CN-121984812-A - Non-uniform quantization method, device, equipment and storage medium based on density perception clustering

CN121984812ACN 121984812 ACN121984812 ACN 121984812ACN-121984812-A

Abstract

The application discloses a non-uniform quantization method, a device, equipment and a storage medium based on density perception clustering, which relate to the technical field of network communication; the method comprises the steps of carrying out clustering division on tap coefficients according to preset quantization bit width to obtain a plurality of target clusters, determining quantization parameters corresponding to the target clusters according to the dynamic range of coefficients in the target clusters, wherein the quantization parameters comprise scaling factors and zero point offset parameters, and carrying out independent uniform quantization on the tap coefficients in the target clusters according to the quantization parameters to obtain recovery coefficients. According to the application, the coefficients are adaptively divided into a plurality of clusters based on the density perception clusters, and then each cluster is independently and uniformly quantized, so that the high-precision quantization of equalizer coefficients under extremely low bit is completed, and the storage efficiency of coefficient quantization and equalizer performance are improved.

Inventors

CHEN CANCAN
XU ZHAOPENG
LIU YUE
WU QI
WEI JINLONG
HE ZHIXUE
JIANG YUAN

Assignees

鹏城实验室

Dates

Publication Date: 20260505
Application Date: 20260126

Claims (10)

1. A non-uniform quantization method based on density-aware clustering, the method comprising: acquiring tap coefficients of an equalizer; clustering and dividing the tap coefficients according to a preset quantization bit width to obtain a plurality of target clusters; determining quantization parameters corresponding to the target cluster according to the dynamic range of coefficients in the target cluster, wherein the quantization parameters comprise a scaling factor and a zero offset parameter; and independently and uniformly quantizing the tap coefficients in the target cluster according to the quantization parameters to obtain recovery coefficients.
2. The method of claim 1, wherein the step of clustering the tap coefficients according to a preset quantization bit width to obtain a plurality of target clusters comprises: determining clustering parameters according to a preset quantization bit width and a nearest neighbor strategy, wherein the clustering parameters comprise a target density threshold value and a density radius; performing density core identification and initial clustering on the tap coefficients according to the clustering parameters to obtain core clusters and outlier noise points; and carrying out cluster distribution on the outlier noise points according to the core cluster to obtain a plurality of target clusters.
3. The method of claim 2, wherein the step of performing density core identification and initial clustering on the tap coefficients according to the clustering parameters to obtain core clusters and outlier noise points comprises: Determining a local neighborhood of the tap coefficient according to the density radius; When the number of tap coefficients contained in the local neighborhood is greater than or equal to the target density threshold, taking the tap coefficients as density cores; classifying tap coefficients which can reach the density core density into the same set to obtain a core cluster; tap coefficients not included in the core cluster are taken as outlier noise points.
4. The method of claim 2, wherein the step of clustering the outlier noise points according to the core cluster to obtain a plurality of target clusters comprises: respectively calculating the average value of all tap coefficients in each core cluster to obtain the mass center of the core cluster; Calculating Euclidean distance from the outlier noise point to the centroid; And distributing the outlier noise points to a core cluster corresponding to the centroid with the minimum Euclidean distance, and determining each core cluster after the outlier noise points are distributed as a target cluster.
5. The method of claim 2, wherein the step of determining the cluster parameters according to a preset quantization bit width and nearest neighbor strategy comprises: determining a target density threshold according to a preset quantization bit width and a preset relation function; Calculating the nearest distance from each tap coefficient to the target density threshold according to a nearest neighbor analysis strategy to obtain a distance set; And determining the density radius according to the distance set.
6. The method of claim 1, wherein the step of determining the quantization parameter corresponding to the target cluster according to the dynamic range of the intra-cluster coefficients in the target cluster comprises: obtaining the maximum value and the minimum value in all tap coefficients in the target cluster; Calculating a scaling factor corresponding to the target cluster according to the maximum value, the minimum value and a preset quantization bit width; and calculating a zero point offset parameter corresponding to the target cluster according to the scaling factor and the maximum value.
7. The method of claim 1, wherein the step of independently uniformly quantizing the tap coefficients within the target cluster according to the quantization parameter to obtain recovery coefficients comprises: mapping tap coefficients in the target cluster into an integer representation range corresponding to a preset quantization bit width according to the scaling factor and the zero offset parameter to obtain a quantized integer value; and performing inverse quantization calculation according to the quantized integer value, the scaling factor and the zero offset parameter to obtain a recovery coefficient.
8. A non-uniform quantization device based on density-aware clustering, the device comprising: the data acquisition module is used for acquiring tap coefficients of the equalizer; The clustering division module is used for carrying out clustering division on the tap coefficients according to a preset quantization bit width to obtain a plurality of target clusters; the quantization processing module is used for determining quantization parameters corresponding to the target cluster according to the dynamic range of the coefficients in the target cluster, wherein the quantization parameters comprise a scaling factor and a zero offset parameter; and the uniform quantization module is used for independently and uniformly quantizing the tap coefficients in the target cluster according to the quantization parameters to obtain recovery coefficients.
9. A density-aware clustering-based non-uniform quantization device, characterized in that the device comprises a memory, a processor and a computer program stored on the memory and executable on the processor, the computer program being configured to implement the steps of the density-aware clustering-based non-uniform quantization method according to any of claims 1 to 7.
10. A storage medium, characterized in that the storage medium is a computer readable storage medium, on which a computer program is stored, which computer program, when being executed by a processor, implements the steps of the non-uniform quantization method based on density-aware clustering according to any one of claims 1 to 7.

Description

Non-uniform quantization method, device, equipment and storage medium based on density perception clustering Technical Field The present application relates to the field of network communications technologies, and in particular, to a method, an apparatus, a device, and a storage medium for non-uniform quantization based on density sensing clustering. Background With the rapid development of applications such as cloud computing, artificial intelligence, and big data, the demand for transmission bandwidth by data center interconnects continues to rise. In short-range optical communications, intensity modulated direct detection systems have become a mainstream solution because of their low cost, low power consumption, and ease of integration. To meet the increasing transmission capacity, the system needs to transmit with higher baud rate signals. However, under high speed conditions, the signal will suffer severe impairments, causing complex channel distortions, limited by the bandwidth bottlenecks of the optoelectronic device, fiber dispersion, and inherent nonlinear effects of the device. To compensate for such impairments, nonlinear equalization techniques based on digital signal processing are widely used. The Volterra nonlinear equalizer can effectively simulate and compensate nonlinear effects in an optical fiber communication system, and is one of key technologies for realizing high-precision signal recovery. In order to ensure compensation accuracy, the tap coefficients of the conventional Volterra equalizer are usually represented and calculated by high-precision floating point numbers (such as 32-bit floating points). However, floating point operations involve complex hardware logic and require a large amount of memory space to hold coefficients. For a high-order equalizer with hundreds of taps, this high-precision implementation may cause huge computational complexity and storage overhead, and is difficult to be applied in a data center optical module with limited cost and power consumption. To reduce hardware overhead, low bit quantization is proposed as an effective strategy to compress memory requirements and simplify circuitry by reducing the number of bits represented by coefficients. Current mainstream quantization methods include uniform quantization and non-uniform quantization. The method is difficult to combine the dynamic range and the quantization precision under the low bit condition because the tap coefficients of the Volterra equalizer are irregularly distributed, the quantization error of a dense area is increased when the step size is too large, and the dynamic range is easily truncated when the step size is too small. To overcome this problem, prior studies have proposed methods such as automatic grouping non-uniform quantization, by ordering and uniformly grouping coefficients, configuring different quantization strategies for different groups. However, the method adopts the grouping with fixed size, does not carry out self-adaptive division according to the probability density distribution of the coefficient, has weak sensing capability on complex distribution with obvious peak value and long tail, and is easy to cause overlarge or undersize data span in the group, thereby reducing quantization efficiency, still having larger quantization error under extremely low bit and limiting system performance. Therefore, how to improve the storage efficiency and equalizer performance of coefficient quantization under low bit conditions is a problem to be solved. Disclosure of Invention The application mainly aims to provide a non-uniform quantization method, device, equipment and storage medium based on density perception clustering, which aim to solve the technical problems of how to improve the storage efficiency of coefficient quantization and equalizer performance under the condition of low bit. In order to achieve the above object, the present application provides a non-uniform quantization method based on density perception clustering, the method comprising: acquiring tap coefficients of an equalizer; clustering and dividing the tap coefficients according to a preset quantization bit width to obtain a plurality of target clusters; determining quantization parameters corresponding to the target cluster according to the dynamic range of coefficients in the target cluster, wherein the quantization parameters comprise a scaling factor and a zero offset parameter; and independently and uniformly quantizing the tap coefficients in the target cluster according to the quantization parameters to obtain recovery coefficients. In an embodiment, the step of clustering the tap coefficients according to a preset quantization bit width to obtain a plurality of target clusters includes: determining clustering parameters according to a preset quantization bit width and a nearest neighbor strategy, wherein the clustering parameters comprise a target density threshold value and a density radius; perfor