An embedded electromagnetic pulse precursor early warning and characteristic code protection method and system

Abstract

The application discloses an embedded electromagnetic pulse precursor early warning and characteristic code protection method and system, belongs to the technical field of electromagnetic compatibility and active protection, and comprises the following steps: extracting the time-frequency characteristics of a precursor wave signal, predicting threat parameters of a main pulse according to a mapping relationship determined according to a time-frequency coupling physical model; extracting multi-dimensional physical characteristics of a system damage response and compressing and encoding the multi-dimensional physical characteristics into electromagnetic damage characteristic codes, so as to search a knowledge base to obtain protection strategy parameters; executing the protection strategy by using fixed-point numbers, and constructing a weight updating process into a difference-inclusion system model containing a quantization noise disturbance term, and deriving a sufficient condition for making the weight index converge by means of Lyapunov stability theory analysis. The method avoids the sampling bottleneck of a high-speed ADC, realizes nanosecond-level early warning, realizes experience reuse under KB-level storage through characteristic code compression, and simultaneously guarantees that the fixed-point weight updating is stably convergent under quantization noise, so that the protection strategy has real-time performance and robustness.

Description

Technical Field

[0001] This invention relates to the field of electromagnetic compatibility and active protection technology, and in particular to an embedded electromagnetic pulse precursor warning and signature protection method and system. Background Technology

[0002] Modern precision-guided components, electromechanical servo drives, and avionics systems for unmanned aerial vehicles (UAVs) often employ nanoscale digital signal processors (DSPs) or field-programmable gate arrays (FPGAs) as their cores. These processors or FPGAs operate at voltages ≤1.8V and have logic level tolerances of only a few hundred millivolts, significantly increasing their sensitivity to transient electromagnetic interference (EMI). Threats such as high-voltage switch discharges and lightning strikes have main pulse leading edges on the nanosecond scale and peak field strengths exceeding tens of kilovolts per meter. When these surge through gap coupling, they induce common-mode overvoltage, leading to processor resets, data rollovers, and ultimately, actuator malfunctions. More critically, to control costs, these systems employ a single-redundancy architecture, often with less than 2KB of static storage, supporting only fixed-point operations and demanding power consumption. This creates a fundamental contradiction with the high-bandwidth sampling, large-capacity storage, and floating-point computation required for electromagnetic pulse protection.

[0003] Meanwhile, existing electromagnetic pulse protection technologies have the following defects: (1) Passive shielding and filtering technologies rely on metal cavities and passive networks, but the cavity resonant frequency is lower than the main frequency band of electromagnetic pulses, which can easily lead to enhancement effects. The high-frequency parasitic parameters of LC filters cause the insertion loss to deteriorate, and the protection parameters are fixed, resulting in problems such as over-design or under-design. (2) Although transient voltage suppressors have a response up to the sub-nanosecond level, they can only suppress energy impacts and cannot repair soft errors in storage. Moreover, after multiple impacts, the clamping voltage drifts, and it is only used as a passive remedy for local energy suppression. (3) Active cognitive protection introduces machine learning, which has the following defects: First, it relies on full sampling of analog-to-digital converters (ADCs). Nanosecond-level signals require ultra-high sampling rates. The link delay of sampling, quantization and processing has far exceeded the pulse action time, causing the protection decision to always lag behind the damage and can only be remedied after the fact. Second, the parameters of deep learning models exceed the single redundant storage capacity, and the weight update competes for control resources. Third, incremental learning lacks the theory of fixed-point stability, and rounding errors lead to iterative instability.

[0004] In summary, the existing technology has not solved the following three interdependent theoretical and engineering problems: (1) Threat prior acquisition problem: Traditional methods filter out precursor waves and fail to find that the two originate from the same discharge event and have an analyzable time-frequency coupling relationship. (2) Single-redundancy knowledge compactness problem: Traditional index learning relies on a large number of parameters, which the single-redundancy system cannot store. (3) Finite word length stability problem: Existing robust control does not consider fixed-point quantized noise. Summary of the Invention

[0005] The purpose of this invention is to overcome the problems of the prior art and provide an embedded electromagnetic pulse precursor early warning and signature code protection method and system.

[0006] The objective of this invention is achieved through the following technical solution: an embedded electromagnetic pulse precursor early warning and signature code protection method, which includes the following steps:

[0007] The precursor wave signal of the electromagnetic pulse is received, the time-frequency characteristics of the precursor wave signal are extracted in the analog domain, and the threat parameters of the main pulse are predicted based on the mapping relationship between the time-frequency characteristics of the precursor wave signal and the main pulse determined by the time-frequency coupling physical model.

[0008] The damage response signal generated by the protection system under the action of the main pulse is collected. Multidimensional physical features are extracted from the damage response signal and compressed and encoded into electromagnetic damage feature code. The electromagnetic damage feature code is used as an index to retrieve matching protection strategy parameters from the knowledge base.

[0009] A protection strategy defined by the fixed-point number execution protection strategy parameters is adopted, and the update process of the weight vector in the protection strategy is constructed as a differential inclusion system model containing quantization noise disturbance terms. The Lyapunov stability theory is used to analyze the differential inclusion system model and derive the sufficient condition for the exponential convergence of the weight vector.

[0010] In one example, the extraction of the time-frequency features of the precursor wave signal in the analog domain includes:

[0011] The precursor wave signal is input into a subthreshold detector circuit operating in the subthreshold region of a transistor and logarithmically compressed to obtain a pulse width signal.

[0012] The autocorrelation function of the pulse width signal is calculated, and the instantaneous frequency change rate of the precursor wave signal is extracted as the time-frequency feature.

[0013] In one example, a lightweight temporal network model is used to predict the threat parameters of the main pulse, including its arrival time, energy, and spectral centroid, based on the mapping relationship between the time-frequency characteristics of the precursor signal and the main pulse determined by the time-frequency coupled physical model.

[0014] In one example, the lightweight temporal network model is a fixed-point network, and forward inference is implemented through shift and accumulation operations.

[0015] In one example, the extraction of multidimensional physical features and their compression encoding into electromagnetic damage feature codes includes:

[0016] The multidimensional physical features are normalized and quantized at fixed points, and then concatenated and mapped using a perfect hash function to generate a fixed-length binary code, thus obtaining the electromagnetic damage feature code.

[0017] In one example, the knowledge base employs a two-level retrieval structure consisting of a Bloom filter and a perfect hash table.

[0018] In one example, the sufficient condition is a joint constraint on the spectral radius of the gain matrix in the weight update dynamics defined by the difference containing the system model and the condition number of the positive definite matrix in the Lyapunov function.

[0019] It should be further noted that the technical features corresponding to the above examples can be combined or replaced to form new technical solutions.

[0020] This invention also includes an embedded electromagnetic pulse precursor early warning and signature protection system, which has the same inventive concept as the method formed by any or more of the above examples, and the system includes:

[0021] The precursor sensing and prediction unit is used to receive the precursor wave signal of the electromagnetic pulse, extract the time-frequency characteristics of the precursor wave signal in the analog domain, and predict the threat parameters of the main pulse based on the mapping relationship between the time-frequency characteristics of the precursor wave signal and the main pulse determined by the time-frequency coupling physical model.

[0022] The index decision unit is used to collect the damage response signal generated by the protection system under the action of the main pulse, extract multi-dimensional physical features based on the damage response signal and compress and encode them into electromagnetic damage feature code, and use the electromagnetic damage feature code as an index to retrieve matching protection strategy parameters from the knowledge base.

[0023] The execution unit is used to execute the protection strategy defined by the protection strategy parameters using fixed-point number, and to construct the update process of the weight vector in the protection strategy as a differential inclusion system model containing quantization noise disturbance terms. The differential inclusion system model is analyzed using Lyapunov stability theory to derive sufficient conditions for the exponential convergence of the weight vector.

[0024] In one example, the precursor sensing and prediction unit includes:

[0025] A slot-coupled antenna is used to receive precursor wave signals;

[0026] The subthreshold detection circuit, connected to the slot-coupled antenna, operates in the transistor subthreshold region and is used to perform logarithmic compression on the precursor wave signal to obtain a pulse width signal.

[0027] The data processing unit is used to calculate the autocorrelation function of the pulse width signal and extract the instantaneous frequency change rate of the precursor wave signal as the time-frequency feature. A lightweight time-series network model is deployed on the data processing unit to calculate the mapping relationship between the time-frequency features of the precursor wave and the threat parameters of the main pulse, thereby predicting the threat parameters, including the arrival time, energy and spectral centroid of the main pulse.

[0028] In one example, the execution unit includes a parameter fixing module that stores sufficient conditions for the exponential convergence of the weight vector. The sufficient conditions are joint constraints on the spectral radius of the gain matrix and the condition number of the positive definite matrix in the Lyapunov function in the weight update dynamics defined by the difference containing the system model.

[0029] It should be further noted that the technical features corresponding to the above system examples can be combined or replaced to form new technical solutions.

[0030] Compared with the prior art, the beneficial effects of the present invention are:

[0031] 1. In one example, by extracting the time-frequency features of the precursor wave and predicting the threat parameters of the main pulse based on the time-frequency coupled physical model, the power consumption and delay bottlenecks of traditional high-bandwidth ADC sampling are avoided, and nanosecond-level early warning before the arrival of the main pulse is achieved; by compressing and encoding multidimensional damage features into compact electromagnetic damage feature codes, the storage resource occupation is significantly reduced, and experience reuse under KB-level storage is achieved; by constructing a differential inclusion system model containing quantization noise through fixed-point weight updates and performing Lyapunov stability analysis, the system oscillation problem caused by quantization noise in fixed-point operations is solved, ensuring the real-time performance and robustness of the protection control strategy.

[0032] 2. In one example, by deploying a lightweight temporal network model to perform physical mapping, the complex physical law calculation is transformed into efficient forward inference, which solves the nonlinear problems that may be encountered when directly solving the physical model, while ensuring prediction accuracy and computational efficiency.

[0033] 3. In one example, a fixed-point network is used for forward inference, eliminating the reliance on floating-point operations and enabling complex neural networks to be deployed on microprocessors that only support fixed-point operations.

[0034] 4. In one example, a two-level retrieval structure consisting of a Bloom filter and a perfect hash table is used. This achieves constant microsecond-level retrieval latency while ensuring an extremely low false positive rate. This meets the requirements of high real-time control systems for protection decision speed and ensures the immediacy of the overall system response.

[0035] 5. In one example, by integrating three major modules—the precursor sensing and prediction unit, the index decision unit, and the execution unit—the precursor warning and signature protection method is solidified into an independently operable protection system. This system fully includes a complete set of functions from signal sensing and intelligent decision-making to stable control. Attached Figure Description

[0036] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. The accompanying drawings are provided to provide a further understanding of the present application and constitute a part of the present application. The same reference numerals are used in these drawings to denote the same or similar parts. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation of the present application.

[0037] Figure 1 A flowchart illustrating a method provided as an example of the present invention;

[0038] Figure 2 This is a schematic diagram of a slot-coupled antenna structure provided as an example of the present invention;

[0039] Figure 3 This is a schematic diagram of a subthreshold detection circuit provided as an example of the present invention;

[0040] Figure 4 This is a schematic diagram of a lightweight temporal network model structure provided as an example of the present invention;

[0041] Figure 5 A flowchart of electromagnetic damage feature extraction and encoding provided as an example of the present invention;

[0042] Figure 6 This is a flowchart of an example of the electromagnetic damage feature code retrieval provided by the present invention;

[0043] Figure 7 This is a flowchart of the feature code and weight update process provided as an example of the present invention;

[0044] Figure 8 A method framework diagram provided for a preferred example of the present invention;

[0045] Figure 9 This is a schematic diagram of the task timing and priority of a protection system provided as an example of the present invention.

[0046] In the diagram: 1-Precursor wave; 2-Metal box; 3-Gap. Detailed Implementation

[0047] The technical solution 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.

[0048] Furthermore, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0049] This invention aims to establish a complete and cost-effective electromagnetic pulse protection method and system, and to achieve engineering verification on a resource-constrained embedded platform. Resource constraints refer to a static storage capacity of approximately 2KB (implemented based on FPGA registers, excluding off-chip SRAM), an incremental power consumption budget of less than 20mW, and a computing core that only supports fixed-point number format. The embedded platform includes a single-redundant digital control system such as a precision guidance component controller, electromechanical servo drivers, and avionics systems for UAV platforms.

[0050] The fundamental goal of this invention is to solve three interdependent theoretical and engineering problems: (1) the problem of acquiring threat prior information. Traditional digital sampling methods face the dual failure of bandwidth and power consumption in the face of deep nanosecond-level pulse fronts. It is necessary to establish an analog domain feature compression and physical layer prediction model to complete the effective inference of threat parameters before energy injection. (2) the problem of knowledge compactness under single-redundancy architecture. Large-capacity model storage is not feasible under resource constraints. It is necessary to construct a low-dimensional manifold embedding and perfect hash index mechanism to realize the knowledge reuse and incremental update of experience across attack cycles. (3) the third, the problem of control stability under finite word length conditions. After the rounding error of fixed-point operation is superimposed with electromagnetic interference, the traditional gradient update law will inevitably become unstable. It is necessary to reconstruct the update mechanism and derive an explicit stability criterion.

[0051] In one example, such as Figure 1 As shown, an embedded electromagnetic pulse precursor early warning and signature protection method includes the following steps:

[0052] S1: Receives the precursor wave signal of an electromagnetic pulse (EMP), extracts the time-frequency characteristics of the precursor wave signal in the analog domain, and predicts the threat parameters of the main pulse based on the mapping relationship between the time-frequency characteristics of the precursor wave signal and the main pulse determined by the time-frequency coupling physical model.

[0053] Specifically, an extremely weak precursor wave leaks before the main pulse of an electromagnetic pulse arrives. Traditional methods filter out this precursor wave as noise. However, the precursor wave and the main pulse are generated by the same discharge event and have a strict physical relationship. Figure 2 As shown, this example creates a slit 3 in the metal box 2 of the controller as a dedicated electromagnetic stethoscope for the precursor wave 1, filtering out other frequency signals, thus predicting its arrival before the main pulse explodes.

[0054] Furthermore, according to the theory of electromagnetic wave propagation in dispersive media, the precursor wave and the main pulse originate from the same discharge event, and their propagation process satisfies:

[0055] ;

[0056] in, Indicates electric field strength; Spatial coordinates representing the direction of propagation; Indicates time; This represents the free permeability, with a value of 4π × 10⁻ 7 H / m; It represents the relative permeability of the medium (dimensionless), which is approximately 1 in the range of 0.1-3 GHz; This represents the frequency-dependent complex permittivity; Indicating dielectric conductivity, this invention focuses on the metal casing. ; This represents the current density of the discharge source, and its spectral centroid determines the energy distribution of the main pulse.

[0057] Dielectric constant Within the characteristic frequency band, it can be approximated as The linear function of the precursor wave is used to derive its instantaneous frequency. With the centroid of the main pulse energy spectrum The mapping relationship is as follows:

[0058] ;

[0059] in, The scaling factor for the analytical mapping; The rate of change of the instantaneous angular frequency of the precursor wave over time; This is the constant term for the analytical mapping. The coefficients of this linear relationship... The unique determination of the dispersive properties of the medium forms the physical basis for the realizability of characteristic compression in the analog domain. The arrival time of the energy center of the main pulse, as a wave packet, is determined by its spectrum. The weighted average group delay is used to predict the threat parameters of the main pulse, including the centroid of the main pulse energy spectrum and the arrival time.

[0060] S2: Collect the damage response signal generated by the protection system under the action of the main pulse, extract multi-dimensional physical features based on the damage response signal and compress and encode them into electromagnetic damage feature codes, and use the electromagnetic damage feature codes as indexes to retrieve matching protection strategy parameters from the knowledge base.

[0061] After each EMP event, the power rail voltage v(t) and PWM error are recorded. Three-dimensional time-series features are extracted from the actuator hysteresis τ(t). The voltage sag energy rate is calculated using a trapezoidal integral expression:

[0062] ;

[0063] in, Voltage sag energy rate; The duration of the voltage drop; This is the rated voltage of the power rail; This represents the instantaneous voltage of the power rail after the EMP event. The calculation of the trapezoidal integral is implemented by the coprocessor through hardware acceleration, with a constant computation cycle of 30 instruction cycles.

[0064] Furthermore, the spectral centroid is derived using a sliding discrete Fourier transform (DFT) recursive formula:

[0065]

[0066] In the formula, For the first Time of the first The sliding DFT results at each frequency point (i.e., the frequency domain components in the process of calculating the centroid of the spectrum). For the first Time of the first Sliding DFT results at each frequency point; For the first The PWM error signal at any given time; For the first The PWM error signal at any given time; The sliding window length is N=32 in this example, corresponding to a 50ns sliding window, which matches the typical EMP duration; For frequency point index (value range is 0, 1, ..., N-1); This is the phase rotation factor for the sliding DFT. Each update requires only two complex additions, avoiding redundant calculations. Point transformation. Furthermore, at each time step... Based on the updated spectral components We obtain the centroid SC of the spectrum at the current moment.

[0067] Three-dimensional physical characteristics: power rail voltage sag energy rate By compressing and encoding the centroid SC of the pulse width modulation error spectrum and the actuator hysteresis time τ, an electromagnetic damage feature code can be obtained, which is used to uniquely identify the system damage mode caused by an EMP attack. The electromagnetic damage feature code is then used as a search key to retrieve information from a knowledge base, and the protection strategy parameters most suitable for the current electromagnetic damage feature code are obtained based on historical experience.

[0068] S3: The protection strategy defined by the fixed-point number execution protection strategy parameters is adopted, and the update process of the weight vector in the protection strategy is constructed as a differential inclusion system model containing quantization noise disturbance terms. The Lyapunov stability theory is used to analyze the differential inclusion system model and derive the sufficient condition for the exponential convergence of the weight vector.

[0069] Specifically, in the fixed-point computing environment of an embedded system, an adaptive protection strategy is stably executed. The iterative update process of the weights (i.e., control parameters) is modeled as a differential inclusion system model containing bounded quantization noise terms, thus accurately reflecting the errors introduced by finite-precision computation. Then, Lyapunov stability theory is used to analyze the differential inclusion system model, deriving the mathematical conditions that guarantee the eventual stability and rapid convergence (exponential convergence) of the weight update process. Further, based on these mathematical conditions, the required control parameters are solved, and applied to the hardware execution unit of the protection system. This achieves adaptive, real-time, and stable active electromagnetic pulse protection control on a resource-constrained platform with computational noise.

[0070] This invention establishes a closed theoretical framework for threat perception and protection execution. This framework does not rely on high-bandwidth analog-to-digital converters or floating-point arithmetic units. It achieves effective sub-nanosecond electromagnetic pulse suppression through three mutually coupled methodological layers: First, by utilizing the time-frequency coupling characteristics of electromagnetic wave dispersion channels, the instantaneous frequency features of the precursor wave are extracted in the analog signal processing domain to construct a physical mapping model with the energy parameters of the main pulse, bypassing the bandwidth-power consumption bottleneck of traditional digital sampling. Second, based on manifold learning, the high-dimensional temporal electromagnetic damage trajectory is compressed into a compact binary electromagnetic damage feature code space. Through knowledge indexing and reuse, the inherent contradiction of exhausting storage resources in a single-redundancy system is resolved. Third, for environments with dual disturbances of fixed-point rounding errors and electromagnetic interference, a weight update mechanism is established, and an explicit Lyapunov stability criterion including quantization noise is derived, theoretically guaranteeing the exponential convergence of weight iteration under finite word length conditions.

[0071] In one example, the time-frequency features of the precursor wave signal are extracted in the analog domain, including:

[0072] The precursor wave signal is input into a subthreshold detector circuit operating in the subthreshold region of a transistor and logarithmically compressed to obtain a pulse width signal.

[0073] The autocorrelation function of the pulse width signal is calculated, and the instantaneous frequency change rate of the precursor wave signal is extracted as the time-frequency feature.

[0074] Specifically, such as Figure 3 As shown, the subthreshold detection circuit uses an SST308 N-channel MOSFET with a gate bias voltage V. gs =0.3V, drain load resistance R d =10kΩ, achieving a logarithmic compression dynamic range of 60dB. The subthreshold detection circuit logarithmically compresses the precursor wave signal into a pulse-width signal. The mathematical essence of logarithmic compression is mapping an exponentially decaying field strength to a linearly measurable time quantity, satisfying:

[0075] ;

[0076] in, Indicates the decay time constant; This represents the initial field strength.

[0077] Furthermore, a data processing unit such as an FPGA is used to calculate the short-time autocorrelation of the pulse width signal (pulse sequence). This calculation can be completed at a 200MHz clock, without the need to store the original waveform, and the memory usage is reduced to a few registers within a nanosecond time window.

[0078] This example uses a subthreshold detection circuit to perform logarithmic compression on the precursor wave signal and calculates the instantaneous frequency change rate through the autocorrelation function. It bypasses the high-power, high-latency analog-to-digital converter sampling stage in traditional solutions, and can complete the extraction of key features in the analog domain with extremely low power consumption (mW level), meeting the power consumption and real-time requirements of embedded platforms.

[0079] In one example, a lightweight temporal network model is used to predict the threat parameters of the main pulse, including its arrival time, energy, and spectral centroid, based on the mapping relationship between the time-frequency characteristics of the precursor signal and the main pulse determined by the time-frequency coupled physical model.

[0080] Specifically, the lightweight timing network model (timing reversible network) deployed in a data processing unit such as an FPGA employs 8-bit symmetric quantization and a scaling factor. , This represents the weight parameters of the lightweight temporal network model. The network's quantization-aware training is completed on a PC, and its loss function is designed as the weighted mean square error of the predicted triples, expressed as:

[0081] .

[0082] In the formula, L is the loss function; , , These are the weighting coefficients for the time difference, energy, and characteristic frequency (or center frequency) terms, respectively. , These are the predicted time difference and the actual time difference, respectively. , These are the predicted energy and the actual energy, respectively. , These are the predicted characteristic frequency (or center frequency) and the actual characteristic frequency (or center frequency), respectively, used to calculate the mean square error.

[0083] Examples of the present invention To emphasize the peak prediction weight, the training dataset consists of standard waveforms from GJB151B-2013 and 500–1000 measured pulse current injection data. After iteration, the model accuracy loss is <3%.

[0084] Preferably, such as Figure 4 As shown, the lightweight temporal network model is a fixed-point network, consisting of sequentially connected blocks 1, 2, 3, 4, and 5. The network's receptive field has a porosity of 2. i Incremental implementation covers the dispersion delay difference between the precursor wave and the main pulse. In this example, the total receptive field = 3 + 7 + 15 + 31 + 63 = 119 ns > the dispersion delay difference Δt (20 - 80 ns). Forward inference in this example model is implemented through shift and accumulation operations, expressed as:

[0085] ;

[0086] in, The output of the reasoning process; For the first Each weighting coefficient; For the first One shift number of bits; For the first There are one input variable. This formula represents converting multiplication into shifting and addition, with a single-channel convolution delay of no more than 8 clock cycles.

[0087] During forward inference, after each block calculates and outputs its result, it immediately releases the intermediate activation value register it occupies, rather than retaining it until the end of the calculation. When subsequent blocks require the above intermediate data results, the required activation values ​​can be dynamically recalculated in reverse based on the stored network weights and inputs. The static storage resources required for the overall model operation are significantly reduced to approximately 8.3KB. This memory footprint allows the weights and runtime state of the entire network to be implemented entirely by the FPGA's on-chip registers, without relying on external SRAM. This enables efficient and low-power deployment of complex neural network functions on embedded platforms.

[0088] In one example, such as Figure 5 As shown, multidimensional physical features are extracted and compressed into electromagnetic damage feature codes, including:

[0089] The multidimensional physical features are normalized and quantized at fixed points, and then concatenated and mapped using a perfect hash function to generate a fixed-length binary code, thus obtaining the electromagnetic damage feature code.

[0090] Specifically, the power rail voltage sag energy rate The pulse width modulation error spectrum centroid SC and actuator hysteresis time τ are normalized and quantized using 16-bit methods, and concatenated into a 48-bit integer vector. A perfect hash function constructed based on the Compact Hashing with Displacement (CHD) algorithm maps the 48-bit integer vector to a 128-bit binary space to generate a 128-bit fixed-length binary code, namely: the electromagnetic damage feature code. Its mathematical expression is: This achieves unique identification of damage patterns across attack cycles through electromagnetic damage signature codes. The hash function in this example is:

[0091]

[0092] in, For hash index (the storage location of the final mapping); For intermediate hash functions; , The random coefficients of the hash function; The modulus is a large prime number (used to ensure the uniformity of the hash function); 1024 is the size of the hash table, and the final index value ranges from 0 to 1023. In this invention, it is set... , For a pre-generated 61-bit odd number, ensure the uniformity of the first-level hash.

[0093] In this example, the power rail voltage sag energy rate is measured. The pulse width modulation error spectrum centroid SC and actuator lag time τ are normalized and fixed-point quantized to generate fixed-length binary codes, so that the electromagnetic damage feature code has both minimal storage overhead and fast matching potential, solving the problem that single-redundancy systems cannot store large parameter models.

[0094] In one example, the knowledge base employs a two-level retrieval structure consisting of a Bloom filter and a perfect hash table. Specifically, such as... Figure 6As shown, the electromagnetic damage signature uses a two-level index structure: the first level is a Bloom filter to determine if there is a historical record (corresponding to the protection strategy of the electromagnetic damage signature). If not, a new electromagnetic damage pattern is created; if so, a second search is performed. The false positive rate of the entire judgment process is ≤1%, which can quickly eliminate unrecorded attacks. The second level is perfect hash direct addressing, which determines whether the original index value calculated by the hash function in the first level has a collision. If so, a second-level hash is performed through a pre-stored offset table to resolve the collision until there is zero collision, then the corrected address is obtained, and the corresponding protection strategy is obtained; if no collision occurs, the address is directly obtained, and then the corresponding protection strategy is obtained. This example of a two-level search structure design achieves a constant search latency of 18 instruction cycles, corresponding to a deterministic search time of 90 nanoseconds on a 200MHz main frequency (5ns per cycle) FPGA platform.

[0095] Preferably, the knowledge base employs an incremental update strategy for continuous learning. Specifically, such as... Figure 6 As shown, incremental updates follow the Lyapunov trigger condition: the oldest record is replaced only when the L2 distance between the new electromagnetic damage feature code and its nearest neighbor in the database is greater than a threshold (e.g., threshold θ=0.2). This strategy avoids frequent jitter in the knowledge base, the replacement operation takes <5μs, and does not encroach on the PWM interrupt response time.

[0096] In one example, based on the protection strategy parameters, a fixed-point execution protection strategy is adopted, and the update process of its weight vector is constructed as a differential inclusion system model containing quantization noise perturbation terms. Specifically, as shown... Figure 7 As shown, the weight update law is expressed in the fixed-point Q15 format as follows:

[0097]

[0098] In the formula, For the first The control weight value of the protection strategy at any given time; For the first Control weight values ​​for the constant protection strategy; Update the gain for the weights (positive number, controlling the update step size); Error signal The sign function (takes values ​​of 1, -1, or 0, representing the direction of the error); Error signal The square root of the absolute value (used to adaptively adjust the update step size; the larger the error, the larger the step size, thus accelerating convergence). This represents the disturbance term (the difference includes uncertainties or external disturbances characterized in the system model). This update law constitutes a saturation-hysteresis composite control mechanism. Among them, To quantize noise; As for the upper bound of noise, they jointly clarified the constraints on quantization noise and the quantitative calculation logic of the upper bound of quantization noise during the weight update process under the Q15 format: among which, The quantization noise introduced by Q15 fixed-point arithmetic indicates that its value is strictly constrained to... Within the interval, to ensure that the perturbation is controllable, key constraints are provided for the Lyapunov stability analysis of the difference-containing system model; while As the upper bound of the amplitude of quantization noise, the noise upper bound calculation formula includes 2. -15 It is the inherent quantization step size of the Q15 format (determined by the format characteristics of 15 decimal places). It is the positive definite matrix used in Lyapunov stability analysis. The condition number clearly establishes the relationship between quantization noise characteristics and Q15 quantization accuracy and Lyapunov matrix characteristics.

[0099] Furthermore, construct piecewise Lyapunov functions:

[0100] ;

[0101] This formula is a piecewise Lyapunov function constructed using a difference-containing system model (adapted to the Q15 format weight update law), used to analyze the stability of the weight update process. Wherein, For piecewise Lyapunov functions, it is the core functional for determining the stability of dynamic processes; It is a positive definite matrix, and its positive definiteness is the basis for the Lyapunov function to satisfy the stability condition; These are preset positive weighting coefficients; The saturation function reflects the nonlinear characteristics of the saturation-hysteresis composite control mechanism. The sufficient condition for the exponential convergence of the weight vector is a joint constraint on the spectral radius of the gain matrix and the condition number of the positive definite matrix in the Lyapunov function during the weight update dynamics defined by the difference-containing system model.

[0102] By performing a difference operation on the constructed piecewise Lyapunov function along the system state update trajectory, we obtain:

[0103] ;

[0104] in, Indicates the first The difference of the Lyapunov function at time t; A positive definite matrix The smallest eigenvalue; Q15 format weight vector 2-norm; A positive definite matrix The induced norm; For matrix The largest eigenvalue; This is the energy decay term. , All of these are disturbance terms caused by quantization noise. By constraining the magnitude relationship between the attenuation term and the disturbance term, it can be further deduced that the dynamic model of weight update can still satisfy the stability condition that ΔV(t) is eventually negative under the action of quantization noise, ensuring that the weight update process converges to a stable state.

[0105] Defined in this invention .when And when ρ(K) < 0.85, the cross term satisfies:

[0106] ;

[0107] Where ρ(K) is the gain matrix. spectral radius, Depend on , The positive coefficient is determined by the parameter. Therefore, Negative definite, the difference includes the system model, which converges exponentially to the radius. The proof forms the basis for fixed-point robust updates in the neighborhood of .

[0108] Furthermore, the gain matrix Obtained by solving linear matrix inequalities offline:

[0109] ;

[0110] The inequality is solved numerically using the RobustControl Toolbox in MATLAB software to obtain the gain matrix. spectral radius Matrix condition number It satisfies theoretical constraints.

[0111] This invention defines The core parameters are precisely constrained by solving the Linear Matrix Inequality (LMI) offline, and the specific values ​​are... , (Diagonal gain matrix, each diagonal element corresponds to the update step size of different weights), spectral radius And satisfy and The key constraint condition. The spectral radius constraint threshold of 0.85 is not empirically chosen, but rather derived from the quantization noise characteristics of the Q15 fixed-point format, representing the critical convergence value of the difference-containing system model. When the quantization noise... When this threshold is met, it ensures that the weight update process still satisfies the Lyapunov stability condition under quantization noise perturbation; conversely, when... Theoretical analysis and numerical simulations show that the weight error diverges after 100 updates, verifying the necessity and scientific validity of the constraint. The quantitative design, constraint basis, and solution method of the above parameters constitute one of the core technical features of this invention to ensure robust weight updates under the Q15 format, providing key parameter support for the real-time stable control of the electromagnetic pulse active protection system.

[0112] Combining the above examples yields preferred examples of the present invention, such as... Figure 8 As shown, the method includes the following steps:

[0113] S10: Receives the precursor wave signal of the electromagnetic pulse, inputs the precursor wave signal into a subthreshold detector circuit operating in the subthreshold region of a transistor for logarithmic compression to obtain the pulse width signal; calculates the autocorrelation function of the pulse width signal and extracts the instantaneous frequency change rate of the precursor wave signal as a time-frequency feature; uses a lightweight time-series network model based on the mapping relationship between the time-frequency features of the precursor wave signal and the main pulse determined by the time-frequency coupling physical model to predict the threat parameters of the main pulse;

[0114] S20: Acquires the damage response signal generated by the protection system under the action of the main pulse, and extracts multi-dimensional features based on the damage response signal: power rail voltage sag energy rate. The pulse width modulation error spectrum centroid SC and actuator hysteresis time τ are used to normalize and quantize the multidimensional physical features. After splicing, a fixed-length binary code is generated by mapping through a perfect hash function to obtain the electromagnetic damage feature code. Using the Hamming space electromagnetic damage feature code as an index, the matching protection strategy parameters are retrieved from the knowledge base with a two-level retrieval structure composed of a Bloom filter and a perfect hash table.

[0115] S30: A protection strategy defined by the fixed-point execution protection strategy parameters is adopted, and the update process of the weight vector in the protection strategy is constructed as a differential inclusion system model containing quantization noise disturbance terms. The differential inclusion system model is analyzed using Lyapunov stability theory to derive sufficient conditions for the exponential convergence of the weight vector. The sufficient conditions are the joint constraints on the spectral radius of the gain matrix and the condition number of the positive definite matrix in the Lyapunov function in the dynamic weight update defined by the differential inclusion system model. Based on the sufficient conditions, the control parameters that meet the requirements are solved, and the control parameters are applied to the hardware execution unit of the protection system. In this way, adaptive, real-time and stable active electromagnetic pulse protection control is achieved on a resource-constrained platform with computational noise.

[0116] like Figure 9As shown in the preferred example above, the pulse width modulation (PWM) control interrupt has zero preemption characteristics, allowing it to interrupt low-priority tasks at any time to ensure the immediacy of the core control loop. Step S10, the precursor detection, is given the highest priority, with its worst-case response time strictly constrained to ≤5 nanoseconds, ensuring rapid perception of threat signals. Step S20, the signature retrieval task, is set to medium priority, with a deterministic execution time of 30 nanoseconds, allowing for rapid damage pattern matching after precursor detection. Step S30, the weight update task, is given the lowest priority, with an execution time of 8 nanoseconds, used for adaptive strategy fine-tuning. Considering the timing of each task, after completing precursor detection and signature retrieval, there is still sufficient time margin before the main pulse energy arrives. This margin is much greater than the execution requirements of background tasks such as weight update, thus ensuring at the hardware level that the complete closed loop of the multi-layer protection response can be reliably completed before the main pulse causes damage, achieving seamless connection from precursor wave warning to main pulse suppression.

[0117] This invention proposes a three-tiered cascaded protection architecture consisting of a precursor perception layer, an index decision layer, and a robust execution layer. The theoretical basis of this architecture is based on three core models: (1) Time-frequency coupling model: There is a linear mapping relationship between the instantaneous frequency change rate of the precursor wave and the centroid of the energy spectrum of the main pulse, and this mapping relationship is uniquely determined by the dispersive channel propagation equation. (2) Manifold embedding model: The intrinsic dimension of the high-dimensional damage trajectory does not exceed eight dimensions, and it can be mapped to a 128-dimensional Hamming space through random projection while maintaining the neighborhood structure. (3) Differential inclusion system model: The fixed-point number update law can be expressed as a differential inclusion system model containing quantization noise, and its stability is jointly determined by the Lyapunov matrix condition number and the spectral radius of the gain matrix.

[0118] Based on the above preferred example method, the beneficial effects of the method of the present invention are shown in Table 1:

[0119] Table 1. Comparison of Beneficial Effects of the Invention Methods: An Example Table

[0120]

[0121] This invention achieves multi-dimensional technological breakthroughs in the field of active electromagnetic pulse protection through the deep integration of electromagnetic propagation theory, control stability analysis, and embedded engineering design. All effects have clear theoretical support and verifiable engineering indicators, as detailed below:

[0122] (1) Breakthrough in threat perception timeliness. Step S10 uses a slot-coupled antenna with a resonant frequency of 2.45 GHz as the detection front end, paired with a subthreshold detection circuit with a gate-source voltage of Vgs=0.3V. The instantaneous frequency change rate of the EMP precursor wave is directly extracted in the analog domain without the need for analog-to-digital conversion (ADC). At the same time, a time-frequency coupling model is established based on the dispersive medium propagation equation, and the linear mapping relationship between the instantaneous frequency change rate and the main pulse energy is derived. This allows for the prediction of the threat parameters of the main pulse, avoiding the power consumption and delay bottlenecks of traditional high-bandwidth ADC sampling, and achieving advance prediction before the arrival of the main pulse, with a protection lead time ≥50 ns. The system's operating power consumption is only 8 mW, which is 99.2% lower than the existing ADC sampling scheme, perfectly adapting to low-power embedded scenarios such as missile-borne and airborne systems. More specifically, exemplary test data on the relationship between precursor wave detection and main pulse arrival time are shown in Table 2:

[0123] Table 2. Example Test Data on the Relationship between Precursor Wave Detection and Main Pulse Arrival Time

[0124]

[0125] In Table 2, the test conditions used were obtained using the measured pulse current injection method, with a rise time t. r =2ns, pulse width t w =50ns, repetition frequency 1kHz, a total of 500 injections were performed, the mean prediction bias was 1.8ns, the standard deviation was 1.2ns, which met the prediction performance.

[0126] (2) Knowledge representation compactness innovation. Step S20 extracts the power rail voltage sag energy rate. The feature vectors of the pulse width modulation error spectrum centroid SC and the actuator hysteresis time τ are mapped into 128-bit fixed-length binary electromagnetic damage feature codes using the CHD perfect hash algorithm. This feature mapping process satisfies the distance-preserving condition of the Johnson-Lindenstraus lemma, ensuring that the core discriminative information of the three-dimensional physical features is not lost in the low-dimensional encoding space. The collision-free characteristic of the CHD hash algorithm guarantees the uniqueness of the feature codes, providing a theoretical basis for accurate damage pattern identification. This feature encoding method significantly reduces storage resource consumption; a single feature code requires only 128 bits (<20 bytes), and the total storage of 1024 electromagnetic damage feature codes is <2KB, fully adaptable to the static RAM resources of a single-redundancy system. Compared to deep learning-based feature representation methods, the parameter size is compressed by 16384 times, and the feature matching delay is ≤1μs, meeting real-time identification requirements.

[0127] (3) Proof of stability of fixed-point update. Step S30 constructs a difference inclusion system model under the Q15 fixed-point number scheme, and obtains the gain matrix by solving the linear matrix inequality offline. Its spectral radius This satisfies the theoretical convergence constraints. The theoretical support for the above update process is as follows: constructing a piecewise Lyapunov function, deriving a stability criterion including quantization noise, and when the positive definite matrix... condition number At that time, the weight update law is within ±1 LSB, and the quantization noise is... The lower exponential convergence to the radius The neighborhood of ε is used to solve the system oscillation problem caused by quantization noise in fixed-point arithmetic. The weight error converges to the neighborhood of ε=0.03, and the convergence speed is improved by 2-3 orders of magnitude compared with the traditional LMS algorithm, ensuring the real-time performance and robustness of the protection and control strategy.

[0128] Furthermore, this invention uses FPGA resource reuse design, the core incremental hardware cost is lower than that of traditional transient voltage suppressor solutions, and the volume increment is close to zero; and the theoretical model is platform independent and can be ported to various fixed-point arithmetic embedded systems without the need to reconstruct the hardware architecture.

[0129] The present invention also includes an embedded electromagnetic pulse precursor early warning and signature protection system, which includes a precursor sensing and prediction unit, an indexing decision unit and an execution unit.

[0130] The system comprises several components: a precursor sensing and prediction unit, an indexing and decision unit, and an execution unit. The former receives the precursor wave signal of the electromagnetic pulse (EMP), extracts its time-frequency characteristics in the analog domain, and predicts the threat parameters of the main pulse based on the mapping relationship between these characteristics and the main pulse determined by the time-frequency coupling physical model. The latter collects the damage response signal generated by the protection system under the main pulse, extracts multi-dimensional physical features from the damage response signal, compresses and encodes them into an electromagnetic damage feature code, and uses this feature code as an index to retrieve matching protection strategy parameters from a knowledge base. The execution unit executes the protection strategy defined by the protection strategy parameters using fixed-point numbering. It constructs the update process of the weight vector in the protection strategy as a differential inclusion system model containing quantized noise disturbance terms. Lyapunov stability theory is used to analyze the differential inclusion system model, deriving sufficient conditions for the exponential convergence of the weight vector. Preferably, the execution unit solves for the required control parameters based on the sufficient conditions and applies these parameters to the hardware execution unit of the protection system, thereby achieving adaptive, real-time, and stable active EMP protection control on a resource-constrained platform with computational noise.

[0131] In one example, the precursor sensing and prediction unit includes a slot-coupled antenna, a subthreshold detection circuit, and a data processing unit.

[0132] Among them, such as Figure 2As shown, a slot 3 is formed in the metal box 2 of the controller to create a slot-coupled antenna for receiving the preceding wave 3. Optionally, a slot is cut into the side wall of a 50mm×50mm×2mm 6061 aluminum alloy shielding box, with a slot length L=12mm, a slot width W=0.5mm, and a slot depth D=2mm (penetrating the box wall). Furthermore, the resonant frequency of the slot-coupled antenna is designed to be within the linear range of dispersion characteristics to ensure the monotonicity of the frequency-amplitude response. Based on the transmission line model... ,in The equivalent dielectric constant is used to determine the resonant frequency. It falls within the linear range (2-3 GHz) of the dispersion characteristics of aerospace aluminum casings. At 2.45 GHz, the insertion loss S21 = -6.2 dB, and the precursor wave detection sensitivity ≤ -65 dBm.

[0133] like Figure 3 As shown, the subthreshold detection circuit is connected to the slot-coupled antenna and operates in the transistor's subthreshold region to perform logarithmic compression on the precursor wave signal, obtaining a pulse width signal. Furthermore, the subthreshold detection circuit uses an SST308 N-channel MOSFET with a gate bias voltage V0. gs =0.3V, drain load resistance R d =10kΩ, achieving a logarithmic compression dynamic range of 60dB.

[0134] Furthermore, the data processing unit is used to calculate the autocorrelation function of the pulse width signal and extract the instantaneous frequency change rate of the precursor wave signal as a time-frequency feature. Preferably, a lightweight time-series network model is deployed on the data processing unit to calculate the mapping relationship between the time-frequency features of the precursor wave and the threat parameters of the main pulse, thereby predicting the threat parameters, including the arrival time, energy, and spectral centroid of the main pulse. At a 200MHz clock, the lightweight time-series network model performs autocorrelation calculation on the pulse sequence and extracts the instantaneous frequency change rate of the pulse sequence. Figure 4 As shown, this example lightweight temporal network model includes a 5-layer dilated convolutional network with a receptive field of 119ns, which can cover the delay (20-80ns) from the precursor wave to the main pulse. It is used to perform reversible calculations, releases the registers after the calculation is completed, occupies only 8.3K of memory, uses 8-bit symmetric quantization, and the model's forward inference is achieved through shift and accumulation operations. The single-path delay is ≤8 clock cycles, and it predicts the arrival time, energy, and spectral centroid of the output main pulse.

[0135] In one example, after each EMP event, the index decision unit sets the power rail voltage sag energy rate. The three data points—pulse width modulation error spectrum centroid SC and actuator hysteresis time τ—are packaged into a 128-bit damage fingerprint. A perfect hash algorithm is used for protection strategy retrieval, and 1024 empirical data points are stored in 2KB space. The retrieval time is less than 1μs. Updates are only performed when the difference between a new damage pattern and an old empirical data point is greater than 0.2, thus achieving incremental learning and ensuring the stability of data updates.

[0136] In one example, to address the issue of weight update divergence caused by the accumulation of fixed-point rounding errors, the execution unit constructs the weight vector update process in the protection strategy as a difference-inclusive system model containing quantization noise perturbation terms. Quantization noise is treated as a bounded perturbation, and a Lyapunov function is used to prove that as long as the gain matrix... If the spectral radius is less than 0.85, the system converges exponentially, and the weighted error converges to the neighborhood of ε=0.03, with the error not exceeding 3%.

[0137] In one example, the execution unit includes a parameter fixing module that stores sufficient conditions for the exponential convergence of the weight vector. These sufficient conditions are joint constraints on the spectral radius of the gain matrix and the condition number of the positive definite matrix in the Lyapunov function within the weight update dynamics defined by the difference-containing system model. By storing parameters that satisfy the stability sufficient conditions through the parameter fixing module, the core parameters such as the gain matrix of the weight update law are physically ensured to always satisfy the spectral radius and condition number constraints during online operation, thus improving the reliability and robustness of the protection system in long-term operation in complex electromagnetic environments.

[0138] The above detailed embodiments are a description of the present invention. It should not be considered that the specific embodiments of the present invention are limited to these descriptions. For those skilled in the art, several simple deductions and substitutions can be made without departing from the concept of the present invention, and all of these should be considered to fall within the protection scope of the present invention.

Claims

1. An embedded electromagnetic pulse precursor early warning and signature code protection method, characterized in that, Includes the following steps: The precursor wave signal of the electromagnetic pulse is received, and its time-frequency characteristics are extracted in the analog domain. A lightweight time-series network model is used to predict the threat parameters of the main pulse, including the arrival time, energy and spectral centroid of the main pulse, based on the mapping relationship between the time-frequency characteristics of the precursor wave signal and the main pulse determined by the time-frequency coupling physical model. The damage response signal generated by the protection system under the action of the main pulse is collected. Multidimensional physical features are extracted from the damage response signal and compressed and encoded into electromagnetic damage feature codes. The electromagnetic damage feature codes are used as indexes to retrieve matching protection strategy parameters from the knowledge base. The extraction of multidimensional physical features and compression and encoding into electromagnetic damage feature codes includes: normalizing and quantizing the multidimensional physical features, concatenating them and generating a fixed-length binary code through a perfect hash function to obtain the electromagnetic damage feature codes. The knowledge base employs a two-level retrieval structure consisting of a Bloom filter and a perfect hash table; A protection strategy defined by the fixed-point number execution protection strategy parameters is adopted, and the update process of the weight vector in the protection strategy is constructed as a differential inclusion system model containing quantization noise disturbance terms. The differential inclusion system model is analyzed using Lyapunov stability theory, and sufficient conditions for the exponential convergence of the weight vector are derived. The sufficient conditions are the joint constraints on the spectral radius of the gain matrix and the condition number of the positive definite matrix in the Lyapunov function in the weight update dynamics defined by the differential inclusion system model.

2. The embedded electromagnetic pulse precursor early warning and signature protection method according to claim 1, characterized in that, The extraction of the time-frequency features of the precursor wave signal in the analog domain includes: The precursor wave signal is input into a subthreshold detector circuit operating in the subthreshold region of a transistor and logarithmically compressed to obtain a pulse width signal. The autocorrelation function of the pulse width signal is calculated, and the instantaneous frequency change rate of the precursor wave signal is extracted as the time-frequency feature.

3. The embedded electromagnetic pulse precursor early warning and signature protection method according to claim 1, characterized in that, The lightweight temporal network model is a fixed-point network, and forward inference is achieved through shift and accumulation operations.

4. An embedded electromagnetic pulse precursor early warning and signature code protection system, characterized in that, include: The precursor sensing and prediction unit is used to receive the precursor wave signal of the electromagnetic pulse, extract the time-frequency characteristics of the precursor wave signal in the analog domain, and use a lightweight time-series network model to predict the threat parameters of the main pulse, including the arrival time, energy and spectral centroid of the main pulse, based on the mapping relationship between the time-frequency characteristics of the precursor wave signal and the main pulse determined by the time-frequency coupling physical model. The indexing decision unit is used to collect the damage response signal generated by the protection system under the action of the main pulse, extract multi-dimensional physical features from the damage response signal and compress and encode them into electromagnetic damage feature codes, and use the electromagnetic damage feature codes as indexes to retrieve matching protection strategy parameters from the knowledge base; the extraction of multi-dimensional physical features and compression and encoding into electromagnetic damage feature codes includes: normalizing and quantizing the multi-dimensional physical features, concatenating them and generating fixed-length binary codes through a perfect hash function to obtain the electromagnetic damage feature codes; the knowledge base adopts a two-level retrieval structure composed of a Bloom filter and a perfect hash table. The execution unit is used to execute the protection strategy defined by the protection strategy parameters using fixed-point number, and to construct the update process of the weight vector in the protection strategy as a differential inclusion system model containing quantization noise perturbation terms. The Lyapunov stability theory is used to analyze the differential inclusion system model and derive the sufficient condition for the exponential convergence of the weight vector. The execution unit includes a parameter fixing module that stores sufficient conditions for the exponential convergence of the weight vector. The sufficient conditions are joint constraints on the gain matrix spectral radius and the condition number of the positive definite matrix in the Lyapunov function in the weight update dynamics defined by the difference system model.

5. The embedded electromagnetic pulse precursor early warning and signature protection system according to claim 4, characterized in that, The precursor sensing and prediction unit includes: A slot-coupled antenna is used to receive precursor wave signals; The subthreshold detection circuit, connected to the slot-coupled antenna, operates in the transistor subthreshold region and is used to perform logarithmic compression on the precursor wave signal to obtain a pulse width signal. The data processing unit is used to calculate the autocorrelation function of the pulse width signal and extract the instantaneous frequency change rate of the precursor wave signal as a time-frequency feature; a lightweight time-series network model is deployed on the data processing unit.

Images