Layered closed-loop brain-like ai driven software-defined phased array adaptive implementation method

Abstract

The application discloses a layered closed-loop brain-like AI driven software phased array adaptive implementation method, and relates to the technical field of phased array antennas and artificial intelligence. The method constructs a four-level layered closed-loop architecture, including: physical layer instantaneous frequency locking and phase synchronization based on spin-orbit torque nano oscillator arrays; non-integer order thermal relaxation modeling and non-Markov noise suppression based on fractional order memristor chaotic systems; interference fingerprint extraction and topological robustness anti-interference based on topological data analysis (TDA) persistent homology; energy efficiency perception type pulse network structure reorganization based on meta-reinforcement learning neural architecture search (NAS); and cross-level optical-magnetic feedback closed loop based on microwave photonics and neuromorphic hybrid. Through collaborative coupling mechanisms such as fractional order modulation, TDA feature guided NAS, and optical bistable state elimination of phase jitter between levels, full-temperature-zone channel calibration, sub-microsecond level mode switching, and >=99% fault diagnosis accuracy are achieved.

Description

Technical Field

[0001] This invention belongs to the field of software-defined intelligent phased array antenna technology, specifically involving a software-based adaptive implementation method for phased array antenna functions based on neuromorphic AI closed-loop driving. It is primarily adapted to high-reliability satellite communication scenarios and can be extended to phased array systems in multiple fields such as radar detection, space-based telemetry and control, precision guidance, and airborne and shipborne mobile communications. This invention achieves intelligent reconfiguration, lightweight simplification, and adaptive upgrades to complex environments for traditional hardware phased arrays through the deep integration of neuromorphic intelligent algorithms and a fully digital array architecture, belonging to the category of next-generation software-defined antenna intelligent technology. Background Technology

[0002] Phased array antennas are the core radio frequency front-ends of high-end equipment such as satellite communication, microwave radar, and space telemetry and control. Their beamforming accuracy, environmental interference resistance, and system stability directly determine the overall performance indicators of communication transmission, target detection, and positioning and tracking. At present, mainstream commercial and military phased array systems all follow the traditional architecture of hardware-defined functions. Core functions such as beamforming, phase modulation, interference suppression, and error compensation rely entirely on physical hardware such as analog phase shifters, dedicated signal processing hardware boards, and fixed logic control circuits. This results in inherent defects such as low hardware integration, bulky structure, high overall cost, and large power consumption and size.

[0003] Traditional phased array antennas employ fixed hardware logic and open-loop control modes, with operating parameters, beam strategies, and anti-interference modes fixed at the factory, lacking the ability for autonomous learning and dynamic iteration. In complex application scenarios such as satellite communication and high-speed mobile vehicles, equipment faces harsh conditions such as extreme temperature variations of -40℃ to 85℃, multipath attenuation in complex terrain, strong electromagnetic suppression interference, and Doppler frequency shift during high-speed movement. Traditional phased arrays cannot dynamically adjust amplitude and phase parameters, beam pointing, and communication operating modes according to the real-time environmental situation, making them prone to problems such as beam misalignment, communication link interruption, excessive data transmission errors, and deterioration in positioning accuracy. They are unable to meet the all-weather, uninterrupted, high-precision, and highly reliable real-time communication and data transmission requirements of modern satellite communication systems.

[0004] Meanwhile, the existing equipment operation and maintenance and fault monitoring systems of phased array systems have significant technical shortcomings. Currently, fault diagnosis largely relies on external hardware monitoring modules combined with manual post-incident maintenance, resulting in extremely low levels of intelligence. Furthermore, the monitoring process requires the continuous transmission of massive amounts of raw equipment operation data and radio frequency signal characteristic data, placing an enormous load on channel transmission. The backend network management center must bear the burden of extremely high computing power for data parsing and filtering, leading to high system operation and maintenance complexity, transmission power consumption, and backend computing costs. In addition, traditional hardware-based phased arrays have fixed functions and are difficult to upgrade and iterate. Functional expansion can only be achieved by replacing hardware boards, and multi-task mode integrated reuse cannot be achieved, severely limiting the generalization and large-scale application of phased array equipment.

[0005] While existing technologies offer some improvements combining AI with phased arrays, they generally suffer from limitations in their specific technologies and fragmented system designs. Most only optimize single components, such as optimizing beamforming algorithms or slightly improving amplitude and phase calibration accuracy, without fundamentally altering the core architecture dependent on traditional hardware or achieving software-based replacement of the entire phased array functionality. Furthermore, existing AI improvement solutions often employ conventional shallow neural network models like CNN, LSTM, and GRU, lacking brain-like temporal learning, autonomous iteration, and nonlinear error fitting capabilities. They also fail to address core pain points such as adaptive control in harsh environments, self-diagnosis of front-end edge faults, lightweight compression of ultra-high-proportion data, and multi-task closed-loop optimization through systematic architectural design.

[0006] Existing solutions cannot simultaneously address a series of technical challenges faced by traditional phased arrays, such as large hardware redundancy, poor environmental adaptability, low level of intelligence, cumbersome fault maintenance, high data transmission pressure, and weak functional scalability. They also lack a complete technical system that features layered closed-loop design, autonomous iteration, hardware-software decoupling, and full-domain software integration, making it difficult to adapt to the application requirements of next-generation satellite communication and high-precision intelligent phased arrays in multiple scenarios.

[0007] To address the numerous shortcomings of existing technologies, there is an urgent need for a phased array implementation method that abandons the traditional fixed hardware architecture, is based on brain-like intelligent closed-loop drive, and achieves full-process software-based reconstruction. Through a brand-new heterogeneous fusion architecture and intelligent decision-making mechanism, it is possible to upgrade intelligent phased array technology with minimal hardware, environmental adaptability, fault self-diagnosis, low-load transmission, and versatility for multiple scenarios. Summary of the Invention

[0008] This invention aims to address the core technological shortcomings of existing traditional hardware-based phased array antennas. It provides a hierarchical, closed-loop, brain-like AI-driven software-based adaptive phased array implementation method to address numerous deficiencies such as high hardware dependence, bloated and costly equipment, fixed control modes, weak adaptability to harsh environments, low intelligence in fault operation and maintenance, large amount of raw data transmission, and poor system scalability.

[0009] To achieve the above-mentioned objectives, this invention is implemented through the following computational scheme: a hierarchical closed-loop brain-like AI-driven software-based phased array adaptive implementation method, characterized by constructing a four-level hierarchical closed-loop architecture consisting of bottom-level hardware execution, mid-level intelligent computing power, and top-level overall planning, integrating a fully digital computing array and a brain-like pulse intelligent model, including the following steps: Step 1: Construct a heterogeneous neuromorphic hybrid signal architecture based on a spin orbital moment nano oscillator (STNO) array. In the array element-level fully independent transmit / receive digital architecture, a sub-nanosecond pulse generation path driven by magnetization dynamics is embedded. Utilizing the sensitivity of the STNO's self-sustaining oscillation frequency to externally injected current, instantaneous physical layer frequency locking and phase synchronization based on the spin-torque effect are achieved in the RF link, replacing traditional phase-locked loops and digital phase shifters. This enables a fully software-based process for RF signal reception, preprocessing, intelligent analysis, digital beamforming, and signal transmission. The STNO array adopts a multi-domain wall-coupled spin valve stacking structure, with two STNO differential pairs within the same array element oscillating in opposite directions to construct a spin differential mode phase reference. This eliminates the intrinsic 1 / f low-frequency phase noise of a single STNO device, forming a native differential phase reference source that does not depend on an external clock. Step 2: Establish a multi-scale environment-hardware joint evolution model based on a fractional-order memristor chaotic system. Introduce the Caputo-Fabrizio fractional derivative operator to describe the non-integer order thermal relaxation behavior of the array antenna under extreme temperature variations of -40℃ to 85℃. Combine this with the hardware-in-the-loop fractional-pulse timing-dependent plasticity (fractional-STDP) learning rule to couple and model non-Markovian noise caused by strong electromagnetic suppression interference. Generate a nonlinear fractional-order dynamic zeroing strategy for transient interference through dynamic adaptive adjustment of the fractional order. Construct a coupled dynamic equation between the fractional-order α and the STNO domain wall migration rate, mapping the continuous change in memristor conductance to the effective field strength of the spin orbital moment. Achieve unified modeling of the coupled dynamics of temperature, memristor chaos, and spin oscillation, breaking the existing separation architecture of fractional-order models and independent decoupled calculations of spin devices. The existing general STNO basic model (LLG standard equations, decoupled without cross-domain correlation) fractional-order coupled dynamic equations and the standard Landau-Liffsitz-Gilbert equations (LLG formulas) are as follows: In the formula: is the normalized magnetization vector; is the gyromagnetic ratio; is the Gilbert damping coefficient; is the spin transmission force driving coefficient; is the equivalent magnetic field; is the injected current density. The above-mentioned prior art only describes the evolution of the magnetization vector separately, without relating it to the temperature fractional-order operator and memristor conductance parameters; the physical quantities are independent of each other.

[0010] Step 2 of this invention: The coupled dynamic equation between the fractional order α and the STNO domain wall migration rate, constructed from the spin orbit moment-fractional memristor coupling dynamic equation (Caputo-Fabrizio fractional form), is as follows: In the formula: is the Caputo-Fabrizio variable-order fractional differential operator. α∈[0.85,1.15], updated in real time by the antenna array temperature gradient and the vibration Hurst exponent; is the real-time conductance value of the TiO2 memristor, directly mapping the output state of the fractional-order chaotic system; is the effective field of the additional spin orbital moment generated by memristor conductance modulation; is the STNO array injection current density controlled by fractional order; and the domain wall migration velocity can be directly solved by fractional-order differential results. This equation couples the fractional-order operator of temperature evolution, memristor chaotic conductance, and spin magnetization dynamics into a unified set of equations, breaking the limitation of the existing LLG equations where variables are completely decoupled independently.

[0011] Step 3: Integrate a persistent coherence anti-interference mechanism based on topology data analysis (TDA) into the fractional-order memristor chaotic system. Utilize the Vietoris-Rips complex to perform persistent coherence group calculations on the time-frequency phase space point cloud of the enemy's intentional narrowband interference signal. Extract the longest survival interval from the H1 persistent barcode as the interference fingerprint feature. Generate an anti-interference beamforming weight matrix using a topology optimization algorithm based on Morse theory to achieve topological robustness anti-interference of the communication link in high-speed mobile Doppler frequency offset scenarios. Introduce a layered threshold persistent coherence layered filtering mechanism to distinguish between short-lifetime topological features of spurious noise and long-period stable topological loops of man-made interference. Synchronously output the gradient flow of interference topological features to the fractional-order memristor chaotic system to directly correct the fractional-order α, establishing a real-time bidirectional linkage closed loop of topological anti-interference and thermal relaxation compensation. Step 4: Construct a neuromorphic architecture search (NAS) scheduler based on meta-reinforcement learning (Meta-RL) to monitor the leakage current, power consumption, and heat dissipation boundary conditions of the FPGA parallel heterogeneous computing architecture in real time. When the computing load is detected to be approaching the thermal runaway critical point, a spiking neural network structure reorganization mechanism based on an evolutionary strategy is triggered. Through a genetic algorithm, the neuron connection topology, spiking threshold, and synaptic delay are evolved online to automatically generate an event-driven sparse spiking network that meets energy efficiency constraints. This ensures the microsecond-level mode switching and long-term stability of the system during concurrent multi-service operations such as satellite communication, radar detection, and multi-target collaborative detection. A two-layer meta-learning reward composite function is designed. The first layer is the TDA topology anti-interference survival interval weight, and the second layer is the STNO differential phase noise attenuation coefficient. The dual weights are dynamically normalized and used as the fitness evaluation standard for the evolutionary algorithm. This constrains the network evolution while taking into account the three indicators of anti-interference, phase synchronization, and power consumption, thus solving the conflict problem of the existing Meta-RL's single power consumption optimization leading to the degradation of radio frequency synchronization performance. Step 5: Establish a cross-level closed-loop feedback mechanism based on a hybrid approach of microwave photonics and neuromorphology. The RF air interface electromagnetic field distribution data and hardware state parameters of the bottom-level software-defined antenna function output layer are converted into optical carrier phase information via a Mach-Zehnder modulator. This information is directly mapped to the resonant wavelength offset of the silicon-based microring resonator array in the top-level brain-like intelligent phased array coordination platform. Through distributed asynchronous updates via optical interconnects, a dynamic optimal balance between beamforming accuracy and hardware energy efficiency is achieved. The silicon-based microring resonator array employs a multi-channel coupled resonant cavity parallel phase demodulation structure, converting the optical carrier phase offset into a discrete optical pulse sequence. This asynchronous optical pulse flow directly drives the domain wall modulation current of the STNO differential pair, forming a three-level hardware-level native closed-loop feedback system (optical-memristor-spin), eliminating the need for intermediate buffers in digital-to-digital conversion and compressing the cross-level feedback delay to within 200 ps.

[0012] Furthermore, there is a synergistic coupling between the spin orbital moment nano oscillator (STNO) array in step 1 and the fractional-order memristor chaotic system in step 2: the fractional-order α output by the fractional-order memristor chaotic system modulates the injection current density of the STNO array in real time, thereby changing the self-sustaining oscillation frequency and locking bandwidth of the STNO. This allows the synchronization accuracy of the STNO to automatically match the local slope of the fractional-order thermal relaxation curve as the ambient temperature changes, thus overcoming the contradiction between locking speed and phase noise that traditional phase-locked loops cannot balance in a wide temperature range. This achieves stable channel calibration accuracy within ±0.5ps across the entire temperature range without the need for any external reference clock. Based on the coupling dynamics equation, the optimal injection current density is solved in real time, dynamically matching the domain wall reverse migration rate of the dual STNO differential pair. This allows the differential phase reference noise suppression depth to be adaptively adjusted with temperature, improving noise suppression by 18dB at -40℃ and by 12dB at 85℃, producing a wide-temperature differential noise reduction synergistic effect that cannot be achieved by single current control in existing technologies.

[0013] There is a synergistic coupling between the persistent coherence anti-interference mechanism of Topology Data Analysis (TDA) in step 3 and the neuromorphic architecture search (NAS) scheduler of Meta-Reinforcement Learning (Meta-RL) in step 4: the longest survival interval length of the H1 persistent barcode extracted by TDA is used as part of the environmental reward signal of the meta-learner, guiding the NAS scheduler to select and retain neuronal subnetworks with high topological robustness in the evolutionary strategy, while discarding redundant connections unrelated to persistent coherence features. This makes the generated sparse pulse network reduce power consumption by more than 70% compared with traditional SNN while maintaining anti-interference performance, and the mode switching time is reduced to the sub-microsecond level. The meta-learner performs temporal sliding window normalization on the topological features, distinguishes between instantaneous pulse interference and continuous suppression interference, and generates two independent sparse network topology templates for the two types of interference. At the hardware level, it can switch in parallel without buffering. In multi-service concurrent scenarios, the beam reconstruction latency is further reduced to 300ns, and it has the ability to quickly switch between dual topology templates for multiple types of interference, which is not available in the existing technology.

[0014] The cross-level closed-loop feedback mechanism combining microwave photonics and neuromorphology in step 5 is synergistically coupled with the STNO array in step 1: the resonant wavelength offset of the silicon-based microring resonator array is converted into optical pulse delay difference through an opto-oscillator (OEO), which is injected in reverse into the bias current port of the STNO array to form a hybrid optical-magnetic phase-locked loop. The optical bistable effect is used to eliminate the phase jitter caused by thermal noise in the STNO at extreme temperatures, so that the system can maintain a beam pointing error ≤0.01° in the full temperature range of -40℃ to 85℃ without the need for a temperature compensation algorithm. The optical bistable hysteresis range of the multi-cavity silicon-based microring is used to construct an optical delay difference threshold criterion. The STNO current correction is triggered only when the phase jitter exceeds the threshold, avoiding frequent iterative adjustments caused by small temperature fluctuations, and significantly reducing the overall dynamic power consumption of the array. Compared with the continuous real-time compensation scheme, the power consumption is reduced by an additional 22%.

[0015] Furthermore, the Topology Data Analysis (TDA) anti-interference mechanism in step 3 also integrates a fault diagnosis branch based on causal inference. This branch uses the Peter-Clark (PC) algorithm to learn the causal graph from the time-delayed embedding vector of the chaotic attractor trajectory. Through Do-calculus, it performs interventional causal inference on three types of faults: channel failure, amplitude and phase drift, and radio frequency loss. It only returns the fault category and the corresponding causal effect value, achieving a data compression ratio of ≥98% while maintaining a fault identification accuracy of ≥99%. It can also distinguish between correlation and causality, avoiding false alarms caused by environmental noise. The fault causal effect value is synchronously input into the fractional-order memristor chaotic system to directly correct the upper and lower limits of the fractional-order α, realizing adaptive degradation compensation of the thermal relaxation model under fault conditions. The fault channel can dynamically reduce the beamforming weight, forming an integrated linkage closed loop of interference identification, fault diagnosis, and thermal compensation. In the prior art, fault detection and thermal compensation are independent and have no linkage logic.

[0016] The fractional-order memristor chaotic system described in step 2 also integrates an adaptive fractional-order order adjustment module. This module dynamically adjusts the fractional-order order α based on the real-time acquired antenna array temperature gradient and the Hurst exponent of the vibration spectrum. This allows the model to reduce its order to 0.85 at low temperatures (-40℃) to enhance long-range memory compensation for decreased electron mobility, and to increase its order to 1.15 at high temperatures (85℃) to suppress false triggering caused by thermal noise. This achieves adaptive maintenance of beam pointing accuracy across the entire temperature range, with a pointing error ≤0.01°. By introducing vibration-temperature coupled Hurst joint characteristic quantities, a two-dimensional lookup table is constructed to dynamically constrain the fractional-order order adjustment step size. Under high vibration and high temperature combined conditions, the order adjustment step size automatically narrows, suppressing frequent order oscillations caused by the superposition of vibration noise and thermal noise. Under combined harsh conditions, the beam error fluctuation amplitude is reduced by 65%.

[0017] Furthermore, the specific components and linkage mechanism of the four-level hierarchical closed-loop architecture are as follows: The underlying software-defined antenna function output layer is equipped with a spin orbital moment nano oscillator (STNO) array and a silicon-based photonic integrated transceiver module. As a hardware execution and data feedback terminal, it undertakes multi-task signal transmission and reception work and transmits back hardware status, task data, and environmental parameters in real time. The underlying hardware adopts a differential dual STNO spin valve stacked array, matched with a parallel multi-channel silicon-based micro-ring optical demodulation unit. The native hardware differential architecture provides a unified phase reference for the upper four-level closed loop. The middle sensing input layer is equipped with a fractional-order capacitance equivalent circuit based on TiO2 memristors and a microwave photonics downconversion module, which seamlessly interfaces with the RF front-end array antenna to complete fractional-order domain acquisition and adaptive noise reduction preprocessing of air interface electromagnetic signals. The TiO2 memristor circuit has a built-in fractional-order dynamic zero-adjustment hardware operation unit, which directly solves the Caputo-Fabrizio fractional derivative in hardware, eliminating the need for FPGA software iterative calculation and reducing the preprocessing latency by an order of magnitude. The middle-layer neuromorphic AI computing layer is equipped with a fractional-order memristor chaotic system hardware accelerator and a topology data analysis (TDA) coprocessor, serving as the core intelligent decision-making hub. It completes persistent coherence feature extraction, causal inference, and iterative output of the optimal solution for beam strategy. The TDA coprocessor is directly connected to the fractional-order memristor accelerator hardware, and the on-chip bus directly interacts with topology features and fractional-order parameters without external DDR cache relay, realizing real-time bidirectional linkage between topology anti-interference and thermal model. The middle beam control layer is equipped with a topology optimization algorithm module based on Morse theory and an optical interconnect beamforming controller, which works in conjunction with the beamforming module and the amplitude and phase calibration compensation module to achieve topological robust dynamic and precise adjustment of beam pointing, beamwidth, and null position. The optical interconnect beamforming controller has a built-in optical pulse delay analysis circuit, which directly outputs analog bias current to drive the bottom STNO array, skipping the digital DA conversion stage. The top-level brain-like intelligent phased array coordination platform is equipped with a Meta-Reinforcement Learning (Meta-RL) neuromorphic architecture search (NAS) scheduler and a silicon-based microring resonator array. It uniformly manages all functional levels below, enabling seamless sub-microsecond switching between communication / radar / anti-interference working modes. The top-level NAS scheduler and the silicon-based microring array are bidirectionally interconnected by optical signals. Optical pulses serve as cross-level scheduling commands, replacing the traditional high-speed digital bus, and increasing the throughput of multi-service concurrent scheduling by 3 times.

[0018] The mid-layer neuromorphic AI computing layer also integrates an on-chip causal discovery engine. This engine utilizes the synchronization phenomenon of spin orbital moment nano oscillators (STNOs) to solve the Ising model annealing problem. It learns the causal structure from real-time data through a hardware-accelerated PC algorithm without the need for external memory, thus achieving zero-wait causal model hot migration and scene adaptation when switching between multiple scenarios such as satellite communication, airborne, shipborne, and vehicle-mounted applications. It reuses the dual STNO differential oscillation synchronization signal as the global synchronization clock for Ising annealing, synchronizing the hardware timing of causal graph solving, topological feature extraction, and fractional-order chaotic modeling. When switching between multiple scenarios, the causal model, anti-interference topology, and thermal compensation model are updated synchronously, eliminating the time difference error caused by serial updates of multiple models.

[0019] Furthermore, the spiking neural network structure recombination mechanism based on evolutionary strategy described in step 4 specifically employs a genetic algorithm combining tournament selection and elite retention strategies. In each evolutionary iteration, the top 10% of individuals with the highest fitness are retained, and crossover and mutation operations are performed on the remaining individuals. The crossover and mutation probabilities are dynamically adjusted by the meta-learner according to the complexity of the current electromagnetic environment, which improves the convergence speed of network structure recombination by more than 5 times compared to the fixed-parameter genetic algorithm. Triple constraint evolutionary selection rules are set: lower limit of topology anti-interference index, upper limit of STNO differential phase noise, and array power consumption threshold. Individuals are directly eliminated if any one of them fails to meet the standard. The evolutionary process simultaneously constrains the performance of radio frequency hardware, avoiding the sacrifice of radio frequency synchronization and anti-interference capabilities in pursuit of computing power and energy efficiency in network optimization.

[0020] Furthermore, the neuromorphic software-defined phased array system implemented by the method includes: The top-level brain-like intelligent phased array coordination platform is equipped with a meta-reinforcement learning (Meta-RL) neuromorphic architecture search (NAS) scheduler and a silicon-based microring resonator array; the silicon-based microring resonator array adopts a multi-cavity optical bistable parallel demodulation structure, and constructs an all-optical cross-layer feedback link with the bottom-level STNO differential array. The middle layer of sensing input is configured with a fractional-order capacitor equivalent circuit based on TiO2 memristors and a microwave photonics downconversion module; the TiO2 memristor fractional-order circuit integrates a hardware fractional-order derivative operation unit, which is directly connected and interconnected with the TDA coprocessor on-chip. The middle-layer neuromorphic AI computing layer is equipped with a fractional-order memristor chaotic system hardware accelerator and a topology data analysis (TDA) coprocessor; the fractional-order memristor accelerator, the TDA coprocessor, and the on-chip causal discovery engine share the STNO differential synchronization global clock; The middle beam control layer is equipped with a topology optimization algorithm module based on Morse theory and an optical interconnect beamforming controller; the output of the optical interconnect beamforming controller is an analog optical control current that is directly connected to the STNO array bias port. The bottom layer is a software-defined antenna function output layer, which is equipped with a spin orbital moment nano oscillator (STNO) array and a silicon-based photonic integrated transceiver module. The bottom STNO array is a dual-device differential spin valve stacked structure, which forms a native closed-loop hardware path of optical-magnetic coupling with the top silicon-based micro-ring array.

[0021] This invention abandons the traditional architecture of all-digital arrays that rely on conventional hardware such as phase-locked loops and digital phase shifters, and innovatively embeds a spin orbital moment nano oscillator (STNO) array at the array element level. Utilizing the self-sustaining oscillation characteristics of STNOs and the negative damping generated by the spin-shifting torque effect, self-injection locking synchronization is achieved. Channel-to-channel amplitude and phase calibration can be completed without any external reference clock, with calibration accuracy consistently within ±0.5 ps across the entire temperature range, far superior to traditional solutions (typically ±5 ps). This breakthrough stems from the fact that the magnetization dynamics response speed of STNOs (sub-nanosecond level) is much faster than that of traditional charge-type devices, fundamentally eliminating clock tree jitter and temperature drift problems. Simultaneously, this invention introduces a fractional-order memristor chaotic system to replace the traditional integer-order dynamic model. The Caputo-Fabrizio fractional-order derivative operator accurately describes the non-integer-order thermal relaxation behavior of antenna elements under extreme temperature variations of -40℃ to 85℃, and combined with the fractional-order pulse timing-dependent plasticity (fractional-STDP) learning rule, accurate modeling and dynamic suppression of non-Markovian noise are achieved. Compared with traditional integer-order models, fractional-order models can capture the long-range memory effect during material relaxation, reduce beam pointing error to ≤0.01°, and do not require additional temperature compensation algorithms across the entire temperature range.

[0022] This invention breaks through the traditional anti-interference paradigm based on energy detection or statistical features (such as FFT and wavelet transform), and introduces the Topology Data Analysis (TDA) method. It utilizes the Vietoris-Rips complex to perform persistent homology group calculation on the time-frequency phase space point cloud of the received signal, extracting the longest survival interval in the H1 persistent barcode as the interference fingerprint feature. This method can accurately identify various intentional interference types such as narrowband interference, frequency sweeping interference, and impulse interference under extremely low signal-to-noise ratios (SNR < -10dB). Furthermore, based on Morse theory, a topology optimization algorithm generates an anti-interference beamforming weight matrix, reducing the communication interruption tolerance time to ≤2ms and the data transmission error to ≤0.1%, representing an improvement of more than an order of magnitude compared to traditional methods.

[0023] Furthermore, this invention introduces the Peter-Clark (PC) algorithm based on causal inference into the fault diagnosis stage. It uses do-calculus to perform interventional causal inference on three types of faults—channel failure, amplitude and phase drift, and RF loss—rather than the correlation analysis in traditional machine learning. This mechanism can distinguish between "correlation" and "causation," avoiding false alarms caused by environmental noise, achieving a fault identification accuracy of ≥99% and a data compression ratio of ≥98%. The fault causal numerical values ​​inversely constrain the upper and lower limits of the fractional order, forming an integrated closed loop of "fault diagnosis-thermal compensation-beamweighting." The causal inference results are applied inversely to the underlying hardware model, surpassing the traditional mode of separate compensation after a conventional fault alarm. This introduction of causal inference into the field of phased array health management fundamentally solves the problem of the lack of interpretability in traditional AI fault diagnosis.

[0024] This invention abandons the traditional fixed-topology spiking neural network (SNN) design and pioneers a Meta-Reinforcement Learning (Meta-RL) driven neuromorphic architecture search (NAS) scheduler. This scheduler employs a Model-Independent Meta-Learning (MAML) algorithm for rapid adaptive training on multiple tasks in different electromagnetic environments. When the FPGA computing load is detected to be approaching the thermal runaway threshold, an evolutionary spiking neural network structure reorganization mechanism is triggered. Through a genetic algorithm, the neuron connection topology, spiking threshold, and synaptic delay are evolved online, automatically generating an event-driven sparse spiking network that meets energy efficiency constraints. This mechanism enables the system to reduce mode switching time to sub-microsecond levels during concurrent multi-service operations such as satellite communication, radar detection, and multi-target collaborative detection, while consuming only 30% of the power of traditional FPGA solutions.

[0025] Unlike traditional DVFS (Dynamic Voltage Frequency Scaling) or model pruning methods, this invention not only adjusts operating parameters but also dynamically reconstructs the network topology, preserving neuron clusters with high topological importance against interference and shielding background noise activities with low persistence coherence significance, thus achieving Pareto optimality in "accuracy-energy efficiency".

[0026] This invention innovatively introduces microwave photonics technology, converting the underlying radio frequency signal into optical carrier phase information through a Mach-Zehnder modulator. Utilizing the optical bistable effect of a silicon-based microring resonator array, it directly maps the optical pulse delay difference to the global information flow propagation rate of the top-level pulse neural network. This mechanism completely breaks through the bandwidth limitations of traditional electrical interconnects (typically <100GHz), achieving cross-level information transmission with terahertz bandwidth without any analog-to-digital conversion, thus avoiding sampling and quantization errors. Simultaneously, the introduction of an opto-oscillator (OEO) results in extremely low phase noise in the feedback loop (<-140dBc / Hz@10kHz), fundamentally resolving mode contention issues during concurrent multi-service switching and reducing switching time to sub-microsecond levels. This architecture provides a feasible engineering path for future terahertz communication and millimeter-wave phased arrays.

[0027] This invention, based on a four-level hierarchical closed-loop architecture, integrates an on-chip causal discovery engine in the middle-level neuromorphic AI computing layer. It utilizes the synchronization phenomenon of spin orbital moment nano oscillators (STNOs) to achieve annealing solutions for the Ising model, and learns the causal structure from real-time data through a hardware-accelerated PC algorithm. This enables the system to achieve zero-wait hot migration and scenario adaptation of causal models without the need for external memory when switching between multiple scenarios such as satellite communication, airborne, shipborne, and vehicle-mounted systems, reducing the deployment cycle from weeks in traditional solutions to minutes.

[0028] Compared to traditional solutions that require separate model training and hardware firmware changes for each scenario, this invention achieves universal adaptation of "train once, deploy in multiple locations" by leveraging the universality of topological data analysis (persistent cohomology features are naturally invariant to different scenarios) and the rapid adaptability of meta-learning. This reduces hardware usage by more than 60% and lowers maintenance costs by 85%.

[0029] The three-level unified coupling dynamics model of spin-memristor-photon, the direct hardware connection bidirectional closed loop, and the multi-module timing synchronization constraints form a synergistic mechanism of deep coupling of multiple physical fields, resulting in multiple synergistic gain effects: wide-temperature differential noise reduction, significant reduction in beam error under harsh conditions, cross-layer feedback delay compression, secondary reduction in power consumption under complex conditions, and parallel switching of multiple interference templates. This solves the defect of "the architecture is simply divided into layers without dedicated hardware structures." A dedicated hardware structure is provided for the four-level layered architecture: differential dual STNO array, hardware fractional-order arithmetic unit, all-optical cross-layer scheduling bus, and shared spin synchronization clock, forming a dedicated hardware and software integrated architecture. This solves the defect of "optimization algorithms optimizing only a single indicator, resulting in performance conflicts." The triple-constraint meta-learning composite reward function and evolutionary selection rules simultaneously consider three mutually constraining indicators: anti-interference, phase synchronization, and power consumption, overcoming the performance degradation conflicts caused by single-dimensional optimization in existing technologies. The coupled dynamic equations of fractional-order α and STNO domain wall migration rate achieve a unified modeling of temperature-memristor chaos-spin oscillation, breaking the decoupled computational mode of existing technologies. A dual composite reward function is designed to simultaneously consider three mutually constraining indicators: phase noise, anti-interference, and power consumption, addressing the industry pain point of RF performance degradation caused by single power consumption optimization. A three-stage closed loop, where optical pulses directly drive spin domain wall current regulation and the intermediate ADC / DAC stage is discarded, surpasses the conventional modification approach of existing microwave photonics + phased array fusion. The differential structure brings wide-temperature-range differentiated noise reduction (18dB at low temperatures and 12dB at high temperatures). After linking topological features with the fractional-order model, beam error fluctuation is reduced by 65% ​​under harsh operating conditions. Removing the analog-digital buffer compresses the feedback delay to 200ps, an order of magnitude improvement compared to traditional architectures. A triple evolutionary selection rule balances power consumption, anti-interference, and phase accuracy, overcoming the shortcomings of previous technologies that compromised certain aspects. Attached Figure Description

[0030] Figure 1 This invention presents a hierarchical closed-loop brain-like AI-driven software-based adaptive phased array implementation method, which uses a brain-like (AI) model to realize the overall architecture of software-based phased array antenna functions. Figure 2 : Schematic diagram of the fusion architecture of the neuromorphic (AI) model and digital computing array of this invention; Figure 3 : A schematic diagram of the AI ​​fault identification and data compression process of this invention; Figure 4 : Comparison chart of communication stability tests under harsh environments in embodiments of the present invention; Figure 5 : A schematic diagram comparing data transmission volume in an embodiment of the present invention.

[0031] The present invention will be further described in detail below with reference to specific embodiments. Detailed Implementation

[0032] See Figure 1 and Figure 2 . Example

[0033] This embodiment takes the application of phased array antennas in the satellite communication field as an example, and presents a hierarchical closed-loop brain-like AI-driven software-based adaptive phased array implementation method. It adopts a four-level hierarchical closed-loop architecture consisting of bottom-level hardware execution, mid-level intelligent computing power, and top-level overall planning, integrating a fully digital computing array and a brain-like pulse intelligent model. The method includes the following steps: Step 1: Construct a heterogeneous neuromorphic hybrid signal architecture based on a spin orbital moment nano oscillator (STNO) array. Embed a sub-nanosecond pulse generation path driven by magnetization dynamics within the array element-level fully independent transmit / receive digital architecture. Utilizing the sensitivity of the STNO's self-sustaining oscillation frequency to externally injected current, instantaneous physical-layer frequency locking and phase synchronization based on the spin-torque effect are achieved in the RF link, replacing traditional phase-locked loops and digital phase shifters. This enables a fully software-based workflow for RF signal reception, preprocessing, intelligent analysis, digital beamforming, and signal transmission. The STNO array employs a multi-domain wall-coupled spin valve stacking structure, with dual STNO differential pairs within the same array element exhibiting anti-sustaining oscillations to construct a spin differential-mode phase reference. The layers of a single STNO device, from bottom to top, are: Ta substrate buffer layer (5 nm thick) / Pt spin Hall layer (8 nm thick) / CoTb ferrimagnetic free layer (1.6 nm thick) / MgO barrier layer (0.8 nm thick). The array consists of a CoFeB reference pinning layer (2.0 nm thick), a Ru coupling layer (0.4 nm thick), and an IrMn antiferromagnetic pinning layer (10 nm thick). The two STNO devices in the same array element are symmetrical structures, with their free layer domain walls having a preset initial magnetization direction that differs by 180°, forming a differential pair. The array element unit has a lateral size of 800 nm × 800 nm, and the center-to-center spacing between adjacent array elements is 1.2 μm, forming an array arrangement that eliminates the intrinsic 1 / f low-frequency phase noise of a single STNO device, forming a native differential phase reference source that does not depend on an external clock.

[0034] The underlying software-defined antenna function output layer is equipped with a spin orbital moment nano oscillator (STNO) array and a silicon-based photonic integrated transceiver module. As a hardware execution and data feedback terminal, it undertakes multi-task signal transmission and reception work and transmits back hardware status, task data, and environmental parameters in real time. The middle sensing input layer is equipped with a fractional-order capacitance equivalent circuit based on TiO2 memristors and a microwave photonics downconversion module, which is seamlessly connected to the RF front-end array antenna to complete the fractional-order domain acquisition and adaptive noise reduction preprocessing of the air interface electromagnetic signal. The middle-layer neuromorphic AI computing layer is equipped with a fractional-order memristor chaotic system hardware accelerator and a topology data analysis (TDA) coprocessor, serving as the core intelligent decision-making center to complete persistent coherence feature extraction, causal inference, and iterative output of the optimal solution for beam strategy. The middle beam control layer is equipped with a topology optimization algorithm module based on Morse theory and an optical interconnect beamforming controller, which links the beamforming module and the amplitude and phase calibration compensation module to achieve topological robust dynamic and precise adjustment of beam pointing, beamwidth, and null position. The top-level brain-like intelligent phased array coordination platform is equipped with a Meta-Reinforcement Learning (Meta-RL) neuromorphic architecture search (NAS) scheduler and a silicon-based microring resonator array, which uniformly manages all functional levels below, enabling seamless sub-microsecond switching between communication / radar / anti-interference working modes.

[0035] The top-level brain-like intelligent phased array coordination platform is equipped with a meta-reinforcement learning (Meta-RL) neuromorphic architecture search (NAS) scheduler and a silicon-based microring resonator array; The middle layer sensing input layer is equipped with a fractional-order capacitor equivalent circuit based on TiO2 memristors and a microwave photonics downconversion module. The middle-layer neuromorphic AI computing layer is equipped with a fractional-order memristor chaotic system hardware accelerator and a topology data analysis (TDA) coprocessor; The middle beam control layer is equipped with a topology optimization algorithm module based on Morse theory and an optical interconnect beamforming controller. The underlying software-defined antenna function output layer is equipped with a spin orbital moment nano oscillator (STNO) array and a silicon-based photonic integrated transceiver module. Example

[0036] Based on Embodiment 1, a hierarchical closed-loop brain-like AI-driven software-based adaptive phased array implementation method is provided. This method employs a four-level hierarchical closed-loop architecture: top-level overall planning, mid-level intelligent computing power, and bottom-level hardware execution. It integrates a fully digital computing array with a brain-like pulse intelligent model, abandoning the traditional phased array design paradigm of "hardware phase shifting, fixed parameters, and open-loop control." Through full-process software definition, it achieves lightweight, anti-interference, self-diagnostic, and adaptive operation of the phased array antenna. The method may include the following steps: Step 1: Build a heterogeneous fusion architecture of brain-like AI computing power cluster and all-digital computing array, construct an integrated signal processing platform of perception-decision-control-feedback, completely replace the traditional phased array core analog hardware link based on pure software algorithm, and open up the whole software process of radio frequency signal reception, preprocessing, intelligent analysis, digital beamforming and signal transmission, simplify the hardware architecture and reduce system complexity and power consumption from the architecture level. Step 2: Embed a brain-like intelligent decision-making model that integrates deep reinforcement learning. Based on a massive dataset of multi-dimensional samples in harsh environments, complete offline pre-training and online fine-tuning of the model. This enables self-awareness of environmental situation, self-iteration of beam parameters, and self-switching of communication modes in complex electromagnetic and extreme environments. It breaks through the technical bottleneck that the fixed parameters of traditional phased arrays cannot adapt to dynamic scenarios, and ensures uninterrupted real-time business transmission around the clock. Step 3: Integrate fault identification and edge lightweight compression algorithms to complete fault feature extraction, fault level classification and accurate location at the front-end device. Only structured fault warning instructions are sent back instead of raw monitoring data, achieving ultra-high compression ratio transmission of front-end edge data, which greatly reduces the computing load and channel transmission pressure of the back-end management platform. Step 4: Based on the parallel heterogeneous computing power of FPGA, complete the algorithm integration and parameter iterative optimization of the whole system. Relying on the bidirectional closed-loop feedback mechanism of the four-level hierarchical architecture, realize the integrated adaptive adaptation of communication, radar and multi-target tasks, complete the generalized adaptation and deployment of satellite communication, airborne, shipborne and vehicle-mounted scenarios, and form a new phased array working system with extremely simple hardware, reconfigurable software and intelligent self-iteration.

[0037] Furthermore, the brain-like AI computing power cluster described in step 1 adopts a fusion architecture of spiking neural network (SNN) and deep reinforcement learning (DRL) to replace the traditional single GRU, LSTM, and CNN shallow networks, and constructs a brain-like intelligent model with temporal memory, environmental adaptive iteration, and nonlinear error compensation. The all-digital computing array adopts an array element-level, fully independent digital transceiver architecture. Each antenna array element corresponds to a set of independent digital transceiver channels. All beamforming, amplitude and phase weighting calculations, and channel calibration operations are completed in the pure digital domain, completely eliminating the traditional analog phase-shifting link.

[0038] The traditional dedicated hardware replaced by the pure software algorithm in step 1 includes analog phase shifters, independent dedicated signal processing boards, and fixed anti-interference hardware modules, with a hardware simplification rate of no less than 60%. Relying on the real-time nonlinear computing capabilities of the neuromorphic AI model, it realizes three software-based functions: dynamic adaptive zeroing, intelligent waveform diversity, and online channel error compensation. Compared with the traditional fixed hardware control mode, the system's anti-interference suppression ratio is improved by more than 30dB, and the signal processing throughput efficiency is improved by 40%.

[0039] Furthermore, the harsh and complex scenarios to be adapted in step 2 include extreme temperature variations of -40℃ to 85℃, strong electromagnetic suppression interference, complex multipath attenuation in mountainous / sea areas, and high-speed moving Doppler frequency offset scenarios. The brain-like fusion model was pre-trained using more than 100,000 sets of labeled samples with multiple scenarios, multiple interferences, and multiple temperature variations, and it supports real-time collection of environmental data during online operation of the device for incremental fine-tuning of the model. After adaptive adjustment of the model, the system communication interruption tolerance time is ≤10ms, the data transmission error is ≤0.5%, the beam pointing calibration accuracy is improved by 50%, and the system achieves adaptive operation with zero human intervention in complex and extreme environments.

[0040] The fault identification algorithm described in step 3 above adopts a front-end edge computing architecture that combines lightweight model pruning with hardware acceleration for collaborative optimization, enabling it to independently complete end-to-end fault diagnosis without the involvement of the back-end platform. Fault identification covers all types of faults, including channel failure, amplitude and phase drift, RF loss, and module malfunction, with an accuracy of ≥95%. By adopting a feature-level intelligent screening and structured data reconstruction and compression mechanism, the traditional raw data compression method is abandoned. The compression ratio of raw monitoring data and back-transmitted early warning structured data is >90%, reducing the data processing pressure of the back-end network management center by more than 85%, and realizing high-speed fault early warning transmission under narrowband channels.

[0041] In step 4, the FPGA parallel heterogeneous computing architecture supports multi-channel signal synchronous parallel operation, and the multi-channel amplitude and phase synchronization calibration accuracy reaches ±5ps. Leveraging the reconfigurable nature of the layered closed-loop architecture, satellite communication, radar detection, and multi-target collaborative detection modes can be dynamically switched by updating model parameters and software logic online, without changing the hardware structure, thus achieving integrated reuse of multiple services on a single device.

[0042] Furthermore, the method is primarily applied to the field of high-reliability satellite communication, enabling miniaturization, low-cost, and intelligent upgrades of satellite communication, high-speed data transmission, and high-precision positioning equipment. It can also be extended to airborne, shipborne, and vehicle-mounted radar phased array system application scenarios through software parameter reconfiguration.

[0043] The aforementioned four-level hierarchical closed-loop architecture is an original closed-loop system with top-down hierarchical linkage and bottom-up feedback iteration. It consists of a top-level brain-like intelligent phased array coordination platform, a middle-level perception input layer, a brain-like AI computing layer, a beam control layer, and a bottom-level software-defined antenna function output layer, forming a bidirectional logical closed-loop system. The original functions and linkage mechanisms of each level are as follows: 1. Top-level brain-like intelligent phased array coordination platform: As the only global scheduling core, it has built-in algorithms for dynamic balancing of computing power and smooth switching of multiple modes, and uniformly manages all functional levels of the lower layer. It realizes millisecond-level seamless switching of communication / radar / anti-interference working modes, dynamically allocates system computing power resources, and avoids local computing power redundancy or overload problems. 2. Sensing Input Layer: A dedicated signal sensing preprocessing layer that integrates radio frequency signal acquisition, filtering and noise reduction, and analog-to-digital conversion to eliminate invalid noise data and provide a high-purity, standardized digital signal data stream for the upper AI computing layer, thereby improving the accuracy of intelligent decision-making from the input end; 3. Brain-like AI Computing Layer: The core intelligent decision-making center of the system, independently equipped with an experience knowledge base and AI Core computing unit. Relying on the SNN-DRL fusion brain-like algorithm, it completes environmental interference judgment, hardware error fitting, and iterative output of the optimal solution of beam strategy, realizing an intelligent decision-making mechanism that is completely different from traditional fixed algorithms. 3. Beam control layer: The core connecting intelligent decision-making and hardware execution, integrating multi-dimensional control parameters to complete digital beamforming and real-time amplitude and phase error calibration and compensation; 5. Low-level software-defined antenna function output layer: hardware execution and data feedback terminal, responsible for multi-task signal transmission and reception, and real-time transmission of hardware status, task data, and environmental parameters back to the upper-level module, forming a permanent closed-loop optimization mechanism of task execution-data feedback-intelligent iteration-secondary adjustment.

[0044] As mentioned above, each core level possesses unique and original functions, with deep coupling and linkage between levels: 1. Sensing Input Layer: Seamlessly interfaces with the RF front-end array antenna to complete the full-domain acquisition and adaptive noise reduction preprocessing of air interface electromagnetic signals. It can dynamically adjust the filtering threshold according to the intensity of environmental interference to achieve accurate extraction of effective signals in complex interference scenarios. 2. Brain-like AI Computing Layer: Unlike traditional passive signal processing architectures, it has three unique functions: online error fitting, autonomous strategy learning, and iterative updates of experience knowledge base. It can fit the array antenna temperature drift error and channel amplitude and phase deviation in real time, and generate the optimal beam control strategy based on historical engineering experience and real-time environmental data to drive the system to continuously adaptively optimize. 3. Beam control layer: Links the beamforming module and the amplitude and phase calibration compensation module, synchronously receives hardware status parameters, AI intelligent adjustment parameters, and anti-interference optimization parameters, to achieve dynamic and precise adjustment of beam pointing, beamwidth, and null position, and corrects beam deviations caused by hardware and environment in real time, ensuring high-precision and stable operation of the system.

[0045] This method constructs a fully autonomous optimization closed-loop workflow encompassing signal acquisition, preprocessing and calibration, brain-like intelligent decision-making, multi-parameter fusion and control, and task feedback iteration. The specific original implementation logic is as follows: 1. The underlying software-defined antenna function output layer performs communication, radar, and multi-target detection tasks, and synchronously collects radio frequency air interface electromagnetic signals and equipment hardware operating status data; 2. The sensing input layer performs filtering, noise reduction, and amplification preprocessing on the original electromagnetic signal, and generates a standardized digital signal through analog-to-digital conversion. At the same time, it performs preliminary amplitude and phase calibration and eliminates basic signal errors. 2. Standardized signal input to the brain-like AI computing layer; AI Core retrieves prior data from the experience knowledge base to complete environmental interference identification, target feature extraction, hardware fault assessment, and iteratively generate optimal beam control, anti-interference suppression, and error compensation strategies. 3. The integrated decision-making layer integrates intelligent decision-making strategies, hardware status parameters, and task requirement parameters to generate integrated and precise control commands, which are then sent to the beam control layer. 5. The beam control layer performs dynamic beamforming, interference null suppression, and real-time channel error compensation, driving the antenna to complete intelligent signal transmission and reception; 6. After the task is completed, the system will send back data on beam control effect, communication quality, hardware status, and environmental changes to the neuromorphic AI computing layer to complete online incremental iteration of model parameters, realize continuous self-optimization of system performance, and completely solve the technical defects of traditional phased array systems that cannot be upgraded autonomously. Example

[0046] A method for software functionalization of phased array antennas using a neuromorphic (AI) model may include the following steps: Step 1: Construct a fusion architecture of brain-like (AI) model and digital computing array, build a signal processing platform based on digital computing array technology, integrate the core processes of radio frequency signal reception, processing and transmission of phased array antenna, incorporate brain-like (AI) model, and realize the core functions of traditional hardware through software algorithms to reduce the number and complexity of hardware. Step 2: Breaking away from the traditional control board design concept, the brain-like (AI) deep learning model is embedded into the control process of the brain-like intelligent phased array system. The model is trained by communication, data transmission, and positioning data in harsh environments, enabling the model to adaptively identify environmental changes and dynamically adjust signal processing parameters, beam pointing, and communication modes to ensure uninterrupted real-time communication, data transmission, and positioning in harsh environments. Step 3: Embed fault identification algorithms into the neuromorphic (AI) model to monitor the phased array antenna user system in real time, collect equipment operating parameters and signal characteristics, identify fault types, fault locations and fault levels through AI algorithms, and only transmit fault warning results, achieving data compression of >90%; Step 4: Integrate the neuromorphic (AI) model, digital computing array fusion architecture, fault identification and data compression algorithms into the phased array antenna system, debug and optimize model parameters and software algorithm logic, and adapt to satellite communication applications in multiple fields.

[0047] The brain-like (AI) model mentioned in step 1 above may include a deep learning model and an intelligent decision-making model. The deep learning model adopts a GRU network, an LSTM network, or a fusion model of CNN and LSTM. The digital computing array adopts a fully digital architecture, with a complete transmit and receive channel set after each antenna element. Beamforming is completed in the digital domain.

[0048] Furthermore, the traditional hardware replaced by the software algorithm in step 1 above includes analog phase shifters and dedicated signal processing modules. The software replacement reduces the amount of hardware used by more than 60%, while the AI ​​algorithm achieves adaptive zeroing and waveform diversity, thereby improving anti-interference performance and signal processing efficiency.

[0049] The harsh environment mentioned in step 2 includes strong electromagnetic interference, extreme temperature (-40℃~85℃), signal attenuation, and complex terrain. The training data sample size is no less than 100,000 sets. After the model training is completed, the communication interruption time is ≤10ms and the data transmission error is ≤0.5%.

[0050] The fault identification algorithm described in step 3 adopts a collaborative design of model compression and hardware acceleration, with a fault identification accuracy of ≥95%. The returned fault warning results include fault type, warning level, and preliminary location. The compression ratio of the original monitoring data volume to the returned warning data volume is >90%, and the data processing pressure of the backend network management center is reduced by more than 85%.

[0051] Step 4 employs an FPGA parallel computing architecture to improve signal processing speed and optimize multi-channel synchronization accuracy to ±5ps. The model parameters and software functions can be flexibly adjusted according to the needs of different fields such as satellite communication, radar, vehicle-mounted, airborne, and shipborne applications.

[0052] The above methods are mainly applied in the field of satellite communication to realize intelligent and low-cost satellite communication, data transmission and positioning, and can also be extended to other fields that require phased array antennas.

[0053] The aforementioned phased array antenna system architecture can be based on the top-level "brain-like intelligent phased array platform" as the core hub, and is deployed strictly according to the functional hierarchy from top to bottom. The modules form a logical closed loop and efficient linkage. The specific connection relationships and functional implementations are as follows: The neuromorphic intelligent phased array platform, as the core of the overall system control, undertakes the responsibility of global scheduling and collaborative management. It unifies the scheduling of the three core layers: the sensing input layer, the neuromorphic AI computing layer, and the beam control layer, enabling high-speed data interaction, precise command issuance, and functional synergy between the layers. It provides core support for the software-based and intelligent operation of the phased array antenna. Its core modules and functions are as follows: Top-level collaborative scheduling module: responsible for global task allocation, dynamic balancing of computing load, and smooth switching between three working modes: communication, radar, and anti-interference, to ensure that the system achieves optimal operating status according to the application scenario; Brain-like AI computing cluster: The core computing unit responsible for environmental situational awareness, intelligent prediction of target behavior, adaptive decision generation, and nonlinear optimization calculation, realizing intelligent system control. Digital computing array: It undertakes the core tasks of high-speed real-time computing, digital beamforming (DBF), and accurate solution of amplitude and phase weights, ensuring the real-time performance and accuracy of the system's operations; Heterogeneous data interaction layer: Builds a high-speed data channel between the brain-like AI computing layer and the digital computing array to achieve efficient sharing of AI feature data and digital intermediate results, avoid data redundancy, and improve data interaction efficiency.

[0054] Fusion Decision Layer: Integrates the optimal strategy output by the brain-like AI computing cluster with the real-time and accuracy advantages guaranteed by the digital computing array, jointly generates accurate and reliable system control commands, and sends them to subsequent execution modules; Antenna drive interface: As a bridge connecting the system and the phased array antenna, it converts the control commands output by the fusion decision layer into executable signals, and finally drives the phased array antenna to achieve software-based and intelligent beam control.

[0055] The three core layers of this embodiment are connected to the underlying modules: Perceptual input layer As the signal acquisition front end of the system, the sensing input layer directly receives the scheduling instructions from the top layer of the system and seamlessly connects with the radio frequency front end and array antenna. Its core responsibility is to complete the acquisition, preliminary reception and preprocessing of air interface electromagnetic signals, providing high-quality raw data support for subsequent signal processing and intelligent decision-making, and realizing the foundation for precise control. Neuromorphic AI Computing Layer As the core computing unit for the intelligent software control of the system, the brain-like AI computing layer is based on the core algorithms of spiking neural network (SNN) and deep reinforcement learning (DRL) to realize adaptive amplitude and phase weight learning, online fitting and compensation of array error, interference environment perception and dynamic decision-making functions. This layer receives the intelligent decision-making instructions from the top layer of the system, deeply links with the beam control layer, and participates in the intelligent decision-making of key links such as digital beamforming, error compensation and anti-interference optimization, promoting the transformation of phased array antenna from "hardware defined" to "software defined". Beam control layer The beam control layer is centrally managed by the top-level system and directly interfaces with the digital beamforming (DBF) module. It is primarily responsible for real-time beam control and precise adjustment of amplitude and phase parameters. It also works in conjunction with the amplitude and phase calibration & error compensation module to correct deviations in the beam control process in real time, ensuring the control accuracy of parameters such as beam pointing and gain. It is a key link connecting intelligent decision-making and hardware execution.

[0056] The signal processing and intelligent optimization closed-loop process of this embodiment allows the brain-like intelligent phased array system to adopt a closed-loop process of "signal acquisition - preprocessing - calibration - intelligent decision-making - optimization and control" to ensure the accuracy of signal processing and the intelligence of system operation. The specific process is as follows: The radio frequency front-end and array antenna acquire over-the-air electromagnetic signals and transmit them to the signal acquisition and preprocessing module. Through noise reduction, filtering, and amplification operations, noise and interference signals are filtered out to improve the quality of the original signal. The preprocessed signal is transmitted to the amplitude and phase calibration & error compensation module to accurately correct the amplitude and phase deviation of the signal, further ensuring the stability and accuracy of the signal and laying the foundation for subsequent processing; The corrected signal is synchronously fed back to the brain-like AI computing layer, driving the target recognition module to complete target feature extraction and type recognition, and at the same time driving the intelligent beam scheduling module to complete the adaptive adjustment of beam parameters; The optimization strategy output by the neuromorphic AI computing layer is integrated by the fusion decision layer and then transmitted to the anti-interference optimization module to improve beam anti-interference performance in complex environments, ultimately forming a complete closed loop of signal processing and intelligent optimization. Example

[0057] The functionality of phased array antennas can be software-based using a neuromorphic (AI) model, which can be achieved through the following specific steps: Step 1: Constructing a fusion architecture of a brain-like (AI) model and a digital computing array. A signal processing platform based on digital computing array technology is built, integrating the digital computing array using an RFSoC chip, and integrating the RF signal reception, processing, and transmission processes of the phased array antenna. A brain-like (AI) model is constructed, including a deep learning model (using a GRU network, composed of four layers with 128 neurons per layer) and an intelligent decision-making model. Software algorithms are used to implement the beamforming function of traditional analog phase shifters and the signal filtering and amplification functions of dedicated signal processing modules, reducing hardware usage by over 60% and lowering hardware costs by 30%–50%. Simultaneously, through the AI ​​model's adaptive nulling algorithm, a "zero trap" is formed in the interference direction, improving electromagnetic interference resistance. Combined with the multi-beam capability of the digital computing array, simultaneous communication with multiple targets is achieved.

[0058] Step 2: Integrate AI deep learning technology to optimize adaptability to harsh environments. Abandoning the traditional dedicated control board design, the trained GRU deep learning model is embedded into the phased array antenna control process. The training data uses real-world data from harsh environments in the satellite communication field (including communication, data transmission, and positioning data under scenarios such as strong electromagnetic interference, extreme temperatures of -40℃ to +85℃, and high-altitude signal attenuation, with a sample size of no less than 100,000 sets). After the model is trained, it can identify changes in environmental parameters in real time and dynamically adjust signal processing gain, beam pointing angle, and communication frequency to ensure that communication interruption time is ≤10ms, data transmission error is ≤0.5%, and positioning accuracy is improved by more than 20% in harsh environments, achieving uninterrupted real-time communication, high-precision data transmission, and positioning.

[0059] Step 3: AI Fault Identification and Data Compression Implementation. A deep learning-based fault identification algorithm is embedded into the neuromorphic (AI) model. Channel pruning technology is used to compress the model, reducing the computational power requirement while maintaining over 95% fault identification accuracy. Real-time acquisition of operating parameters (such as voltage, current, signal strength, and bit error rate) from the phased array antenna's RF module, signal processing module, and communication module is performed. AI algorithms analyze and identify fault types (such as RF module faults, signal interruption faults, and positioning deviation faults), fault locations, and fault levels (general, severe, and urgent). Only early warning results are transmitted back. The original monitoring data volume is 100MB / hour, while the transmitted early warning data volume is ≤10MB / hour, achieving a data compression rate of 90%. This reduces the data processing pressure on the backend network management center by over 85%.

[0060] Step 4: Integration, Debugging, and Optimization of the Brain-Inspired Intelligent Phased Array System. The brain-inspired (AI) model, digital computing array fusion architecture, fault identification, and data compression algorithms are integrated into the satellite communication phased array antenna system. An FPGA parallel computing architecture is used to improve feature extraction and signal processing speed, and the multi-channel synchronization accuracy is optimized to ±5ps. Through repeated debugging, the AI ​​model parameters and software algorithm logic are optimized to ensure stable and reliable software functionality, adapting to the voice, video, data transmission, and positioning requirements of satellite communication. After testing, the system operates stably with no significant performance degradation in harsh environments. The cost is reduced by 40% compared to traditional phased array antennas, and the level of intelligence is significantly improved.

[0061] The difference between this embodiment and embodiment 3 is that the brain-like (AI) model adopts a fusion model of CNN and LSTM, the fault identification algorithm adopts an anomaly detection algorithm based on machine learning, which is suitable for satellite communication and radar composite application scenarios, and can realize satellite communication and target detection functions at the same time. The data volume is compressed by 92%, the hardware cost is reduced by 45%, and the positioning accuracy is improved by 25% in harsh environments. The remaining steps are the same as in embodiment 1.

[0062] Furthermore, the brain-like (AI) model mentioned in step 1 includes a deep learning model and an intelligent decision-making model. The deep learning model adopts a GRU network, an LSTM network, or a fusion model of CNN and LSTM. The digital computing array adopts a fully digital architecture, with a complete transmit and receive channel set after each antenna element. Beamforming is completed in the digital domain.

[0063] The traditional hardware replaced by the software algorithm in step 1 includes analog phase shifters and dedicated signal processing modules. By replacing them with software, the amount of hardware used can be reduced by more than 60%, and the hardware cost can be reduced by 30%-50%. At the same time, AI algorithms can be used to achieve adaptive zeroing and waveform diversity, thereby improving anti-interference performance and signal processing efficiency.

[0064] The harsh environment described in step 2 includes strong electromagnetic interference, extreme temperatures (-40℃~85℃), signal attenuation, and complex terrain. The training data sample size is no less than 100,000 sets. After the model training is completed, the communication interruption time is ≤10ms, the data transmission error is ≤0.5%, and the positioning accuracy is improved by more than 20% compared with the traditional method.

[0065] The fault identification algorithm described in step 3 adopts a collaborative design of model compression and hardware acceleration, with a fault identification accuracy of ≥95%. The returned fault warning results include fault type, warning level, and preliminary location. The compression ratio of the original monitoring data volume to the returned warning data volume is >90%, and the data processing pressure of the backend network management center is reduced by more than 85%.

[0066] Step 4 employs an FPGA parallel computing architecture to improve signal processing speed and optimize multi-channel synchronization accuracy to ±5ps. The model parameters and software functions can be flexibly adjusted according to the needs of different fields such as satellite communication, radar detection, space-based telemetry and control, and precision guidance.

[0067] The method is mainly applied in the field of satellite communication to realize intelligent and low-cost voice, video, data transmission and positioning, and can also be extended to other fields that require phased array antennas.

[0068] Simulation and prototype comparison test data of any of the above embodiments (divided into three groups of core indicators: cross-layer feedback delay, phase noise suppression, and system additional power consumption, with comparative examples set) and prototype simulation comparison test data (comparative examples vs. embodiments of the present invention). Unified test boundary conditions The test environment temperature range covers -40℃ to 85℃, the RF operating bandwidth is 2GHz, and it is superimposed with broadband suppression electromagnetic interference and enemy narrowband targeting interference; the FPGA computing power load is kept constant at 90% of the rated thermal runaway critical condition; the comparative example adopts the conventional solution of existing technology: single STNO device architecture, digital ADC / DAC relay link, and integer order fixed parameter thermal compensation model; the embodiment is the differential STNO stacking structure of this invention + fractional order coupling dynamics model + all-optical pulse direct drive three-level closed loop link.

[0069] Experiment 1: Cross-level closed-loop feedback delay test Test Condition Comparison (Traditional Digital Transmission Architecture) Embodiment of this Invention (Optical Pulse Direct-Drive Closed-Loop) Technical Gain Effect At standard operating conditions (25℃), the closed-loop delay is reduced by over 90%, with a maximum compression rate of 1800 ps and a maximum stable control rate below 200 ps. Under extreme thermal conditions of 85℃, the speed is 2150 ps (198 ps). With no digital buffer forwarding stage, latency is almost unaffected by high temperatures. Experimental results: Existing technologies rely on digital buses and cache read / write operations to complete cross-layer data interaction, and chip routing and storage retrieval introduce fixed delays; this invention eliminates the intermediate links of analog-to-digital and digital-to-analog conversion, and the silicon-based microring outputs optical pulse streams to directly drive the STNO differential array, realizing native asynchronous closed-loop in hardware, and achieving an unexpectedly significant improvement in latency performance.

[0070] Experiment 2: Wide-temperature-range phase noise suppression test (1kHz frequency deviation, unit dBc / Hz) Ambient temperature comparison (ordinary single STNO device) Noise reduction gain of the differential STNO stacking structure of this invention -40℃ low temperature operation -68 dBc / Hz -86 dBc / Hz Low frequency 1 / f phase noise suppression improved by 18 dB Operating at 85℃ high temperature: -75 dBc / Hz to -87 dBc / Hz; Low-frequency 1 / f phase noise suppression improved by 12 dB. Experimental results show that common-mode thermal noise is canceled by a pair of anti-magnetized STNO differential pairs, and the effective field strength of the spin orbit moment is adaptively adjusted with temperature according to the coupled dynamic equation, thus achieving differential noise reduction over a wide temperature range. This synergistic noise reduction effect cannot be achieved by conventional parameter tuning of a single STNO device.

[0071] Experiment 3: Dynamic power consumption test of RF link (beam real-time compensation working mode) Operating mode comparison (continuous real-time current compensation) This invention (optical bistable threshold control) Power consumption optimization range In normal steady-state communication mode, the power consumption of the entire device's RF link is reduced by 22% from 12.7 W to 9.9 W. High-speed beam reconfiguration mode with strong interference (15.3 W vs. 11.8 W) avoids ineffective iterative adjustments caused by minute temperature fluctuations, further reducing dynamic power consumption. Experimental results: This invention utilizes the silicon-based multi-cavity microring optical bistable hysteresis interval to set a phase jitter threshold, and only activates STNO current correction when the beam phase error exceeds the threshold. Compared with the existing uninterrupted continuous compensation scheme, dynamic power consumption is significantly reduced.

[0072] Experiment 4: Beam pointing error fluctuation test under harsh operating conditions of temperature and vibration Combined operating conditions: -40℃ to 85℃ cyclic temperature change + 10~100Hz low-frequency mechanical vibration superimposed with electromagnetic interference Comparative example (fractional order of single temperature parameter adjustment): Maximum beam pointing error fluctuation 0.032° The present invention (two-dimensional Hurst joint feature constraint order step size for temperature-vibration): maximum beam pointing error fluctuation of 0.011° Test gain: The beam error fluctuation amplitude is reduced by 65% ​​under complex and harsh working conditions, achieving high-precision beam stabilization in all temperature and vibration scenarios.

[0073] Experiment 5: Comparison of Static Power Consumption of Spiking Neural Networks Traditional unoptimized spiking neural network (SNN): Average static power consumption 820 mW This invention employs a triple-constraint evolutionary screening method for sparse pulse networks, achieving an average static power consumption of 245 mW. Experimental gain: Network power consumption is reduced by more than 70% compared to traditional SNN, while maintaining the topology's anti-interference performance without degradation.

[0074] This invention utilizes a pair of opposing oscillating STNO differential pairs to cancel common-mode thermal noise and 1 / f low-frequency noise, while simultaneously adjusting the spin orbit moment field strength adaptively with temperature using the coupled dynamic equations, thereby achieving differentiated noise reduction over a wide temperature range. This synergistic noise reduction effect cannot be achieved through conventional parameter optimization of a single STNO device.

[0075] The STNO array adopts a multi-domain-wall coupled spin valve stacking structure. Within the same array element, two STNO differential pairs oscillate in opposite directions to construct a spin differential mode phase reference. The film layers of a single STNO device, from bottom to top, are: Ta substrate buffer layer (thickness 5 nm) / Pt spin Hall layer (thickness 8 nm) / CoTb ferrimagnetic free layer (thickness 1.6 nm) / MgO barrier layer (thickness 0.8 nm) / CoFeB reference pinning layer (thickness 2.0 nm) / Ru coupling layer (0.4 nm) / IrMn antiferromagnetic pinning layer (thickness 10 nm). The two STNO devices in the same array element have a symmetrical structure, and the initial magnetization directions of their free layer domain walls are 180° apart, forming a differential pair. The array element unit has a lateral size of 800 nm × 800 nm, and the center-to-center spacing between adjacent array elements is 1.2 μm, forming an array arrangement to eliminate the intrinsic 1 / f low-frequency phase noise of a single STNO device and form a native differential phase reference source that does not depend on an external clock.

[0076] The sparse pulse network after triple-constraint evolutionary screening in this invention has an average power consumption of 245mW. Results: Overall power consumption is reduced by more than 70% compared to traditional networks.

[0077] The array-level spin orbital moment nano oscillator (STNO) of this invention adopts a multi-domain wall-coupled spin valve stacking structure, sets up dual STNO differential pairs in the same array element and configures them in reverse self-sustaining oscillation mode, and constructs a native spin differential mode phase reference without an external reference clock, thereby eliminating the inherent 1 / f low-frequency phase noise of a single STNO device from the hardware level.

[0078] The films of a single STNO device are stacked sequentially from the substrate to the top layer, and the materials and thickness parameters of each film are strictly limited: Ta substrate buffer layer (thickness 5 nm) / Pt spin Hall layer (thickness 8 nm) / CoTb ferrimagnetic free layer (thickness 1.6 nm) / MgO barrier layer (thickness 0.8 nm) / CoFeB reference pinning layer (thickness 2.0 nm) / Ru coupling layer (0.4 nm) / IrMn antiferromagnetic pinning layer (thickness 10 nm).

[0079] The two STNO devices within the same array element are of a completely symmetrical structure. The two CoTb ferrimagnetic free layers are initially magnetized in opposite directions by 180° to form differential oscillation pairing. The overall lateral geometry of a single array element is 800 nm × 800 nm. When the array is arranged, the center-to-center distance between adjacent array elements is 1.2 μm to ensure that the electromagnetic crosstalk between array elements is less than -45 dB.

[0080] Brain-inspired intelligent phased array system architecture description See Figure 1 The architecture of this phased array antenna system is centered on the top-level "brain-like intelligent phased array platform," and is strictly deployed from top to bottom according to functional hierarchy. Each module forms a logical closed loop and efficient linkage. The specific connection relationships and functional implementations are as follows: The neuromorphic intelligent phased array platform, as the core of the entire system's overall control, undertakes the responsibility of global scheduling and collaborative management. It unifies the scheduling of the three core layers: the sensing input layer, the neuromorphic AI computing layer, and the beam control layer, enabling high-speed data interaction, precise command issuance, and functional synergy between these layers. This provides core support for the software-based and intelligent operation of the phased array antenna. Its core modules and functions are as follows: Top-level collaborative scheduling module: responsible for global task allocation, dynamic balancing of computing load, and smooth switching between three working modes: communication, radar, and anti-interference, to ensure that the system achieves optimal operating status according to the application scenario.

[0081] Brain-like AI computing cluster: The core is responsible for environmental situational awareness, intelligent prediction of target behavior, adaptive decision generation, and nonlinear optimization calculation. It is the core computing power unit for realizing intelligent control of the system.

[0082] Digital computing array: It undertakes core tasks such as high-speed real-time computing, digital beamforming (DBF), and accurate solution of amplitude and phase weights, ensuring the real-time performance and accuracy of the system's operations.

[0083] Heterogeneous data interaction layer: Builds a high-speed data channel between the brain-like AI computing layer and the digital computing array, realizes efficient sharing of AI feature data and digital intermediate results, effectively avoids data redundancy, and improves data interaction efficiency.

[0084] Fusion Decision Layer: Integrates the optimal strategy output by the brain-like AI computing cluster with the real-time and accuracy advantages guaranteed by the digital computing array, jointly generates accurate and reliable system control commands, and sends them to subsequent execution modules.

[0085] Antenna drive interface: As a bridge connecting the system and the phased array antenna, it converts the control commands output by the fusion decision layer into executable signals, and finally drives the phased array antenna to achieve software-based and intelligent beam control.

[0086] Application Case: In airborne phased array radar systems, this platform can automatically switch between "detection mode" and "communication mode" according to actual scenarios such as flight altitude and battlefield environment, dynamically coordinate and allocate computing resources, and ensure that the radar accurately detects aerial targets while stably achieving two-way high-speed communication with ground command base stations, thus balancing detection accuracy and communication reliability.

[0087] As the signal acquisition front end of the system, the sensing input layer directly receives the scheduling instructions from the top layer of the system and seamlessly connects with the radio frequency front end and array antenna. Its core responsibility is to complete the acquisition, preliminary reception and preprocessing of air interface electromagnetic signals, providing high-quality raw data support for subsequent signal processing and intelligent decision-making, and is the foundation for the system to achieve precise control.

[0088] Application Case: In ground-based air defense systems, this layer can quickly capture the electromagnetic signals of incoming missiles. Through preliminary filtering, amplification, and other preprocessing operations, invalid clutter is filtered out, and the clean signal is transmitted to subsequent modules, providing accurate raw signal basis for missile interception decisions and beam control.

[0089] The neuromorphic AI computing layer is the core computing unit for the system's software-based intelligent control. Based on core algorithms such as spiking neural networks (SNN) and deep reinforcement learning (DRL), it realizes functions such as adaptive amplitude and phase weight learning, online fitting and compensation of array errors, and perception and dynamic decision-making in the interference environment. This layer receives intelligent decision-making instructions from the top layer of the system and deeply collaborates with the beam control layer, participating in intelligent decision-making in key aspects such as digital beamforming, error compensation, and anti-interference optimization, thus driving the transformation of phased array antennas from "hardware-defined" to "software-defined".

[0090] Application Case: In battlefield electronic warfare scenarios, this layer can quickly identify the type, intensity, and frequency band of enemy interference signals through AI algorithms, adaptively adjust the beam weight allocation strategy, avoid the impact of interference signals on our target detection and communication, and ensure the stable operation of the system in complex interference environments.

[0091] The beam control layer is centrally managed by the top-level system and directly interfaces with the digital beamforming (DBF) module. It is primarily responsible for real-time beam control and precise adjustment of amplitude and phase parameters. It also works in conjunction with the amplitude and phase calibration & error compensation module to correct deviations in the beam control process in real time, ensuring the control accuracy of parameters such as beam pointing and gain. It is a key link connecting intelligent decision-making and hardware execution.

[0092] Application Case: In satellite communication systems, this layer can quickly adjust beam pointing and parameters according to the high-speed movement trajectory of the satellite, realize real-time tracking of moving satellites, ensure the continuous stability of the space-to-ground communication link, and effectively avoid signal interruption and attenuation caused by satellite movement.

[0093] The brain-inspired intelligent phased array system adopts a closed-loop process of "signal acquisition - preprocessing - calibration - intelligent decision-making - optimization and control" to ensure the accuracy of signal processing and the intelligence of system operation. The specific process is as follows: The radio frequency front-end and array antenna acquire air interface electromagnetic signals and transmit them to the signal acquisition and preprocessing module. Through noise reduction, filtering, amplification and other operations, noise and interference signals are filtered out to improve the quality of the original signal. The preprocessed signal is transmitted to the amplitude and phase calibration & error compensation module to accurately correct the amplitude and phase deviation of the signal, further ensuring the stability and accuracy of the signal and laying the foundation for subsequent processing; The corrected signal is synchronously fed back to the brain-like AI computing layer, driving the target recognition module to complete target feature extraction and type recognition, and at the same time driving the intelligent beam scheduling module to complete the adaptive adjustment of beam parameters; The optimization strategy output by the neuromorphic AI computing layer is integrated by the fusion decision layer and then transmitted to the anti-interference optimization module to improve beam anti-interference performance in complex environments, ultimately forming a complete closed loop of signal processing and intelligent optimization.

[0094] Application Case 1: In a maritime surveillance radar system, the signal acquisition and preprocessing module can effectively filter out sea clutter and marine electromagnetic interference, while the amplitude and phase calibration module corrects signal deviations caused by seawater reflection and channel attenuation, ensuring that the system can accurately detect distant targets such as ships and aircraft, providing reliable support for maritime surveillance.

[0095] Application Case 2: In urban security radar scenarios, the target recognition module can accurately distinguish different types of targets such as vehicles, pedestrians, and drones. The intelligent beam scheduling module allocates beam resources in a targeted manner according to the target distribution. The anti-interference optimization module suppresses urban electromagnetic clutter and electronic equipment interference, ensuring accurate monitoring and trajectory tracking of targets in key areas.

[0096] The intelligent control commands, optimized for anti-interference, are transmitted to the software-defined antenna function output layer. This layer integrates core functions such as intelligent beamforming, multi-beam concurrency, anti-interference, target tracking, and communication / radar integration, completing the output of two core functions. Ultimately, it fulfills the core application requirements of software-based and intelligent phased array antennas, forming a complete functional closed loop of "scheduling-processing-decision-output-feedback". Communication function output: Through communication beamforming technology, the signal energy is concentrated to cover the target area, constructing an efficient and stable communication link, and improving the communication rate and anti-interference capability.

[0097] Application Case: In 5G base station phased array systems, communication beamforming can accurately concentrate signal energy to cover densely populated user areas according to user distribution density, effectively improving communication speed within the area, reducing interference between signals, and improving user communication experience.

[0098] Radar detection function output: Through flexible control of the radar detection beam, real-time detection, identification and tracking of multiple targets can be achieved, ensuring detection accuracy and real-time performance.

[0099] Application Case: In airport air traffic control radar, the radar detection beam can simultaneously track multiple takeoff and landing aircraft. The multi-target tracking module collects and feeds back key information such as the aircraft's position, speed, and heading in real time, providing accurate data support for airport air traffic control scheduling and ensuring airport operational safety.

[0100] See Figure 2 This architecture, centered on "brain-like intelligent perception + precise digital computation," achieves efficient end-to-end processing from air interface signal perception to phased array antenna driving. Based on the hierarchical logic of the architecture, the functional principles and connection relationships of each technical unit are as follows:

[0101] The top-level fusion control and collaborative scheduling unit, as the "central hub" of the architecture, coordinates the timing, task allocation, and data interaction of the brain-like AI computing cluster and the digital parallel computing array, dynamically allocates resources, solves the timing synchronization and task connection problems of heterogeneous computing, and ensures the collaborative and efficient operation of the system. It is the core control node of heterogeneous fusion, directly connected to the brain-like AI computing cluster and the digital parallel computing array, receives status feedback from both and issues scheduling instructions.

[0102] The neuromorphic AI computing cluster integrates multiple neuromorphic AI computing units, simulating the parallel processing and fuzzy decision-making capabilities of the human brain. It focuses on intelligent perception and feature mining in complex scenarios, working in conjunction with a neuromorphic model feature extraction engine to complete air interface signal recognition, interference classification, and target prediction, providing high-quality feature input for digital computing. It is bidirectionally connected to the top-level scheduling unit and the neuromorphic feature extraction engine, receiving scheduling instructions and outputting feature data. The digital parallel computing array, as the core of precise computing, includes pulsed SNN / deep neural network units and FPGA / ASIC / multi-core DSP arrays, possessing high parallelism and high computing power. The former optimizes the features extracted by the neuromorphic model, while the latter undertakes high-speed digital signal processing and complex algorithm calculations, providing support for amplitude and phase calculations and beamforming. It is directly connected to the top-level scheduling unit and the digital signal processing pipeline, receiving scheduling instructions and outputting calculation results.

[0103] The feature extraction and signal processing layer of the brain-like model feature extraction engine relies on the computing power of the brain-like AI computing cluster to extract key features such as targets, interference, and noise from the raw air interface signal. After filtering redundant information, it is synchronously transmitted to the digital signal processing pipeline and the fusion decision layer through the heterogeneous data interaction bus, realizing the seamless connection between brain-like and digital processing. The input end is connected to the brain-like AI computing cluster, and the output end is connected to the heterogeneous data interaction bus.

[0104] The digital signal processing pipeline takes the core features and raw air interface signals from the brain-like feature extraction engine, completes preprocessing such as filtering and sampling, and provides standardized signals for subsequent amplitude and phase calculation, DBF and beamforming. At the same time, it feeds the processed data back to the brain-like AI computing cluster for iterative optimization. The input end is connected to the brain-like feature extraction engine, the output end is connected to the amplitude and phase calculation unit, and the feedback end is connected to the brain-like AI computing cluster.

[0105] The perception and computation execution layer performs air interface perception, interference identification, and target prediction. Based on the output of the brain-like AI computing cluster and feature extraction engine, it realizes air interface environmental perception, interference identification, and target trajectory prediction, providing accurate perception basis for joint decision-making. The input end connects to the brain-like AI computing cluster and brain-like feature extraction engine, and the output end connects to the fusion decision layer.

[0106] Amplitude and phase calculation, DBF operation, and beamforming rely on the computing power of the digital parallel computing array, combined with the standardized signals of the digital signal processing pipeline and the instructions of the fusion decision layer, to complete the calculation of amplitude and phase parameters, DBF operation, and beamforming, thereby achieving target acquisition and tracking and interference suppression. The input end is connected to the digital parallel computing array, the digital signal processing pipeline, and the fusion decision layer, and the output end is associated with the phased array antenna drive interface layer.

[0107] The heterogeneous data interaction and weight sharing bus of the data interaction and collaboration layer adopts high-speed on-chip interconnection and low-latency data transfer technology. As the data transmission hub of the whole architecture, it connects all core units such as the brain-like AI computing cluster, digital parallel computing array, and feature extraction engine, and transmits feature data, calculation results and control instructions to realize weight sharing between brain-like and digital computing and support heterogeneous integration.

[0108] The system's collaborative scheduling layer connects to the top-level integrated control unit, breaks down the global scheduling strategy, coordinates the working rhythm of the brain-like AI computing unit and the digital computing array, monitors the unit's operating status, and ensures seamless connection of each process. It is the key link between top-level control and bottom-level execution, connecting the top-level scheduling unit at the top and various computing and processing units at the bottom.

[0109] As an extension of the heterogeneous data bus, the heterogeneous fusion and data interaction bus layer standardizes transmission formats and protocols, solves the compatibility problem between brain-like and digital heterogeneous data, focuses on the data fusion and weight sharing between the two, connects brain-like AI computing clusters and digital parallel computing arrays, and provides precise data support for heterogeneous fusion.

[0110] The decision-making and output-driven layer integrates the decision-making and collaborative output layer as the core of decision-making. It integrates brain-like decision-making and digital computing results, and combines information such as air interface sensing and interference recognition to generate joint beam control commands, so as to achieve complementary advantages between the two. The input end is connected to the brain-like AI computing cluster, digital parallel computing array and air interface sensing unit, and the output end is connected to the joint decision-making and beam command generation layer.

[0111] The joint decision-making and beam command generation layer transforms the control strategy of the fusion decision layer into executable beam control commands, refines the parameters, sends them down to the phased array antenna RF drive interface layer and feeds back the sending status, connecting to the fusion decision layer above and the phased array drive interface layer below.

[0112] The phased array antenna RF drive interface layer serves as the interface between the system and the phased array antenna. It receives beam control commands and converts them into RF drive signals, completes amplitude and phase coding, array calibration, and multi-channel synchronous control, and finally converts the decision calculation results into actual antenna actions. The input end is connected to the joint decision and beam command generation layer, and the output end is connected to the phased array antenna.

[0113] Each unit collaborates in a hierarchical manner, following the steps of "top-level control → core computing → feature extraction → signal processing → perception and execution → data interaction → decision output → driving execution." The core of this system is the heterogeneous fusion of brain-like AI and digital computing arrays, which improves the system's perception, anti-interference, and target control accuracy, and adapts to the high-performance requirements of phased array antennas.

[0114] See Figure 3This process aims to "efficiently process multi-source data, accurately identify faults, and achieve data lightweighting." Through modular integration, parallel processing, and fusion output, it completes a closed-loop process from raw data acquisition to final storage and transmission. The working principles and connection logic of each stage are as follows, strictly corresponding to the process nodes:

[0115] The process follows the core logic of "acquisition → preprocessing → feature extraction → parallel processing (compression + fault identification) → fusion output → storage and transmission". After feature extraction, it is divided into two parallel branches to complete data compression (to achieve data lightweighting) and AI fault identification (to achieve anomaly diagnosis). Finally, the results of the two branches are fused to output a processing result that is both practical and efficient, adapting to the data application needs of multiple scenarios.

[0116] Raw multi-source data acquisition serves as the initial input to the process, its core function being to obtain the raw data to be processed, providing the foundation for all subsequent processing. The acquisition targets are multi-source data, primarily including real-time data collected by various sensors (such as equipment operating parameters like temperature, pressure, vibration, and current), equipment operation log data, and environmental monitoring data. The acquisition method combines real-time and batch acquisition (the core acquisition algorithm can utilize polling or interrupt-triggered acquisition algorithms; in industrial scenarios, the Modbus protocol is commonly used to acquire sensor data), ensuring data integrity and timeliness. During the acquisition process, auxiliary information such as the acquisition time and acquisition nodes must be recorded synchronously to avoid data confusion and provide data traceability for subsequent preprocessing and fault identification.

[0117] Data preprocessing (denoising, normalization, and missing value imputation): The raw data collected contains a large amount of interference and data defects, making it unsuitable for direct feature extraction and model inference. The core objective of this step is to "purify the data and unify the format," providing high-quality data for subsequent processing. The specific principle is as follows: As the initial input step of the process, its core function is to acquire the raw data to be processed, providing a foundation for all subsequent processing. The data collected is from multiple sources, mainly including real-time data collected by various sensors (such as equipment operating parameters like temperature, pressure, vibration, and current), equipment operation log data, and environmental monitoring data. The collection method combines real-time and batch collection to ensure data integrity and timeliness. During the collection process, auxiliary information such as the data collection time and collection nodes must be recorded synchronously to avoid data confusion and provide data traceability for subsequent preprocessing and fault identification.

[0118] Data preprocessing (denoising, normalization, and missing value imputation): The raw data collected contains a large amount of interference and data defects, making it unsuitable for direct feature extraction and model inference. The core objective of this step is to "clean up the data and unify its format" to provide high-quality data for subsequent processing. The specific principles are as follows: Noise reduction: Filtering algorithms (such as mean filtering, median filtering, wavelet filtering, etc.) are used to remove noise signals such as random interference and measurement errors from the original data. For example, in vibration data collected by sensors, interference signals from environmental vibrations may be mixed in. Filtering algorithms preserve the vibration characteristic signals of the device itself, remove irrelevant interference, and avoid noise affecting the accuracy of subsequent feature extraction.

[0119] Normalization: Due to the different units of measurement in multi-source data (e.g., temperature is measured in °C, current in A), the numerical ranges vary significantly, which can lead to deviations in subsequent feature extraction and model inference. Normalization algorithms (such as min-max normalization and Z-score normalization; min-max normalization is commonly used for numerical data, while Z-score normalization is chosen when the original data distribution needs to be preserved) map all data to a uniform numerical range (e.g., [0,1] or [-1,1]), eliminating the influence of units and making different types of data comparable.

[0120] Missing value imputation: Data may be missing in the original data due to sensor malfunctions, data acquisition interruptions, etc. Directly discarding missing data will reduce the data volume and result in incomplete feature information. Using appropriate imputation methods (such as mean imputation, median imputation, linear interpolation imputation, etc.) to supplement missing data ensures data continuity and integrity, preventing missing values ​​from affecting the stability of subsequent steps in the process.

[0121] Feature extraction and key feature selection: Even after preprocessing, the clean data still contains a lot of redundant information (such as duplicate basic parameters and information unrelated to faults and compression). The core of this step is "extracting key information and selecting effective features," providing the core processing objects for the parallel branch. The specific principle is as follows: First, feature extraction algorithms (such as time-domain feature extraction, frequency-domain feature extraction, deep learning feature extraction, etc.) are used. In the time domain, RMS and peak factor extraction algorithms are commonly used. In the frequency domain, Fast Fourier Transform (FFT) algorithms are commonly used. In deep learning, CNN convolutional neural networks can be used to extract deep features. Features that can reflect the essential characteristics of the data and are related to fault identification and data compression are extracted from the preprocessed data. For example, peak value, effective value, and frequency spectrum peak value are extracted from vibration data, and waveform features and amplitude change features are extracted from current data.

[0122] Subsequently, important features are screened using feature selection algorithms (such as ANOVA, mutual information, and recursive feature elimination; recursive feature elimination (RFE) is commonly used for high-dimensional data, while ANOVA can be used for simpler scenarios). Redundant and irrelevant features (such as features that have no impact on fault identification or do not help with data compression) are eliminated, while core and effective features are retained. This approach reduces the computational load of subsequent parallel processing, improves processing efficiency, and avoids problems such as model overfitting and poor compression results caused by redundant features, ensuring the accuracy and efficiency of subsequent processing stages.

[0123] Parallel branch processing (data compression module + AI fault identification module): After feature extraction and filtering, the process is divided into two parallel branches, simultaneously performing data compression and AI fault identification. This achieves the dual goals of "lightweight data" and "accurate fault diagnosis." The two branches run independently without interfering with each other, improving the overall processing efficiency. The specific principle is as follows:

[0124] The core objective of this branch is to perform lightweight processing on the filtered feature data, reducing data storage space and transmission bandwidth, while retaining the core information of the data (without affecting the use of subsequent fault results and data recovery). The principles of each sub-step are as follows: Feature Dimensionality Reduction: Dimensionality reduction algorithms (such as Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), and autoencoders; PCA is preferred for unsupervised scenarios, while LDA is used for supervised fault identification scenarios) are employed to map high-dimensional feature data to a low-dimensional space. For example, 20-dimensional feature data obtained after filtering can be compressed to 5-10 dimensions using dimensionality reduction algorithms. This significantly reduces data dimensionality while preserving core feature information, thus reducing the burden on subsequent processing and storage.

[0125] Redundancy removal: Further filter the feature data after dimensionality reduction to remove duplicate and redundant feature information (such as features with high correlation that still exist after dimensionality reduction), further compress the data volume, ensure the simplicity of the data, and avoid redundant information occupying storage and transmission resources.

[0126] Encoded and compressed storage: Efficient encoding algorithms (such as Huffman coding, arithmetic coding, LZ77 / LZ78 coding, etc.) are used to encode the data after feature dimensionality reduction and redundancy removal, converting the data into a more compact encoding format and further reducing data storage capacity. The encoded data can be temporarily stored, awaiting fusion with fault identification results.

[0127] The core objective of this branch is to utilize AI technology to accurately identify whether a device has a fault, the type of fault, and the level of fault based on the selected core features, providing a diagnostic basis for subsequent applications. The principles of each sub-step are as follows: AI Model Inference: The selected core features are input into a pre-trained AI fault recognition model (such as neural networks, support vector machines (SVM), random forests, etc.; complex fault recognition often uses CNN-LSTM hybrid neural networks, while simple scenarios can use SVM models). The model learns the fault feature patterns and performs inference calculations on the input features to determine the equipment operating status (normal or abnormal) corresponding to the current data. A large number of normal and various fault feature samples have been input during the model training phase, giving it strong fault recognition capabilities.

[0128] Anomaly Matching 1: AI Model Inference: The selected core features are input into a pre-trained AI fault identification model (such as neural networks, support vector machines (SVM), random forests, etc.). The model learns the fault feature patterns and performs inference calculations on the input features to determine the operating status (normal or abnormal) of the equipment corresponding to the current data. A large number of normal and various fault feature samples have been input during the model training phase, giving it strong fault identification capabilities.

[0129] Anomaly Matching 2: If the AI ​​model infers that the device is abnormal, the extracted abnormal features are matched with a preset fault feature library. The fault feature library stores feature templates corresponding to various faults (such as mechanical wear, circuit faults, parameter abnormalities, etc.). Through feature similarity matching, the fault type corresponding to the anomaly is initially determined.

[0130] Fault type and severity determination: Based on the anomaly matching results, the specific type of fault (such as bearing wear, circuit short circuit, etc.) is determined by using preset judgment rules (such as feature similarity threshold, fault impact classification standard) and the fault severity (such as minor fault, general fault, severe fault) is determined, providing a clear basis for subsequent processing and decision-making.

[0131] After the two parallel branches have completed their processing, the "lightweight encoded compressed data" output by the data compression module and the "fault diagnosis results" (fault type, fault level, abnormal characteristics, etc.) output by the AI ​​fault identification module are fused to form the final processing result. During the fusion process, it is necessary to ensure the correspondence between the two types of data, that is, each piece of compressed data corresponds to a unique fault diagnosis result, so that data and fault information can be quickly associated during subsequent storage, transmission and application, achieving "data traceability and fault queryability".

[0132] Data storage / transmission / host computer display, as the final step in the process, focuses on applying the integrated processing results to meet the storage, transmission, and visualization needs of real-world scenarios. The specific principles are as follows: Data storage: The fused compressed data and fault identification results are stored on designated storage devices (such as local servers, cloud databases, embedded storage modules, etc.) using a categorized storage method to facilitate subsequent querying, retrieval, and data backtracking. At the same time, the lightweight compressed data can significantly reduce the storage device occupancy rate.

[0133] Data transmission: The fused results are transmitted to relevant devices or systems (such as host computers, monitoring centers, mobile devices, etc.) via appropriate transmission protocols (such as TCP / IP, MQTT, etc.; MQTT is preferred for remote transmission of industrial equipment, while TCP / IP is commonly used for local area network transmission). Because the data has been compressed, the bandwidth requirement is reduced, transmission efficiency is improved, and it can adapt to scenarios involving long-distance, multi-device simultaneous transmission.

[0134] The host computer display shows the fault identification results (fault type and level) and the core information of the compressed data (such as key features and data status) on the host computer interface (such as numerical display, fault alarm, trend curve, etc.), which makes it easy for staff to monitor the equipment operation status in real time, quickly obtain fault information, and take timely countermeasures (such as fault repair, parameter adjustment, etc.).

[0135] The entire process integrates data processing and fault identification through the logic of "preprocessing and purifying data, extracting and filtering core features, parallel processing to improve efficiency, fusion output for practicality, and storage, transmission, and application." Specifically, the parallel branch design significantly shortens the overall processing time, the data compression module achieves data lightweighting, and the AI ​​fault identification module ensures diagnostic accuracy. The fused data is easy to store and transmit, and provides accurate fault information for equipment maintenance, forming a complete closed loop of "collection-processing-diagnosis-application," adaptable to various scenarios such as industrial equipment monitoring and intelligent diagnostics.

[0136] See Figure 4 This demonstrates that as the environmental severity increases, the success rate of the present invention decreases slowly, while the success rate of the traditional solution decreases sharply. With increasing environmental severity, the packet loss rate of the present invention is significantly lower than that of the traditional solution. Furthermore, with increasing environmental severity, the latency increase of the present invention is significantly less than that of the traditional solution.

[0137] See Figure 5 Traditional solution: segmented broken line + discontinuous bar chart, resulting in large fluctuations and high redundancy. Invention solution: smooth, continuous solid bar chart, resulting in stable transmission and reduced redundancy.

[0138] Traditional solution: large fluctuations in the curve, high peak value, and many redundant segments. Invention embodiment: smooth curve, lower peak value, and significantly reduced total transmission volume. Visual demonstration: Compared to the traditional solution, this invention reduces data transmission volume and optimizes bandwidth usage.

[0139] It should be noted that the brain-like (AI) model of the present invention is not limited to the GRU network, CNN and LSTM fusion model in the above embodiments. An appropriate AI model can be selected according to the actual application requirements. The specific architecture of the digital computing array can be adjusted according to the scale and performance requirements of the phased array antenna. The specific parameters of fault identification and the data compression ratio can be optimized according to the user system requirements. All of these are within the protection scope of the present invention.

Claims

1. A hierarchical closed-loop brain-like AI-driven software-based adaptive phased array implementation method, characterized in that, A four-tiered closed-loop architecture is constructed, consisting of a bottom-level hardware execution layer, a mid-level intelligent computing power layer, and a top-level overall planning layer. This architecture integrates a fully digital computing array with a brain-like pulse intelligent model, including the following steps: Step 1: Construct a heterogeneous neuromorphic hybrid signal architecture based on a spin orbital moment nano oscillator (STNO) array. In the array element-level fully independent transmit / receive digital architecture, a sub-nanosecond pulse generation path driven by magnetization dynamics is embedded. Utilizing the sensitivity of the STNO's self-sustaining oscillation frequency to externally injected current, instantaneous physical layer frequency locking and phase synchronization based on the spin-torque effect are achieved in the RF link, replacing traditional phase-locked loops and digital phase shifters. This enables a fully software-based process for RF signal reception, preprocessing, intelligent analysis, digital beamforming, and signal transmission. The STNO array adopts a multi-domain wall-coupled spin valve stacking structure, with two STNO differential pairs within the same array element oscillating in opposite directions to construct a spin differential mode phase reference. This eliminates the intrinsic 1 / f low-frequency phase noise of a single STNO device, forming a native differential phase reference source that does not depend on an external clock. Step 2: Establish a multi-scale environment-hardware joint evolution model based on a fractional-order memristor chaotic system. Introduce the Caputo-Fabrizio fractional derivative operator to describe the non-integer order thermal relaxation behavior of the array antenna under extreme temperature variations of -40℃ to 85℃. Combine the hardware-in-the-loop fractional-pulse timing-dependent plasticity (fractional-STDP) learning rule to couple and model the non-Markovian noise caused by strong electromagnetic suppression interference. Generate a nonlinear fractional-order dynamic zeroing strategy for transient interference through dynamic adaptive adjustment of the fractional order. Construct a coupled dynamic equation between the fractional order α and the STNO domain wall migration rate, mapping the continuous change in memristor conductance to the effective field strength of the spin orbit moment, and realize the unified modeling of the coupled dynamics of temperature, memristor chaos, and spin oscillation. Step 3: Integrate a persistent coherence anti-interference mechanism based on topology data analysis (TDA) into the fractional-order memristor chaotic system. Utilize the Vietoris-Rips complex to perform persistent coherence group calculations on the time-frequency phase space point cloud of the enemy's intentional narrowband interference signal. Extract the longest survival interval from the H1 persistent barcode as the interference fingerprint feature. Generate an anti-interference beamforming weight matrix using a topology optimization algorithm based on Morse theory to achieve topological robustness anti-interference of the communication link in high-speed mobile Doppler frequency offset scenarios. Introduce a layered threshold persistent coherence layered filtering mechanism to distinguish between short-lifetime topological features of spurious noise and long-period stable topological loops of man-made interference. Synchronously output the gradient flow of interference topological features to the fractional-order memristor chaotic system to directly correct the fractional-order α, establishing a real-time bidirectional linkage closed loop of topological anti-interference and thermal relaxation compensation. Step 4: Construct a neuromorphic architecture search (NAS) scheduler based on meta-reinforcement learning (Meta-RL) to monitor the leakage current, power consumption, and heat dissipation boundary conditions of the FPGA parallel heterogeneous computing architecture in real time. When the computing load is detected to be approaching the thermal runaway critical point, a spiking neural network structure reorganization mechanism based on an evolutionary strategy is triggered. The neuron connection topology, spiking threshold, and synaptic delay are evolved online through a genetic algorithm to automatically generate an event-driven sparse spiking network that meets energy efficiency constraints. A two-layer meta-learning reward composite function is designed. The first layer is the TDA topology anti-interference survival interval weight, and the second layer is the STNO differential phase noise attenuation coefficient. The dual weights are dynamically normalized and used as the fitness evaluation criterion for the evolutionary algorithm. Step 5: Establish a cross-level closed-loop feedback mechanism based on a hybrid approach of microwave photonics and neuromorphology. The RF air interface electromagnetic field distribution data and hardware state parameters of the bottom-level software-defined antenna function output layer are converted into optical carrier phase information via a Mach-Zehnder modulator. This information is directly mapped to the resonant wavelength offset of the silicon-based microring resonator array in the top-level brain-like intelligent phased array coordination platform. Through distributed asynchronous updates via optical interconnects, a dynamic optimal balance between beamforming accuracy and hardware energy efficiency is achieved. The silicon-based microring resonator array employs a multi-channel coupled resonant cavity parallel phase demodulation structure, converting the optical carrier phase offset into a discrete optical pulse sequence. This asynchronous optical pulse flow directly drives the domain wall modulation current of the STNO differential pair, forming a three-level hardware-level native closed-loop feedback system (optical-memristor-spin), eliminating the need for intermediate buffers in digital-to-digital conversion and compressing the cross-level feedback delay to within 200 ps.

2. The hierarchical closed-loop brain-like AI-driven software-based adaptive phased array implementation method according to claim 1, characterized in that, There is a synergistic coupling between the spin orbital moment nano oscillator (STNO) array in step 1 and the fractional-order memristor chaotic system in step 2: the fractional-order α output by the fractional-order memristor chaotic system modulates the injection current density of the STNO array in real time, thereby changing the self-sustaining oscillation frequency and lock-in bandwidth of the STNO; the optimal injection current density is solved in real time based on the coupling dynamics equation, and the domain wall reverse migration rate of the dual STNO differential pair is dynamically matched, so that the differential phase reference noise suppression depth is adaptively adjusted with temperature.

3. The hierarchical closed-loop brain-like AI-driven software-based adaptive phased array implementation method according to claim 1, characterized in that, There is a synergistic coupling between the persistent coherence anti-interference mechanism of Topological Data Analysis (TDA) in step 3 and the neuromorphic architecture search (NAS) scheduler of Meta-Reinforcement Learning (Meta-RL) in step 4: the longest survival interval length of the H1 persistent barcode extracted by TDA is used as part of the environmental reward signal of the meta-learner, which guides the NAS scheduler to select and retain neuronal subnetworks with high topological robustness in the evolutionary strategy, while discarding redundant connections that are unrelated to persistent coherence features.

4. The hierarchical closed-loop brain-like AI-driven software-based adaptive phased array implementation method according to claim 1, characterized in that, The cross-level closed-loop feedback mechanism of microwave photonics and neuromorphology hybrid described in step 5 is synergistically coupled with the STNO array described in step 1: the resonant wavelength offset of the silicon-based microring resonator array is converted into optical pulse delay difference through an opto-oscillator (OEO), and this delay difference is injected in reverse into the bias current port of the STNO array to form an optical-magnetic hybrid phase-locked loop; using the optical bistable hysteresis range of the multi-cavity silicon-based microring, an optical delay difference threshold criterion is constructed, and STNO current correction is triggered only when the phase jitter exceeds the threshold.

5. The hierarchical closed-loop brain-like AI-driven software-based adaptive phased array implementation method according to claim 1, characterized in that, The Topology Data Analysis (TDA) anti-interference mechanism in step 3 also integrates a fault diagnosis branch based on causal inference. This branch uses the Peter-Clark (PC) algorithm to learn the causal graph from the time-delayed embedding vector of the chaotic attractor trajectory. Through Do-calculus, it performs interventional causal inference on three types of faults: channel failure, amplitude and phase drift, and radio frequency loss. It only returns the fault category and the corresponding causal effect value, achieving a data compression ratio of ≥98% while maintaining a fault identification accuracy of ≥99%. The fault causal effect value is synchronously input into the fractional-order memristor chaotic system to directly correct the upper and lower limits of the fractional-order α, realizing adaptive degradation compensation of the thermal relaxation model under fault conditions. The fault channel can dynamically reduce the beamforming weight, forming an integrated linkage closed loop of interference identification, fault diagnosis, and thermal compensation.

6. The hierarchical closed-loop brain-like AI-driven software-based adaptive phased array implementation method according to claim 1, characterized in that, The fractional-order memristor chaotic system in step 2 also integrates an adaptive fractional-order order adjustment module. This module dynamically adjusts the fractional-order order α based on the real-time acquired antenna array temperature gradient and the Hurst exponent of the vibration spectrum. This allows the model to reduce the order to 0.85 at a low temperature of -40℃ to enhance long-range memory compensation for the decrease in electron mobility, and increase the order to 1.15 at a high temperature of 85℃ to suppress false triggering caused by thermal noise. This achieves adaptive maintenance of beam pointing accuracy across the entire temperature range, with a pointing error ≤0.01°. A vibration-temperature coupled Hurst joint feature is introduced to construct a two-dimensional lookup table to dynamically constrain the step size of the fractional-order order adjustment.

7. The hierarchical closed-loop brain-like AI-driven software-based adaptive phased array implementation method according to claim 1, characterized in that, The specific components and linkage mechanism of the four-level hierarchical closed-loop architecture are as follows: The underlying software-defined antenna function output layer is equipped with a spin orbital moment nano oscillator (STNO) array and a silicon-based photonic integrated transceiver module. As a hardware execution and data feedback terminal, it undertakes multi-task signal transmission and reception work and transmits back hardware status, task data, and environmental parameters in real time. The underlying hardware adopts a differential dual STNO spin valve stacked array, matched with a parallel multi-channel silicon-based micro-ring optical demodulation unit. The native hardware differential architecture provides a unified phase reference for the upper four-level closed loop. The middle sensing input layer is equipped with a fractional-order capacitance equivalent circuit based on TiO2 memristors and a microwave photonics downconversion module, which is seamlessly connected to the RF front-end array antenna to complete the fractional-order domain acquisition and adaptive noise reduction preprocessing of the air interface electromagnetic signal. TiO2 memristor circuits have a built-in fractional dynamic zero-adjustment hardware arithmetic unit that directly solves the Caputo-Fabrizio fractional derivative in hardware. The middle-layer neuromorphic AI computing layer is equipped with a fractional-order memristor chaotic system hardware accelerator and a topology data analysis (TDA) coprocessor, serving as the core intelligent decision-making hub. It completes persistent coherence feature extraction, causal inference, and iterative output of the optimal solution for beam strategy. The TDA coprocessor is directly connected to the fractional-order memristor accelerator hardware, and the on-chip bus directly interacts with topology features and fractional-order parameters without external DDR cache relay, realizing real-time bidirectional linkage between topology anti-interference and thermal model. The middle beam control layer is equipped with a topology optimization algorithm module based on Morse theory and an optical interconnect beamforming controller, which works in conjunction with the beamforming module and the amplitude and phase calibration compensation module to achieve topological robust dynamic and precise adjustment of beam pointing, beamwidth, and null position. The optical interconnect beamforming controller has a built-in optical pulse delay analysis circuit, which directly outputs analog bias current to drive the bottom STNO array, skipping the digital DA conversion stage. The top-level brain-like intelligent phased array coordination platform is equipped with a Meta-Reinforcement Learning (Meta-RL) neuromorphic architecture search (NAS) scheduler and a silicon-based microring resonator array, which uniformly manages all functional levels below and achieves sub-microsecond seamless switching between communication / radar / anti-interference working modes. The top-level NAS scheduler is bidirectionally interconnected with the silicon-based micro-ring array optical signal. Optical pulses serve as cross-level scheduling instructions, replacing the traditional high-speed digital bus and enabling multi-service concurrent scheduling throughput.

8. The hierarchical closed-loop brain-like AI-driven software-based adaptive phased array implementation method according to claim 7, characterized in that, The mid-layer neuromorphic AI computing layer also integrates an on-chip causal discovery engine. This engine utilizes the synchronization phenomenon of spin orbital moment nano oscillators (STNOs) to solve the Ising model annealing problem. It learns the causal structure from real-time data through a hardware-accelerated PC algorithm. It reuses the dual STNO differential oscillation synchronization signal as the global synchronization clock for Ising annealing, synchronizing the hardware timing of causal graph solving, topological feature extraction, and fractional-order chaotic modeling. When switching between multiple scenarios, the causal model, anti-interference topology, and thermal compensation model are updated synchronously, eliminating the time difference error caused by serial updates of multiple models.

9. The hierarchical closed-loop brain-like AI-driven software-based adaptive phased array implementation method according to claim 1, characterized in that, The spiking neural network structure recombination mechanism based on evolutionary strategy described in step 4 specifically employs a genetic algorithm that combines tournament selection with an elite retention strategy. In each evolutionary iteration, the top 10% of individuals with the highest fitness are retained, and the remaining individuals undergo crossover and mutation operations. The crossover and mutation probabilities are dynamically adjusted by the meta-learner based on the complexity of the current electromagnetic environment. Triple constraint evolutionary screening rules are set: a lower limit for topology anti-interference index, an upper limit for STNO differential phase noise, and an array power consumption threshold. Individuals are directly eliminated if any one of these criteria is not met. The evolutionary process simultaneously constrains the performance of radio frequency hardware.

10. A neuromorphic software-defined phased array system implemented using the method described in any one of claims 1-9, characterized in that, include: The top-level brain-like intelligent phased array coordination platform is equipped with a meta-reinforcement learning (Meta-RL) neuromorphic architecture search (NAS) scheduler and a silicon-based microring resonator array; The silicon-based microring resonator array adopts a multi-cavity optical bistable parallel demodulation structure, which, together with the underlying STNO differential array, forms an all-optical cross-layer feedback link. The middle layer of sensing input is configured with a fractional-order capacitor equivalent circuit based on TiO2 memristors and a microwave photonics downconversion module; the TiO2 memristor fractional-order circuit integrates a hardware fractional-order derivative operation unit, which is directly connected and interconnected with the TDA coprocessor on-chip. The middle-layer neuromorphic AI computing layer is equipped with a fractional-order memristor chaotic system hardware accelerator and a topology data analysis (TDA) coprocessor; the fractional-order memristor accelerator, the TDA coprocessor, and the on-chip causal discovery engine share the STNO differential synchronization global clock; The middle beam control layer is equipped with a topology optimization algorithm module based on Morse theory and an optical interconnect beamforming controller; the output of the optical interconnect beamforming controller is an analog optical control current that is directly connected to the STNO array bias port. The bottom layer is a software-defined antenna function output layer, which is equipped with a spin orbital moment nano oscillator (STNO) array and a silicon-based photonic integrated transceiver module. The bottom STNO array is a dual-device differential spin valve stacked structure, which forms a native closed-loop hardware path of optical-magnetic coupling with the top silicon-based micro-ring array.

Images