Intelligent management method and system for signal control process of printed circuit board

Abstract

The invention relates to the technical field of printed circuit board signal management, and discloses an intelligent management method and system for a printed circuit board signal control process, and the method comprises the steps: obtaining the noise data of a target channel of a target management object and all adjacent channels in real time, analyzing the noise data of all channels, generating a frequency domain collaborative management domain, and carrying out the frequency domain collaborative management domain; the target management object is a server printed circuit board; the method comprises the following steps: S1, analyzing noise data to generate a frequency domain collaborative management domain, S2, mapping transient current into an N * N matrix and establishing a two-layer graph convolutional network, and S4, acquiring load data in a 100ms window, processing by using a two-way long-short-term memory network of two hidden layers, and adapting to signal transmission of a server printed circuit board, so as to avoid misjudgment caused by crosstalk. Coupling of a strong electromagnetic field is suppressed, and the signal management effect and the transmission stability are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of printed circuit board signal management, and particularly relates to a printed circuit board signal control process intelligent management method and system. BACKGROUND

[0002] At present, in the actual management process of printed circuit board signal control, it is often relied on traditional neural networks, such as 3-4 layers of full connection neural networks or single signal dedicated convolutional neural networks, to realize control management: a training set is constructed by collecting historical amplitude and phase data of a single signal and corresponding control response data, and the neural network is trained to make it learn the mapping relationship between single signal features and control parameters; in actual application, the amplitude and phase data of the current single signal are collected in real time and input into the trained neural network, and the control instruction is output by the network, and then the management execution instruction is realized to achieve dynamic management of the single signal transmission process, so as to maintain the transmission stability of the single signal.

[0003] However, the above management method still has the following defects when it is used for high-density server printed circuit boards: in order to meet the requirements of multiple interfaces and high bandwidth, the DDR5 signal (transmission rate ≥ 4800MT / s) and PCIe5.0 high-speed interface signal (channel rate ≥ 32GB / s) between CPU and memory need to be transmitted synchronously, and the frequency of such signals is all ≥ 10GHz; and due to the limitation of the space layout of the mainboard, the spacing between adjacent signal traces is often compressed to less than 0.2mm, which easily causes strong electromagnetic field coupling effect, such as the transient current change of a signal inducing interference current in adjacent traces. The traditional neural network relied on by the existing system only supports time domain or frequency domain feature extraction of a single signal, such as the single signal dedicated convolutional neural network which can only perform convolution operation on the phase fluctuation of a single signal, and cannot construct multi-dimensional coupling feature mapping of adjacent signal current change, interference current and target signal distortion, and thus cannot capture the correlation among the three, and finally the signal distortion caused by crosstalk is misjudged by the neural network as the regular fluctuation of a single signal, and the control instruction output cannot specifically suppress crosstalk, resulting in poor control management effect of signal crosstalk. SUMMARY

[0004] In view of the deficiencies of the prior art, the present application provides a printed circuit board signal control process intelligent management method and system, which solves the above problems.

[0005] The above technical purpose of the present application is realized by the following technical scheme: A printed circuit board signal control process intelligent management method, comprising: Step S1, real-time acquisition of target channel and all adjacent channels of target management object noise data, analysis of all channel noise data, generation of frequency domain collaborative management domain, target management object is server printed circuit board; Step S2, real-time acquisition of transient current of all channels in the frequency domain collaborative management domain, mapping of the transient current of each channel to an N×N inter-channel energy correlation matrix, construction of a 2-layer graph convolution network based on a graph neural network, taking channels as nodes and element weights of the inter-channel energy correlation matrix as edge weights, and generation of a target channel energy influence coefficient; Step S3, real-time acquisition of impedance data of all channels in the frequency domain collaborative management domain, fusion of the target channel energy influence coefficient and the impedance data through a neural network, and generation of a domain energy management factor representing the degree of energy influence on the target channel; Step S4, real-time acquisition of load data of all channels in the frequency domain collaborative management domain, continuous collection of T windows with 100ms as a collection window, and arrangement of the load data as time series data, input of the time series data and the energy management factor into a 2-layer hidden layer bidirectional long short-term memory network for analysis, and output of a pre-signal management vector containing amplitude and phase through a fully connected layer; Step S5, waveform remodeling of the original signal of the target channel according to the pre-signal management vector, and generation of an anti-interference reconstructed signal; Step S6, eye diagram analysis of the anti-interference reconstructed signal through a convolutional neural network, extraction of key feature indicators through a 5-layer CNN with 2-layer max pooling layers, comparison of the key feature indicators with an eye diagram quality threshold, and generation of an eye diagram quality management coefficient; Step S7, continuous tracking of the time-varying trajectory of the eye diagram quality management coefficient, calculation of its stability, comparison of the stability with a stability management boundary, and derivation of a management instruction.

[0006] Further, analysis of the noise data of all channels to generate a frequency domain collaborative management domain, including: Analysis of the noise data, calculation of the amplitude mutation synchronization rate and phase drift time difference of the noise of each adjacent channel and the noise of the target channel, and generation of a noise dynamic coupling degree; Real-time acquisition of the working environment temperature of the target management object and the signal reflection times of the wiring of each adjacent channel, calculation of the medium loss increment and reflection loss superposition value of the noise of each adjacent channel during the transmission to the target channel based on the working environment temperature and the signal reflection times, and generation of a noise time-varying transmission loss rate; Analysis of the amplitude fluctuation period of each adjacent channel noise, correlation analysis of the noise amplitude fluctuation period corresponding to each frequency point and the current signal transmission period of the target channel, and derivation of the period fit degree of each frequency point; The resonant interference effective frequency point threshold is obtained by calculating the periodic coincidence degree and the noise time-varying transmission loss rate; Based on the resonant interference effective frequency point threshold, the noise dynamic coupling degree and the noise time-varying transmission loss rate are analyzed to obtain the frequency domain collaborative management domain.

[0007] Further, the transient current of each channel is mapped to an N*N inter-channel energy correlation matrix, a two-layer graph convolution network is constructed based on a graph neural network, the channels are taken as nodes, and the element of the inter-channel energy correlation matrix is taken as the edge weight, to generate the target channel energy influence coefficient, including: The current amplitude of all channels in the frequency domain collaborative management domain is obtained in real time; The occurrence frequency of the transient current peak value per unit time and the real-time change amplitude of the current amplitude of each channel are calculated, and the occurrence frequency and the real-time change amplitude are fused to generate a current transient characteristic value; The wiring length of each channel on the target management object and the medium filling thickness between adjacent channels are collected in real time, and the current transient characteristic value is combined to calculate the loss ratio of the current radiation energy in the medium during transmission, to generate the energy cross-channel transmission loss coefficient of the two channels; The total number of channels in the frequency domain collaborative management domain is set as N, N is taken as the dimension to construct an N*N matrix, and the product of the energy cross-channel transmission loss coefficient between each two channels and the transient current of the two channels is taken as the element value of the corresponding position of the matrix, to generate an inter-channel energy correlation matrix, and the element of the inter-channel energy correlation matrix is the initial edge weight, which is used to reflect the correlation degree of the radiation energy between the channels; Based on the two-layer graph convolution network structure of the graph neural network, each channel is taken as a node, the channel signal frequency correlation degree of each node with other nodes is calculated, and the channel signal frequency correlation degree is multiplied by the corresponding element of the inter-channel energy correlation matrix to generate a correction factor; The initial edge weight is corrected based on the correction factor to obtain the modified node correlation weight; The node correlation weight is calculated based on the two-layer graph convolution network to obtain the feature vector of each node, the correlation degree of the feature vector of the target channel node with the feature vectors of all other nodes is calculated, and the inter-channel energy management coefficient is generated by collaborative analysis combined with the current transient characteristic value; The inter-channel energy management coefficient of all adjacent channels to the target channel is fused with the energy cross-channel transmission loss coefficient of the corresponding channel to generate the target channel energy influence coefficient.

[0008] Further, the target channel energy influence coefficient and the impedance data are fused through a neural network to generate a domain energy management factor representing the degree of energy influence on the target channel, including: Based on the impedance data of all channels, the impedance values are divided into three dynamic intervals of high resistance, medium resistance and low resistance, the switching frequency and weighted fluctuation amplitude of impedance in different dynamic intervals per unit time are calculated, the switching frequency and weighted fluctuation amplitude are fused to generate an impedance dynamic classification coefficient; The target channel energy influence coefficient and the impedance dynamic classification coefficient are associated to calculate an energy-impedance adaptive attenuation coefficient; The dynamic energy tolerance reference value is generated by analyzing the impedance dynamic classification coefficient and the dynamic interval of the impedance, and the difference between the energy-impedance adaptive attenuation coefficient and the dynamic tolerance reference value is calculated to generate a dynamic fusion feature value; Based on the lightweight neural network, the dynamic fusion feature value is calculated, and the energy-impedance adaptive attenuation coefficient is calculated to obtain an intra-domain energy management factor.

[0009] Further, the load data is collected continuously for T windows with 100ms as a collection window, and is arranged as time series data. The time series data and the energy management factor are input into a bidirectional long short-term memory network with two hidden layers for analysis. A pre-signal management vector containing amplitude and phase is output through a fully connected layer, including: The load data of all channels in the frequency domain cooperative management domain is classified according to the disturbance intensity to the signal amplitude and phase, the mutation influence weight of each level feature is calculated, and a load disturbance classification weight is generated; Based on the energy management factor in the domain, the window length is dynamically adjusted with 100ms as the basic collection window to generate a window dynamic configuration set containing the adjusted window length, the window mutation feature threshold and the window energy correlation weight; Based on the adjusted window length, the load data of T windows is continuously collected, the effective load features of each window are selected based on the window mutation feature threshold, the effective load features are weighted according to the load disturbance classification weight, and are concatenated in the order of collection time. The window energy correlation weight corresponding to each window is labeled to generate a load energy correlation sequence.

[0010] Further, the load data is collected continuously for T windows with 100ms as a collection window, and is arranged as time series data. The time series data and the energy management factor are input into a bidirectional long short-term memory network with two hidden layers for analysis. A pre-signal management vector containing amplitude and phase is output through a fully connected layer, including: Based on the load energy correlation sequence, the coincidence degree of strong disturbance features between adjacent windows and the change rate of window energy correlation weights in adjacent windows are analyzed. The product of the coincidence degree and the change rate is taken as a time series correlation correction weight. The correlation strength of adjacent window data in the load energy correlation sequence is adjusted based on the time series correlation correction weight to generate a time series correlation enhancement vector; The time series correlation enhanced vector is analyzed based on a bidirectional long short-term memory network with 2 layers of hidden layers, and the load disturbance grading weight is taken as the attention weight of the intermediate layer of the bidirectional long short-term memory network to obtain a bidirectional time series attention matrix representing time series dependence and feature attention; The bidirectional time series attention matrix is input into a fully connected layer, the output dimension is adjusted in combination with the window energy correlation weight in the window dynamic configuration set, and an amplitude and phase initial vector is obtained; The adjustment requirement of the signal amplitude and the compensation requirement of the phase by the load mutation are analyzed based on the time series correlation enhanced vector, the amplitude compensation accuracy and the phase compensation accuracy are calculated, and an amplitude and phase consistent calibration factor is generated; The amplitude and phase initial vector is calibrated according to the amplitude and phase consistent calibration factor, and a pre-signal management vector containing amplitude and phase is generated.

[0011] Further, according to the pre-signal management vector, the original signal of the target channel is waveform reshaped to generate an anti-interference reconstructed signal, including: The instantaneous fluctuation characteristics of the original signal of the target channel are collected in real time, the synchronous change of the instantaneous fluctuation characteristics of the original signal and the time series correlation enhanced vector is calculated, and a signal fluctuation-load correlation degree is generated; The signal fluctuation-load correlation degree and the pre-signal management vector are combined and calculated to generate an amplitude and phase dynamic compensation coefficient; Based on the resonant interference effective frequency point threshold, the overlap degree of each frequency band in the original signal and the resonant interference effective frequency point is analyzed, and the interference-signal dual-band suppression weight is generated in combination with the signal fluctuation-load correlation degree; The amplitude and phase of the original signal are corrected based on the amplitude and phase dynamic compensation coefficient, and then the frequency band overlapping with the resonant interference effective frequency point in the original signal is attenuated according to the interference-signal dual-band suppression weight, to generate an anti-interference reconstructed signal.

[0012] Further, the anti-interference reconstructed signal is analyzed by a convolutional neural network to generate an eye diagram quality management coefficient, including: The anti-interference reconstructed signal of the target channel is drawn into an eye diagram, and the correlation degree of different regions in the eye diagram and the interference residue is analyzed in combination with the interference-signal dual-band suppression weight to generate an eye diagram interference correlation factor; The basic data of the eye diagram is input into a 5-layer CNN with 2 layers of max-pooling layers, and the eye diagram interference correlation factor is taken as the attention weight embedded into the 3rd layer of the CNN, so that the CNN focuses on the region with high correlation degree with the interference residue in the eye diagram to extract features and generate a weighted eye diagram feature set; The amplitude and phase consistent calibration factor and the anti-interference reconstructed signal are analyzed to generate a dynamic eye diagram quality threshold; The deviation rate of the weighted eye diagram feature set from the dynamic eye diagram quality threshold is calculated, and the deviation rate is adjusted according to the eye diagram interference correlation factor to generate an eye diagram quality management coefficient.

[0013] Further, the time-varying trajectory of the eye diagram quality management coefficient is continuously tracked, the stability thereof is calculated, and the stability is compared with a stability management boundary to obtain a management instruction, including: The eye diagram quality management coefficient is continuously collected at preset time intervals, the numerical variation trajectory thereof is recorded, and the fluctuation amplitude and frequency of the eye diagram quality management coefficient per unit time are calculated in combination with the eye diagram interference correlation factor to generate a coefficient time sequence fluctuation degree; A fluctuation index is calculated according to the coefficient time sequence fluctuation degree and the amplitude-phase consistent calibration factor to generate a dynamic stability evaluation value; An adaptive stability boundary value is generated based on the dynamic eye diagram quality threshold to adapt to the current anti-interference scene; The dynamic stability evaluation value is compared with the adaptive stability boundary value to obtain a management instruction.

[0014] Further, an intelligent management system for a printed circuit board signal control process applies the above-described intelligent management method for a printed circuit board signal control process, and includes: An analysis unit is configured to acquire noise data of a target channel and all adjacent channels of a target management object in real time, analyze the noise data of all channels, and generate a frequency domain cooperative management domain, wherein the target management object is a server printed circuit board. An energy management unit is configured to acquire transient currents of all channels in the frequency domain cooperative management domain in real time, map the transient currents of the channels into an N×N inter-channel energy correlation matrix, construct a two-layer graph convolution network based on a graph neural network, take the channels as nodes and the element of the inter-channel energy correlation matrix as edge weights, and generate a target channel energy influence coefficient. A fusion management unit is configured to acquire impedance data of all channels in the frequency domain cooperative management domain in real time, fuse the target channel energy influence coefficient and the impedance data through a neural network, and generate a domain energy management factor representing the degree of energy influence on the target channel. A load management unit is configured to acquire load data of all channels in the frequency domain cooperative management domain in real time, collect the load data in T windows with a 100ms window as a collection window, and arrange the load data into time sequence data, input the time sequence data and the energy management factor into a two-layer hidden layer bidirectional long short-term memory network for analysis, and output a pre-signal management vector containing amplitude and phase through a fully connected layer. A reconstruction unit is configured to reshape the original signal of the target channel according to the pre-signal management vector to generate an anti-interference reconstructed signal. The extraction unit is used for eye diagram analysis on the anti-interference reconstructed signal through a convolutional neural network, a 5-layer CNN with 2 layers of maximum pooling layers is used to extract key feature indicators, and the key feature indicators are compared with eye diagram quality thresholds to generate an eye diagram quality management coefficient; The signal management unit is used for continuously tracking the time-varying trajectory of the eye diagram quality management coefficient, calculating the stability thereof, comparing the stability with a stability management boundary, and obtaining a management instruction.

[0015] In summary, the present application has the following advantages: By integrating the amplitude mutation synchronization rate, phase shift time difference, working environment temperature, signal reflection times and other parameters of the noise data in step S1, a frequency domain cooperative management domain is generated, the overall coverage of the multi-channel noise correlation characteristics is realized, and the one-sidedness of the traditional method of only focusing on single noise is avoided. Then, in step S2, an N*N channel energy correlation matrix is constructed, the energy cross-channel transmission loss is quantified by combining the wiring length, medium filling thickness and current transient characteristic value of each channel, and the channel signal frequency correlation is calculated by relying on a 2-layer graph convolution network to correct the node correlation weight, and finally the target channel energy influence coefficient is generated. The strong electromagnetic field coupling effect caused by the adjacent signal wiring spacing ≤0.2mm can be accurately captured, the energy interference of the adjacent channel transient current change on the target channel is quantified, the energy management factor in the domain is generated by fusing the impedance data in step S3, and the multi-dimensional coupling feature mapping is further improved by processing the load time sequence data through the bidirectional long short-term memory network with 2 layers of hidden layers in step S4. The signal distortion caused by crosstalk is completely avoided from being misjudged as the regular fluctuation of the single signal, and the high-frequency scene of the DDR5 signal high-speed interface signal from CPU to memory can be perfectly adapted, and the accuracy and effectiveness of the signal crosstalk control are improved.

[0016] The energy interference degree of the multi-channel is determined by the frequency domain cooperative management domain generated by the analysis unit and the target channel energy influence coefficient calculated by the energy management unit, the management unit is further refined by fusing the impedance dynamic classification and the energy influence coefficient, and the load management unit dynamically adjusts the acquisition window length and constructs the time sequence correlation enhancement vector through the bidirectional long short-term memory network and the attention weight to ensure the accurate extraction of the load disturbance characteristics. Based on the pre-signal management vector, the anti-interference reconstructed signal is generated by amplitude-phase dynamic compensation and resonance interference frequency band attenuation, which directly and specifically suppresses the interference caused by strong electromagnetic field coupling. The extraction unit and the signal management unit output the continuous management instruction or the alarm instruction in real time through the 5-layer CNN eye diagram analysis and stability comparison, avoiding the problems of lagging control instruction and insufficient specificity in the traditional system. In the high-density scene with adjacent wiring spacing ≤0.2mm and signal frequency ≥10GHz, the transmission stability of the signal is stably maintained, the signal distortion rate caused by crosstalk is reduced, and the signal control and management effect of the server printed circuit board is improved. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the steps of an intelligent management method for signal control process on a printed circuit board according to the present invention. Figure 2 This is a schematic diagram of an intelligent management system for the signal control process of a printed circuit board according to the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] refer to Figure 1 and Figure 2 A method for intelligent management of signal control processes on printed circuit boards, comprising: Step S1: Real-time acquisition of noise data for the target channel and all adjacent channels of the target management object. The noise data includes amplitude, phase, frequency of occurrence, and trend of change. Analyze the noise data of all channels to generate a frequency domain collaborative management domain. The target management object is the server printed circuit board. Step S2: Real-time acquisition of transient currents of all channels in the frequency domain collaborative management domain; mapping the transient currents of each channel to an N×N channel energy correlation matrix; constructing a 2-layer graph convolutional network based on a graph neural network; using channels as nodes and the elements of the channel energy correlation matrix as edge weights to generate the target channel energy influence coefficient. Step S3: Acquire impedance data of all channels in the frequency domain collaborative management domain in real time. Impedance data includes impedance value, impedance fluctuation amplitude, etc.; fuse the target channel energy influence coefficient with the impedance data through a neural network to generate an in-domain energy management factor that represents the degree of energy influence on the target channel. Step S4: Real-time acquisition of load data for all channels within the frequency domain collaborative management domain. Load data includes: load current value, current step rate of change, signal transmission rate, rate switching frequency, etc. The load data is collected in 100ms increments for T consecutive windows and organized into time-series data. A two-layer bidirectional long short-term memory network with hidden layers is constructed. The time-series data and energy management factors are input into the bidirectional long short-term memory network for analysis. A pre-signal management vector containing amplitude and phase is output through a fully connected layer, where T represents the number of windows. Step S5: Based on the pre-signal management vector, reshape the waveform of the original signal of the target channel to generate an anti-interference reconstruction signal; Step S6, eye diagram analysis is performed on the anti-interference reconstructed signal by the convolutional neural network, 5-layer CNN with 2-layer max pooling layer is used to extract key feature indicators, and the key feature indicators are compared with eye diagram quality threshold to generate an eye diagram quality management coefficient; Step S7, the time-varying trajectory of the eye diagram quality management coefficient is continuously tracked, the stability is calculated, and the stability is compared with a stability management boundary to obtain a management instruction.

[0020] In step S1, noise data of the target channel and all adjacent channels is acquired in real time to generate a frequency domain collaborative management domain, breaking through the limitation of single signal processing. Then, in step S2, an N*N channel energy correlation matrix and a 2-layer graph convolution network are constructed to accurately capture the coupling characteristics of the transient current of adjacent channels, avoiding misjudgment of crosstalk as a single regular fluctuation. In step S3, impedance data is fused, and in step S4, load time series data is processed by a 2-layer hidden layer bidirectional long short-term memory network. Then, in step S5, an anti-interference reconstructed signal is generated based on waveform remodeling. Thus, in step S6 and step S7, analysis and comparison are performed to output a targeted control instruction, which can effectively suppress the strong electromagnetic field coupling effect and ensure the transmission stability of the signal.

[0021] In one case of the embodiment, noise data of all channels is analyzed to generate a frequency domain collaborative management domain, including: The noise data is analyzed to calculate the amplitude mutation synchronization rate and phase drift time difference of the noise of each adjacent channel and the noise of the target channel to generate a noise dynamic coupling degree. Specifically, for the preprocessed noise data, 5ms is taken as a fixed sliding window, and the noise amplitude is sampled at intervals of 0.1ms. The first window starts from 0-5ms, and the first three windows are 0-5ms, 5-10ms, and 10-15ms. The mean value of the amplitude of all sampling points in these three windows is calculated, and each subsequent window is checked sample by sample. If the absolute value of the difference between the amplitude at a certain time and the mean value exceeds 2 times the mean value, the time is marked as an amplitude mutation point and the timestamp is recorded. After traversing all windows, the total number of mutation points of the target channel and the number of adjacent channel mutation points within ±1ms of the target channel mutation points are counted. The number is divided by the total number of mutation points to obtain the amplitude mutation synchronization rate. For the pretreated noise data, high-frequency interference is filtered out by an 8th order Butterworth low-pass filter (cut-off frequency is 100 kHz), and the amplitude sequence is resampled at 0.1 ms intervals; the amplitude sequence is subjected to Hilbert transform: a 9-tap FIR Hilbert filter is constructed, and the amplitude sequence is input into the Hilbert filter to obtain an orthogonal signal sequence; the amplitude sequence and the orthogonal signal sequence are combined to obtain an analytical signal; the amplitude angle of the complex amplitude of the analytical signal is calculated, and the amplitude angle is the instantaneous phase, forming a phase sequence that changes with time; the phase difference between adjacent sampling points is calculated, and if the absolute value of the difference exceeds 5°, i.e. the rate of change exceeds 50° / ms, the time is marked as a phase drift point and the time stamp is recorded; All the phase drift points of the adjacent channels are arranged in time sequence to form a sequence as the adjustment object, and the adjustment range is limited between moving forward 10 ms to moving backward 10 ms, the amplitude of each adjustment is 0.1 ms, and the number of time differences between the phase drift points of the adjacent channels and the target channel within 1 ms is counted each time the adjustment is made, which is the matching number, after repeated adjustment, the adjustment amplitude with the most matching number is found, which is the phase drift time difference; The phase drift time difference is converted into a correlation coefficient of 0-1, and the critical value is 10 ms, the initial coupling value is obtained by multiplying the amplitude mutation synchronization rate by the correlation coefficient and then taking the square root, if both the amplitude mutation synchronization rate and the correlation coefficient are greater than 0.5, it means that the consistency is strong, then the correction factor is 1.2, if any one is greater than 0.5 and the other is less than 0.5, it means that the consistency is weak, then the correction factor is 0.8, and the rest is the correction factor 1.0, the initial coupling value is multiplied by the correction factor, and the result is normalized to 0-1, which is the noise dynamic coupling degree, which is updated every 5 ms.

[0022] The working environment temperature of the target management object and the signal reflection times of the wiring of each adjacent channel are obtained in real time, and based on the noise dynamic coupling degree, the medium loss increment and the reflection loss superposition value of each adjacent channel in the process of noise transmission to the target channel are calculated based on the working environment temperature and the signal reflection times, and the noise time-varying transmission loss rate is generated, which specifically includes: taking 25℃ as the reference temperature, the reference loss correction coefficient at this temperature is 1, the difference between the real-time environment temperature and the reference temperature is increased or decreased by 1℃, the reference loss correction coefficient is increased or decreased by 0.02, the increased or decreased reference loss correction coefficient is multiplied by the reference temperature and then multiplied by the noise dynamic coupling degree to obtain the medium loss increment; The signal reflection times are divided into three grades: the first grade is 1-3 times, the single loss weight is 1.0, the second grade is 4-6 times, the single loss weight is 1.2, and the third grade is 7 times and above, the single loss weight is 1.5, and the total weight is the number of times multiplied by the corresponding grade weight; Real-time acquisition of signal propagation speed and signal frequency of the printed circuit board, divide the signal propagation speed by (2 times the signal frequency), get the half wavelength of the signal in the wiring, divide the wiring length by the half wavelength, get the electrical length ratio; 0.05dB is multiplied by the electrical length ratio to get the single reference loss; multiply the single reference loss by the total weight of the corresponding signal reflection times, then multiply by the coupling degree to get the reflection loss superposition value; Subtract the temperature of the last 5ms from the current 5ms collected temperature, and divide by 0.005 seconds to get the temperature change rate. If the temperature change rate is greater than 2℃ / s, the average of the medium loss increment + reflection loss superposition value of the current and the last period is taken as the noise time-varying transmission loss rate. If the temperature change rate is ≤2℃ / s, the medium loss increment + reflection loss superposition value of the last 5 periods is selected, and the average of the 5 sums is calculated as the noise time-varying transmission loss rate. Update every 5ms.

[0023] Analyze the amplitude fluctuation period of the noise of each adjacent channel, correlate the noise amplitude fluctuation period corresponding to each frequency point with the current signal transmission period of the target channel, and obtain the period fit degree of each frequency point, which specifically includes: dividing the noise of the adjacent channel into frequency points at 1kHz intervals, and analyzing each frequency point data using autocorrelation algorithm: set the lag time range to 0-10ms, calculate the autocorrelation coefficient under different lag times, and select the time interval corresponding to the peak value of the autocorrelation coefficient as the fluctuation period of the noise amplitude of the frequency point; Real-time acquisition of the signal transmission period of the target channel, for each frequency point, the length of the waveform overlap between the noise period and the signal transmission period is counted, such as 5ms noise period and 100ns transmission period overlap 1ms, divide the overlap length by the signal transmission period to get the matching ratio, multiply the matching ratio by the noise dynamic coupling degree, and normalize the result to 0-1, which is the period fit degree of the frequency point.

[0024] Calculate the period fit degree and noise time-varying transmission loss rate to obtain the resonant interference effective frequency point threshold, which specifically includes: for each frequency point, subtract the noise time-varying transmission loss rate and multiply by the period fit degree to get the resonant interference intensity of the frequency point; Statistically analyze the resonant interference intensity of all frequency points, and select the 75% quantile value as the resonant interference effective frequency point threshold. The frequency points exceeding the threshold are the resonant interference effective frequency points.

[0025] Based on the resonant interference effective frequency point threshold, the noise dynamic coupling degree and the noise time-varying transmission loss rate are analyzed to obtain a frequency domain cooperative management domain, which specifically includes: for each adjacent channel, which is the interference source channel, at each frequency point in the resonant interference effective frequency point range, the noise dynamic coupling degree of the channel is multiplied by (1-noise time-varying transmission loss rate) to obtain the interference intensity of the channel to the target channel at the frequency point; for each adjacent channel, the maximum value of the interference intensity of all frequency points in the effective frequency point range is taken as the channel interference weight of the channel; the adjacent channels with channel interference weight greater than 0.5 and the corresponding resonant interference effective frequency points are retained, and these channels and frequency points jointly form the frequency domain cooperative management domain.

[0026] By calculating the amplitude mutation synchronization rate and phase shift time difference of the adjacent channel and the target channel to generate the noise dynamic coupling degree, and combining the working environment temperature and the signal reflection times to calculate the noise time-varying transmission loss rate, and simultaneously associating the noise amplitude fluctuation period of each frequency point with the target signal transmission period to obtain the period fit degree, the adjacent channels are screened, the strong electromagnetic field coupling characteristics among multiple channels are fully captured, the signal distortion caused by crosstalk is avoided to be misjudged as single regular fluctuation, and the high-frequency scene of DDR5 signal and PCIe5.0 high-speed interface signal can be targetedly adapted, and the precision and effectiveness of signal crosstalk control are improved.

[0027] In one case of the embodiment, the transient current of each channel is mapped to an N×N channel energy correlation matrix, a 2-layer graph convolution network is constructed based on a graph neural network, the channels are taken as nodes, and the elements of the channel energy correlation matrix are taken as edge weights, and a target channel energy influence coefficient is generated, including: The current amplitudes of all channels in the frequency domain cooperative management domain are acquired in real time; The occurrence frequency of the transient current peak value and the real-time change amplitude of the current amplitude per unit time of each channel are calculated, the occurrence frequency and the real-time change amplitude are fused, and a current transient characteristic value is generated, specifically including: taking 10ms as a unit time, sampling the transient current according to 0.1ms, calculating the mean and standard deviation of the transient current, and setting the sampling points exceeding the mean+3 times the standard deviation as peak values, and the peak value number as the occurrence frequency; the difference between the maximum and minimum current amplitudes is calculated in each window with a 1ms sliding window, and the mean of all window differences is taken as the real-time change amplitude; the real-time change amplitude weight is set to 0.6, and the occurrence frequency weight is set to 0.4, the occurrence frequency and the real-time change amplitude are multiplied by the corresponding weights respectively, and then added, and the result is normalized to 0-1, which is the current transient characteristic value; Among them, the real-time change amplitude directly reflects the energy fluctuation intensity of the transient current, is the core factor to determine the size of the channel radiation energy, and is more critical to the subsequent construction of the inter-channel energy correlation matrix, so the weight is biased high at 0.6, and the frequency of the peak value only reflects the fluctuation frequency, which is related to the transient characteristics but the influence is less than the amplitude, so the weight is biased low at 0.4.

[0028] The wiring length of each channel on the target management object and the medium filling thickness between adjacent channels are collected in real time, and the loss proportion of the current radiation energy when transmitted in the medium is calculated in combination with the current transient characteristic value, to generate the energy cross-channel transmission loss coefficient of the two channels, which specifically includes: the wiring length refers to the complete signal transmission line length of each channel itself, that is, the physical wiring length from the signal transmitting end (such as the chip pin) of the channel to the signal receiving end (such as the load interface); the medium filling thickness is the actual thickness of the insulating medium between adjacent channels, and the smaller the thickness, the easier the energy to cross-channel transmission; set the wiring length to correspond to a basic loss of 0.02 per 1 cm, and set the medium filling thickness to correspond to a basic loss of 0.03 per 0.1 mm; add the basic loss corresponding to the wiring length and the medium filling thickness, and multiply by the current transient characteristic value, and normalize the result to 0-1, which is the energy cross-channel transmission loss coefficient of the two channels.

[0029] The total number of channels in the frequency domain cooperative management domain is set as N, N is taken as the dimension to construct an N*N matrix, and the product of the energy cross-channel transmission loss coefficient between each two channels and the transient current of the two channels is taken as the element value of the corresponding position of the matrix to generate the inter-channel energy correlation matrix. The elements of the inter-channel energy correlation matrix are the initial edge weights, which are used to reflect the correlation degree of the radiation energy between channels, and specifically include: setting the total number of channels in the frequency domain cooperative management domain as N, and constructing a blank matrix of N rows and N columns, each row and each column of the matrix corresponds to a channel; for any two channels, multiply the energy cross-channel transmission loss coefficient of the two channels by the transient current of the first channel, then multiply by the transient current of the second channel, and fill the calculated result into the corresponding position of the matrix; all two-by-two combinations of channels are calculated and filled in this way, and finally the inter-channel energy correlation matrix is formed. Each element in the matrix is an initial edge weight, which corresponds to two channels and directly reflects the correlation strength of the radiation energy between the two channels.

[0030] In one case of the embodiment, the transient current of each channel is mapped to an N*N inter-channel energy correlation matrix, a 2-layer graph convolution network is constructed based on a graph neural network, the channel is taken as a node, and the element of the inter-channel energy correlation matrix is taken as an edge weight to generate a target channel energy influence coefficient, which further includes: The 2-layer graph convolution network structure based on the graph neural network takes each channel as a node, calculates the channel signal frequency correlation degree of each node with other nodes, and then multiplies the channel signal frequency correlation degree with the corresponding element of the inter-channel energy correlation matrix to generate a correction factor, specifically including: setting a 2-layer graph convolution network, the first layer GCN takes the sequence of the resonant interference effective frequency point combination of each channel as input, adopts a 1D convolution kernel with a size of 1*3, and slidingly convolves to extract the local correlation features of each frequency point and adjacent frequency points, and outputs a 128-dimensional frequency feature vector of each channel; the second layer GCN takes the channels as nodes and the elements of the inter-channel energy correlation matrix as initial edge weights, for any two nodes, the 128-dimensional feature vector of node 1 is multiplied by the initial edge weight, and then is added to the 128-dimensional feature vector of node 2 element by element, and the addition result is mapped to 0-1, which is the channel signal frequency correlation degree of the two channels; then the channel signal frequency correlation degree is multiplied by the corresponding element of the inter-channel energy correlation matrix, and the result is normalized to 0-1, which is the correction factor.

[0031] The initial edge weight is corrected based on the correction factor to obtain a corrected node correlation weight, specifically including: for any two channels, the initial edge weight of the two channels in the inter-channel energy correlation matrix is multiplied by the corresponding correction factor to obtain the corrected node correlation weight.

[0032] The node correlation weight is calculated based on the 2-layer graph convolution network to obtain the feature vector of each node, the correlation degree of the feature vector of the target channel node and the feature vectors of all other nodes is calculated, and the current transient characteristic value is combined for collaborative analysis to generate an inter-channel energy management coefficient, specifically including: in the 2-layer GCN, taking the corrected node correlation weight as the edge weight, for each node, the 128-dimensional frequency features of the node and all adjacent nodes are weighted and summed according to the corresponding dimensions, for example, the first dimension is the first dimension feature of the node itself plus the first dimension feature of the adjacent node 1 multiplied by the correlation weight, so the new characteristic value of the first dimension of the node can be obtained, and the remaining 127 dimensions are calculated in the same way, finally forming the final 128-dimensional feature vector of the node; the correlation degree between the node of the target channel and the nodes of other adjacent channels is calculated by using the cosine similarity algorithm, and the value range of the correlation degree is 0-1; the correlation degree is multiplied by the smallest current transient characteristic value in the two channels, and the result is normalized to 0-1, which is the inter-channel energy management coefficient.

[0033] By sampling the transient current at a time unit of 10 ms with an interval of 0.1 ms, combining the real-time change amplitude and the peak value occurrence frequency to generate the current transient characteristic value, and then integrating the wiring length and the dielectric filling thickness to calculate the energy cross-channel transmission coefficient, a matrix reflecting the correlation strength of the radiation energy between channels is constructed, which can accurately capture the energy coupling relationship between adjacent channels caused by a wiring spacing of less than 0.2 mm, avoid the signal distortion caused by crosstalk from being misjudged as a regular fluctuation, and through the generated energy management coefficient between channels, the degree of influence of other channels on the target channel is quantified, the strong electromagnetic field coupling effect is targetedly suppressed, the problem of non-targeted control instructions in the traditional method is solved, the stability of high-frequency signal transmission is effectively guaranteed, and the signal crosstalk control and management effect is improved.

[0034] In one case of the embodiment, the target channel energy influence coefficient and the impedance data are fused through a neural network to generate a domain energy management factor representing the degree of influence of energy on the target channel, including: Based on the impedance data of all channels, the impedance values are divided into three dynamic intervals of high impedance, medium impedance and low impedance, the switching frequency and the weighted fluctuation amplitude of the impedance in different dynamic intervals within a unit time are calculated, the switching frequency and the weighted fluctuation amplitude are fused to generate an impedance dynamic classification coefficient, specifically including: for all channels, the impedance values are divided into intervals: low impedance < 50 Ω, medium impedance 50-200 Ω, high impedance > 200 Ω, the impedance is sampled at a time unit of 10 ms and an interval of 0.1 ms; the number of impedance interval switching of each channel is counted as the switching frequency; the high impedance, medium impedance and low impedance fluctuation amplitudes are respectively assigned weights of 0.5, 0.3 and 0.2; the difference between the maximum impedance value and the minimum impedance value of all sampling points in each interval is calculated, which represents the fluctuation amplitude in the interval; the difference value of each interval is multiplied by the corresponding weight of the interval, and then the calculation results of each interval are added to obtain the weighted fluctuation amplitude; the switching frequency is multiplied by 0.4+ the weighted fluctuation amplitude is multiplied by 0.6, and then the result is normalized to 0-1, which is the impedance dynamic classification coefficient; Wherein, the weighted fluctuation amplitude directly reflects the actual change size of the impedance value, the larger the impedance amplitude fluctuation, the more intense the energy transmission efficiency change of the channel, which will directly change the degree of influence of energy on the target channel, for example, a large fluctuation in the high impedance interval will increase the energy loss, therefore, the weighted fluctuation amplitude is the core dimension to determine the degree of energy influence, so it is assigned a high weight of 0.6; while the switching frequency only reflects the frequency of impedance interval switching, and does not directly determine the change amplitude of the energy transmission efficiency, it is an auxiliary embodiment of the impedance dynamics, so it is assigned a low weight of 0.4.

[0035] The target channel energy influence coefficient is associated with the impedance dynamic classification coefficient to calculate an energy-impedance adaptive attenuation coefficient, specifically including: multiplying the target channel energy influence coefficient and the impedance dynamic classification coefficient to obtain a basic adaptive value; for the basic adaptive value, if the impedance currently belongs to a high impedance interval, the basic adaptive value is added by 0.1, the medium impedance interval is not added or subtracted, and if the impedance currently belongs to a low impedance interval, the basic adaptive value is subtracted by 0.1; finally, the final basic adaptive value is normalized to 0-1, which is the energy-impedance adaptive attenuation coefficient.

[0036] The impedance dynamic classification coefficient and the dynamic interval of the impedance are analyzed to generate a dynamic energy tolerance reference value, the difference between the energy-impedance adaptive attenuation coefficient and the dynamic tolerance reference value is calculated to generate a dynamic fusion feature value, specifically including: according to the impedance interval, if the impedance is a high impedance interval, the dynamic energy tolerance reference value = impedance dynamic classification coefficient x 0.8; if the impedance is a medium impedance interval, the dynamic energy tolerance reference value = impedance dynamic classification coefficient x 1; if the impedance is a low impedance interval, the dynamic energy tolerance reference value = impedance dynamic classification coefficient x 1.2; subtract the dynamic energy tolerance reference value from the energy-impedance adaptive attenuation coefficient, and normalize the obtained difference value to 0-1, which is the dynamic fusion feature value.

[0037] Based on the light neural network, the dynamic fusion feature value is calculated, and the energy-impedance adaptive attenuation coefficient is calculated to obtain the domain energy management factor, specifically including: inputting the dynamic fusion feature value into a 1-layer hidden layer (32 neurons) of a light neural network, and outputting an intermediate feature value of 0-1 through a ReLU activation function; multiplying the intermediate feature value and the energy-impedance adaptive attenuation coefficient to obtain a preliminary fusion value; if the current impedance is a high impedance interval, the preliminary fusion value + 0.05 = the corrected fusion value; if the current impedance is a medium impedance interval, the preliminary fusion value is directly used as the corrected fusion value; if the current impedance is a low impedance interval, the preliminary fusion value - 0.05 = the corrected fusion value; the corrected fusion value is normalized to 0-1, which is the domain energy management factor.

[0038] By dividing the impedance value into low impedance, medium impedance and high impedance intervals, combining the switching frequency and the weighted fluctuation amplitude to generate the impedance dynamic classification coefficient, and then through associated calculation to obtain the energy-impedance adaptive attenuation coefficient, and based on the dynamic energy tolerance reference value to generate the dynamic fusion feature value, after processing by a 1-layer hidden layer light neural network, the domain energy management factor is generated, which can accurately quantify the degree of influence of the target channel by energy, adapt to the high-frequency scene of DDR5 signal and PCIe5.0 high-speed interface signal, avoid the crosstalk caused by strong electromagnetic field coupling from being misjudged as regular fluctuation, make the control command more targeted, and significantly improve the signal crosstalk control and management effect.

[0039] In one case of the embodiment, the load data is collected as a 100 ms collection window, T windows are collected continuously, and are arranged as time series data. The time series data and energy management factors are input into a bidirectional long short-term memory network with two hidden layers for analysis. A pre-signal management vector containing amplitude and phase is output through a fully connected layer, including: For the load data of all channels in the frequency domain cooperative management domain, the disturbance intensity to the signal amplitude and phase is classified, the mutation influence weight of each level feature is calculated, and the load disturbance classification weight is generated. Specifically, the load current value, current step change rate, signal transmission rate, and rate switching frequency of the load data are classified into three levels according to the disturbance intensity to the amplitude and phase. When the current step change rate is > 5A / μs, the rate switching frequency is > 100kHz, the load current value is > 20A, and the signal transmission rate is > 10Gbps, it is high disturbance. When the current step change rate is between 1-5A / μs, the rate switching frequency is between 50-100kHz, the load current value is between 5-20A, and the signal transmission rate is between 1-10Gbps, it is medium disturbance. When the current step change rate is < 1A / μs, the rate switching frequency is < 50kHz, the load current value is < 5A, and the signal transmission rate is < 1Gbps, it is low disturbance. For high, medium and low disturbance, the disturbance values are set to 1.0, 0.5 and 0.2 respectively. The load current value weight is set to 0.3, the current step change rate weight is set to 0.3, the signal transmission rate weight is set to 0.2, and the rate switching frequency weight is set to 0.2. The weights of each load parameter are multiplied by the corresponding disturbance value and then added. The result is normalized to 0-1, which is the load disturbance classification weight of each channel.

[0040] With 100 ms as the basic collection window, combined with the energy management factor in the domain, the window length is dynamically adjusted to generate a window dynamic configuration set containing the adjusted window length, the window internal mutation feature threshold and the window energy correlation weight. Specifically, the energy management factor in the domain is divided into three grades. When the energy management factor in the domain is > 0.7, it indicates high energy impact, and the 100 ms basic length is increased by 50 ms to obtain the adjusted window length. When the energy management factor in the domain is between 0.3-0.7, it indicates medium energy impact, and the 100 ms is directly used as the adjusted window length. When the energy management factor in the domain is < 0.3, it indicates low energy impact, and the 100 ms basic length is reduced by 50 ms to obtain the adjusted window length. Each load parameter is multiplied by 0.01 to obtain the window internal mutation feature threshold of each load parameter. The average of the sum of the energy management factor in the domain and the load disturbance classification weight is calculated, which is the window energy correlation weight. The adjusted window length, the window internal mutation feature threshold and the window energy correlation weight are combined to obtain the window dynamic configuration set.

[0041] Based on the adjusted window length, T windows of load data are continuously collected, the effective load features of each window are screened based on the window internal mutation feature threshold, the effective load features are weighted according to the load disturbance grading weight, and the windows are connected in sequence according to the collection time sequence, the window energy correlation weight corresponding to each window is marked, and a load energy correlation sequence is generated. Specifically, T windows of load data are collected according to the adjusted window length; for each window, if the difference between the maximum value and the minimum value of the parameter in the load data is greater than the window internal mutation feature threshold of the corresponding parameter, the effective load feature is retained, otherwise it is excluded; the value of each retained effective load feature is multiplied by the load disturbance grading weight corresponding to the channel to obtain the weighted effective load feature value, and then all the weighted effective load feature values are connected in sequence according to the collection time, each window is marked with its window energy correlation weight, and a load energy correlation sequence is formed.

[0042] In one case of the embodiment, the load data is collected in T windows with 100ms as one collection window, and is arranged as time sequence data. The time sequence data and the energy management factor are input into the bidirectional long short-term memory network with two hidden layers for analysis, and the pre-signal management vector containing amplitude and phase is output through the full connection layer. In addition, the following information is also included: Based on the load energy correlation sequence, the coincidence degree of strong disturbance features between adjacent windows and the change rate of window energy correlation weights in adjacent windows are analyzed, the product of the coincidence degree and the change rate is taken as the time sequence correlation correction weight, the correlation strength of adjacent window data in the load energy correlation sequence is adjusted based on the time sequence correlation correction weight, and a time sequence correlation enhancement vector is generated. Specifically, for the load energy correlation sequence, if the weighted effective load feature value is greater than 0.5, the strong disturbance feature of the window is marked, otherwise no marking is performed; for the two adjacent windows, the number of strong disturbance features common to the two windows is counted, the number is divided by the total number of strong disturbance features of the two windows, and the result is normalized to 0-1, which is the coincidence degree; for the window energy correlation weights of adjacent windows, the absolute value of the difference between the window energy correlation weight of the next window and the window energy correlation weight of the previous window is calculated, the absolute value is divided by the window energy correlation weight of the previous window, and the result is normalized to 0-1, which is the change rate; the coincidence degree is multiplied by the change rate to obtain the time sequence correlation correction weight. For each pair of adjacent windows in the load energy correlation sequence, the adjacent windows are the previous window and the subsequent window, and the weighted effective load feature values of the previous window remain unchanged; for the subsequent window, each strong disturbance feature value is multiplied by the time sequence correlation correction weight, and the non-strong disturbance feature value remains unchanged; the weighted effective load features of all windows in the sequence are connected in sequence according to the time sequence to form an ordered sequence, which is the time sequence correlation enhancement vector.

[0043] The time series correlation enhanced vector is analyzed based on a two-layer hidden layer bidirectional long short-term memory network, and the load disturbance grading weight is taken as the attention weight of the intermediate layer of the bidirectional long short-term memory network, so as to obtain a bidirectional time series attention matrix representing time series dependence and feature attention. Specifically, the time series correlation enhanced vector is input into the two-layer hidden layer bidirectional LSTM intermediate layer with 64 neurons in each layer, which outputs a forward hidden state vector and a backward hidden state vector (each with 64 dimensions) of each window. The load disturbance grading weight corresponding to each window is selected, and the load disturbance grading weight is multiplied by the forward hidden state vector and the backward hidden state vector of the window, respectively, to obtain the weighted forward and backward hidden state vectors. The weighted forward hidden state vector and the backward hidden state vector of each window are spliced, with a total of 128 dimensions, and then arranged in the order of collection time, wherein T windows correspond to T rows, forming a T*128-dimensional bidirectional time series attention matrix.

[0044] The bidirectional time series attention matrix is input into a fully connected layer, and the window energy correlation weight in the window dynamic configuration set is combined to adjust the output dimension, so as to obtain an amplitude-phase initial vector. Specifically, the T*128-dimensional bidirectional time series attention matrix is input into a fully connected layer, and is first mapped into a 32-dimensional feature vector of each window. Then, the window energy correlation weight corresponding to each window is multiplied by the 32-dimensional vector of the window, to obtain a weighted 32-dimensional vector. The weighted 32-dimensional vectors of all windows are spliced in time sequence, and then input into a fully connected layer with 16 neurons, to output a 2-dimensional vector (1-dimensional amplitude and 1-dimensional phase), which is the amplitude-phase initial vector.

[0045] The adjustment requirement of the signal amplitude and the compensation requirement of the phase by the load mutation are analyzed based on the time series correlation enhanced vector, the amplitude compensation accuracy and the phase compensation accuracy are calculated, and an amplitude-phase consistent calibration factor is generated. Specifically, from the time series correlation enhanced vector, the current step change rate and the load current value of the weighted effective load feature are classified as amplitude-related features; the rate switching frequency and the signal transmission rate corresponding to the weighted effective load feature are classified as phase-related features; and the load mutation is identified from these features: it has been defined that the feature with a weighted value greater than 0.5 is a strong disturbance feature, which represents the load mutation. For example, if the weighted value of the current step change rate in a window is 0.6, it belongs to the amplitude-related load mutation feature, and if the weighted value of the rate switching frequency is 0.7, it belongs to the phase-related load mutation feature. Set the safety threshold of amplitude, phase respectively 0.3; the maximum value of amplitude related mutation characteristics-0.3, get the amplitude adjustment demand; the maximum value of phase related mutation characteristics-0.3, get the phase adjustment demand; 1 divided by amplitude adjustment demand, then the result is normalized to 0-1, get the amplitude compensation accuracy, if the amplitude adjustment demand≤0, then the amplitude compensation accuracy is set to 1; 1 divided by phase adjustment demand, then the result is normalized to 0-1, get the phase compensation accuracy, if the phase adjustment demand≤0, then directly set the phase compensation accuracy to 1; the mean value of amplitude compensation accuracy and phase compensation accuracy after normalization to 0-1 as amplitude and phase consistent calibration factor.

[0046] According to the amplitude and phase consistent calibration factor, the amplitude and phase initial vector is calibrated to generate a pre-signal management vector containing amplitude and phase, specifically including: from the amplitude and phase initial vector, split out 1-dimensional amplitude initial value and 1-dimensional phase initial value; the amplitude initial value is multiplied by the amplitude and phase consistent calibration factor to obtain the calibrated amplitude value; the phase initial value is multiplied by the amplitude and phase consistent calibration factor to obtain the calibrated phase value; the calibrated amplitude value and the calibrated phase value are combined in the order of amplitude first and phase second to form a 2-dimensional vector, which is the pre-signal management vector containing amplitude and phase.

[0047] Through the generated load disturbance hierarchical weight, combined with the in-domain energy management factor to dynamically adjust the acquisition window length, the effective load characteristics are screened to form a load energy correlation sequence, and finally analyzed by a two-layer hidden layer bidirectional long short-term memory network, combined with the amplitude and phase consistent calibration factor to generate a pre-signal management vector containing amplitude and phase, which can capture the load and timing coupling relationship in high-frequency scenarios, avoid the crosstalk caused by strong electromagnetic field coupling from being misjudged as regular fluctuations, and improve the signal crosstalk control management effect.

[0048] In one case of the embodiment, according to the pre-signal management vector, the original signal of the target channel is waveform reshaped to generate an anti-interference reconstructed signal, including: The instantaneous fluctuation characteristics of the original signal of the target channel are collected in real time, the synchronous change of the original signal instantaneous fluctuation characteristics and the timing correlation enhancement vector is calculated, and the signal fluctuation-load correlation degree is generated, specifically including: collecting the instantaneous fluctuation characteristics of the original signal of the target channel with 0.1ms as the sampling interval, the instantaneous fluctuation characteristics including amplitude fluctuation and phase shift; using a 10ms sliding window, the Pearson correlation coefficient of the instantaneous fluctuation characteristics and the timing correlation enhancement vector in the window is calculated, and then the Pearson correlation coefficient is normalized to 0-1, and the result is the signal fluctuation-load correlation degree.

[0049] The signal fluctuation-load correlation degree and the pre-signal management vector are combined to calculate amplitude-phase dynamic compensation coefficients, specifically including: for the calibrated amplitude value and the calibrated phase value in the pre-signal management vector; the threshold of the signal fluctuation-load correlation degree is set to 0.5; when the signal fluctuation-load correlation degree > 0.5, it indicates that the correlation degree is high, and the basic adjustment coefficient is 0.2; when the signal fluctuation-load correlation degree ≤ 0.5, it indicates that the correlation degree is low, and the basic adjustment coefficient is 0.1; the signal fluctuation-load correlation degree + the corresponding basic adjustment coefficient is multiplied by the calibrated amplitude value to obtain the amplitude dynamic compensation coefficient; the signal fluctuation-load correlation degree + the corresponding basic adjustment coefficient is multiplied by the calibrated phase value to obtain the phase dynamic compensation coefficient; the amplitude dynamic compensation coefficient and the phase dynamic compensation coefficient are combined in the order of amplitude first and phase second, which is the amplitude-phase dynamic compensation coefficient.

[0050] Based on the resonant interference effective frequency point threshold, the overlap degree of each frequency band in the original signal and the resonant interference effective frequency point is analyzed, and the signal fluctuation-load correlation degree is combined to generate interference-signal dual-band suppression weight, specifically including: dividing the frequency band of the original signal according to 1 kHz interval, counting the number of resonant interference effective frequency points in each frequency band, and dividing the number by the total number of resonant interference effective frequency points to obtain the overlap degree of each frequency band in the original signal and the effective frequency point; the overlap degree of each frequency band is multiplied by (1+signal fluctuation-load correlation degree), and then the result is normalized to 0-1, which is the interference-signal dual-band suppression weight of the frequency band.

[0051] Based on the amplitude-phase dynamic compensation coefficient, the amplitude and phase of the original signal are corrected, and then the frequency band overlapping with the resonant interference effective frequency point in the original signal is attenuated according to the interference-signal dual-band suppression weight to generate an anti-interference reconstructed signal, specifically including: sampling the original signal at 0.1 ms interval, multiplying the instantaneous amplitude of the original signal by the amplitude dynamic compensation coefficient, and then adding the instantaneous phase of the original signal to the phase dynamic compensation coefficient to obtain the amplitude and phase corrected signal; the amplitude and phase corrected signal is decomposed into 1 kHz interval frequency bands by FFT (Fast Fourier Transform), and each frequency band corresponds to a signal component, such as 0.5V for the 1 kHz frequency band component, which represents the signal strength; if the frequency band belongs to the resonant interference effective frequency point, the signal component of the frequency band is multiplied by (1 minus the interference-signal dual-band suppression weight of the corresponding frequency band); if the frequency band does not belong to the resonant interference effective frequency point, the signal component of the frequency band remains unchanged; all adjusted frequency band components are recombined by inverse FFT (Inverse Fast Fourier Transform) to obtain the anti-interference reconstructed signal.

[0052] The instantaneous fluctuation characteristics of the original signal, such as amplitude fluctuation and phase drift, are collected at an interval of 0.1 ms, the Pearson correlation coefficient of the time sequence associated enhancement vector is calculated through a 10 ms sliding window, the signal fluctuation-load correlation degree is obtained, then the amplitude-phase dynamic compensation coefficient is generated, and the interference-signal dual-band suppression weight is calculated in combination with the correlation degree by dividing the frequency band at 1 kHz. Finally, the anti-interference reconstructed signal is generated through amplitude-phase correction and interference frequency band attenuation, avoiding the misjudgment of crosstalk caused by strong electromagnetic field coupling as a regular fluctuation, and improving the signal anti-interference performance.

[0053] In one case of the embodiment, the eye diagram analysis of the anti-interference reconstructed signal is performed through the convolutional neural network, the 5-layer CNN with 2-layer max-pooling layer is used to extract the key feature index, and the eye diagram quality management coefficient is generated by comparing the key feature index with the eye diagram quality threshold, including: The eye diagram of the anti-interference reconstructed signal of the target channel is drawn, and the interference-signal dual-band suppression weight is combined to analyze the correlation degree of different regions in the eye diagram and the interference residual, and the eye diagram interference correlation factor is generated, specifically including: sampling the signal amplitude at an interval of 0.1 ms according to the transmission period of the anti-interference reconstructed signal, superimposing the sampling values of 50 continuous periods in the same amplitude-time coordinate system to form an eye diagram containing eye opening and cross point, wherein the eye opening represents the distance between the upper and lower eyelids, and the cross point represents the zero-crossing point of the signal; the eye opening corresponds to the frequency band of 1-10 kHz, and the average value of the interference-signal dual-band suppression weight of the frequency band is calculated; the cross point corresponds to the frequency band of 10-20 kHz, and the average value of the interference-signal dual-band suppression weight of the frequency band is calculated; the eye opening value is multiplied by the average value of the corresponding weight, and the cross point value is multiplied by the average value of the corresponding weight, and the sum is normalized to 0-1, which is the eye diagram interference correlation factor.

[0054] The basic data of the eye diagram includes the pixel distribution and gray scale features of the eye diagram; The basic data of the eye diagram is input into the 5-layer CNN with 2-layer max-pooling layer, and the eye diagram interference correlation factor is embedded into the 3rd layer of the CNN as the attention weight, so that the CNN focuses on the region with high correlation degree with the interference residual in the eye diagram to extract the features and generate the weighted eye diagram feature set, specifically including: the eye diagram is a 256x256 pixel image, each pixel corresponds to a position coordinate, the brightness value (0-255) of each pixel is divided by 255 and normalized to 0-1 to obtain a 256x256 pixel value matrix; the number of pixels of different gray scales in the eye diagram is counted, and the pixel value matrix is combined after the number of pixels is normalized to 0-1, forming an input data of 3x256x256, wherein 3 represents 3 channels: pixel brightness, gray scale distribution, and position correlation; the input data is input into the 5-layer CNN. In the first convolutional layer, the input data is scanned with a 3x3 convolutional kernel to extract primary local features of the eye diagram, such as edges, light-dark boundaries, etc., such as eye diagram contours, intersection edges, etc., output multiple feature maps, and reduce the dimension of the multiple feature maps through 2x2 max pooling to retain the strongest features in each region, such as the most obvious edges, while reducing the amount of data and strengthening key information. Then in the second convolutional layer, also with a 3x3 convolutional kernel on the reduced feature map, more convolutional kernels are used to extract more complex combined features, such as eye opening width, intersection density, etc. After dimension reduction through the 2x2 max pooling layer, it is output to the third layer, which is the middle convolutional layer. At this time, the eye diagram interference correlation factor is directly multiplied by each element of the feature map of the third layer to strengthen the features of the high correlation region and weaken the features of the low correlation region. In the fourth convolutional layer, based on the strengthened features, interference correlation features are extracted, such as eye diagram differences corresponding to different frequency band interferences. In the fifth convolutional layer, a 1x1 convolutional kernel is used to compress the feature dimension to retain the most core interference-eye diagram correlation features. Finally, the weighted eye diagram feature set is output, which contains the vector of all key interference-related features.

[0055] The amplitude-phase consistent calibration factor and the anti-interference reconstructed signal are analyzed to generate a dynamic eye diagram quality threshold, which specifically includes: sampling the anti-interference reconstructed signal at 0.1 ms intervals, calculating the standard deviation of the amplitudes of the consecutive 50 sampling points, dividing the standard deviation by the average amplitude of the 50 points to obtain the amplitude fluctuation coefficient, which ranges from 0 to 1; multiplying the amplitude fluctuation coefficient by the amplitude-phase consistent calibration factor to obtain the anti-interference-calibration synergy coefficient; normalizing the average of the anti-interference-calibration synergy coefficient and the interference-signal dual-band suppression weight after addition to 0-1, which is the dynamic eye diagram quality threshold.

[0056] The deviation rate of the weighted eye diagram feature set and the dynamic eye diagram quality threshold is calculated, and then the eye diagram quality management coefficient is generated according to the eye diagram interference correlation factor, which specifically includes: taking the average of all elements of the weighted eye diagram feature set as the feature mean; calculating the difference between the feature mean and the dynamic eye diagram quality threshold, dividing the absolute value of the difference by the dynamic eye diagram quality threshold to obtain the original deviation rate; multiplying the original deviation rate by (1+eye diagram interference correlation factor), and then normalizing the result to 0-1, which is the eye diagram quality management coefficient.

[0057] By sampling the transmission period of the anti-interference reconstructed signal, superimposing 50 cycles to generate an eye diagram containing eye opening and crossing points, combining the interference-signal dual-band suppression weight mean to generate an eye diagram interference correlation factor, and then inputting the eye diagram basic data into a 5-layer CNN, embedding the factor in the third layer to focus on the high interference correlation area, extracting the core interference-eye diagram correlation characteristics to form a weighted eye diagram feature set, and combining the amplitude and phase consistent calibration factor with the amplitude fluctuation coefficient of the anti-interference reconstructed signal to generate a dynamic eye diagram quality threshold, calculating the deviation rate and adjusting the eye diagram quality management coefficient to accurately capture the interference residue in high-frequency scenarios and avoid misjudgment of crosstalk distortion as regular fluctuations.

[0058] In one case of the present embodiment, the time-varying trajectory of the eye diagram quality management coefficient is continuously tracked, and its stability is calculated. The stability is compared with the stability management boundary to obtain a management instruction, including: The eye diagram quality management coefficient is continuously collected at a preset time interval, the value change trajectory is recorded, and the fluctuation amplitude and frequency of the eye diagram quality management coefficient per unit time are calculated in combination with the eye diagram interference correlation factor to generate a coefficient time sequence fluctuation degree, specifically including: setting a 5ms preset interval, collecting the eye diagram quality management coefficient once every 5ms, continuously collecting 20 times, and recording the time sequence trajectory of 20 eye diagram quality management coefficients; the maximum value in the time sequence trajectory is divided by the minimum value to obtain an amplitude ratio; the absolute value of the difference between the eye diagram quality management coefficients collected adjacent twice in the time sequence trajectory is counted, and when the absolute value exceeds 0.05, it is recorded once. The number of all records meeting the conditions is taken as the fluctuation frequency; the amplitude ratio and the fluctuation frequency are multiplied by the eye diagram interference correlation factor respectively, added and calculated, and the mean value is normalized to 0-1, which is the coefficient time sequence fluctuation degree.

[0059] According to the coefficient time sequence fluctuation degree and the amplitude and phase consistent calibration factor, a fluctuation index is calculated, and a dynamic stability evaluation value is generated, specifically including: multiplying the coefficient time sequence fluctuation degree by (1 minus the amplitude and phase consistent calibration factor) to obtain a modified actual fluctuation index; subtracting the actual fluctuation index from 1, and then normalizing the result to 0-1, which is the dynamic stability evaluation value.

[0060] Based on the dynamic eye diagram quality threshold, a stable boundary interval adapted to the current anti-interference scene is determined, and an adaptive stable boundary value is generated, specifically including: taking the dynamic eye diagram quality threshold as a reference, the lower limit is taken as the threshold minus 0.05, and the upper limit is taken as the threshold plus 0.05, forming a stable boundary interval; the upper limit value and the lower limit value of the stable boundary interval are added and divided by 2 to obtain a median value, which is normalized to 0-1, which is the adaptive stable boundary value.

[0061] The management instruction is specifically: if the dynamic stability evaluation value is greater than or equal to the adaptive stability boundary value, it indicates that the current signal quality is stable, and a continue management instruction is generated; if the dynamic stability evaluation value is less than the adaptive stability boundary value, it indicates that the current signal quality is unstable, and an alarm instruction is generated.

[0062] The dynamic stability evaluation value is obtained by continuously collecting the eye diagram quality management coefficient 20 times at a preset interval of 5 ms, generating the coefficient time sequence fluctuation degree through the amplitude ratio and the fluctuation frequency in combination with the eye diagram interference correlation factor, and then correcting the amplitude and phase consistent calibration factor. The stable boundary interval is constructed based on the dynamic eye diagram quality threshold value, and the adaptive stability boundary value is determined to capture the signal instability caused by strong electromagnetic field coupling and avoid misjudgment as single-channel conventional fluctuation. The management precision and effectiveness of signal crosstalk control are improved in the adaptive DDR5 signal scenario, so as to generate a continue management instruction or an alarm instruction.

[0063] In one case of the embodiment, an intelligent management system for a printed circuit board signal control process is applied to the intelligent management method for the printed circuit board signal control process, and includes: The analysis unit is configured to acquire noise data of a target channel and all adjacent channels of a target management object in real time, analyze the noise data of all channels, and generate a frequency domain cooperative management domain. The target management object is a server printed circuit board. The energy management unit is configured to acquire transient currents of all channels in the frequency domain cooperative management domain in real time, map the transient currents of the channels into an N*N channel energy correlation matrix, construct a 2-layer graph convolution network based on a graph neural network, take the channels as nodes and the element of the channel energy correlation matrix as edge weight, and generate a target channel energy influence coefficient. The fusion management unit is configured to acquire impedance data of all channels in the frequency domain cooperative management domain in real time, fuse the target channel energy influence coefficient and the impedance data through a neural network, and generate a domain energy management factor representing the energy influence degree of the target channel. The load management unit is configured to acquire load data of all channels in the frequency domain cooperative management domain in real time, collect T windows with 100 ms as a collection window, and arrange the load data into time sequence data. The time sequence data and the energy management factor are input into a 2-layer hidden layer bidirectional long short-term memory network for analysis, and a pre-signal management vector containing amplitude and phase is output through a full connection layer. The reconstruction unit is configured to perform waveform remodeling on the original signal of the target channel according to the pre-signal management vector, and generate an anti-interference reconstructed signal. The extraction unit is used for eye diagram analysis on the reconstructed signal against disturbance by a convolutional neural network, adopts a 5-layer CNN with 2 layers of maximum pooling layer to extract key feature indicators, and compares the key feature indicators with eye diagram quality thresholds to generate an eye diagram quality management coefficient; The signal management unit is used for continuously tracking a time-varying trajectory of the eye diagram quality management coefficient, calculating a stability of the eye diagram quality management coefficient, comparing the stability with a stability management boundary, and obtaining a management instruction.

[0064] Although the embodiments of the present application have been shown and described, it is to be understood that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the present application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent management of signal control processes on printed circuit boards, characterized in that, include: Step S1: Real-time acquisition of noise data of the target channel and all adjacent channels of the target management object; analysis of noise data of all channels to generate a frequency domain collaborative management domain; the target management object is the server printed circuit board. Step S2: Real-time acquisition of transient currents of all channels in the frequency domain collaborative management domain; mapping the transient currents of each channel to an N×N channel energy correlation matrix; constructing a 2-layer graph convolutional network based on a graph neural network; using channels as nodes and the elements of the channel energy correlation matrix as edge weights to generate the target channel energy influence coefficient. Step S3: Acquire impedance data of all channels in the frequency domain collaborative management domain in real time, and fuse the target channel energy influence coefficient with the impedance data through a neural network to generate an energy management factor in the domain that represents the degree of energy influence on the target channel. Step S4: Real-time acquisition of load data for all channels within the frequency domain collaborative management domain. The load data is collected in T consecutive windows of 100ms each, and then organized into time series data. By constructing a two-layer bidirectional long short-term memory network with hidden layers, the time series data and energy management factors are input into the bidirectional long short-term memory network for analysis. The pre-signal management vector containing amplitude and phase is output through a fully connected layer. Step S5: Based on the pre-signal management vector, reshape the waveform of the original signal of the target channel to generate an anti-interference reconstruction signal; Step S6: Eye diagram analysis is performed on the anti-interference reconstruction signal using a convolutional neural network. A 5-layer CNN with 2 max pooling layers is used to extract key feature indicators and compare them with the eye diagram quality threshold to generate eye diagram quality management coefficients. Step S7: Continuously track the time-varying trajectory of the eye diagram quality management coefficient, calculate its stability, compare the stability with the stability management boundary, and derive management instructions.

2. The intelligent management method for printed circuit board signal control process according to claim 1, characterized in that, Noise data from all channels is analyzed to generate a frequency domain collaborative management domain, including: The noise data is analyzed to calculate the amplitude abrupt change synchronization rate and phase drift time difference between the noise of each adjacent channel and the target channel, and the noise dynamic coupling degree is generated. The system acquires the operating ambient temperature of the target management object and the number of signal reflections in each adjacent channel cabling in real time. Based on the noise dynamic coupling degree, and based on the operating ambient temperature and the number of signal reflections, it calculates the superposition value of the dielectric loss increment and reflection loss during the noise transmission process from each adjacent channel to the target channel, and generates the noise time-varying transmission loss rate. Analyze the amplitude fluctuation period of noise in each adjacent channel, and perform correlation analysis between the noise amplitude fluctuation period corresponding to each frequency point and the current signal transmission period of the target channel to obtain the period matching degree of each frequency point; The effective frequency threshold of the resonant interference is obtained by calculating the periodic fit and the time-varying noise transmission loss rate. Based on the effective frequency threshold of resonant interference, the dynamic coupling degree of noise and the time-varying transmission loss rate of noise are analyzed to obtain the frequency domain collaborative management domain.

3. The intelligent management method for signal control process of printed circuit boards according to claim 1, characterized in that, The transient currents of each channel are mapped to an N×N inter-channel energy correlation matrix. A two-layer graph convolutional network is constructed based on a graph neural network, using channels as nodes and elements of the inter-channel energy correlation matrix as edge weights to generate the target channel energy influence coefficient, including: Real-time acquisition of current amplitude values ​​for all channels within the frequency domain collaborative management domain; Calculate the frequency of occurrence of transient current peak value and the real-time change of current amplitude for each channel per unit time, and fuse the occurrence frequency and real-time change amplitude to generate transient current characteristic value; Real-time acquisition of wiring length of each channel on the target management object and dielectric filling thickness between adjacent channel wiring, combined with current transient characteristic value, to calculate the loss ratio of current radiation energy when it is transferred in the medium, and generate energy cross-channel loss coefficient between the two channels. The total number of channels in the frequency domain collaborative management domain is set to N. An N×N matrix is ​​constructed with N as the dimension. The energy transfer coefficient between each pair of channels and the product of the transient current of the two channels are used as the element value of the corresponding position in the matrix to generate the energy correlation matrix between channels. The elements of the energy correlation matrix between channels are the initial edge weights, which are used to reflect the degree of correlation of the radiated energy between channels. Based on a graph neural network, a two-layer graph convolutional network structure is used, with each channel as a node. The correlation degree of channel signal frequency between each node and other nodes is calculated, and then the correlation degree of channel signal frequency frequency is multiplied by the corresponding elements of the energy correlation matrix between channels to generate a correction factor. The initial edge weights are corrected based on the correction factor to obtain the corrected node association weights; The node association weights are calculated based on a 2-layer graph convolutional network to obtain the feature vectors of each node. The correlation degree between the feature vector of the target channel node and the feature vectors of all other nodes is calculated. Combined with the transient current feature values, a collaborative analysis is performed to generate the inter-channel energy management coefficient. The inter-channel energy management coefficients of all adjacent channels to the target channel are fused with the energy cross-channel transfer loss coefficients of the corresponding channels to generate the energy influence coefficient of the target channel.

4. The intelligent management method for signal control process of printed circuit boards according to claim 3, characterized in that, The target channel energy influence coefficient and impedance data are fused using a neural network to generate an in-domain energy management factor representing the degree of energy influence on the target channel, including: Based on the impedance data of all channels, the impedance values ​​are divided into three dynamic ranges: high impedance, medium impedance, and low impedance. The switching frequency and weighted fluctuation amplitude of the impedance in different dynamic ranges per unit time are calculated. The switching frequency and weighted fluctuation amplitude are then fused to generate impedance dynamic classification coefficients. The target channel energy influence coefficient and impedance dynamic classification coefficient are correlated and calculated to generate an energy-impedance matching attenuation coefficient. The dynamic impedance grading coefficient and the dynamic range of impedance are analyzed to generate a dynamic energy tolerance benchmark value. The difference between the energy-impedance adaptation attenuation coefficient and the dynamic tolerance benchmark value is calculated to generate a dynamic fusion characteristic value. The dynamic fusion feature value is calculated based on a lightweight neural network, and the energy-impedance adaptation attenuation coefficient is combined to calculate the in-domain energy management factor.

5. The intelligent management method for printed circuit board signal control process according to claim 4, characterized in that, The load data was collected in 100ms increments for T consecutive windows and then processed into time-series data. A two-layer bidirectional long short-term memory (LSTM) network was constructed, and the time-series data and energy management factors were input into the LSM network for analysis. A fully connected layer outputs a pre-signal management vector containing amplitude and phase, including: For the load data of all channels in the frequency domain collaborative management domain, the perturbation intensity of the signal amplitude and phase is classified according to the perturbation intensity of the perturbation data. The weight of the abrupt change of each level of feature is calculated to generate the load perturbation classification weight. Based on a 100ms acquisition window, the window duration is dynamically adjusted in conjunction with the domain energy management factor to generate a dynamic window configuration set that includes the adjusted window duration, the threshold for mutation features within the window, and the window energy correlation weight. Load data for T consecutive windows is collected based on the adjusted window duration. Effective load features for each window are filtered based on the threshold of mutation features within the window. The effective load features are perturbation-weighted according to the load perturbation classification weight and concatenated in the order of collection time. The window energy association weight corresponding to each window is labeled to generate a load energy association sequence.

6. The intelligent management method for signal control process of printed circuit boards according to claim 5, characterized in that, The load data was collected in 100ms increments for T consecutive windows and then processed into time-series data. A two-layer bidirectional long short-term memory (LSTM) network was constructed, and the time-series data and energy management factors were input into the LSM network for analysis. A fully connected layer outputs a pre-signal management vector containing amplitude and phase, which also includes: Based on the load energy correlation sequence, the overlap of strong perturbation features between adjacent windows and the rate of change of window energy correlation weights in adjacent windows are analyzed. The product of overlap and rate of change is used as the temporal correlation correction weight. Based on the temporal correlation correction weight, the correlation strength of adjacent window data in the load energy correlation sequence is adjusted to generate a temporal correlation enhancement vector. The temporal correlation enhancement vector is analyzed based on a two-layer bidirectional long short-term memory network. The load perturbation hierarchical weights are used as the attention weights of the intermediate layer of the bidirectional long short-term memory network to obtain a bidirectional temporal attention matrix representing temporal dependence and feature attention. The bidirectional temporal attention matrix is ​​input into the fully connected layer, and the output dimension is adjusted by combining the window energy correlation weights in the window dynamic configuration set to obtain the amplitude and phase initial vector. Based on time-series correlation enhancement vector analysis, the load mutation affects the signal amplitude adjustment requirement and the phase compensation requirement. The amplitude compensation accuracy and phase compensation accuracy are calculated, and an amplitude-phase consistency calibration factor is generated. The amplitude and phase initial vectors are calibrated based on the amplitude-phase consistency calibration factor to generate a pre-signal management vector containing amplitude and phase.

7. The intelligent management method for signal control process of printed circuit boards according to claim 6, characterized in that, Based on the pre-signal management vector, the original signal of the target channel is reshaped to generate an anti-interference reconstruction signal, including: The instantaneous fluctuation characteristics of the original signal of the target channel are acquired in real time, and the synchronous changes of the instantaneous fluctuation characteristics of the original signal and the time-series correlation enhancement vector are calculated to generate the signal fluctuation-load correlation degree. The amplitude and phase dynamic compensation coefficients are generated by combining the signal fluctuation-load correlation degree and the anticipatory signal management vector. Based on the effective frequency threshold of resonant interference, the degree of overlap between each frequency band in the original signal and the effective frequency of resonant interference is analyzed. Combined with the signal fluctuation-load correlation, interference-signal dual-band suppression weights are generated. The amplitude and phase of the original signal are corrected based on the amplitude and phase dynamic compensation coefficient. Then, the frequency bands in the original signal that overlap with the effective frequency points of the resonant interference are attenuated according to the interference-signal dual-band suppression weight to generate an anti-interference reconstructed signal.

8. The intelligent management method for signal control process of printed circuit boards according to claim 7, characterized in that, Eye diagram analysis of the reconstructed signal against perturbation is performed using a convolutional neural network. A five-layer CNN with two max-pooling layers is used to extract key feature indicators, which are then compared with eye diagram quality thresholds to generate eye diagram quality management coefficients, including: Eye diagrams are plotted on the anti-interference reconstructed signal of the target channel, and the correlation between different regions in the eye diagram and interference residues is analyzed by combining interference-signal dual-band suppression weights, generating eye diagram interference correlation factors. The basic data of the eye diagram is input into a 5-layer CNN with 2 max pooling layers. The eye diagram interference correlation factor is embedded as an attention weight into the 3rd layer of the CNN, so that the CNN focuses on extracting features from the regions in the eye diagram that are highly correlated with the interference residue, and generates a weighted eye diagram feature set. The amplitude-phase consistency calibration factor and the disturbance rejection reconstruction signal are analyzed to generate a dynamic eye diagram quality threshold. The deviation rate between the weighted eye diagram feature set and the dynamic eye diagram quality threshold is calculated, and then the deviation rate is adjusted according to the eye diagram interference correlation factor to generate the eye diagram quality management coefficient.

9. The intelligent management method for signal control process of printed circuit boards according to claim 8, characterized in that, Continuously track the time-varying trajectory of the eye diagram quality management coefficient, calculate its stability, compare the stability with the stability management boundary, and derive management instructions, including: Eye diagram quality management coefficients are continuously collected at preset time intervals, their numerical change trajectory is recorded, and the fluctuation amplitude and frequency of eye diagram quality management coefficients per unit time are calculated in combination with eye diagram interference correlation factors to generate coefficient time series volatility. The volatility index is calculated based on the coefficient time-series volatility and the amplitude-phase consistency calibration factor to generate a dynamic stability assessment value. Based on the dynamic eye diagram quality threshold, a stable boundary interval suitable for the current anti-disturbance scenario is determined, and an adaptive stable boundary value is generated. The dynamic stability assessment value is compared with the adaptive stability boundary value to obtain management instructions.

10. A printed circuit board signal control process intelligent management system, applied to the printed circuit board signal control process intelligent management method according to any one of claims 1-9, characterized in that, include: The analysis unit is used to acquire noise data of the target channel and all adjacent channels of the target management object in real time, analyze the noise data of all channels, and generate a frequency domain collaborative management domain. The target management object is the server printed circuit board. The energy management unit is used to acquire the transient current of all channels in the frequency domain collaborative management domain in real time, map the transient current of each channel into an N×N channel energy correlation matrix, construct a 2-layer graph convolutional network based on graph neural network, and generate the target channel energy influence coefficient with channels as nodes and the elements of the channel energy correlation matrix as edge weights. The fusion management unit is used to acquire impedance data of all channels in the frequency domain collaborative management domain in real time, and fuse the target channel energy influence coefficient with the impedance data through a neural network to generate an intra-domain energy management factor that represents the degree of energy influence on the target channel. The load management unit is used to acquire load data of all channels in the frequency domain collaborative management domain in real time. The load data is acquired in 100ms increments for T consecutive windows and then organized into time series data. The time series data and energy management factors are input into the bidirectional long short-term memory network with two hidden layers for analysis. The pre-signal management vector containing amplitude and phase is output through the fully connected layer. The reconstruction unit is used to reshape the waveform of the original signal of the target channel according to the pre-signal management vector to generate an anti-interference reconstruction signal. The extraction unit is used to perform eye diagram analysis on the anti-interference reconstructed signal through a convolutional neural network. It uses a 5-layer CNN with 2 max pooling layers to extract key feature indicators and compares them with the eye diagram quality threshold to generate eye diagram quality management coefficients. The signal management unit is used to continuously track the time-varying trajectory of the eye diagram quality management coefficient, calculate its stability, compare the stability with the stability management boundary, and derive management instructions.

Images