Modeling method of ICIM-CI model based on particle swarm optimization

Abstract

The present application relates to the ICIM-CI model modeling field, specifically relates to a kind of ICIM-CI model modeling method based on particle swarm optimization.The modeling method mainly includes input signal segmentation based on particle swarm optimization;Output signal recombination;Segmented Volterra kernel function identification.The present application effectively describes the nonlinear behavior of chip by segmented Volterra series, and optimizes the segmented threshold by using particle swarm optimization algorithm, which significantly improves the accuracy and versatility of the model.The present application overcomes the problems of traditional ICIM-CI model in the modeling process, such as linear hypothesis, difficulty in accurately describing the nonlinear response of MADIC under electromagnetic interference, and lack of sensitivity quantitative analysis, output quantitative information and parameterized cascade simulation under multiple criteria synthesis.The present application is suitable for ICIM-CI model modeling.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of ICIM-CI (Integrated Circuit Immunity Model-Conducted Immunity) model modeling, and particularly relates to an ICIM-CI model modeling method based on particle swarm optimization. BACKGROUND

[0002] Integrated circuits are the core of electronic devices and systems, and the electromagnetic immunity of the integrated circuits directly affects the electromagnetic compatibility of the entire device and the safety of the system operation. Analog and digital mixed chips are integrated circuits that have both nonlinear and linear characteristics, and are very sensitive to electromagnetic interference. Small changes in circuit topology can also cause significant changes in the electromagnetic immunity of the device or the entire system. Therefore, predicting the impact of analog and digital mixed chips on the operation under radio frequency interference and establishing an accurate chip conducted susceptibility model have become key electromagnetic compatibility issues.

[0003] However, the accurate conducted immunity characteristics of the chip cannot be obtained from the simulation function model provided by the chip manufacturer, because the function model only focuses on the function and logic of the chip and does not include the response characteristics and stability of the chip under external electromagnetic interference and other immunity characteristics.

[0004] If the conducted immunity of the chip is not accurately modeled, the sensitivity of the circuit cannot be correctly defined, which will affect the electromagnetic compatibility design and circuit layout of the circuit. To solve this problem, it is urgent to study the chip conducted susceptibility modeling method, especially the "black box modeling". Because the "black box modeling" does not require knowledge of the circuit structure or design of the chip, it only needs to model the input and output characteristics of the chip being modeled, so the chip user does not need to understand the internal structure of the circuit to establish the model. That is, the chip user can obtain the model through the test results without understanding the internal circuit structure of the chip.

[0005] In previous studies, black box models such as statistical models, envelope domain models, and artificial neural networks, etc. were used, but these models are more concerned with the failure of the chip at a certain frequency point, but there are some problems and deficiencies in the wide frequency band, model universality, etc.

[0006] In recent years, ICIM-CI model based on PDN (Passive Distribution Network) module and IB (Immunity Behavior) module has obvious advantages in circuit simulation, which can provide a chip broadband anti-interference simulation model. However, since the model needs linear assumption when building IB, which is contrary to the nonlinear effect of analog-digital hybrid chip after suffering electromagnetic interference, which also leads to the difference between the model and the actual measurement results. SUMMARY

[0007] The present application aims to overcome the shortcomings of the prior art, and provides an ICIM-CI model modeling method based on particle swarm optimization, which can provide accurate nonlinear output response.

[0008] The present application adopts the following technical solutions to achieve the above-mentioned purposes, and provides an ICIM-CI model modeling method based on particle swarm optimization, which comprises:

[0009] The IB module of the chip is established by introducing the segmented Volterra series, and the particle swarm optimization algorithm is used to reduce the order of the segmented Volterra series, specifically including:

[0010] S1, input signal segmentation based on particle swarm optimization:

[0011] S2, output signal recombination;

[0012] S3, segmented Volterra kernel function identification.

[0013] Further, step S1 specifically includes:

[0014] The input signal is decomposed according to the power size, and the threshold value of the input power is confirmed, and the threshold value is expressed as follows:

[0015] Γ={λ1,λ2,…,λ s},λ1<λ2<…<λ s ;

[0016] Where λ i ,i=1,2,…,s is the threshold value of the input signal power amplitude, and s is the number of set threshold values;

[0017] The input signal x(t) is decomposed into s+1 sub-signals according to the following formula after the set threshold value:

[0018]

[0019] The decomposed input sub-signals are sampled in discrete time, and the input sub-signals are identical in time dimension, and the sampling point number is n, and the decomposed input sub-signals are represented by a (s+1)×n matrix:

[0020]

[0021] In the decomposition process, the initial value of the particle in the optimization space is randomly set according to the target function space, the individual extreme value p best and the group extreme value g best are updated, and the optimal value in the target function space is searched:

[0022] Let p num be the total number of particles, D be the particle dimension, when the maximum number of iterations max iter is reached, the iteration is stopped, v a (t) is the iteration speed of the a-th particle at t, x a (t) is the position of the a-th particle in the target function space at t, pbest a (t) is the individual extreme value of the a-th particle at t.

[0023] v a (t)=[v a1 (t),v a2 (t),…,v aD (t)] T , x a (t)=[x a1 (t),x a2 (t),…,x aD (t)] T ;

[0024] The updating method of the particle in the target function space is as follows:

[0025] v a+1 (t+1)=ωv a (t)+c1a1(pbest a (t)-x a (t))+c2a2(gbest a (t)-x a (t));

[0026] x a+1 (t+1)=x a (t)+v a+1 (t+1);

[0027] Wherein, ω is an inertia weight coefficient, representing the influence of the last iteration on the present iteration. c1 is an individual learning factor, and c2 is a group learning factor. a1 and a2 are random numbers.

[0028] Further, step S2 specifically comprises:

[0029] After completing the power threshold segmentation, the output is weighted combined to obtain a complete output signal;

[0030] The output signal corresponding to each segment of the input sub-signal is as follows:

[0031]

[0032] In the formula, y i (·) is the output sub-signal corresponding to the i-th segment of the input sub-signal, K i is the order of the Volterra series corresponding to the i-th segment of the input sub-signal, M i is the memory length of the i-th segment of the input sub-signal;

[0033] The output signal is obtained by comprehensively processing each segment of the output sub-signal as follows:

[0034]

[0035] Further, step S3 specifically comprises:

[0036] The segmented kernel function adopts the following formula to predict the output waveform:

[0037]

[0038] Where X pre_i , i = 1, 2, …, s + 1 is the input matrix obtained by multiplying the i-th input sub-signal and the memory time delay, is the segmented kernel function.

[0039] The beneficial effects of the present application are:

[0040] The present application effectively describes the nonlinear behavior of the chip through the segmented Volterra series, and optimizes the segmented threshold value by using the PSO algorithm, which significantly improves the precision and universality of the model. The model overcomes the problems of the traditional ICIM-CI model, such as linear assumption in the modeling process, difficulty in accurately describing the nonlinear response of the MADIC (Mixed Analog-Digital Integrated Circuit) under electromagnetic interference, and insufficient sensitivity quantitative analysis, output quantization information and parameterized cascade simulation under multiple criteria synthesis. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 is the ICIM-CI model modeling method flowchart based on particle swarm optimization provided by the embodiment of the present application;

[0042] Figure 2 is a schematic diagram of an ICIM-CI model structure provided by an embodiment of the present application.

[0043] Figure 3 is an optimization flowchart of a particle swarm optimization algorithm provided by an embodiment of the present application.

[0044] Figure 4 is a normalized mean square error convergence curve diagram of particle swarm optimization provided by an embodiment of the present application.

[0045] Figure 5 is a comparison curve diagram of sensitivity prediction and actual output result in the case of output waveform offset 5% or off provided by an embodiment of the present application.

[0046] Figure 6 is a comparison curve diagram of sensitivity prediction and actual output result in the case of output waveform offset 10% or off provided by an embodiment of the present application.

[0047] Figure 7 is a comparison curve diagram of model output and actual output result when the chip is normally working (interference at 90MHz injection-10dBm) provided by an embodiment of the present application.

[0048] Figure 8 is a comparison curve diagram of model output and actual output when the chip output appears slight disturbance (interference at 90MHz injection 10dBm) provided by an embodiment of the present application.

[0049] Figure 9 is a comparison curve diagram of model output and actual output result when the chip output appears serious disturbance (interference at 90MHz injection 25dBm) provided by an embodiment of the present application.

[0050] Figure 10 is a comparison curve diagram of model output and actual output result when the chip output is seriously disturbed and the waveform is seriously distorted (interference at 900MHz injection 25dBm) provided by an embodiment of the present application.

[0051] Figure 11 is a comparison curve diagram of sensitivity prediction result and actual output result of the ICIM-CI (PSVIB) modeling method provided by an embodiment of the present application. DETAILED DESCRIPTION

[0052] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application.

[0053] The ICIM-CI model of the present application is composed of a PDN module and an IB module, and the model structure is as followsFigure 2 are shown.

[0054] The PDN describes the linear behavior of the chip, the IB describes the behavior of the chip in-band and out-of-band frequency response, and the information output by the monitoring port describes the response of the chip to the interference signal applied to the disturbance injection port. The IB contains parameters such as frequency, transmission power, and changes in the monitoring port. IB data is obtained by using pass / fail or fail sensitivity test criteria.

[0055] There are two modeling methods for IB. The first method is to directly obtain IB binary data using go-no-go sensitivity test criteria, and the second method is to extract fault detection data using fail or fail test and then use sensitivity test criteria to predict the pass or fail result of the device. Compared with the first method, the second method can provide output data such as power, voltage, etc. by introducing additional monitoring and analysis steps to predict faults, and then applying sensitivity criteria to evaluate whether these data meet the fault condition. Since additional analysis steps are introduced, the number of tests is indirectly reduced. At the same time, it provides a more flexible means for fault analysis, which is the same as the improvement direction of the present application.

[0056] When using the second method, a linear fitting equation needs to be established to obtain fault detection simulation data. That is, at each frequency point, record the interference injection power / voltage, frequency, and sensitivity performance index parameters, and establish a linear fitting equation. The model finally represents a linearized sensitivity prediction model expressed in a netlist form, and does not include the output data of the monitoring port quantization. However, in fact, when a mixed analog-digital chip shows sensitivity, it usually shows nonlinearity, which is different from the linear assumption of the model. This leads to the difference between the prediction result of the model and the actual measurement result.

[0057] In order to improve the nonlinearity of the ICIM-CI model and increase the quantized waveform of the model output for immunity analysis, the present application introduces a segmented Volterra series to establish the IB module of the chip, and uses a particle swarm optimization algorithm to reduce the order of the segmented Volterra series, further improving the accuracy of the model.

[0058] The Volterra series is a functional series used to describe nonlinear systems, which is based on the input signal and the output signal to establish a nonlinear system. The Volterra time domain kernel and the frequency domain kernel both have corresponding physical meanings. The time domain kernel corresponds to the high-order impulse response function of the system, and the frequency domain kernel corresponds to the generalized frequency response function of the system. When the Volterra series is used to describe a nonlinear system, it can be described in the following form:

[0059]

[0060] where x(t), y(t)∈R, x(t) is the input signal of the system, y(t) is the output signal of the system, h n (τ1,τ2,…,τ n ) is the nth order Volterra time domain kernel corresponding to the described nonlinear system. The above is subjected to multidimensional Fourier transformation, and the solution in the frequency domain can also be obtained.

[0061] When actually using the Volterra series to describe a nonlinear system, a discrete model is often used, and the order and memory length are truncated, and the order is N and the memory length is M. The nonlinear system is described, and the corresponding Volterra series model is as follows:

[0062]

[0063] Where g(m1)… is the kernel function obtained by sampling the time domain kernel with time T, and its specific representation is as follows:

[0064] g(m1)=h1(m1T)·T

[0065] g(m1,m2)=h2(m1T,m2T)·T 2

[0066] g(m1,m2,…,m N )=h N (m1T,m2T,…,m N T)·T N

[0067] For a strong nonlinear system, because a higher order and memory length need to be used for fitting, the number of Volterra kernel functions will increase exponentially. To solve this problem, the present application proposes a piecewise Volterra series method applied in the modeling process of the memristor, which can realize the order reduction of the kernel function by using the piecewise Volterra series. For the application of chip conduction sensitivity modeling problem, the input power can be segmented, each segmented signal corresponds to a low-order Volterra kernel function, and then the total output signal is obtained by synthesizing and weighting the output signals. This processing method not only reduces the complexity of the model, but also maintains high modeling accuracy. The piecewise Volterra series modeling process mainly includes: input signal segmentation, output signal recombination, and identification of piecewise Volterra kernel function.

[0068] However, the segmented Volterra series modeling method does not have a concentrated definition for the selection of the power threshold. Generally, it is necessary to constantly try and blindly set the power threshold according to the input waveform, which can cause the increase of system error, thereby reducing the universality of the model. Therefore, in order to solve this problem, the power threshold is intelligently segmented by using a particle swarm optimization algorithm to optimize the segmented Volterra series, so as to ensure the accuracy in the process of reducing the order of the segmented Volterra and further improve the universality and repeatability of the method in the modeling process of the IB module, thereby being more suitable for quickly building the IB module.

[0069] As Figure 1 shown, the specific technical solutions are as follows.

[0070] Segmentation of the input signal based on the particle swarm optimization:

[0071] The input signal is decomposed according to the power size, and the threshold value of the input power is confirmed, and the threshold value is expressed as follows:

[0072] Γ={λ1,λ2,…,λ s},λ1<λ2<…<λ s ;

[0073] Wherein λ i ,i=1,2,…,s is the threshold value of the input signal power amplitude, and s is the number of set threshold values;

[0074] The input signal x(t) is decomposed into s+1 sub-signals according to the following formula after the set threshold value:

[0075]

[0076] The decomposed input sub-signals are discretely sampled, and the input sub-signals are the same in the time dimension. Assuming that the number of sampling points is n, the decomposed input sub-signals are represented by a (s+1)×n matrix:

[0077]

[0078] In the decomposition process, the initial value of the particle in the optimization space is randomly set according to the target function space, and the individual extreme value p best and the group extreme value g best are updated to find the optimal value in the target function space:

[0079] Assuming that p num is the total number of particles, D is the dimension of the particle, the iteration is stopped when the maximum number of iterations max iter is reached, v a (t) is the iteration speed of the a-th particle at t, and x a(t) is the position of the a-th particle in the objective function space at time t, pbest a (t) is the individual extremum of the a-th particle at time t;

[0080] v a (t) = [v a1 (t), v a2 (t), …, v aD (t)] T , x a (t) = [x a1 (t), x a2 (t), …, x aD (t)] T ;

[0081] pbest a (t) = [p a1 (t), p a2 (t), …, p aD (t)] T ;

[0082] gbest = [g1, g2, …, g d ] T ;

[0083] The updating mode of the particle in the objective function space is shown in the following formula:

[0084] v a+1 (t+1) = ωv a (t) + c1a1(pbest a (t) - x a (t)) + c2a2(gbest a (t) - x a (t));

[0085] x a+1 (t+1) = x a (t) + v a+1 (t+1);

[0086] wherein ω is an inertia weight coefficient, representing the influence of the last iteration on the present iteration. c1 is an individual learning factor, and c2 is a group learning factor. a1 and a2 are random numbers.

[0087] According to the above analysis, the optimization process of the particle swarm optimization algorithm is shown in Figure 3 .

[0088] S1, set the objective function space for the particle swarm optimization algorithm;

[0089] S2, initialize parameters such as population size and inertia weight;

[0090] S3, calculating the function value of each particle;

[0091] S4, updating the individual optimal value and the global optimal value;

[0092] S5, updating the speed and position of the particle;

[0093] S6, recalculating the function value of each particle;

[0094] S7, setting the iteration termination condition, if the termination condition is not reached, returning to step S5, if the termination condition is reached, outputting the optimal solution.

[0095] Output signal recombination:

[0096] After completing the power threshold segmentation, the output is weighted combined to obtain the complete output signal;

[0097] The output signal corresponding to each segment of the input sub-signal is as follows:

[0098]

[0099] In the formula, y i (·) is the output sub-signal corresponding to the i-th segment of the input sub-signal, K i is the order of the Volterra series corresponding to the i-th segment of the input sub-signal, M i is the memory length of the i-th segment of the input sub-signal;

[0100] Each segment of the output sub-signal is processed as follows to obtain the output signal:

[0101]

[0102] Segmented Volterra kernel function identification:

[0103] The segmented kernel function is used to predict the output waveform as follows:

[0104]

[0105] Where X pre_i , i = 1, 2, …, s + 1 is the input matrix obtained by multiplying the i-th input sub-signal and the memory time delay, is the segmented kernel function.

[0106] The scheme of the present application will be further described below in combination with specific embodiments.

[0107] The modeling work of SF702A AMC analog switch chip, SJ96F174MC type four-way differential driver, etc. is carried out below to verify the universality of the modeling method.

[0108] Modeling example of SF702A type analog switch:

[0109] SF702A type analog switch (Analog Switch) is selected as the test and modeling object. SF702A type analog switch is a semiconductor device for analog signal control and transmission. It is a single-pole single-throw (SPST) analog switch with wide bandwidth (3dB bandwidth greater than or equal to 110MHz), low power consumption (less than or equal to 1uA) and low on-resistance (less than or equal to 6Ω). It uses a 5-pin ceramic leadless chip carrier (LCC05B) package, supports signal switching based on control voltage level, and can provide reliable signal routing in various applications. Its basic function is to select the on-off mode between the analog input end S and the analog output end D under the control of the switch control end IN. When the IN port is low, the S port and the D port are on; when the IN port is high, the S port and the D port are off. The symbol table of the corresponding lead port is shown in Table 6. According to the corresponding port relationship and basic function, a conduction sensitive modeling test board is designed.

[0110] Table 6 SF702A analog switch lead port function symbol table

[0111] Lead number Symbol Function Lead number Symbol Function 1 D Analog output 4 IN Switch control 2 S Analog input 5 <![CDATA[V DD ]]> Power supply 3 GND Ground

[0112] Modeling process: according to the specific power size and port characteristics, select power meter and power amplifier and other equipment. When extracting the PSVIB module, set the normal working condition of the analog switch as the analog input frequency f = 1kHz, the peak-to-peak value V PP = 0.5V of the sine wave, and the switch control end IN port is set to low V = 0.8V. At this time, the corresponding analog switch is in the on state, the S port and the D port are on, and the analog output port should be the output frequency f = 1kHz, the peak value V PP = 0.5V of the sine wave. Inject radio frequency interference into the port that needs to be built into the model, build PSVIB using the scheme provided by the present application, and build PDN model using the active integrated device PDN vector fitting modeling method. Finally, the ICIM-CI (PSVIB) model is generated.

[0113] The verification result is that during the process of applying electromagnetic interference to all ports of the analog switch, the IN port responsible for controlling the switch state is the most sensitive to EMI. Therefore, the IN port is taken as an example for detailed analysis. When a sinusoidal interference f = 90 MHz is injected into the analog switch, when the interference power reaches P = 10 dBm, at this time, the output of the analog switch has obvious waveform distortion. This shows that the chip becomes sensitive to the interference signal at 10 dBm. Since the traditional ICIM-CI modeling method cannot generate the output waveform in the time domain, at this point, the IB module is built by three methods (original Volterra series, segmented Volterra series, and PSVIB), and the corresponding ICIM-CI model is established, and the output waveform obtained by the actual DPI test is compared to verify the accuracy of the PSVIB method compared with other methods.

[0114] According to the requirements of the IEC62433 standard, the DPI test is carried out on the chip, and the time domain measurement results obtained are compared with the time domain results generated by the three models to verify the accuracy of the model. When the interference is injected at 90 MHz and 10 dBm frequency, the output of the chip shows obvious nonlinearity. At this time, when the IB model is built using the original Volterra series, when the order N = 3 and the storage length M = 3 are set, it captures some nonlinear behavior, but the error is still large, and the normalized mean square error (NMSE) is -9.7755 dB. After using a higher order (N = 5) and a longer memory length (M = 4), the model is improved, and the normalized mean square error is reduced to -13.4332 dB, but the error is still large, and 1362 kernel functions are needed to identify. When the IB model is built using the segmented Volterra series, the input signal is divided into intervals {-0.15, 0, 0.15}, and the order and storage length of each segment (N = 3, M = 3) are kept the same, at this time, the model significantly improves the ability to capture nonlinear behavior, and the NMSE reaches -17.7621 dB. Based on the segmented Volterra model, the optimal power threshold is automatically obtained by using the PSO algorithm optimization, and the model accuracy is further improved, and the NMSE is -19.9152 dB, at this time, the optimal solution of the power segmentation threshold is {0.2195, -0.1287, 0.2693}, and the normalized mean square error convergence of the solving process is as shown in Figure 4 The number of time domain kernel functions identified by different Volterra series models and the normalized mean square error are shown in Table 1.

[0115] Table 1 Number of time domain kernel functions identified by different Volterra series models and NMSE

[0116]

[0117]

[0118] It can be found that in the ICIM-CI model of IB module constructed by different methods, the model error of traditional Volterra series model is relatively high even if high order and long memory depth are used, and the number of time-domain kernel to be identified is also large. However, when the piecewise Volterra series is introduced, the model can more accurately describe the strong nonlinear characteristics, thereby significantly improving the model accuracy. Further, by applying particle swarm optimization to segment the power threshold, it can be observed that the model accuracy is significantly improved again.

[0119] In order to verify the accuracy of the ICIM-CI (PSVIB) model in a wide frequency range, the sensitivity curves between the measured data, the traditional ICIM-CI and the ICIM-CI (PSVIB) model proposed in this paper are compared in the frequency range of 1MHz-1GHz. In the actual test, the interference is injected into the IN port of the chip, and the highest injection power is 25dBm. In the sensitivity evaluation of the circuit board or system, the user needs to quantitatively design the electromagnetic compatibility parameters through systemization, and set a reasonable sensitivity standard. In this invention, first of all, the sensitivity curve is drawn when the output waveform shift of 5% or off is used as the sensitivity test standard (Table 3). At the same time, according to Table 2, the sensitivity prediction curve of the traditional ICIM-CI model is drawn. For the ICIM-CI (PSVIB) model that has been built, the sensitivity prediction curve under the test standard (criteria) is also drawn according to Table 2. Finally, the three curves are compared (see Figure 5 ). The results show that compared with the traditional ICIM-CI modeling method, the results obtained by ICIM-CI (PSVIB) are more close to the sensitivity curve obtained by test in a wide frequency range. Among them, using the traditional ICIM-CI modeling method, the overall NMSE of the model is-18.3521dB, and using the ICIM-CI (PSVIB) modeling method, the overall NMSE of the model is-25.6765dB, which is increased by 7.3dB. This shows that the ICIM-CI (PSVIB) method has higher accuracy in simulating the sensitivity of the chip to electromagnetic interference. In addition, when using the traditional ICIM-CI model according to the sensitivity test standard (criteria) in Table 3, four prediction netlists need to be built and judged comprehensively, which means that the traditional ICIM-CI method must perform four independent modeling tasks to obtain the final result. In contrast, ICIM-CI (PSVIB) only needs one modeling task to obtain the same result, because it can make a judgment according to the output waveform. This means that the significant reduction in the number of modeling iterations directly leads to at least 75% improvement in efficiency.

[0120] Table 2 SF702A MC analog switch sensitivity test criteria

[0121] Characteristic quantity Limit value Maximum value Greater than 0.275 V or less than 0.1 V Minimum value Less than -0.275 V or greater than -0.1 V

[0122] Table 3 SJ96F174MC type four-way differential driver lead port function symbol table

[0123]

[0124]

[0125] In order to verify the accuracy of the model generated by ICIM-CI(PSVIB) after adjusting the test criteria, the sensitivity curve of the ICIM-CI(PSVIB) model of the chip is drawn when the output waveform offset of 10% or off is used as the sensitivity test criteria (Table 4). At this time, due to the change of the criteria, the original ICIM-CI model needs to be rebuilt. Similarly, the sensitivity curve of the two models is drawn according to the new test criteria and compared with the test results, as shown in Figure 6 It can be found that the overall NMSE of the model is -11.0895 dB using the traditional ICIM-CI modeling method, and the overall NMSE of the model is -24.429 dB using the ICIM-CI(PSVIB) modeling method. ICIM-CI(PSVIB) can still maintain high prediction accuracy and does not need to redevelop the model.

[0126] Table 4 SJ96F174MC sensitivity test criteria

[0127] Characteristic quantity Limit value Maximum value Greater than or equal to 2.5 V or less than or equal to 1.5 V Minimum value Greater than or equal to -1.5 V or less than or equal to -2.5 V Average value Greater than or equal to 0.2 V or less than or equal to -0.2 V

[0128] Table 5 shows the performance comparison of the two modeling methods in different scenarios, including model error, complexity of handling multiple criteria, and whether the criteria need to be re-modeled. It can be seen that the ICIM-CI(PSVIB) model is superior to the traditional ICIM-CI model in multiple dimensions.

[0129] Firstly, from the perspective of model error, in single-point analysis, the traditional ICIM-CI model cannot output single-point waveforms and can only provide partial monitoring parameter points, while the ICIM-CI(PSVIB) model can output accurate waveforms with an error of -25.6765 dB. In wide-band conductive sensitivity prediction modeling, the error of the traditional ICIM-CI model is -18.3521 dB, while the error of the ICIM-CI(PSVIB) model is even smaller, at -25.6765 dB. This shows that the ICIM-CI(PSVIB) model can more accurately predict the impact of electromagnetic interference on chips in both single-point analysis and wider frequency ranges, and is closer to actual test results.

[0130] Secondly, in terms of complexity in handling multiple criteria, the ICIM-CI(PSVIB) model is more convenient. The ICIM-CI(PSVIB) model can directly provide detailed actual waveform output and set criteria, which directly simplifies the comprehensive processing of multiple criteria. The traditional ICIM-CI model needs to build multiple prediction models when dealing with complex criteria, which greatly increases the modeling difficulty and computational load.

[0131] Finally, the ICIM-CI(PSVIB) model has the significant advantage of not needing to be re-modeled when changing criteria. Users can directly adjust the criterion standard according to their needs without the need to rebuild the model, greatly improving flexibility and efficiency. The traditional ICIM-CI model needs to be re-modeled when changing criteria, which limits its adaptability in practical applications.

[0132] Table 5 Performance comparison of two modeling methods in different scenarios

[0133]

[0134] SJ96F174MC four-channel differential driver modeling example:

[0135] This section verifies the universality of the ICIM-CI(PSVIB) modeling method using the SJ96F174MC four-channel differential driver. SJ96F174MC is a single integrated circuit manufactured using 0.6 μm CMOS technology, which is a data transmission interface circuit for EIA-422A / 485 communication standards. Its main function is to convert input TTL logic signals into high-speed, high-drive-capability differential signals to achieve long-line transmission. The device can achieve tri-state output through the enable control terminal. Each enable terminal controls the output of two drivers, with the advantages of fast speed and strong driving capability. The device uses a 20-pin ceramic LCC package. In addition, Table 3 shows the port functions and symbols of the chip.

[0136] Modeling process: set the normal working condition as the 1st driver input 1A port input: frequency is 1 kHz, peak-to-peak value is 4V pp , offset 0V DC square wave. The power supply voltage Vcc=5V, the EN1 / 2 enable port is 3V high level. At this time, the 1Y port and the IN1 port are turned on, the 1Y port normally outputs the square wave with the frequency of 1 kHz, the peak-to-peak value of 4Vpp and the offset of 0VDC. The disturbance is injected into the EN1 / 2 enable port, and the time domain waveforms of the 1A port input and the 1Y port output are collected. The ICIM-CI(PSVIB) model is built by using the theory in the 2nd chapter.

[0137] By comparing the sensitivity curve obtained from the DPI measurement with the time domain output result of the established ICIM-CI(PSVIB) model Figure 7 - Figure 10 , it can be found that in the case of injecting different frequency and different injected power disturbances into the EN1 / 2 enable port, the time domain output result of the model generated by using the ICIM-CI(PSVIB) modeling method is highly consistent with the actual measurement result. The model can accurately provide the quantitative waveform output of the chip in the normal working condition and in the case of disturbance. Even if the output waveform is severely disturbed (see Figure 9 and Figure 10 , the output result of the model can still be accurately consistent with the measurement result. This shows that the model has good accuracy and universality in the time domain output result under different disturbance frequencies and different disturbance injection powers.

[0138] When the sensitivity test standard of the chip is set according to the circuit or system design requirements, the ICIM-CI(PSVIB) model is taken as an example, the sensitivity prediction curve of the ICIM-CI(PSVIB) model in the frequency range of 10MHz-1GHz is drawn (see Figure 11 ) and compared with the actual test sensitivity curve. It can be seen that the ICIM-CI(PSVIB) method proposed in this paper still performs well in the wide frequency range, which shows that the model has high accuracy and good universality in the frequency domain. Due to the same memory length setting, errors occur in the high frequency band. Subsequently, the automatic adjustment method of the memory length will be carried out to improve this problem.

[0139] The ICIM-CI(PSVIB) modeling method exhibits good accuracy and universality in both time domain and frequency domain by combining the time domain output results and the frequency domain wideband prediction results. Under high frequency band and high power injection conditions, there is a slight difference between the model and the measurement results. This is due to the increase in high frequency noise level, which requires a longer memory length to accurately characterize, while the uniform memory length setting is adopted to maintain the modeling efficiency, thereby introducing modeling errors. Currently, efforts are being made to optimize the program and the model, and to explore advanced methods to further improve the overall modeling accuracy of the wideband.

[0140] In summary, the ICIM-CI model IB module is improved by introducing a particle swarm optimization-based slice Volterra series improved black box modeling method, ICIM-CI(PSVIB). This method uses the particle swarm optimization slice Volterra series to improve the accuracy of the output waveform and the nonlinearity of the model, and significantly improves the modeling efficiency under multiple sensitivity criteria. The improved model effectively solves the limitations of the traditional ICIM-CI model in describing the nonlinear response of the MADIC under electromagnetic interference (EMI), and improves the ability of the model in sensitivity quantitative analysis, output quantization information and parameterized cascade simulation under multiple sensitivity criteria. Through experiments on analog switches and four-way differential drivers, the results show that the ICIM-CI(PSVIB) model has significantly improved in modeling accuracy, universality and modeling efficiency, with a wideband model accuracy improvement of 7.3dB and a modeling efficiency improvement of 75%.

[0141] The above description is only the preferred embodiment of the present application, and it should be understood that the present application is not limited to the form disclosed herein, and should not be considered as excluding other embodiments, but can be used in various other combinations, modifications and environments, and can be modified within the scope of the concept described herein, by the above teaching or related art or knowledge. Any modification and change made by those skilled in the art without departing from the spirit and scope of the present application shall be within the protection scope of the appended claims of the present application.

Claims

1. The ICIM-CI modeling method based on particle swarm optimization is characterized by: The modeling method comprises: The chip's IB module is established by introducing the piecewise Volterra series and using the particle swarm optimization algorithm to reduce the order of the piecewise Volterra series. Specifically, the following steps are performed: S1, input signal segmentation based on particle swarm optimization; The input signal is decomposed according to the power size to determine the input power threshold. The threshold is expressed as follows: C={λ1,λ2,…,λ s },λ1<λ2<…<λ s ; where λ i ,i=1,2,…,s is the threshold of the input signal power amplitude, and s is the number of thresholds set; The input signal x(t) passes through the set threshold and is decomposed into s+1 sub-signals according to the following formula: The decomposed input sub-signals are sampled in discrete time. The input sub-signals have the same time dimension. Assuming the number of sampling points is n, the decomposed input sub-signals are represented by a (s+1)×n matrix: In the decomposition process, the initial value of the particle in the optimization space is randomly set according to the objective function space, and the particle initial value is set according to the individual extreme value p. best and the group extreme value g best Update its own speed and position, and find the optimal value in the objective function space: Let p num is the total number of particles, D is the particle dimension, when the maximum number of iterations max is reached iter Stop iteration when v a (t) is the iteration speed of the a-th particle at time t, x a (t) is the position of the ath particle in the objective function space at time t, pbest a (t) is the individual extreme value of the a-th particle at time t; v a (t)=[v a1 (t),v a2 (t),…,v aD (t)] T ,x a (t)=[x a1 (t),x a2 (t),…,x aD (t)] T ; The update method of particles in the objective function space is as follows: v a+1 (t+1)=ωv a (t)+c1a1(pbest a (t)-x a (t))+c2a2(gbest a (t)-x a (t)); x a+1 (t+1)=x a (t)+v a+1 (t+1); Among them, ω is the inertia weight coefficient, which represents the influence of the previous iteration on the current iteration, c1 is the individual learning factor, c2 is the group learning factor, and a1 and a2 are random numbers; S2, output signal recombined; After completing the power threshold segmentation, the output is weighted and combined to obtain the complete output signal; The output signal corresponding to each input sub-signal is shown in the following formula: Where y i (·) is the output sub-signal corresponding to the input sub-signal of the i-th segment, K i is the order of the Volterra series corresponding to the input sub-signal of the i-th segment, M i The memory length of the input sub-signal for the i-th segment; Each output sub-signal is processed by the following formula to obtain the output signal: S3, piecewise Volterra kernel function identification; The piecewise kernel function uses the following formula to predict the output waveform: where X pre_i ,i=1,2,…,s+1 is the input matrix obtained by multiplying the i-th input sub-signal and the memory delay, is a piecewise kernel function.

Images