtual nodes of the additional state calculation sublayers at different time steps and the initial input signal. That is, the state value of the initial input signal of size M×T can be extracted from the input layer to construct the complementary matrix. The state values of the additional state calculation sublayers at different time steps are added to the complementary matrix to further improve the performance, which means that at the time point t, the distorted signal S t When performing equalization, if the distorted signal S before or after the i-th time step t±i If the balance effect is greater, S t±i The different state calculation sublayers corresponding to TDRC are added to the complementary matrix C as additional state matrices and used together for S t balance.
[0043] In constructing the complementary matrix C, all the above components are arranged as follows Figure 4The complementary matrix C of size N×T is obtained by splicing and combining the exponential dimensions shown. The constructed complementary matrix C is used to alleviate the performance loss caused by mask reduction. By combining the rich dynamic characteristics between different layers of the reserve pool and the additional time step state value, the performance is improved and the channel equalization speed of the reserve pool calculation is improved. The complementary matrix C represents the state inside the reserve pool and is used for the training and prediction process in the readout stage.
[0044] In the embodiment, the signal post-processing unit is further used to combine the complementary matrix C with the readout matrix W of size L×N out The multiplication realizes channel equalization and obtains the equalized signal Y. The readout matrix is parameter optimized before being applied. During the optimization, the parameters of the readout matrix are adjusted by comparing the difference between the equalized signal output by the signal post-processing unit and the original signal.
[0045] like Figure 5 As shown, the embodiment also provides a method for calculating a delay-type reserve pool for high-speed channel equalization, comprising the following steps:
[0046] S1, using a signal preprocessing unit to perform mask processing on the distorted signal obtained after the original signal is transmitted through the optical fiber channel to obtain an initial input signal of the time-delay type depth reserve pool;
[0047] S2. Calculate the state value of each position based on the initial input signal using a time-delayed deep reserve pool, the state value corresponds to the amplitude or phase of the photoelectric domain, and read the state of the virtual node through a readout matrix;
[0048] S3. Use a signal post-processing unit to multiply the constructed complementary matrix with the readout matrix to achieve channel equalization and obtain an equalized signal.
[0049] In the above method and system, the state layer of the constructed time-delay deep reserve pool is composed of multiple state calculation sublayers connected in sequence, has a certain depth structure, and combines the deep structure with the complementary matrix, so that the rich dynamic characteristics inside the reserve pool can be fully utilized in the complementary matrix, thereby effectively reducing the size of the mask, reducing the amount of data generated, and increasing the channel equalization speed while minimizing performance damage, ultimately achieving a balance between speed and accuracy.
[0050] The specific implementation methods described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the protection scope of the present invention.
Claims
1. A time-delay reserve pool computing system for high-speed channel equalization, It is characterized in that include: A signal preprocessing unit is used to perform mask processing on the distorted signal obtained after the original signal is transmitted through the optical fiber channel to obtain the initial input signal of the time-delayed depth reserve pool, specifically selecting a mask matrix of size M×K and randomly generated to multiply the distorted signal of size K×T to implement mask processing, and convert the distorted signal into an initial input signal of dimension M×T related to the mask matrix; A time-delay deep reserve pool, comprising an input layer, a state layer, and an output layer, wherein the state layer is composed of a plurality of state calculation sublayers connected in sequence to form a deep structure, each state calculation sublayer comprises a nonlinear node and is combined with a time-delay loop composed of a plurality of virtual nodes, and the nonlinear node of the latter state calculation sublayer of two adjacent state calculation sublayers is connected to the virtual node or nonlinear node of the previous state calculation sublayer, the initial input signal is input through the input layer, and nonlinear mapping and interaction are performed through each state calculation sublayer, and the state of the virtual node is read by the output layer through a readout matrix; A signal post-processing unit is used to construct a complementary matrix and multiply the complementary matrix with the readout matrix to achieve channel equalization to obtain an equalized signal, wherein the dimension of the complementary matrix is N×T, and N is greater than M, and the complementary matrix includes the state values of the virtual nodes in all state calculation sublayers, or also includes the state values of the virtual nodes of different state calculation sublayers in additional time steps and at least one of the initial input signals.
2. The time-delay reserve pool computing system for high-speed channel equalization according to claim 1, It is characterized in that After the initial input signal is expanded into a one-dimensional signal, it is encoded into the amplitude or phase of the optical domain and then serially input into the time-delay type deep reserve pool.
3. The time-delay reserve pool computing system for high-speed channel equalization according to claim 1, It is characterized in that Each nonlinear node is used to perform nonlinear mapping on the input signal and map it to a high-dimensional space. The nonlinear mapping is achieved by using a circuit, an optical path, or an optical circuit method, wherein the optical circuit method includes a Mach-Zehnder modulator.
4. The time-delay reserve pool computing system for high-speed channel equalization according to claim 1, It is characterized in that Each delay loop is connected by means of a circuit, an optical path, or an optical circuit, wherein the optical path includes an optical fiber delay line.
5. The time-delay reserve pool computing system for high-speed channel equalization according to claim 1, It is characterized in that Each virtual node corresponds to a section of the delay loop. Specifically, the delay loop is virtually divided into multiple sections according to the requirements of the channel equalization task. Each section corresponds to a virtual node. The encoded information on the virtual node is the state value of the virtual node, which corresponds to the amplitude or phase of the optical domain.
6. The time-delay reserve pool computing system for high-speed channel equalization according to claim 1, It is characterized in that The nodes of two adjacent state calculation sublayers are connected by means of circuits, optical paths, or optical circuits, wherein the optical paths include optical fiber delay lines.
7. The time-delay reserve pool computing system for high-speed channel equalization according to claim 1, It is characterized in that The readout matrix is parameter optimized before being used. During the optimization, the parameters of the readout matrix are adjusted by comparing the difference between the equalized signal output by the signal post-processing unit and the original signal.
8. A delay-type reserve pool calculation method for high-speed channel equalization, It is characterized in that The method adopts the time-delayed reserve pool computing system according to any one of claims 1 to 7, and comprises the following steps: The signal preprocessing unit is used to perform mask processing on the distorted signal obtained after the original signal is transmitted through the optical fiber channel to obtain the initial input signal of the time-delay type depth reserve pool; The time-delayed deep reserve pool is used to calculate the state value of each position based on the initial input signal. The state value corresponds to the amplitude or phase of the photoelectric domain, and the state of the virtual node is read through the readout matrix. The signal post-processing unit is used to multiply the constructed complementary matrix with the readout matrix to achieve channel equalization and obtain an equalized signal.
