A high-speed serial interface signal compensation method based on an adaptive equalization algorithm

Abstract

The application discloses a high-speed serial interface signal compensation method based on an adaptive equalization algorithm, and relates to the technical field of electronic device data transmission.The method comprises a signal preprocessing module, a deep separable convolution lightweight CNN feature extraction module, an adaptive threshold FFE equalization module and an error detection feedback module to form a closed-loop compensation system.The received signal is first denoised, and then the core features of the signal are extracted, and then the equalization weight and threshold are dynamically adjusted in combination with the feature parameters and the real-time signal-to-noise ratio to complete signal equalization, and finally the equalization error and the real-time signal-to-noise ratio are detected to feedback and calibrate the parameters of each module.The application is integrated by using a CMOS process, does not require special hardware, and the feature extraction module does not require training, can dynamically adapt to different signal-to-noise ratio scenes, effectively reduces the equalization bit error rate, improves the signal transmission stability, adapts to different rate high-speed serial interfaces, and is suitable for various electronic data transmission scenes with low signal-to-noise ratio and high-speed transmission.

Description

Technical Field

[0001] This invention relates to the field of electronic device data transmission technology, specifically to a high-speed serial interface signal compensation method based on an adaptive equalization algorithm. Background Technology

[0002] In the field of electronic device data transmission, high-speed serial interfaces, with their advantages of high transmission rate and low interface resource consumption, have become the core component for data interaction in various electronic systems, and are widely used in industrial control, automotive electronics, data centers, and other scenarios. During transmission, high-speed serial signals are affected by various factors such as channel loss, external electromagnetic interference, and changes in the characteristics of the transmission medium, leading to problems such as signal attenuation, distortion, and inter-symbol interference. In low signal-to-noise ratio scenarios, signal distortion and interference are more pronounced, directly affecting the accuracy and stability of data transmission, and thus restricting the performance improvement of the entire electronic system. Adaptive equalization technology, as a key technology for solving the above signal transmission problems, dynamically adjusts equalization parameters to offset the negative impacts of channel loss and interference, becoming the core means of signal compensation for high-speed serial interfaces.

[0003] Currently, adaptive equalization technology based on neural networks has been widely studied and applied in high-speed serial interface signal compensation due to its strong nonlinear impairment compensation capability. Existing equalization schemes based on neural networks typically use convolutional neural networks or hybrid neural network structures for signal feature extraction, and then combine them with feedforward equalizers to achieve signal equalization processing to improve signal transmission quality. However, the feature extraction modules of these existing equalization schemes mostly use traditional convolutional structures, lacking specific anti-interference design for low signal-to-noise ratio scenarios, and their linkage logic with the feedforward equalizer is relatively fixed, unable to dynamically adjust according to channel noise conditions.

[0004] Existing high-speed serial interface equalization technology based on neural networks has a core technical problem: in low signal-to-noise ratio (SNR) scenarios, its feature extraction module has weak anti-interference capability and is prone to misjudging noise signals as valid signal features, leading to deviations in equalization weight calibration. This, in turn, causes a significant increase in the equalization bit error rate, making it impossible to achieve stable and reliable signal compensation. This severely limits the widespread application of high-speed serial interfaces in low SNR, high-speed transmission scenarios. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a high-speed serial interface signal compensation method based on an adaptive equalization algorithm. This method addresses the core problems of existing neural network-based high-speed serial interface equalization technologies in low signal-to-noise ratio scenarios, such as weak anti-interference capability of the feature extraction module, easy misjudgment of noise signals as valid signal features, leading to equalization weight calibration deviation, significantly increased equalization bit error rate, and inability to achieve stable and reliable signal compensation.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a high-speed serial interface signal compensation method based on an adaptive equalization algorithm, comprising a signal preprocessing module, a depthwise separable convolutional lightweight CNN feature extraction module, an adaptive threshold FFE equalization module, and an error detection feedback module. The signal preprocessing module performs noise reduction processing on the input signal of the high-speed serial interface receiving end and then transmits it to the depthwise separable convolutional lightweight CNN feature extraction module. The depthwise separable convolutional lightweight CNN feature extraction module extracts features from the preprocessed signal and outputs feature parameters to the adaptive threshold FFE equalization module. The adaptive threshold FFE equalization module adjusts the equalization weight and equalization threshold according to the feature parameters and real-time signal-to-noise ratio to perform equalization processing on the signal. The error detection and feedback module detects the equalization error and collects the real-time signal-to-noise ratio. It feeds the relevant data back to the depthwise separable convolutional lightweight CNN feature extraction module and the adaptive threshold FFE equalization module to dynamically calibrate the relevant parameters and form a closed-loop compensation.

[0007] Furthermore, the signal preprocessing module employs an RC low-pass filter structure with dynamically adjustable cutoff frequency. The filter resistors are polysilicon resistors from CMOS technology, and the filter capacitors are MOS capacitors. This module filters out high-frequency noise and clutter from the input signal, retaining the core frequency components and transmitting them to the depthwise separable convolutional lightweight CNN feature extraction module. The depthwise separable convolutional lightweight CNN feature extraction module uses a three-layer depthwise separable convolutional structure without pooling or fully connected layers, including an input layer, an intermediate layer, and an output layer. The input layer receives the preprocessed signal, and after three layers of depthwise separable convolution operations, outputs feature parameters to the adaptive threshold FFE equalization module. The depthwise separable convolution operation satisfies the following formula: ,in The output feature parameters of the j-th convolutional layer are... For the j-th layer depth separable convolution kernel, The input signal for the (j-1)th convolutional layer is... Let be the bias parameters of the j-th convolutional layer. This is the depthwise separable convolution operator; The signal feature parameters used to characterize the output signal after the j-th layer convolution processing. Used for feature extraction from input signals. Used to provide the basis for the input signal of the j-th convolution. Used to calibrate the bias of convolution operations and ensure the accuracy of feature extraction.

[0008] Furthermore, the adaptive threshold FFE equalization module adopts a 12-order adaptive tap structure, with taps divided into feedforward taps and feedback taps. The tap weights are jointly determined by the feature parameters output by the depthwise separable convolutional lightweight CNN feature extraction module and the real-time signal-to-noise ratio. Digital logic units are used to store and adjust the weights, without setting analog multipliers. The adaptive threshold FFE equalization module sets three signal-to-noise ratio thresholds, corresponding to three different equalization thresholds and weight adjustment step sizes. Based on the real-time signal-to-noise ratio collected by the error detection feedback module, the corresponding equalization threshold and weight adjustment step size are matched to perform equalization processing on the input signal, and the equalized signal is output to the error detection feedback module.

[0009] Furthermore, the error detection feedback module is implemented using an error comparator and a signal-to-noise ratio (SNR) detection unit. The error comparator compares the equalized signal output from the adaptive threshold FFE equalization module with a standard reference signal to calculate the equalization error. The SNR detection unit calculates the real-time SNR based on the ratio of signal power to noise power, with the sampling frequency matched to the high-speed serial interface signal rate. The error detection feedback module feeds back the calculated equalization error and the acquired real-time SNR to the depthwise separable convolutional lightweight CNN feature extraction module and the adaptive threshold FFE equalization module, respectively, triggering dynamic calibration of relevant parameters. The calculation of the equalization error satisfies the following formula: ,in To balance the error, The equalized signal output by the adaptive threshold FFE equalization module. As a standard reference signal, This is the absolute value operator; Used to characterize the degree of deviation between the equalization signal and the standard reference signal. Used to provide the actual signal data after equalization Used as a reference benchmark for signal equalization effect to ensure the accuracy of equalization error calculation.

[0010] Furthermore, the input layer of the depthwise separable convolutional lightweight CNN feature extraction module uses a single-channel input, a three-dimensional convolutional kernel, a fixed convolutional stride, and a padding method that ensures consistent input and output sizes. The convolutional kernel is implemented using a computational unit in CMOS technology. The intermediate layer receives the output feature map from the input layer, uses a convolutional kernel of the same size as the input layer, has the same convolutional stride as the input layer, and uses the same padding method as the input layer. This is used to enhance the discriminative power of signal features and filter out residual noise interference. The output layer uses a one-dimensional convolutional kernel with no padding, a fixed convolutional stride, and twelve-dimensional output feature parameters. The feature parameter value range is fixed, and the accuracy meets the preset requirements. These parameters are directly used as the initial values ​​of the tap weights of the adaptive threshold FFE equalization module.

[0011] Furthermore, the adaptive threshold FFE equalization module has a fixed number of feedforward and feedback taps, a tap spacing of half the signal symbol period, a fixed tap weight adjustment range, and accuracy that meets preset requirements. Three signal-to-noise ratio (SNR) thresholds are clearly defined, corresponding to extremely low SNR, low SNR, and normal SNR scenarios, respectively. In the extremely low SNR scenario, the equalization threshold is lowered, and the tap weight adjustment step size is increased. In the low SNR scenario, the equalization threshold and the tap weight adjustment step size are set to the median value. In the normal SNR scenario, the equalization threshold is set to the maximum value, and the tap weight adjustment step size is set to the minimum value, ensuring equalization adaptation under different SNR scenarios.

[0012] Furthermore, the error accuracy of the error detection feedback module meets preset requirements. When the equalization error exceeds a preset threshold, it triggers an emergency adjustment of the weights in the adaptive threshold FFE equalization module. The detection accuracy of the signal-to-noise ratio detection unit meets preset requirements, and it outputs the signal-to-noise ratio data of the current channel in real time. The signal fed back to the depthwise separable convolutional lightweight CNN feature extraction module is used to adjust the filtering intensity of the convolution kernel, optimize the feature extraction effect, and reduce noise interference. The signal fed back to the adaptive threshold FFE equalization module is used to adjust the tap weights and equalization threshold to ensure that the equalization error is controlled within a reasonable range.

[0013] Furthermore, the overall workflow of the method includes the following steps: The first step is for the high-speed serial interface receiver to receive the input signal, which then enters the signal preprocessing module for noise reduction and outputs the preprocessed signal. The second step is to input the preprocessed signal into the depthwise separable convolution lightweight CNN feature extraction module. After three layers of depthwise separable convolution operations, the core features of the signal are extracted and twelve-dimensional feature parameters are output. The third step involves inputting the feature parameters into the adaptive threshold FFE equalization module, collecting the real-time signal-to-noise ratio from the error detection feedback module, determining the corresponding equalization threshold and weight adjustment step size, and then having the FFE module perform equalization processing on the signal and output the equalized signal. The fourth step is for the error detection and feedback module to calculate the equalization error, collect the real-time signal-to-noise ratio, and feed the relevant data back to the corresponding module. Fifth, the corresponding module adjusts the relevant parameters based on the feedback data, repeats the above steps, forms a closed-loop working mode, and continuously realizes signal compensation.

[0014] Furthermore, the method utilizes existing CMOS technology for integration, allowing it to be integrated within the SerDes chip without requiring additional special hardware components. The devices employ conventional components found in CMOS technology, including polysilicon resistors, MOS capacitors, digital logic units, error comparators, and signal-to-noise ratio detection units. The process is mature, cost-controllable, and mass production is feasible. The chip package size meets the application requirements of high-speed serial interfaces, adaptable to high-speed serial interfaces of different rates, covering mainstream protocols, and applicable to low signal-to-noise ratio, high-speed transmission scenarios.

[0015] Furthermore, the depthwise separable convolution lightweight CNN feature extraction module requires no training data or training process and can work normally upon startup. By splitting the traditional convolution into channel convolution and spatial convolution through a depthwise separable convolution structure, it reduces the number of parameters and computational load, and improves the anti-interference capability of feature extraction in low signal-to-noise ratio scenarios. The adaptive threshold FFE equalization module works in conjunction with the depthwise separable convolution lightweight CNN feature extraction module, and achieves dynamic calibration through an error detection feedback module to ensure stable equalization results in low signal-to-noise ratio scenarios.

[0016] Compared with existing technologies, this high-speed serial interface signal compensation method based on an adaptive equalization algorithm has the following advantages: I. This invention utilizes a depthwise separable convolution lightweight CNN feature extraction module to perform targeted feature extraction on preprocessed signals. It separates the channel and spatial convolutions of traditional convolutions to improve the anti-interference capability of feature extraction under low signal-to-noise ratio (SNR) conditions. Simultaneously, it integrates an error detection feedback module to achieve dynamic linkage calibration between this module and the adaptive threshold FFE equalization module. This effectively prevents noisy signals from being misjudged as valid signal features, solves the problem of equalization weight calibration deviation, and reduces the equalization bit error rate in low SNR scenarios. This achieves stable and reliable high-speed serial interface signal compensation, improving the accuracy and stability of signal transmission in low SNR, high-speed transmission scenarios.

[0017] Second, this invention integrates various functional modules using conventional CMOS processes and devices, eliminating the need for additional special hardware components. Furthermore, the depthwise separable convolutional lightweight CNN feature extraction module eliminates the need for training data and training processes, allowing it to operate immediately upon power-on. Meanwhile, the adaptive threshold FFE equalization module uses digital logic units to adjust weights, abandoning analog multipliers. This simplifies hardware implementation complexity, controls manufacturing costs, and allows for compatibility with high-speed serial interfaces of different rates and mainstream protocols, enabling mass production. This broadens the application scenarios of the method and improves its adaptability and practicality in actual engineering.

[0018] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0020] Figure 1 This is a system architecture diagram of the present invention; Figure 2 This diagram illustrates the structure and weight adjustment mechanism of the adaptive threshold FFE equalization module of this invention. Detailed Implementation

[0021] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0022] Example 1 The high-speed serial interface signal compensation method based on adaptive equalization algorithm in this embodiment consists of a signal preprocessing module, a depthwise separable convolutional lightweight CNN feature extraction module, an adaptive threshold FFE equalization module, and an error detection feedback module, which together form a closed-loop compensation system. Each module is integrated into the SerDes chip using CMOS technology, without adding any special hardware devices. The integrated devices are all conventional CMOS technology devices, including polysilicon resistors, MOS capacitors, digital logic units, error comparators, and signal-to-noise ratio detection units. It is compatible with high-speed serial interfaces of different rates and mainstream transmission protocols, and can be directly applied to scenarios such as industrial control, automotive electronics, and data centers with low signal-to-noise ratio and high-speed transmission.

[0023] The overall workflow of the method in this embodiment is as follows: signal preprocessing, feature extraction, adaptive equalization, error and signal-to-noise ratio detection, and dynamic parameter calibration, forming a continuous closed-loop signal compensation. The specific implementation steps are as follows: Step 1 Signal Preprocessing The high-speed serial interface receiver receives the externally input serial signal, which is directly transmitted to the signal preprocessing module for noise reduction. The signal preprocessing module employs an RC low-pass filter structure. The filter cutoff frequency can be dynamically adjusted according to the input signal's transmission rate and noise levels. Its filter resistors are polysilicon resistors from CMOS technology, and its filter capacitors are MOS capacitors from CMOS technology. This structure filters out high-frequency noise and clutter from the input signal, retaining the core frequency components of the signal and preventing high-frequency noise from interfering with subsequent feature extraction and equalization. The preprocessed signal is then converted from analog to digital and transmitted to the depthwise separable convolutional lightweight CNN feature extraction module.

[0024] Step 2: Depthwise Separable Convolution Lightweight Feature Extraction The depthwise separable convolution lightweight CNN feature extraction module receives the preprocessed signal and extracts its core features. This module adopts a three-layer depthwise separable convolution structure without pooling layers or fully connected layers. It requires no training data or training process and can work normally right out of the box. By splitting the traditional convolution into channel convolution and spatial convolution, it effectively reduces the number of parameters and computation, while improving the anti-interference ability of feature extraction in low signal-to-noise ratio scenarios.

[0025] This module consists of an input layer, an intermediate layer, and an output layer. The input layer uses a single-channel input, a three-dimensional convolutional kernel, a fixed stride, and padding to ensure consistent input and output sizes. The convolutional kernel is implemented using computational units in CMOS technology. The intermediate layer receives the output feature map from the input layer, uses a convolutional kernel of the same size as the input layer, and has the same stride and padding to enhance the discriminative power of signal features and filter out residual noise interference. The output layer uses a one-dimensional convolutional kernel with no padding and a fixed stride, ultimately outputting twelve-dimensional feature parameters. The feature parameters have a fixed value range, and their accuracy meets the preset requirements for high-speed serial interface signal compensation. These twelve-dimensional feature parameters are directly used as the initial values ​​for the tap weights of the subsequent adaptive threshold FFE equalization module.

[0026] The depthwise separable convolution operation in this module follows the formula:

[0027] In the formula, These are the output feature parameters of the j-th convolution layer, used to characterize the signal feature parameters output after the j-th convolution layer processing; The j-th depth-separable convolution kernel is used for feature extraction from the input signal, and its kernel parameters are implemented by the CMOS process computing unit. For the j-th The input signal of the first convolution layer is used to provide the basis for the input signal of the j-th convolution layer; The bias parameter for the j-th layer convolution is used to calibrate the deviation of the convolution operation and ensure the accuracy of feature extraction. This is a depthwise separable convolution operator that performs the step-by-step operations of channel convolution and spatial convolution in sequence.

[0028] In this embodiment, j takes the values ​​1, 2, and 3, which correspond to the convolution operations of the input layer, intermediate layer, and output layer, respectively. After the three convolution operations are performed sequentially, the output layer obtains a twelve-dimensional J3, which is the final feature parameter.

[0029] Step 3 Adaptive threshold FFE equalization processing The adaptive threshold FFE equalization module receives the twelve-dimensional feature parameters output by the depthwise separable convolutional lightweight CNN feature extraction module, and simultaneously receives the real-time signal-to-noise ratio collected by the error detection feedback module. Based on the feature parameters and the real-time signal-to-noise ratio, it adjusts the equalization weights and equalization thresholds to perform equalization processing on the input signal.

[0030] This module adopts a 12-order adaptive tap structure, with taps divided into feedforward taps and feedback taps. The number of taps is fixed, the tap spacing is half the signal symbol period, the tap weight adjustment range is fixed, and the accuracy meets the preset requirements of high-speed serial interface signal compensation. The tap weights are determined by the above-mentioned 12-dimensional characteristic parameters and the real-time signal-to-noise ratio. Digital logic units are used to realize the storage and dynamic adjustment of weights, without setting up analog multipliers, effectively reducing the complexity of hardware implementation.

[0031] This module sets three signal-to-noise ratio (SNR) thresholds, corresponding to three scenarios: extremely low SNR, low SNR, and normal SNR. These three thresholds are clearly defined. Based on the real-time SNR collected by the error detection feedback module, the module matches the corresponding equalization threshold and weight adjustment step size: In the extremely low SNR scenario, the equalization threshold is lowered, and the tap weight adjustment step size is increased to quickly adapt to the signal characteristics under strong noise; in the low SNR scenario, the equalization threshold and tap weight adjustment step size are set to an intermediate value to achieve a balance between noise suppression and signal fidelity; in the normal SNR scenario, the equalization threshold is set to the maximum value, and the tap weight adjustment step size is set to the minimum value to ensure the accuracy of signal equalization. The module performs equalization processing on the signal according to the matched equalization threshold and weight adjustment step size, and then transmits the equalized signal to the error detection feedback module.

[0032] Step 4: Equalization Error and Real-time Signal-to-Noise Ratio Detection The error detection feedback module consists of an error comparator and a signal-to-noise ratio detection unit. After receiving the equalization signal output by the adaptive threshold FFE equalization module, it synchronously completes the equalization error calculation and real-time signal-to-noise ratio acquisition.

[0033] The error comparator compares the equalized signal with a standard reference signal and calculates the equalization error, which follows the formula:

[0034] In the formula, Equalization error is used to characterize the degree of deviation between the equalization signal and the standard reference signal; The equalization signal output by the adaptive threshold FFE equalization module is used to provide the actual signal data after equalization. It serves as a standard reference signal, a standard signal template for high-speed serial interfaces, used as a reference benchmark for signal equalization effects to ensure the accuracy of equalization error calculations. This is the absolute value operator, which eliminates calculation errors caused by signal phase deviation.

[0035] The signal-to-noise ratio (SNR) detection unit calculates the real-time SNR based on the ratio of signal power to noise power. Its sampling frequency matches the signal rate of the high-speed serial interface, ensuring the real-time performance and accuracy of SNR detection. The detection accuracy meets the preset requirements of high-speed serial interface signal compensation and can output the SNR data of the current channel in real time.

[0036] Step 5: Parameter Dynamic Calibration and Closed-Loop Compensation The error detection feedback module feeds back the calculated equalization error and the acquired real-time signal-to-noise ratio to the depthwise separable convolutional lightweight CNN feature extraction module and the adaptive threshold FFE equalization module, respectively, triggering dynamic calibration of the relevant parameters of the two modules to form a closed-loop compensation mode.

[0037] The equalization error and real-time signal-to-noise ratio are fed back to the feature extraction module of the depthwise separable convolutional lightweight CNN. This is used to adjust the filtering intensity of each depthwise separable convolutional kernel, optimize the feature extraction effect, reduce noise interference, and make the feature parameters output by the module more consistent with the actual signal characteristics of the current channel. The equalization error and real-time signal-to-noise ratio are also fed back to the adaptive threshold FFE equalization module. This is used to adjust the tap weights and equalization threshold in real time. If the equalization error exceeds the preset threshold, it will trigger an emergency adjustment of the module's weights to ensure that the equalization error is always controlled within a reasonable range.

[0038] After parameter calibration is completed, the system repeats steps 1 to 5 above to continuously preprocess, extract features, equalize, and calibrate the input signal of the high-speed serial interface receiver, thereby achieving continuous closed-loop compensation for the high-speed serial interface signal and effectively offsetting the signal attenuation, distortion, and inter-symbol interference caused by channel loss and electromagnetic interference.

[0039] Example 2 Based on Example 1, this embodiment optimizes the dynamic adjustment method of the filter cutoff frequency of the signal preprocessing module and the convolution kernel parameters of the depthwise separable convolution lightweight CNN feature extraction module, further improving the noise reduction effect and feature extraction accuracy in low signal-to-noise ratio scenarios.

[0040] The filter cutoff frequency of the signal preprocessing module is dynamically adjusted in stages based on the real-time signal-to-noise ratio (SNR) fed back by the SNR detection unit: when the real-time SNR is low, the filter cutoff frequency is appropriately reduced to enhance the filtering capability of high-frequency noise; when the real-time SNR is high, the filter cutoff frequency is appropriately increased to avoid over-filtering that could lead to the loss of the core frequency components of the signal. The adjustment of the filter cutoff frequency is automatically completed by the digital logic unit according to the preset SNR-cutoff frequency correspondence, with a fixed adjustment step size to ensure the stability of the filter.

[0041] The depthwise separable convolutional lightweight CNN feature extraction module uses 3D convolutional kernels in its input and intermediate layers. The spatial dimensions of these kernels are adapted to the signal symbol length of the high-speed serial interface, and the channel dimension is single-channel, matching the single-channel input of the input layer. The length of the 1D convolutional kernel in the output layer is adapted to a 12-order adaptive tap structure, ensuring that the 12-dimensional feature parameters of the output can be directly mapped to the initial values ​​of the tap weights. The convolutional kernels of each layer... The bias parameters are fixed calibration values, which are solidified by the CMOS process computing unit during the chip manufacturing stage, eliminating the need for subsequent manual adjustments and further improving the module's ease of use.

[0042] This embodiment improves noise filtering efficiency in low signal-to-noise ratio scenarios, reduces the false positive rate of feature extraction, thereby improving the weight calibration accuracy of the adaptive threshold FFE equalization module and further reducing the bit error rate of the equalized signal.

[0043] Example 3 Based on Embodiment 1, this embodiment replaces the weight adjustment logic of the adaptive threshold FFE equalization module and the feedback frequency of the error detection feedback module, adapting to high-speed serial interfaces of different rates and improving system compatibility. Furthermore, the replacement content of this embodiment does not exceed the scope of the claims and can be flexibly selected according to the actual application scenario.

[0044] The tap weights of the adaptive threshold FFE equalization module are linearly adjusted based on the real-time signal-to-noise ratio (SNR) changes, starting from the initial values ​​(twelve-dimensional feature parameters output by depthwise separable convolution). The adjustment coefficient is positively correlated with the SNR; that is, the higher the real-time SNR, the smaller the weight adjustment coefficient and the smoother the weight change; the lower the real-time SNR, the larger the weight adjustment coefficient and the faster the weight adaptation. The linear adjustment coefficient is determined by the digital logic unit according to the preset SNR-adjustment coefficient correspondence to ensure the rationality of the weight adjustment.

[0045] The feedback frequency of the error detection feedback module is adaptively adjusted according to the signal rate of the high-speed serial interface. The higher the signal rate, the higher the feedback frequency, ensuring the real-time performance of parameter calibration; the lower the signal rate, the lower the feedback frequency, reducing the chip's power consumption. The ratio of feedback frequency to signal rate is a fixed value, which is implemented by the chip and requires no subsequent adjustment.

[0046] This embodiment can be adapted to high-speed serial interfaces with different speeds from gigabit to 10 gigabit. While ensuring signal compensation effect, it effectively reduces the power consumption of the chip and is more suitable for low signal-to-noise ratio scenarios with power consumption requirements, such as automotive electronics and portable industrial control equipment.

[0047] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A high-speed serial interface signal compensation method based on an adaptive equalization algorithm, characterized in that, It includes a signal preprocessing module, a depthwise separable convolutional lightweight CNN feature extraction module, an adaptive threshold FFE equalization module, and an error detection feedback module. The signal preprocessing module performs noise reduction processing on the input signal from the high-speed serial interface receiver and then transmits it to the depthwise separable convolutional lightweight CNN feature extraction module. The depthwise separable convolutional lightweight CNN feature extraction module extracts features from the preprocessed signal and outputs feature parameters to the adaptive threshold FFE equalization module. The adaptive threshold FFE equalization module adjusts the equalization weight and equalization threshold according to the feature parameters and real-time signal-to-noise ratio to perform equalization processing on the signal. The error detection and feedback module detects the equalization error and collects the real-time signal-to-noise ratio. It feeds the relevant data back to the depthwise separable convolutional lightweight CNN feature extraction module and the adaptive threshold FFE equalization module to dynamically calibrate the relevant parameters and form a closed-loop compensation.

2. The high-speed serial interface signal compensation method based on an adaptive equalization algorithm according to claim 1, characterized in that, The signal preprocessing module adopts an RC low-pass filter structure, the filter cutoff frequency can be dynamically adjusted, the filter resistor adopts a polysilicon resistor in CMOS process, and the filter capacitor adopts a MOS capacitor. The signal preprocessing module filters out high-frequency noise and clutter in the input signal, retains the core frequency components of the signal and transmits them to the depth-separable convolutional lightweight CNN feature extraction module. The depthwise separable convolution lightweight CNN feature extraction module adopts a three-layer depthwise separable convolution structure without pooling layers and fully connected layers. It includes an input layer, intermediate layers, and an output layer. The input layer receives the preprocessed signal, and after three layers of depthwise separable convolution operations, outputs feature parameters to the adaptive threshold FFE equalization module. The depthwise separable convolution operation satisfies the following formula: ,in The output feature parameters of the j-th convolutional layer are... For the j-th layer depth separable convolution kernel, The input signal for the (j-1)th convolutional layer is... Let j be the bias parameters of the j-th convolutional layer. This is the depthwise separable convolution operator; The signal feature parameters used to characterize the output signal after the j-th layer convolution processing. Used for feature extraction from input signals. Used to provide the basis for the input signal of the j-th convolution. Used to calibrate the bias of convolution operations and ensure the accuracy of feature extraction.

3. The high-speed serial interface signal compensation method based on an adaptive equalization algorithm according to claim 1, characterized in that, The adaptive threshold FFE equalization module adopts a 12-order adaptive tap structure, with taps divided into feedforward taps and feedback taps. The tap weights are jointly determined by the feature parameters output by the depthwise separable convolutional lightweight CNN feature extraction module and the real-time signal-to-noise ratio. Digital logic units are used to store and adjust the weights, without setting analog multipliers. The adaptive threshold FFE equalization module sets three signal-to-noise ratio thresholds, corresponding to three different equalization thresholds and weight adjustment step sizes. Based on the real-time signal-to-noise ratio collected by the error detection feedback module, the corresponding equalization threshold and weight adjustment step size are matched to equalize the input signal, and the equalized signal is output to the error detection feedback module.

4. The high-speed serial interface signal compensation method based on an adaptive equalization algorithm according to claim 1, characterized in that, The error detection feedback module is implemented using an error comparator and a signal-to-noise ratio (SNR) detection unit. The error comparator compares the equalized signal output from the adaptive threshold FFE equalization module with a standard reference signal to calculate the equalization error. The SNR detection unit calculates the real-time SNR based on the ratio of signal power to noise power, with the sampling frequency matched to the high-speed serial interface signal rate. The error detection feedback module feeds back the calculated equalization error and the acquired real-time SNR to the depthwise separable convolutional lightweight CNN feature extraction module and the adaptive threshold FFE equalization module, respectively, triggering dynamic calibration of relevant parameters. The equalization error is calculated according to the following formula: ,in To balance the error, The equalized signal output by the adaptive threshold FFE equalization module. As a standard reference signal, This is the absolute value operator; Used to characterize the degree of deviation between the equalization signal and the standard reference signal. Used to provide the actual signal data after equalization Used as a reference benchmark for signal equalization effect to ensure the accuracy of equalization error calculation.

5. A high-speed serial interface signal compensation method based on an adaptive equalization algorithm according to claim 2, characterized in that, The depthwise separable convolutional lightweight CNN feature extraction module uses a single-channel input layer with a three-dimensional convolutional kernel, a fixed stride, and padding to ensure consistent input and output sizes. The kernel is implemented using computational units in CMOS technology. The intermediate layer receives the output feature map from the input layer, uses a convolutional kernel of the same size as the input layer, has the same stride, and uses the same padding to enhance the discriminative power of signal features and filter out residual noise interference. The output layer uses a one-dimensional convolutional kernel with no padding, a fixed stride, and twelve-dimensional output feature parameters. The feature parameter values ​​have a fixed range and meet preset accuracy requirements, and are directly used as the initial values ​​for the tap weights of the adaptive threshold FFE equalization module.

6. The high-speed serial interface signal compensation method based on an adaptive equalization algorithm according to claim 3, characterized in that, The adaptive threshold FFE equalization module has a fixed number of feedforward and feedback taps, the tap spacing is half of the signal symbol period, the tap weight adjustment range is fixed, and the accuracy meets the preset requirements. The three signal-to-noise ratio (SNR) thresholds are clearly defined, corresponding to extremely low SNR, low SNR, and normal SNR scenarios, respectively. In the extremely low SNR scenario, the equalization threshold is reduced and the tap weight adjustment step size is increased. In low signal-to-noise ratio scenarios, the equalization threshold is set to the median value, and the tap weight adjustment step size is also set to the median value. In normal signal-to-noise ratio scenarios, the equalization threshold is set to the maximum value, and the tap weight adjustment step size is set to the minimum value to ensure balanced adaptation under different signal-to-noise ratio scenarios.

7. A high-speed serial interface signal compensation method based on an adaptive equalization algorithm according to claim 4, characterized in that, The error accuracy of the error detection feedback module meets the preset requirements. When the equalization error exceeds the preset threshold, the weight of the adaptive threshold FFE equalization module is urgently adjusted. The signal-to-noise ratio (SNR) detection unit meets the preset requirements for detection accuracy and outputs the SNR data of the current channel in real time. The signal fed back to the depthwise separable convolutional lightweight CNN feature extraction module is used to adjust the filtering intensity of the convolution kernel, optimize the feature extraction effect, and reduce noise interference. The signal fed back to the adaptive threshold FFE equalization module is used to adjust the tap weights and equalization threshold to ensure that the equalization error is controlled within a reasonable range.

8. A high-speed serial interface signal compensation method based on an adaptive equalization algorithm according to claim 1, characterized in that, The overall workflow of the method includes the following steps: The first step is for the high-speed serial interface receiver to receive the input signal, which then enters the signal preprocessing module for noise reduction and outputs the preprocessed signal. The second step is to input the preprocessed signal into the depthwise separable convolution lightweight CNN feature extraction module. After three layers of depthwise separable convolution operations, the core features of the signal are extracted and twelve-dimensional feature parameters are output. The third step involves inputting the feature parameters into the adaptive threshold FFE equalization module, collecting the real-time signal-to-noise ratio from the error detection feedback module, determining the corresponding equalization threshold and weight adjustment step size, and then having the FFE module perform equalization processing on the signal and output the equalized signal. The fourth step is for the error detection and feedback module to calculate the equalization error, collect the real-time signal-to-noise ratio, and feed the relevant data back to the corresponding module. Fifth, the corresponding module adjusts the relevant parameters based on the feedback data, repeats the above steps, forms a closed-loop working mode, and continuously realizes signal compensation.

9. A high-speed serial interface signal compensation method based on an adaptive equalization algorithm according to claim 1, characterized in that, The method employs existing CMOS technology for integration, which can be integrated into the SerDes chip without requiring additional special hardware components. The devices used are conventional components in the CMOS process, including polysilicon resistors, MOS capacitors, digital logic units, error comparators, and signal-to-noise ratio detection units. The process is mature, cost-controllable, and can be mass-produced. The chip package size meets the application requirements of high-speed serial interfaces, can be adapted to high-speed serial interfaces of different rates, covers mainstream protocols, and can be applied to low signal-to-noise ratio, high-speed transmission scenarios.

10. A high-speed serial interface signal compensation method based on an adaptive equalization algorithm according to claim 1, characterized in that, The depthwise separable convolutional lightweight CNN feature extraction module requires no training data or training process and can work normally upon startup. By splitting the traditional convolutional channel convolution and spatial convolution through a depthwise separable convolutional structure, it reduces the number of parameters and computational load, and improves the anti-interference capability of feature extraction in low signal-to-noise ratio scenarios. The adaptive threshold FFE equalization module works in conjunction with the depthwise separable convolutional lightweight CNN feature extraction module, and achieves dynamic calibration through an error detection feedback module to ensure stable equalization results in low signal-to-noise ratio scenarios.

Images