A reconfigurable photonic sampling analog-to-digital converter control method, system and apparatus

Abstract

The application belongs to the technical field of integrated circuits, and particularly relates to a reconfigurable photonic sampling analog-digital converter control method, system and device, which comprises the following steps: S1, obtaining a starting safety configuration, performing an initialization self-check, and when the initialization self-check results all satisfy preset safety starting conditions, the initialization self-check is passed; S2, after the initialization self-check is passed, a target analysis frequency band is determined based on a target working interval, an initial sampling configuration is obtained, calibration samples corresponding to each activated channel are collected, and a coarse calibration configuration is obtained; S3, continuously sampling an analog signal to be converted to generate a performance state frame, performing deep learning analysis based on the performance state frame, and outputting an error estimation result; and S4, generating a candidate control vector according to the error estimation result, performing safety constraint auditing on the candidate control vector, taking the candidate control vector that passes the auditing as a system control vector, and performing target parameter soft updating in the system control vector. The application can be used for data conversion circuit design in a high-performance measurement system.

Description

Technical Field

[0001] This invention belongs to the field of integrated circuit technology, specifically relating to a control method, system, and device for a reconfigurable photonic sampling analog-to-digital converter. Background Technology

[0002] Analog-to-digital converters (ADCs), as key components in analog signal processing systems, are widely used in fields such as communications, radar, sensors, and testing and measurement. Especially in the areas of high-speed signal sampling and high-precision quantization, the limitations of traditional ADCs in terms of performance, flexibility, and stability are becoming increasingly apparent.

[0003] Photonic links introduce more complex non-ideal factors than purely electronic links, including: intensity fluctuations and residual jitter of the optical pulse source; bias point drift of the electro-optic modulator (MZM); temperature drift of the center wavelength of the wavelength demultiplexing device; inconsistency in the responsivity of the multiplexed photodetectors (PDs); and gain / bias mismatch between the back-end transimpedance amplifier (TIA) and the analog channel. These factors exhibit a superposition of slow and fast changes during temperature variations and aging, making it difficult for traditional offline one-time calibration methods to maintain the long-term high performance of the system.

[0004] Existing solutions mostly adopt fixed time division multiplexing (TDM), wavelength division multiplexing (WDM) or spectrum slicing architectures and fixed back-end processing parameters. Once the chip is fabricated or the system is built, the trade-offs between its core indicators such as the number of channels, equivalent sampling rate, working mode, bandwidth, resolution and power consumption are locked, which limits the applicable scenarios and performance ratio of the system.

[0005] Most existing systems employ open-loop or simple feedback control strategies, with parameter adjustments often being static or based on finite rules. This makes it impossible for the system to automatically adapt to changes in device aging, environmental disturbances, or input signal characteristics, making it difficult to guarantee long-term stability and optimal performance. Summary of the Invention

[0006] The purpose of this invention is to provide a control method, system, and apparatus for a reconfigurable photon sampling analog-to-digital converter.

[0007] A method for controlling a reconfigurable photon sampling analog-to-digital converter includes the following steps: S1. Obtain the startup security configuration and perform an initialization self-test. If the initialization self-test results meet the preset security startup conditions, the initialization self-test is deemed to have passed, and S2 is executed. S2. After the initial self-test passes, determine the target analysis frequency band based on the target working range; Based on the target application scenario, select the working mode from time interleaving mode, wavelength multiplexing mode and spectrum slicing mode; The initial sampling configuration is determined based on the startup security configuration, working mode and target analysis frequency band. Under the initial sampling configuration, calibration samples corresponding to each active channel are collected from the target analysis frequency band. Based on the calibration samples, the channel error vector corresponding to each channel is constructed. When the corresponding error component meets the preset threshold requirement, the coarse calibration configuration is obtained. S3. Under coarse calibration configuration, the analog signal to be converted is input into the reconfigurable multi-channel photonic sampling module. The optical domain is sampled according to the selected working mode and distributed to M channels. After photoelectric conversion, M electrical signals are obtained. The M electrical signals are then subjected to transimpedance amplification, programmable gain and bias adjustment, ΣΔ oversampling noise shaping and quantization, and digital filtering and decimation downsampling to obtain M digital signals. Digital signals and intermediate monitoring data from each channel are collected and stitched together to form a performance status frame. The performance status frame is then input into a pre-trained deep learning error estimation model, which analyzes and outputs the error estimation results for each channel. S4. Generate candidate control vectors based on the error estimation results, review the candidate control vectors for safety constraints, and use the approved candidate control vectors as system control vectors. Perform soft updates of the target parameters in the system control vectors.

[0008] In S2, the target analysis frequency band is determined based on the target working interval, specifically as follows: When the target working interval is represented by a frequency range [lowest frequency, highest frequency], this frequency range is used as the initial analysis frequency band; When the target working interval is at the center frequency and target bandwidth When indicating, will As the initial analysis frequency band; The target analysis frequency band is obtained by taking the intersection of the initial analysis frequency band and the frequency band that the system hardware can cover.

[0009] The digital signals and intermediate monitoring data from each channel are collected in S3 and spliced ​​together to form a performance status frame. The specific method is as follows: Continuous sampling is performed using a sliding window method. Each window contains N consecutive sampling points, and an overlapping area of ​​M consecutive sampling points is retained between adjacent windows. During window acquisition, a data validity flag is generated according to the sampling clock. The data validity flag is preprocessed, and time domain features, frequency domain features, bitstream statistical features, and channel consistency features are extracted and encapsulated as a performance status frame.

[0010] In S4, candidate control vectors undergo safety constraint review. The approved candidate control vectors are then used as the system control vectors. The specific method is as follows: Based on risk indicators, hard fault flags, and model confidence, candidate control vectors are marked as allowed to execute, restricted to execute, or prohibited to execute: When all risk indicators are below the first safety threshold, the model confidence level meets the standard, and there are no hard faults, it is marked as allowed to execute and the review is approved. When the risk indicator exceeds the second safety threshold, the model confidence level is lower than the standard lower limit, or a hard failure occurs, it is marked as prohibited from execution; Other cases are marked as restricted execution; The approved candidate control vectors are used as system control vectors.

[0011] Before generating the performance status frame in S3, the following steps are also included: Before collecting digital signals and intermediate monitoring data from each channel, a stabilization waiting time is set, and sampling points generated during the stabilization waiting period are discarded.

[0012] In S2, the channel error vector corresponding to each channel is constructed based on the calibration samples, specifically as follows: , in, Indicates the first The relative amplitude error of each channel, Indicates the first DC bias error of each channel Indicates the first The relative delay error of each channel, Indicates the first Optical power error of each channel.

[0013] In S4, candidate control vectors are generated based on the error estimation results. Specifically, based on the error estimation results, combined with the current target application mode, hardware constraints, and stability margin, initial parameters are generated, and the initial parameters are converted into candidate control vectors.

[0014] A reconfigurable photon sampling analog-to-digital converter control system, used to implement the aforementioned reconfigurable photon sampling analog-to-digital converter control method, includes: The initialization module obtains the startup security configuration and performs an initialization self-test. If the initialization self-test results meet the preset security startup conditions, the initialization self-test is deemed to have passed. The coarse registration module determines the target analysis frequency band based on the target working interval after the initial self-test passes. Based on the target application scenario, select the working mode from time interleaving mode, wavelength multiplexing mode and spectrum slicing mode; The initial sampling configuration is determined based on the startup security configuration, working mode and target analysis frequency band. Under the initial sampling configuration, calibration samples corresponding to each active channel are collected from the target analysis frequency band. Based on the calibration samples, the channel error vector corresponding to each channel is constructed. When the corresponding error component meets the preset threshold requirement, the coarse calibration configuration is obtained. The error estimation module, under coarse calibration configuration, inputs the analog signal to be converted into the reconfigurable multi-channel photonic sampling module, performs optical domain sampling according to the selected working mode and distributes it to M channels. After photoelectric conversion, M electrical signals are obtained. The M electrical signals are then subjected to transimpedance amplification, programmable gain and bias adjustment, ΣΔ oversampling noise shaping and quantization, and digital filtering and decimation downsampling to obtain M digital signals. Digital signals and intermediate monitoring data from each channel are collected and stitched together to form a performance status frame. The performance status frame is then input into a pre-trained deep learning error estimation model, which analyzes and outputs the error estimation results for each channel. The update module generates candidate control vectors based on the error estimation results, performs safety constraint review on the candidate control vectors, and uses the approved candidate control vectors as the system control vectors, and performs soft update of the target parameters in the system control vectors.

[0015] The system is applied to a reconfigurable multi-channel photonic sampling analog-to-digital converter. The front end of the reconfigurable multi-channel photonic sampling analog-to-digital converter is configured to support time-interleaving mode, wavelength multiplexing mode, and spectrum slicing mode through its internal optical switch matrix, tunable optical delay unit, and tunable optical filter. It can also dynamically switch between the three modes during operation according to the mode switching instructions in the candidate control vector.

[0016] A reconfigurable photonic sampling analog-to-digital converter control device includes a processor and a memory, wherein the processor executes a computer program stored in the memory to implement the reconfigurable photonic sampling analog-to-digital converter control method.

[0017] The main beneficial effects of this invention are as follows: This invention designs the multi-channel approach (TDM / WDM / spectral slicing) of the photonic sampling front-end as a configurable structure, achieving dynamic reconstruction of the channelized mode through an optical switch matrix, an adjustable optical delay unit, and a programmable filter. Simultaneously, the oversampling rate (OSR) and decimation ratio of the back-end multi-channel Sigma-Delta modulator array can also be adjusted as needed, balancing ultra-wideband acquisition with high-precision quantization.

[0018] This invention introduces deep learning for real-time analysis of the digital output bitstream. Utilizing the powerful nonlinear fitting capabilities of neural networks, it accurately extracts subtle features hidden in the output spectrum, including: gain / bias / delay mismatches in each channel, nonlinear distortion of the optoelectronic link, intermodulation distortion caused by modulator bias point drift, and even slowly varying errors due to temperature variations. Compared to traditional pilot- or correlation-based detection methods, the deep learning model can separate and quantify the coupling effects of multiple errors in complex noise backgrounds, obtaining a more accurate representation of the system state.

[0019] This invention constructs a complete dynamic adaptive control process. This process is driven by a state machine and includes initialization self-test, rapid coarse calibration, steady-state acquisition, performance index monitoring, parameter update decision-making, and stability protection. Based on the error estimate obtained from deep learning analysis, the optical parameters, analog front-end parameters, and digital processing parameters are adjusted in a closed loop.

[0020] To address potential stability issues in high-order Sigma-Delta modulators during parameter adjustment, this invention introduces stability safeguards in the control flow. When the deep learning network detects signs of instability such as excessively large quantizer inputs or excessively high NTF out-of-band gain (OBG), it prioritizes a safe backoff strategy to ensure the system always operates in a stable region and avoids loop divergence. Detailed Implementation

[0021] Example 1 To further understand the content of this invention, the invention will be described in detail with reference to the embodiments.

[0022] This invention relates to a control method for a reconfigurable photonic sampling analog-to-digital converter, applicable to reconfigurable multi-channel photonic sampling Σ-Δ ADC circuits, including an input terminal, a photonic sampling front-end, a Σ-Δ ADC array, and dynamic adaptive control of the digital back-end. The invention includes the following steps: S1. Obtain the startup security configuration and perform an initialization self-test. If the initialization self-test results meet the preset security startup conditions, the initialization self-test is deemed to have passed, and S2 is executed.

[0023] After power-on, the system enters the initialization secure boot phase, acquiring external boot commands, historical stable operation identifiers, hardware status registers, and pre-stored boot configuration tables, and loading a boot security configuration that matches the current boot conditions. .

[0024] Start security configuration This includes a general default security configuration and a mode-associated security configuration. The general default security configuration is used when the external boot command is missing; the mode-associated security configuration is used when the external boot command carries the target working area, target application mode, and priority performance indicators. Specific configurations include one or more of the following: safe optical power range, conservative front-end gain, initial channel mask, initial oversampling rate, initial decimation ratio, modulator safety order, digital synchronization configuration, and startup phase mode switching constraints.

[0025] After loading and starting the security configuration, an initialization self-test is performed. This initialization self-test includes at least the following: sampling clock lock status detection, light source lock status detection, optical power range detection for each channel, electro-optic modulator bias range detection, photodetector, TIA and PGA output unsaturation detection, and channel temperature detection. Modulator initial stability margin detection and back-end digital synchronization link validity detection.

[0026] If all the above test results meet the preset safe startup conditions, the initialization self-test is deemed successful. The current startup security configuration will then be updated. The available channel mask, hardware self-test results, parameter allowable range, and target working interval carried in the external startup command are sent to S2 for subsequent target analysis frequency band determination, initial operating mode selection, and rapid coarse calibration.

[0027] When there are only minor local bias exceedances, minor deviations in optical power of some channels, or minor exceedances in output amplitude of individual channels, and the number of effective channels is not lower than the minimum channel count threshold. If the error occurs, it is determined to be a recoverable startup anomaly. A startup-level safety correction is performed to restore the relevant parameters to the startup safety range, and then the initialization self-test is re-executed.

[0028] When the number of effective channels is not less than However, if some channels are unavailable, the local optoelectronic link margin is insufficient, or the modulator stability margin is insufficient, it is judged as a restricted startup anomaly. A degraded startup method is adopted, including reducing the number of initially active channels, reducing the front-end gain, reducing the optical power, increasing the oversampling rate, restricting mode switching, and loading conservative digital processing parameters, and then re-executing the initialization self-test.

[0029] When the sampling clock is locked, the light source is locked, or the number of effective channels is less than [a certain value] Front-end temperature exceeds severe threshold; critical simulation front-end remains saturated. If the modulator's initial stability margin is severely insufficient or the digital synchronization link is invalid, an unrecoverable startup anomaly is identified. Maintain or revert to the startup safety configuration. The rapid coarse calibration process is prohibited, and the system will enter safe mode or alarm state.

[0030] Through the above processes, S1 completes safe power-on, hardware availability verification, startup security configuration loading, and startup anomaly handling, and provides safe initial conditions for S2's rapid coarse calibration.

[0031] S2. After the initial self-test passes, determine the target analysis frequency band based on the target working range; Based on the target application scenario, select the working mode from time interleaving mode, wavelength multiplexing mode and spectrum slicing mode; The initial sampling configuration is determined based on the startup security configuration, working mode, and target analysis frequency band. Under the initial sampling configuration, calibration samples corresponding to each active channel are collected from the target analysis frequency band. Based on the calibration samples, the channel error vector corresponding to each channel is constructed. When the corresponding error component meets the preset threshold requirements, the coarse calibration configuration is obtained.

[0032] After the initial self-test passes, the system enters the rapid coarse calibration phase. It receives the current startup security configuration output from S1. The system uses available channel masks, hardware self-test results, parameter allowable ranges, and the target operating range carried in the external boot command to determine the target analysis frequency band. The initial sampling configuration is described. The rapid coarse calibration phase is used to perform preliminary alignment of the amplitude, DC bias, time delay, and optical power status of each active channel before the system enters steady-state acquisition.

[0033] The target analysis frequency band is determined based on the target working interval, specifically as follows: When the target working interval is represented by a frequency range [lowest frequency, highest frequency], this frequency range is used as the initial analysis frequency band; When the target working interval is at the center frequency and target bandwidth When indicating, will As the initial analysis frequency band; The initial analysis frequency band is expanded based on the preset protection bandwidth, filter transition bandwidth, or frequency estimation error. The intersection of the initial analysis frequency band and the frequency band that the system hardware can cover is then taken to obtain the target analysis frequency band. .

[0034] When S1 does not output a valid target working range, it enters autonomous sensing mode. Based on Using low front-end gain, safe optical power, high oversampling rate, and limited number of active channels as initial conditions, broadband coarse detection or segmented spectrum scanning is performed on the current input signal.

[0035] Based on the frequency band to be detected With current instantaneous coverage bandwidth Size relationship selects detection method: when Broadband coarse detection method is used; when The segmented spectrum scanning method is used.

[0036] Based on short-time spectrum estimation, power spectrum estimation, or sub-band energy detection results, candidate signal frequency bands with energy exceeding the preset noise floor margin and continuously existing within a short window are selected. The target analysis frequency band is then determined based on the frequency boundaries, guard bandwidth, and system hardware coverage of the candidate signal frequency bands. .

[0037] Determine the target analysis frequency band Then, based on the target application scenario, the operating mode is selected from time interleaving mode, wavelength multiplexing mode, and spectrum slicing mode. Specifically: When the target signal is an ultra-wideband transient signal, select the time interleaving mode; When the target signal is a multi-band parallel signal, select wavelength multiplexing mode; When the target signal is a broadband sparse spectrum signal, select the spectrum slicing mode; When the target signal spectrum distribution is unknown, the spectrum slicing mode is selected first for spectrum sensing, and the wavelength multiplexing mode or time interleaving mode is switched according to the sensing results.

[0038] according to Working mode, current startup security configuration The initial sampling configuration is generated using the available channel mask and the allowed range of parameters.

[0039] The initial sampling configuration includes one or more of the following: number of active channels, optical delay, optical power allocation, center wavelength, filter center frequency, filter bandwidth, PGA gain, channel bias, oversampling rate, decimation ratio, digital filter parameters, phase compensation amount, and reconstruction weight.

[0040] This invention provides differentiated optimal performance solutions for different signal characteristics by setting three switchable photon sampling modes: (1) The time interleaving mode improves the effective sampling rate by M times through M-channel arithmetic delay sampling, achieving ultra-high-speed sampling at the level of hundreds of GSPS, which is suitable for capturing ultra-wideband time domain signals such as radar pulses and transient interference; (2) Wavelength multiplexing mode utilizes the wavelength dimension to achieve fully parallel synchronous sampling of M channels, with each channel quantized independently, avoiding interleaving spurious signals and achieving the highest spurious-free dynamic range (SFDR), which is suitable for high dynamic range scenarios such as multi-carrier communication and MIMO reception; (3) The spectrum slicing mode achieves full coverage of the broadband spectrum with the lowest light source requirements and power consumption through frequency domain sub-band segmentation and digital reconstruction. It can also dynamically shut down idle channels, resulting in optimal power efficiency. It is suitable for broadband sparse signal scenarios such as spectrum monitoring, electronic reconnaissance, and cognitive radio. The three modes can be dynamically switched on the same hardware platform in milliseconds and support multi-mode collaborative workflow. A single system can cover diverse signal acquisition needs from transient to steady state, from narrowband to ultra-wideband, and from low power consumption to high dynamic range.

[0041] The initial sampling configuration parameters differ depending on the mode; they are not a uniform set. For time-interleaved mode, they must include at least the number of channels and the inter-channel delay difference Δτ. For wavelength multiplexing mode, they must include at least the number of channels and the center wavelength λ of each channel. m If it is a spectrum slicing mode, it shall at least include the number of sub-bands and the center frequency and bandwidth of each sub-band, and in combination with the startup security configuration and the target analysis frequency band, determine the specific value of the parameter type set as the initial sampling configuration.

[0042] Regardless of the mode, determining the initial sampling configuration follows a general three-step process: Determine the set of parameter types that need to be configured based on the working mode; Read hardware constraint boundaries based on startup security configuration; The specific values ​​of each parameter are calculated based on the target frequency band and engineering margin.

[0043] Furthermore, this can be understood through the following process: Under Time Interleaving Mode (TDM), the specific process for determining the initial sampling configuration is as follows: Step 1: Read the boot security configuration C safe From this, the highest stable sampling rate f of the single-channel ADC can be obtained. ch Maximum number of allowed active channels M max Safe optical power range [P] opt,min ,P opt,max Front-end gain safety limit G max Oversampling rate safety lower limit OSR min and Oversampling Rate Safety Limit (OSR) max ; Step 2: Obtain the upper frequency f of the target analysis band. max And determine the required equivalent sampling rate f of the system according to the Nyquist sampling theorem. eq ≥2f max And preferably leave a 10% to 20% margin in the project, i.e. f eq =α⋅2f max , where α∈[1.1,1.2]; Step 3: Calculate the minimum number of channels required based on the equivalent sampling rate and the highest stable sampling rate of a single-channel ADC. Let the number of activated channels M = min(M req +M margin M max ), where M margin ≥1 represents the number of reserved redundant channels (for subsequent channel degradation or abnormal rollback), with M being the preferred number. margin Take 1~2; Step 4: Calculate the equivalent sampling period T of the system based on the equivalent sampling rate. eq =1 / f eq Configure the optical delay of each channel according to an arithmetic progression sequence. Where m = 0, 1, ..., M-1; Step 5: According to C safe The safe optical power range [P] opt,min ,P opt,max [, for the optical power P of each channel] mPerform equal and uniform power distribution, and ensure P m ∈[P opt,min ,P opt,max ]; Step 6: Ensure the light source is configured as a single-wavelength ultrashort pulse mode with a pulse width less than 1 ps and a timing jitter less than 20 fs; ensure that all adjustable filters in each channel are uniformly set to wide passband with the same center wavelength and passband width covering... Full frequency band; determine the initial value of PGA gain for each channel as G. m =G init (G) init (Preset conservative value), initial bias value b for each channel m =0; Determine the initial value of the oversampling rate (OSR). init Taken from [OSR] min OSR max The lower value within the range (to reduce initial power consumption) is selected, and the value is lower than the initial value R. init According to OSR init Determine the target output data rate; Step 7: Set the above parameters {M,τ} m ,P m ,λ m =λ c G m ,b m OSR init ,R init The combination of filter configuration and light source configuration forms the initial sampling configuration in time-interleaved mode.

[0044] In wavelength multiplexing (WDM) mode, the specific process for determining the initial sampling configuration is as follows: Step 1: Read the boot security configuration C safe The total number of comb teeth N can be used to obtain the micro-comb light source. comb List of wavelengths for each comb tooth {λ comb ,1,λ comb,2 ,…,λ comb N comb} and the wavelength interval Δλ between adjacent comb teeth FSR (Free spectral range), Maximum allowed number of activated channels M max Safe optical power range [P] opt,min ,P opt,max ] Filter minimum bandwidth BW filter,min (Used to ensure inter-channel crosstalk is below -30dB), Front-end gain safety limit G max Oversampling rate safety lower limit OSR min and Oversampling Rate Safety Limit (OSR) max ; Step 2: Obtain the target analysis frequency band And according to the microcomb wavelength interval Δλ FSR With frequency spacing Conversion relationship between , For the speed of light, determine The number of comb teeth N that can be covered avail The control module is based on The coverage range is determined by selecting a set of candidate comb teeth that fall within the target frequency band from the list of microcomb wavelengths, and then selecting M comb teeth from this candidate set as the working wavelengths λ0, λ1, ..., λ M-1 Where M=min(N) avail M max N comb The wavelengths are arranged in ascending order and the interval between adjacent wavelengths is Δλ. FSR Integer multiples of; Step 3: Assign a corresponding center wavelength λm to each wavelength channel, and determine the center frequency f of the tunable filter for each channel. c,m =c / λ m Filter passband width BW m Set to be greater than the signal modulation bandwidth but less than / 2, to ensure that crosstalk between adjacent channels is suppressed below -30dB; preferably, BW m Take 1.2 to 1.5 times the signal bandwidth; Step 4: According to C safe The safe optical power range [P] opt,min ,P opt,max [and the historical loss pre-compensation values ​​for each wavelength link, to uniformly allocate the initial optical power P to each channel.] m Alternatively, pre-compensation allocation can be performed according to a preset wavelength loss table, and P can be ensured. m ∈[P opt,min ,P opt,max ]; Step 5: Ensure the light source is configured in multi-wavelength synchronous pulse mode, with the optical pulses corresponding to each wavelength being completely synchronized in the time domain; set the initial value of each channel's optical delay unit to a small calibration value τ. m =τ cal,m Used to compensate for time delay deviations caused by differences in optical paths, the initial value of the PGA gain for each channel is set to G. m =G init Initial bias value b for each channel m =0; Determine the initial value of the oversampling rate (OSR). init Taken from [OSR] min OSR max High values ​​within the range (to improve resolution), extract values ​​higher than the initial value R. init Increase accordingly to match the output data rate under high OSR; Step 6: The control module sets the above parameters {M,λ} m BW m P m ,τ m G m ,b m OSR init ,R init The combination of filter configuration and light source configuration forms the initial sampling configuration in wavelength multiplexing mode.

[0045] In spectrum slicing mode, the specific process for determining the initial sampling configuration is as follows: Step 1: The control module reads the boot security configuration C safe Obtain the maximum number of allowed active channels M from it. max Safe optical power range [P] opt,min ,P opt,max ] Filter out-of-band rejection ratio (preferably greater than 25dB), front-end gain safety limit G max Oversampling rate safety lower limit OSR min and Oversampling Rate Safety Limit (OSR) max Minimum subband bandwidth BW sub,min This is used to limit the difficulty in implementing filters caused by excessively narrow subbands; Step 2: Obtain the target analysis frequency band Total bandwidth B total And determine the effective signal based on prior signal information or fast spectrum analysis results. The energy distribution within the signal; if the signal is known prior to be sparsely distributed (e.g., the signal exists only in certain sub-bands), then candidate signal sub-bands are selected based on a preset energy threshold; if the signal is unknown prior to be distributed, then... Divided into M at equal intervals max One pre-segmented sub-band is used as a candidate; Step 3: Based on M max B total BW sub,min Based on the signal energy distribution, determine the number of sub-bands M and the center frequency f of each sub-band. c,m and bandwidth B m The following conditions must be met: each sub-band continuously covers the target frequency band in the frequency domain. An overlap region δB is set between adjacent sub-bands. m (Preferably 5%~15% of the subband bandwidth), each subband bandwidth B m ≥BW sub,min And M≤M max For sparse signals, channels are activated only for sub-bands containing valid signals, and no channels are allocated for other frequency bands. Step 4: Set the corresponding adjustable filter parameters for each active sub-band: center frequency fc,m Passband width B m Out-of-band rejection greater than 25dB, and ensure that no band gaps are generated between the passband and transition band of adjacent sub-band filters; Step 5: According to C safe The safe optical power range [P] opt,min ,P opt,max The optical power of each channel is initially allocated uniformly, P. m And ensure P m ∈[P opt,min ,P opt,max The initial value of the optical delay unit for each channel is set to τ. m =τ cal,m Used to compensate for the time delay differences between the optical path and the filter group, the initial value of the PGA gain for each channel is set to G. m =G init Initial bias value b for each channel m =0; Step 6: The control module determines that the light source is configured as a broadband flat continuous light mode with spectral coverage. Full frequency band; determine the ADC sampling rate f for each channel. s,m Satisfy f s,m ≥2B m Due to the narrower subband bandwidth, the required sampling rate for each channel is significantly lower than in TDM mode, resulting in a lower oversampling rate (OSR). init According to OSR=f s,m / (2B m The extraction ratio R is calculated and determined. init Determine the initial values ​​of the digital reconstruction parameters based on the target output data rate, including the gain normalization factor w for each sub-band. gain,m =1, Phase compensation amount ϕ m =0, the overlapping area splicing weight is preset according to linear or cosine weighting method; Step 7: Set the above parameters {M,f} c,m B m ,P m ,τ m G m ,b m OSR init ,R init ,w gain,m ,ϕ m The combination of filter configuration and light source configuration forms the initial sampling configuration in the spectrum slicing mode.

[0046] Under the initial sampling configuration, from the target analysis frequency band Internally collect calibration samples corresponding to each activated channel. The calibration samples include those for each activated channel. The data includes one or more of the following: bitstream, decimation output, reconstructed output, channel short-time energy, root mean square amplitude, spectral integral energy, DC component, inter-channel cross-correlation results, cross-power spectrum phase slope, time-interleaved spurious characteristics, optical power monitoring value, quantizer input peak value, integrator amplitude, overflow count, and digital synchronization status. Sampling windows exhibiting front-end saturation, channel out-of-synchronization, optical power exceeding limits, modulator overload, or transient changes are removed. Based on calibration samples, a channel error vector is constructed for each channel. Specifically: , in, Indicates the first The relative amplitude error of each channel, Indicates the first DC bias error of each channel Indicates the first The relative delay error of each channel, Indicates the first Optical power error of each channel.

[0047] In one implementation, each error component satisfies the following condition: , , , , in, Indicates the first Each channel is in The amplitude, root mean square value, or spectral integral energy within the range; Indicates the first DC component of each channel; Indicates the first Optical power monitoring output for each channel; , and These represent the effective channel mean, target equilibrium value, or historical stable value of the corresponding parameter, respectively. Indicates the first Index of latency estimates for each channel; Indicates the ideal alignment position; Indicates the current sampling period.

[0048] Based on the channel error vector, the parameters of each channel are corrected in the following order: first, the gain parameter is corrected. With bias parameters Then correct the optical delay or equivalent sampling delay parameters. Finally, the optical power allocation parameters were adjusted. The above order is adopted because gain and bias errors affect the amplitude reference and DC operating point, which in turn affect the reliability of time delay estimation of cross-correlation peak, phase slope, or interleaved spurious characteristics; optical power adjustment affects the operating point of the front-end link and local nonlinearity, so it is preferable to perform the adjustment after the gain, bias, and time delay have initially converged.

[0049] Generate corresponding correction values ​​based on each error component: , , in, , , and These are the step size coefficients for gain, bias, time delay, and optical power corrections, respectively. Each correction is constrained by the allowable range of the parameters output by S1 and is iteratively updated with a preset step size. After each update, calibration samples are reacquired and the updated channel error vector is calculated.

[0050] When the corresponding error component meets the preset threshold requirement, the coarse calibration is successful. The corresponding parameter correction is stopped, the current coarse calibration result is latched, and the result configuration is saved to [the appropriate location / system / etc.]. .

[0051] When the preset maximum number of corrections is reached If the threshold requirement is still not met, the convergence status of each activated channel is determined. For the first... For each channel, the error vector is determined to be... The channels converged.

[0052] , , , , in, , , and These are the amplitude error threshold, DC bias error threshold, time delay error threshold, and optical power error threshold, respectively. When the... If any error component in a channel exceeds the corresponding threshold, the channel is determined to be a non-converged channel.

[0053] Count the number of non-converging channels Number of convergence channels and the remaining channels for target analysis frequency bands The coverage status. Degradation will be executed when the following conditions are met simultaneously: , in, This is the minimum effective channel number threshold.

[0054] During the degradation process, non-converged channels are either masked from the current set of active channels or their weight in the reconstructed output is reduced. A new degradation configuration is then generated based on the remaining converged channels. This degradation configuration includes one or more of the following: reducing the number of active channels, reducing front-end gain, reducing optical power, increasing the oversampling rate, increasing the decimation ratio, limiting the output data rate, freezing mode switching, and disabling channel expansion.

[0055] when At that time, instead of entering a downgraded operation, it will revert to the currently running security configuration. Then, re-execute the initialization self-check.

[0056] After degrading to a lower operating level, a status check is performed to check the system's stability margin and the coverage status of the target analysis frequency band.

[0057] When the remaining convergence channels can cover the target analysis frequency band The rollback is successful when the stability margin meets the following conditions, and the system enters the S3 steady-state data acquisition and performance monitoring phase under the downgraded configuration; otherwise, it rolls back to [previous configuration]. Then, re-execute the initialization self-check.

[0058] , , , , in Input the peak value to the quantizer. For the maximum stable input amplitude, A safety margin factor less than 1 For overflow rate, The maximum allowable overflow rate, For the first The optical power monitoring values ​​of each channel and These are the lower and upper limits of permissible optical power, respectively. This is the channel synchronization flag, which takes a value of 1 when channel synchronization is active.

[0059] S3. Under coarse calibration configuration, the analog signal to be converted is input into the reconfigurable multi-channel photonic sampling module. The optical domain is sampled according to the selected working mode and distributed to M channels. After photoelectric conversion, M electrical signals are obtained. The M electrical signals are then subjected to transimpedance amplification, programmable gain and bias adjustment, ΣΔ oversampling noise shaping and quantization, and digital filtering and decimation downsampling to obtain M digital signals. Digital signals and intermediate monitoring data from each channel are collected and spliced ​​together to form a performance status frame. The performance status frame is then input into a pre-trained deep learning error estimation model, which analyzes and outputs the error estimation results for each channel.

[0060] After the rapid coarse calibration is completed, the converged operating mode, number of active channels, optical delay, optical power allocation, PGA gain, channel bias, oversampling rate, decimation ratio, and digital reconstruction weights are latched as the current operating configuration. In maintaining Without any abrupt changes, the system transitions from rapid coarse calibration to steady-state acquisition. Steady-state acquisition does not inject an independent calibration signal; instead, it uses the current signal to be converted. Under continuous real-world operating conditions, the main acquisition link and the bypass monitoring link are sampled synchronously to obtain performance characteristics of the system under actual operating conditions.

[0061] Before generating the performance status frame, the process also includes: setting a stable waiting time before continuous sampling and discarding sampling points generated during the stable waiting period.

[0062] To avoid the transient response after rapid coarse calibration or mode switching affecting steady-state judgment, a stabilization waiting time is set before entering steady-state acquisition. and discard the previous generation generated during the stable waiting period. One sampling point.

[0063] The According to and sampling period Determined, preferably satisfying: , in, The response time is determined based on the maximum value among the digital filter settling time, ADC stabilization time, optical power stabilization time, and optical delay time, preferably 2 to 3 times the maximum transient response time of the system.

[0064] After the stabilization period ends, start the steady-state data window. During window acquisition, a valid window data flag is generated based on the sampling clock lock flag, channel synchronization flag, front-end saturation flag, and optical power over-limit flag. Specifically, in the steady-state data window The system tracks the number of sampling clock lockouts, channel synchronization failures, front-end saturation events, and optical power exceedances. When none of these counts exceed the corresponding preset threshold, [the system will proceed]. Set to valid, and perform preprocessing, feature extraction, and performance index calculation on the steady-state data window; when any statistical count exceeds the corresponding preset threshold, If the data window is set to invalid, it is discarded, and the next steady-state data window is collected sequentially. Discarded steady-state data windows are not included in the performance status frame. The generation is not used to update the most recent stable configuration. .

[0065] Furthermore, the optical domain sampling and distribution for different operating modes are performed as follows: Under Time Interleaving Mode (TDM): Light source configuration: The microcomb light source is configured in single-wavelength ultrashort pulse mode, with an output repetition frequency of f. s =1 / T s A regular light pulse sequence, T s The sampling interval is defined as follows: pulse width < 1 ps, jitter < 20 fs.

[0066] Sampling operation: The analog signal x(t) to be converted is first loaded onto an electro-optic modulator (MZM) to modulate a continuous optical carrier, generating a modulated optical signal. This modulated optical signal and an ultrashort pulse sequence simultaneously meet in an optical sampling gate (such as a nonlinear optical loop mirror (NOLM) or an electroabsorption modulator (EAM). The sampling gate opens an extremely narrow time window at the instant the pulse arrives, allowing the modulated optical signal to pass through, and outputs an optical pulse whose envelope is proportional to the instantaneous amplitude of x(t) at that moment. Therefore, the output is a discrete optical pulse sequence, with each pulse carrying a sampled value.

[0067] All sampling pulses are of the same wavelength, but are arranged at equal intervals on the time axis.

[0068] Distribution operation: The sampled single-wavelength optical pulse sequence enters the optical switch matrix and is split into M active channels. At this time, each channel receives the exact same pulse sequence—all with an interval of T. s A single-wavelength pulse sequence with no time difference.

[0069] The tunable delay timer for each channel is configured with a different delay amount: m = 1, 2, ..., M; Therefore, Channel 1: Delay 0 (received original pulse sequence, sampling time is 0, T) s 2T s ,…) Channel 2: Delay The same pulse sequence was received, but the overall sequence was shifted backward. The sampling time becomes , , ,…) And so on; Channel M: Delay (M-1)T s / M; Originally, there was a channel at time 0, T s 2T s The sampled sequence is now sampled by M channels at their respective offset times, with an equivalent sampling rate from f s Upgrade to M f s .

[0070] Subsequent operations after distribution to each channel: Adjustable optical attenuator: Each channel independently adjusts its optical power to the optimal range, balancing uneven beam splitting and link loss differences; Tunable filter: All channels are uniformly set to wide passband with the same center wavelength, and only out-of-band noise is filtered out; Photodetector: Each channel independently converts light pulses into electrical signals; Backend ADC: Each channel acquires data synchronously, and the digital domain is reassembled into a complete high-sampling-rate sequence using an interleaving reconstruction algorithm.

[0071] In wavelength multiplexing (WDM) mode: Light source configuration: The micro-comb light source is configured in a multi-wavelength synchronous pulse mode, selecting M equally spaced comb teeth from the frequency comb as the working wavelengths λ1, λ2, ..., λ3. M The light pulses corresponding to each wavelength are completely synchronized in the time domain (arriving at the optical sampling gate simultaneously).

[0072] Sampling gate operation: The input signal x(t) is modulated onto the optical carrier by a broadband electro-optic modulator. The modulated optical signal and the multi-wavelength pulse sequence meet simultaneously in the optical sampling gate. The sampling gate opens at the moment the pulse arrives, allowing the modulated optical signal to pass through. However, since the pulse contains M wavelengths, the sampling gate outputs M optical pulses of different wavelengths, which carry the sampled value of x(t) at the same moment, only with different carrier wavelengths.

[0073] All wavelength channels are sampled at the same time, so it is a true parallel synchronous sampling, without needing to be staggered on the time axis.

[0074] Distribution operation: A sequence of mixed optical pulses carrying M wavelengths enters the optical switch matrix and is split into M channels. At this point, each channel receives optical pulses of mixed wavelengths, including λ1 to λ2. M All wavelength components.

[0075] Completely different from TDM. The tunable filter for each channel is precisely set with a different center wavelength and narrow passband: Channel 1: Center wavelength λ1, narrow passband allows only λ1 to pass through, filtering out other wavelengths; Channel 2: Center wavelength λ2, narrow passband allows only λ2 to pass through; And so on... Channel M: Center wavelength λ M Narrow passband only allows λ M pass; The filter bandwidth is designed to meet the requirement of inter-channel crosstalk <-30dB, ensuring that each channel retains only the sampling pulse corresponding to its own wavelength.

[0076] Subsequent operations after distribution to each channel: Adjustable optical delay unit: Fine adjustments compensate for time delay deviations caused by differences in the optical path, ensuring precise alignment of sampling times for all channels; Tunable optical attenuator: finely adjusted according to the inherent loss of each wavelength to achieve cross-wavelength power balance; Photodetector: Each channel converts signals into electrical signals in parallel; Backend ADC: Parallel acquisition is sufficient, without the need for complex digital interleaving and reconstruction.

[0077] In spectrum slicing mode: Optical sampling: Light source configuration: The micro-comb light source is configured as a broadband flat continuous light mode, outputting continuous light with a flat spectrum that covers the entire frequency band of the signal, and is non-pulsed, CW light.

[0078] Sampling operation: The input signal x(t) is modulated onto the optical carrier by a broadband electro-optic modulator to form a continuously modulated optical signal carrying complete spectral information. Unlike TDM and WDM, the spectrum slicing mode does not use a pulse sampling gate (or the sampling gate is continuously turned on). Instead, it samples the modulated continuous optical signal at a uniform time. The sampling rate is determined by the system clock, which can be understood as the sample-and-hold operation of a conventional ADC.

[0079] Optical signals are continuous waveforms rather than discrete pulses, and their spectral information is continuous and complete.

[0080] Distribution operation: The sampled broadband continuous optical signal enters the optical switch matrix and is distributed to M channels.

[0081] Each channel's adjustable filter is set to a specific center frequency and bandwidth, corresponding to a sub-band of the signal spectrum: Channel 1: Center frequency f c,1 Passband bandwidth B1, extract subband 1; Channel 2: Center frequency f c,2 Passband bandwidth B2, extract subband 2; And so on... Channel M: Center frequency f c,M passband bandwidth B M Extract subband M; The filter bank design ensures that each sub-band continuously covers the target frequency band in the frequency domain, with an overlap area (typically 5% to 15% of the sub-band bandwidth) between adjacent sub-bands for subsequent frequency domain stitching and fusion. Each filter has an out-of-band rejection of >25dB to reduce inter-sub-band leakage.

[0082] The sub-band signals extracted from each channel undergo delay consistency processing (compensating for differences in filter group delays) and power adjustment; Photodetector: Each channel converts sub-band optical signals into electrical signals; Backend ADC: Sampling rate requirements for each channel are significantly reduced (only narrow subbands need to be processed, f s,m ≥2B m ); Digital reconstruction: The digital post-processing module performs amplitude normalization, frequency grid alignment, channel delay correction, mapping to a unified frequency coordinate system, and weighted fusion of overlapping areas on the data of each sub-band, splicing them into a complete broadband spectrum; The analog signal to be converted is continuously sampled to generate a performance status frame. The specific method is as follows: Continuous sampling is performed using a sliding window method. Each window contains α consecutive sampling points, and an overlapping area of ​​length β consecutive sampling points is retained between adjacent windows. During window acquisition, a data validity flag is generated according to the sampling clock. The data validity flag is preprocessed, and time domain features, frequency domain features, bitstream statistical features, and channel consistency features are extracted and encapsulated into a performance status frame.

[0083] Each steady-state data window include 10 consecutive sampling points, with adjacent windows updated using an overlapping sliding method; wherein, the current window retains a length of 1 / 2 * 1 / 3 ... The overlapping sampling region, and The window sliding step size is Feature extraction and performance index calculation are performed for each steady-state data window separately, and the state description results of adjacent windows are smoothly fused by utilizing overlapping regions to improve statistical stability and avoid abrupt changes in state estimation caused by window switching.

[0084] In each steady-state data window Internally, operational data is acquired synchronously through bypass replication and status reading. This operational data includes data from each channel. Modulator output bitstream Channel digital sequence after digital decimation filtering The reconstructed output after interleaving reconstruction, parallel merging, or spectrum splicing The system also includes auxiliary monitoring parameters used to characterize its operating status. These auxiliary monitoring parameters include one or more of the following: optical power, center wavelength, modulator bias, adjustable optical delayer status, light source lock status, channel temperature, PGA / TIA saturation flag, quantizer input peak value, integrator amplitude, overflow count, and digital synchronization flag. The above data is obtained through bypass copying, monitoring ports, or status register readings, with the main data link maintaining continuous conversion and continuous output.

[0085] For different reconfigurable sampling modes, corresponding mode labels and synchronization markers are attached to the steady-state window data. In time-interleaved mode, the digital sequence of each channel is time-marked according to the unified frame counter, channel number, and optical delay configuration; in wavelength multiplexing mode, the center wavelength, synchronization sampling marker, and optical power equalization status of each wavelength channel are recorded; in spectrum slicing mode, the center frequency, bandwidth, and phase compensation label of each sub-band are recorded. Thus, the steady-state window data simultaneously includes the input signal sampling results as well as the optical, analog, and digital domains in the current operating configuration. The synchronous working state under the current conditions.

[0086] Subsequently, the steady-state data window was analyzed. The data within the system undergoes preprocessing. This preprocessing includes DC removal, window function weighting, amplitude normalization, outlier removal, channel time alignment, spectral leakage suppression, and adjustment based on the current operating configuration. One or more of the following processes are involved: scale restoration and invalid channel masking. The invalid channel masking process is used to mask channels that are saturated, out of lock, have excessive optical power, or are abnormally synchronized, and to generate valid channel masks.

[0087] After preprocessing, basic features are extracted. For each channel... bitstream The statistical code stream density, consecutive in-line length, idle tone correlation component, and overload code pattern are used to characterize the modulator quantization state and loop stability. For decimation output, reconstruction output, or spectrum splicing output, the short-time spectrum, power spectral density, subband energy distribution, cross-correlation peak value, and phase slope are calculated to characterize noise, spurious signals, distortion, channel mismatch, and subband splicing state.

[0088] Based on steady-state data window Calculate performance monitoring metrics. These metrics include output performance metrics, channel consistency metrics, optoelectronic link status metrics, and stability risk metrics. Specifically, output performance metrics include one or more of the following: SNR, SNDR, ENOB, SFDR, noise floor, in-band spurious emissions, and output data rate; channel consistency metrics include one or more of the following: gain residual, DC bias residual, cross-correlation peak position, phase difference, group delay residual, and splicing error; optoelectronic link status metrics include one or more of the following: optical power, center wavelength, modulator bias drift, temperature change rate, and light source lock-in status; and stability risk metrics include one or more of the following: quantizer input peak value, integrator amplitude, overflow rate, bitstream density deviation, estimated out-of-band gain (OBG) of NTF, and maximum stable input amplitude (MSA) margin.

[0089] The aforementioned performance monitoring indicators can be obtained through digital signal processing. Fast Fourier Transform, Welch average spectrum estimation, or Goertzel narrowband detection can be used to extract fundamental, harmonic, intermodulation, and spurious components; cross-correlation, cross-power spectrum phase slope, or subband phase fitting can be used to calculate residual time delay and phase error; moving average, exponential weighted average, or trend estimation can be used to obtain slow-varying drift; and short-window peak detection, overflow counting, bitstream statistics, and OBG lookup table estimation can be used to obtain stability risk.

[0090] After obtaining the performance monitoring metrics, the current steady-state data window... Perform a steady-state operation quality assessment and generate corresponding window status markers. The steady-state operation quality assessment is used to determine the current operating configuration. Does it have the qualifications for stable configuration?

[0091] When the output performance index is not lower than the preset performance lower limit, the channel consistency index does not exceed the preset mismatch upper limit, the optoelectronic link status index is within the preset operating range, and the stability risk index does not exceed the preset risk upper limit, the current operating configuration is determined. It is qualified for stable configuration and written to the stable configuration cache, recorded as the most recent stable configuration. The aforementioned This serves as a fallback basis when subsequent parameter updates fail, output performance deteriorates, or stability risks increase.

[0092] When any performance monitoring indicator exceeds the corresponding limit, or when a trend of output performance degradation, channel consistency deterioration, optoelectronic link drift, or reduced stability margin appears in multiple consecutive steady-state data windows, an analysis flag is generated, and the analysis flag is used as the current steady-state data window. Status markers.

[0093] Subsequently, the steady-state data window The preprocessed data, basic characteristics, performance monitoring metrics, status flags, and current running configuration obtained within the system. Encapsulated into a performance status frame according to a preset field format. The performance status frame generation process, wherein the performance status frame... It is used to characterize the overall operating state of the system within the current analysis window and serves as the input for subsequent deep learning networks to perform state recognition, error modeling, and control vector generation.

[0094] The performance status frame Including time-domain feature matrix, frequency-domain feature matrix, code stream statistical features, channel consistency features, and optoelectronic link state vector. Stability risk vector, current running configuration The data may include one or more of the following: historical performance change, channel valid mask, timestamp, and data validity flag. The historical performance change is formed by the difference or trend estimate of the performance index between the current steady-state data window and one or more historical steady-state data windows; the data validity flag is used to characterize whether the current window data meets the conditions of effective acquisition, complete synchronization, and statistical reliability.

[0095] Performance status frame After normalization, fixed-point encoding, or floating-point encoding, the data is transmitted to the deep learning network via on-chip bus, AXI interface, DMA cache, shared memory, or control FIFO. Simultaneously, the performance status frame... It is written to the historical state cache for subsequent performance trend comparison, online model updates, and rollback configuration determination in case of anomalies.

[0096] Receive performance status frame And based on the performance status frame Deep learning fusion analysis is used to analyze the current system status, error sources, performance change trends, and parameter adjustment directions.

[0097] During the input processing phase, the deep learning network performs performance status frames based on the current operating mode, the number of activated channels, the decimation ratio, and the effective channel mask. The various features in the data undergo dimensional alignment and invalid channel masking. Since the number of activated channels, subbands, feature lengths, and effective channel positions may differ under different operating modes, the deep learning network first performs performance state frame analysis before performing state analysis. The various features in the input are subjected to input standardization processing. The input standardization processing includes one or more of the following: feature dimension alignment, invalid channel masking, operating mode identification, and feature format unification.

[0098] Specifically, when the number of currently valid channels is less than the preset number of input channels, the deep learning network pads the corresponding channels with zeros and uses a channel validity mask to indicate that the channel does not participate in valid analysis. When the feature length exceeds the preset input length, the deep learning network truncates the feature according to the target analysis frequency band or the preset sampling range. When the system is in different sampling modes, the deep learning network identifies the current operating mode through a pattern embedding vector or a pattern-specific feature processing unit. After the above processing, all types of features are encoded into state tensors with uniform dimensions and uniform field order. This serves as the input for subsequent deep learning networks.

[0099] The deep learning network includes a feature encoding network, a temporal analysis network, and a multi-task output network. The feature encoding network is used to extract data from the state tensor. Features related to channel mismatch, spurious components, noise floor changes, idle tones, spectral splicing traces, and nonlinear distortion are extracted. The time-series analysis network is used to model the state changes in multiple continuous steady-state data windows to identify trend changes caused by temperature drift, optical power attenuation, modulator bias point drift, slow channel response changes, and device aging. The multi-task output network is used to output error estimation results, performance prediction results, state determination results, and stability risk results under the same input state.

[0100] In one implementation, the feature encoding network employs a one-dimensional convolutional network, a residual network, an attention network, or a combination thereof; the temporal analysis network employs LSTM, GRU, a temporal convolutional network, a Transformer network, or a combination thereof; and the multi-task output network employs a structure with a shared feature layer and multiple task output heads.

[0101] Deep learning networks on state tensors After analysis, the output error estimation vector is determined. Performance prediction vector Running status label Stability risk vector and model confidence Among them, the error estimation vector Used to characterize one or more of the following: gain mismatch, bias mismatch, delay mismatch, phase mismatch, optical power imbalance, center wavelength shift, modulator bias drift, optoelectronic link nonlinearity, and digital splicing weight bias; performance prediction vector. Used to characterize one or more predicted values ​​among SNR, SNDR, ENOB, SFDR, noise floor, output bandwidth, output data rate, power consumption, and response latency; operating status label Used to indicate whether the system is in a stable operating state, awaiting calibration, bandwidth-limited, resolution-limited, power-limited, channel mismatch-dominated, nonlinear distortion-dominated, temperature drift-accumulating, or critically unstable state; stability risk vector. It is used to characterize the risks of quantizer input approaching the maximum stable input amplitude (MSA), NTF out-of-band gain (OBG) exceeding the limit, increased overflow rate, integrator amplitude exceeding the limit, or mode switching transient disturbances.

[0102] S4. Generate candidate control vectors based on the error estimation results, review the candidate control vectors for safety constraints, and use the approved candidate control vectors as system control vectors. Perform soft updates of the target parameters in the system control vectors.

[0103] Candidate control vectors are generated based on the error estimation results. Specifically, based on the error estimation results, combined with the current target application mode, hardware constraints, and stability margin, initial parameters are generated, and the initial parameters are converted into candidate control vectors.

[0104] Furthermore, after obtaining the results of S3, the deep learning network generates candidate control vectors. The candidate control vector Represents the target parameter set or relative to the current running configuration. The set of incremental parameters, including optical control subvectors Simulated front-end control subvector and digital processing control subvector .

[0105] Among them, the optical control sub-vector This includes one or more of the following: operating mode, number of active channels, optical switch matrix connection status, optical power allocation, optical delay, center wavelength, tunable filter center frequency, and tunable filter bandwidth; analog front-end control sub-vector. Including one or more of the following: PGA gain, channel bias, electro-optic modulator bias, modulator order, loop coefficient, sampling frequency, and oversampling rate; digital processing control sub-vector. This includes one or more of the following: decimation ratio, digital filter coefficients, gain normalization factor, channel splicing and reconstruction weights, phase compensation amount, nonlinear compensation coefficients, and channel start / stop mask.

[0106] Candidate control vector The generation employs a combination of network output and policy constraints. The deep learning network first determines the error based on the error estimation vector. Performance prediction vector Running status label Stability risk vector and model confidence Generate initial parameter adjustment suggestions; the control strategy layer then adjusts them according to the current target application mode and current running configuration. The initial parameter adjustment suggestions are converted into candidate control vectors based on performance targets, hardware constraints, stability margins, and parameter switching costs. .

[0107] Specifically, when the error estimation results indicate that gain mismatch or bias mismatch is the main source of error, the candidate control vector... Prioritize adjusting PGA gain, channel offset, and digital gain normalization factor; when delay mismatch, phase mismatch, or spectral splicing error is the main source of error, prioritize adjusting optical delay, digital phase compensation, and reconstruction weight; when optical power imbalance or center wavelength shift is the main source of error, prioritize adjusting optical power allocation, center wavelength, and filter center frequency; when stability risk vector When the system exhibits a critical instability risk, the candidate control vector... Priority should be given to safety actions such as reducing front-end gain, increasing oversampling rate, reducing modulator order, reducing loop gain, or freezing mode switching.

[0108] The control strategy layer also adjusts the control direction based on the target application mode. When the target application mode is bandwidth-first, the candidate control vector... Prioritize control actions such as increasing the number of active channels, optimizing the time interleaving delay sequence, reducing decimation delay, or increasing the output data rate; when the target application mode is resolution-priority, prioritize control actions such as increasing the oversampling rate, increasing the decimation ratio, optimizing the filter coefficients, or reducing the noise floor; when the target application mode is power-priority, prioritize control actions such as closing idle channels, reducing optical power, reducing the local clock frequency, or reducing invalid digital processing load.

[0109] To facilitate subsequent hardware implementation and security auditing, candidate control vectors This also includes one or more of the following: parameter validity bits, update priority, maximum single-step change, expected effective window, execution preconditions, alternative configurations, version number, timestamp, and checksum. For continuously adjustable parameters, candidate control vectors... Provide the target value or incremental value and the maximum single-step change; for discrete switching parameters, provide the switching preconditions and alternative configurations; for parameters with insufficient model confidence, provide indicators to maintain the current value, reduce the adjustment range, or re-acquire steady-state window data.

[0110] Candidate control vectors generated by deep learning networks The data is sent to the dynamic adaptive control state machine. The dynamic adaptive control state machine does not directly send the data. Instead of writing to the hardware execution unit, it is sent to the security protection, soft update, consistency check and rollback execution process.

[0111] The candidate control vectors undergo safety constraint review, and the approved candidate control vectors are used as the system control vectors. The specific method is as follows: Based on risk indicators, hard fault flags, and model confidence, candidate control vectors are marked as allowed to execute, restricted to execute, or prohibited to execute: When all risk indicators are below the first safety threshold, the model confidence level meets the standard, and there are no hard faults, it is marked as allowed to execute and the review is approved. When the risk indicator exceeds the second safety threshold, the model confidence level is lower than the standard lower limit, or a hard failure occurs, it is marked as prohibited from execution; Other cases are marked as restricted execution; The approved candidate control vectors are used as system control vectors.

[0112] After approval, the dynamic adaptive control state machine then sends signals to the photon sampling module, the analog front-end, and... The ADC module and the digital post-processing and reconstruction module issue executable control words.

[0113] Through the above processing, the deep learning network will convert the performance status frame... Transformed into candidate control vectors with clearly defined parameter meanings, adjustment directions, execution priorities, and safety indicators. This provides a quantitative basis for subsequent security audits and closed-loop parameter execution.

[0114] Candidate control vectors output by S4 Perform security constraint audits, incremental parameter updates, and anomaly rollback handling. Receive candidate control vectors. and read the current running configuration. Recent stable configuration Performance status frames Stability risk vector Model confidence Together with the current status information in the hardware status register, a closed-loop execution process of "security audit - soft update - short window retest - anomaly rollback" is formed.

[0115] First, the candidate control vectors are... Perform tiered safety audits and, based on continuous risk indicators, hard failure flags, model confidence levels, and control action types, [the following will be implemented]. Marked as allowed execution, restricted execution, or prohibited execution. The continuous risk indicator includes the quantizer input peak value and... The ratio, overflow rate, integrator amplitude ratio, bitstream density deviation and One or more of the following; the hard fault flags include one or more of the following: front-end continuous saturation, optical power severely exceeding limits, temperature severely exceeding limits, sampling clock lockout, critical channel synchronization failure, insufficient number of effective channels, control vector verification error, and parameter unpruning exceeding limits.

[0116] When all continuous risk indicators are below the corresponding first safety threshold, the model confidence level is not lower than the first confidence threshold, and no hard fault flag appears, the safety protection unit determines... Execution is permitted.

[0117] When at least one continuous risk indicator is between the first safety threshold and the second safety threshold, or when the model confidence level is between the lowest confidence threshold and the first confidence threshold, or If the current stability margin is insufficient, the following actions are considered: increasing gain, increasing optical power, reducing OSR, increasing modulator order, increasing loop gain, rapidly expanding channel number, or mode switching. Under restricted execution, the security protection unit blocks the above actions and prioritizes one or more conservative control actions, such as reducing front-end gain, reducing optical power, increasing OSR, increasing decimation ratio, reducing modulator order, reducing loop gain, freezing mode switching, or reducing the number of active channels.

[0118] When any continuous risk indicator exceeds the corresponding second safety threshold, the model confidence level falls below the minimum confidence threshold, or a hard fault indicator appears, a judgment is made. Disable execution, do not issue parameter update commands, and maintain the current running configuration.

[0119] Control vectors approved by the security safeguards are not written to all hardware registers at once, but are gradually applied through a soft update. The soft update includes loading shadow registers, switching double-buffer parameters, applying frame boundaries, waiting for the protection interval, adjusting the ramp step by step, retesting in a short window, and submitting for confirmation.

[0120] Specifically, the target parameters are first written to the shadow register, spare configuration area, or double-buffered parameter table, while keeping the currently active register unchanged. The corresponding triggering conditions are determined based on the type of the parameter to be updated, and the preferred execution order follows the principle of "digital compensation first, then analog operating point, and finally optical reconstruction and mode switching." For digital filter coefficients, normalization factors, phase compensation amounts, and reconstruction weights, they take effect at the frame boundary between the end of the current data frame or steady-state data window and the beginning of the next frame or window. For PGA gain, channel bias, modulator bias, optical power, optical delay, center wavelength, and filter parameters, the time since the last update is not less than the corresponding guard interval. It takes effect after the previous short window retest is completed; for channel start / stop, optical switch matrix switching and operating mode switching, it takes effect when the backup path is established, the channel is synchronized and effective and the control state machine reaches the preset version switching point.

[0121] Based on the maximum single-step change in the candidate control vectors, the target parameters are decomposed into multiple small-step updates. After each update step is completed, short-window data is collected, and the overflow rate, quantizer input peak value, integrator amplitude, channel consistency index, and output performance index are recalculated. When the retest results meet the safety conditions, the next update is executed until all data is updated. When the retest results do not meet the safety conditions, subsequent updates are stopped and a rollback process is initiated.

[0122] Once all step-by-step updates are complete and the final short-window retest results meet the requirements, the control vector execution is confirmed to be successful, and the configuration after execution is updated to the new, most recently stable configuration. The corresponding performance status frame, control vector, and execution result are written to the historical status cache.

[0123] Maintain at least three levels of rollback configurations, namely the most recently stable configuration. Quick coarse calibration configuration and Startup Security Configuration .in, Configured for stable operation in the most recent update, and used as a priority rollback in case of a failed soft update; The configuration saved after a successful rapid coarse calibration is used for... If system stability cannot be restored, re-establish the channel amplitude, offset, and delay references; This is a low-risk configuration loaded during system power-up to maintain a minimum-risk operating state in the event of severe anomalies.

[0124] If the following issues occur during a soft update: configuration readback failure, performance retest failure, continuously rising overflow rate, quantizer input peak value approaching the maximum stable input amplitude (MSA), OBG exceeding limits, sampling clock lockout, critical channel failure, channel synchronization anomaly, or abnormal model confidence, immediately stop subsequent updates and execute a rollback. During rollback, prioritize using a reverse soft update method to restore to the previous state. If restored to If the system still does not meet the stability condition, then proceed to... Then, re-perform channel calibration; if calibration fails, the number of valid channels is lower than the minimum channel number threshold, or the system still has serious stability risks, then roll back to... Disable aggressive parameter optimization and mode switching, and only maintain the lowest risk data output; if the system still does not meet the stability requirements under this configuration, it will enter a safety alarm state and wait for maintenance.

[0125] Through the aforementioned security safeguards, soft updates, consistency checks, and three-level rollback mechanisms, the candidate control vectors output by deep learning networks can be constrained within physical security boundaries and hardware constraints, avoiding errors caused by model errors, sudden input changes, excessively rapid parameter switching, or hardware response lag. This system addresses loop instability, photoelectric front-end saturation, channel reconfiguration anomalies, and mode switching failures, thereby improving the stability and anomaly recovery capability of the closed-loop control.

[0126] A reconfigurable photon sampling analog-to-digital converter control system, used to implement a reconfigurable photon sampling analog-to-digital converter control method, includes: The initialization module obtains the startup security configuration and performs an initialization self-test. If the initialization self-test results meet the preset security startup conditions, the initialization self-test is deemed to have passed. After the initial self-test passes, the coarse registration module determines the target analysis frequency band based on the target working range, determines the initial sampling configuration based on the startup security configuration and the target analysis frequency band, and collects calibration samples corresponding to each active channel from the target analysis frequency band under the initial sampling configuration. Based on the calibration samples, it constructs the channel error vector corresponding to each channel. When the corresponding error component meets the preset threshold requirements, the coarse calibration configuration is obtained. The error estimation module, under the coarse calibration configuration, continuously samples the analog signal to be converted, generates a performance status frame, and analyzes the performance status frame using deep learning to output the error estimation result; The update module generates candidate control vectors based on the error estimation results, performs safety constraint review on the candidate control vectors, and uses the approved candidate control vectors as the system control vectors, and performs soft update of the target parameters in the system control vectors.

[0127] The system is applied to a reconfigurable multi-channel photonic sampling analog-to-digital converter. The front end of the reconfigurable multi-channel photonic sampling analog-to-digital converter is configured to support time-interleaving mode, wavelength multiplexing mode, and spectrum slicing mode through its internal optical switch matrix, tunable optical delay unit, and tunable optical filter. It can also dynamically switch between the three modes during operation according to the mode switching instructions in the candidate control vector.

[0128] The photonic sampling module receives the analog signal x(t) to be converted at its input and outputs multiple electrical signals after photonic sampling. This module employs a configurable optical network to implement three channelization modes: time-interleaving mode, wavelength multiplexing mode, and spectrum slicing mode. The configurable optical network includes an optical switch matrix, an adjustable optical delay unit, an adjustable optical attenuator, an adjustable optical filter, and a control and monitoring unit, used to dynamically adjust the number of channels M, the inter-channel delay difference Δτ, and the channel center wavelength λ according to control commands. m and optical power distribution P m .

[0129] Time Interleaving Mode (TDM): In TDM mode, the system aims to achieve ultra-high time resolution. The input analog signal x(t) is loaded onto an electro-optic modulator and modulated onto a continuous optical carrier to form a modulated optical signal carrying signal information. Simultaneously, the microcomb light source is configured in single-wavelength ultrashort pulse mode, outputting a series of regular optical pulse sequences with pulse widths less than 1 picosecond and time jitter less than 20 femtoseconds. These pulses serve as the system's sampling clock.

[0130] The modulated optical signal and the sampling pulse meet in the optical sampling gate, completing the time-domain sampling process. The sampled optical pulse sequence then enters the optical switch matrix and is split into M active channels. This is a key step in the TDM mode: the adjustable optical delayer for each channel is set to a different delay value, according to... The arithmetic sequence configuration, where It is the sampling period. This configuration allows signals at the same sampling moment to be assigned to different channels and staggered on the time axis.

[0131] The optical signals from each channel are then power-equalized by adjustable optical attenuators to ensure that the optical power of all channels is within the optimal range. In this mode, the adjustable filters are set to the same center wavelength and wide passband, primarily to filter out out-of-band noise without affecting signal transmission. Finally, the photodetectors in each channel convert the optical signal into an electrical signal. The back-end ADC array acquires these electrical signals using a synchronous clock and reassembles them into a complete high-sampling-rate sequence through a data interleaving algorithm. Throughout the process, the control and monitoring unit monitors the status of each channel in real time, and the controller dynamically adjusts the attenuator settings based on the feedback data to maintain the system's optimal operating condition.

[0132] During the reconstruction process, the main parameters adjusted are M (number of channels) and Δτ (inter-channel delay difference). The effective sampling rate can be directly expanded through parameter adjustments. This method is suitable for capturing ultra-wideband time-domain signals, such as transient phenomena, ultrafast waveform measurements, and radar signal processing.

[0133] Wavelength multiplexing mode (WDM): The core idea of ​​WDM mode is to achieve true parallel sampling using the wavelength dimension. During system initialization, the microcomb light source is configured in multi-wavelength pulse mode, selecting M equally spaced comb teeth from the frequency comb as the working wavelengths, and the optical pulses corresponding to each wavelength are completely synchronized in the time domain. The input signal x(t) is also first modulated onto the optical carrier by an electro-optic modulator, but the modulator here needs to have sufficient bandwidth to cover all working wavelengths.

[0134] After the modulated signal enters the optical sampling gate, it interacts with a multi-wavelength pulse sequence to generate M sampling optical pulses of different wavelengths but carrying the same signal information. These mixed-wavelength optical signals are distributed to various channels through an optical switch matrix. The tunable optical filter of each channel is precisely set to the corresponding center wavelength and narrow passband to achieve wavelength separation, ensuring that crosstalk between channels is suppressed to below -30dB. Different wavelength channels are physically independent and can simultaneously sample different copies of the same signal.

[0135] The single-wavelength signals separated from each channel undergo minor adjustments by an adjustable optical delay unit to compensate for time delay deviations caused by differences in the optical path, ensuring precise alignment of sampling times across all channels. An adjustable optical attenuator is finely adjusted based on the inherent characteristics of each wavelength and monitoring feedback to achieve cross-wavelength power equalization. Finally, a photodetector array converts the optical signals from each wavelength channel into electrical signals in parallel. Because all channel sampling is completely synchronized, backend processing is relatively simple, requiring only parallel acquisition. The system continuously monitors the optical power of each channel through closed-loop control, dynamically adjusting attenuator settings to address the impact of environmental changes and device aging.

[0136] During the reconstruction process, the main adjustments are made to M (number of channels), λm (center wavelength of each channel), and Pm (power allocation). In this mode, each channel operates independently, and the pulse width requirement for the light source is relatively low. It is suitable for scenarios requiring high dynamic range or simultaneous processing of multi-band signals, such as software-defined radio and spectrum sensing.

[0137] Spectrum slicing mode: In spectrum slicing mode, a frequency domain parallel processing strategy is employed, which is particularly suitable for broadband sparse spectrum signals. At the start of operation, the system first performs a rapid spectrum analysis on the input signal x(t) to determine the spectrum slicing scheme: the total bandwidth of the signal is divided into M sub-bands, each corresponding to a different center frequency and bandwidth. The microcomb light source is configured in broadband continuous light mode, outputting a spectrally flat light source covering the entire frequency band of the signal.

[0138] The input signal is modulated onto an optical carrier by a broadband electro-optic modulator, forming a modulated optical signal containing complete spectral information. Unlike TDM and WDM modes, the sampling process here is closer to uniform sampling, with the sampling rate determined by the system clock. The sampled broadband optical signal enters the optical switch matrix and is assigned to M channels. Each channel's tunable filter is set to a specific center wavelength and bandwidth, corresponding to a sub-band of the signal spectrum. The filter bank design ensures that each sub-band continuously covers the target frequency band in the frequency domain, and that predetermined overlap or transition regions can be set between adjacent sub-bands; the out-of-band rejection of each filter is preferably greater than 25dB to reduce inter-sub-band leakage and interference with non-target frequency bands.

[0139] The sub-band signals extracted from each channel undergo delay consistency processing and power adjustment before being converted into electrical signals by a photodetector. Since each channel only needs to process one relatively narrow sub-band signal, the single-channel sampling rate and analog bandwidth requirements of the back-end ADC are significantly reduced. Each sub-band is sampled and quantized by the narrow-band ADC to form the corresponding sub-band digital data.

[0140] In the digital reconstruction stage, the digital post-processing module first performs amplitude normalization, frequency grid alignment, and channel delay correction on the digital data of each sub-band to compensate for gain differences, filter response differences, and optoelectronic link delay differences between different channels. Subsequently, the system maps each sub-band to a unified full-bandwidth frequency coordinate system based on its center frequency, bandwidth, and phase compensation parameters. For overlapping or transitional regions of adjacent sub-bands, the digital reconstruction module uses a weighted fusion method for smooth splicing to avoid amplitude steps, phase abrupt changes, or splicing spurious signals at sub-band boundaries. After frequency domain alignment, phase compensation, and overlapping region fusion, the digital signals of each sub-band are reconstructed into a complete broadband spectrum.

[0141] The system can also dynamically adjust the slicing scheme based on real-time spectrum monitoring results. When a frequency band has dense signals or a complex spectral structure, the control module can reduce the subband width of that band or increase the number of corresponding channels to improve local frequency resolution. When a frequency band has no effective signal for a long period, the corresponding optical channel, ADC channel, or digital processing unit can be shut down to reduce power consumption. Thus, the spectrum slicing mode can achieve a flexible trade-off between broadband coverage, frequency resolution, reconstruction accuracy, and power consumption. Dynamic switching between the three modes on the same hardware platform is achieved through a reconfigurable optical network. When the system receives a mode switching command, it first pauses the current data acquisition and saves the working state. The light source subsystem is reconfigured according to the target mode requirements: TDM mode requires single-wavelength ultrashort pulses, WDM mode requires multi-wavelength synchronous pulses, and spectral slicing mode requires a broadband flat spectrum. The optical switch matrix synchronously adjusts the connection topology to establish an optical path structure suitable for the target mode.

[0142] The programmable devices in each channel are then configured according to the new mode parameters: the tunable filter sets the center wavelength and bandwidth according to mode requirements; the tunable delay line is set with arithmetic delay, calibration delay, or fixed delay according to mode characteristics; and the tunable attenuator is initialized based on pre-stored calibration data. After the system completes rapid calibration, it starts the new mode and performs fine-tuning through monitoring feedback during operation. The entire switching process can be completed within hundreds of milliseconds, and the state saving mechanism allows for rapid recovery of commonly used modes.

[0143] In practical applications, the three modes can work together based on signal characteristics. For example, the system can first use the spectrum slicing mode to quickly analyze the signal's spectral characteristics, identify important frequency bands, and then switch to WDM mode for detailed monitoring of these bands. If a transient event is detected, it can temporarily switch to TDM mode for high-resolution acquisition. This flexible multi-mode collaborative capability enables a single system to meet diverse signal acquisition needs, playing a vital role in fields such as communications, radar, and scientific measurement.

[0144] A reconfigurable photonic sampling analog-to-digital converter control device includes a processor and a memory, wherein the processor executes a computer program stored in the memory to implement a reconfigurable photonic sampling analog-to-digital converter control method.

[0145] Example 1 In this embodiment, the present invention is applied to scenarios involving the acquisition of ultra-wideband radar echoes, transient high-frequency signals, or wideband measurement signals. In these scenarios, the input signal bandwidth is high, and the main objective of the system is to improve the equivalent sampling rate and suppress time-interleaving spurious signals caused by channel gain errors, bias errors, and delay errors. Therefore, the control module preferably uses the time-interleaving mode as the target operating mode.

[0146] When the target input bandwidth Greater than the first bandwidth threshold At this time, the control module invokes the time interleaving mode. Preferably, the first bandwidth threshold... 5GHz is acceptable. For example, when the target input bandwidth... Single-channel stable sampling rate At that time, in order to meet the Nyquist sampling requirements, the system's equivalent sampling rate is... The speed should be no less than 16 GSa / s. The control module determines the required number of active channels according to the following formula:

[0147] for , In this scenario, the required number of channels is calculated to be no less than 7. Therefore, in this embodiment, it is preferable to activate 8 channels to obtain an equivalent sampling rate of approximately 20 GSa / s. The significance of using 8 channels is twofold: firstly, it meets the sampling rate requirements of wideband input; secondly, it reserves redundancy for subsequent channel degradation, local calibration, or abnormal rollback.

[0148] Under the above conditions, the required number of channels is calculated to be no less than 7; therefore, in this embodiment, it is preferable to activate 8 channels. If the single-channel sampling period is... The equivalent sampling interval for 8 channels is:

[0149] No. The initial optical delay of each channel can be set as follows:

[0150] Thus, the eight channels form an equally spaced interleaved sampling sequence on the time axis.

[0151] During the rapid coarse calibration phase, the control module directs each channel into... The signal amplitude before the modulator is adjusted to 60% to 75% of full scale, ensuring both high dynamic range utilization and sufficient transient signal amplitude margin. Subsequently, channel bias and optical delay are corrected to ideally minimize the residual delay. The optical power difference between channels was controlled within ±0.5dB. The calibrated channel number, optical delay sequence, optical power allocation, PGA gain, bias, OSR, decimation ratio, and interleaving / reconstruction weights were recorded as the current operating configuration. It is used for subsequent scale restoration, performance comparison, control amplitude constraints, and anomaly rollback.

[0152] After entering steady-state operation, when the performance status frame When the channel cross-correlation peak deviates from the ideal position, the group delay residual changes continuously in the same direction, and the interleaving spurious signals in the short-time spectrum of the reconstructed output increase, the deep learning module identifies this condition as performance degradation dominated by delay mismatch. At this point, the candidate control vector... This mainly includes fine-tuning the optical delay of each channel, updating the digital phase compensation amount, and interleaving reconstruction weights.

[0153] When performance status frame When the high-frequency subband energy continues to rise and the output bandwidth prediction result approaches the upper limit of the current equivalent sampling rate, the deep learning module determines this condition as bandwidth-limited. At this point, the candidate control vector... This could include increasing the number of active channels, reducing the extraction ratio, updating the interleaving and reconstruction weights, and moderately reducing the OSR when stability allows.

[0154] Before execution, the safety protection unit checks the quantizer input peak value, overflow rate, and OBG. When the quantizer input peak value exceeds 80% of the maximum stable input amplitude (MSA), the safety protection unit blocks actions that reduce OSR, increase PGA gain, or rapidly expand the number of channels, prioritizing actions that reduce front-end gain or maintain or increase OSR. After approval, the control module first updates the channel mask and interleaving reconstruction weights in the digital domain, then adjusts the optical delay step by step or enables newly added channels. After each update, the control module retests interleaving spurious emissions, channel synchronization status, and stability risks through a short window. If the retest passes, the updated configuration is recorded as the most recently stable configuration. If the retest fails, the configuration will revert to the previous configuration.

[0155] Through this embodiment, the system can call the time interleaving mode according to the target bandwidth and single-channel sampling capability under ultra-wideband input conditions, and distinguish between two types of problems, namely bandwidth limitation and time delay mismatch, based on the characteristics in the steady-state window. Thus, it can take control actions to expand the channel or correct the delay respectively, which is beneficial to improve transient capture capability and reduce interleaving spurious emissions.

[0156] Example 2 In this embodiment, the system is applied to narrowband communication reception, high-precision measurement, or high dynamic range spectrum monitoring scenarios. The main requirements in this operating condition are to improve ENOB and SFDR while reducing the noise floor, and the system preferably uses wavelength multiplexing mode.

[0157] When the target input bandwidth Less than the second bandwidth threshold At this time, the control module selects the wavelength multiplexing mode. Preferably, the second bandwidth threshold... A frequency of 200MHz is acceptable. For example, when the target input bandwidth is 100MHz, the control module preferably enables four synchronous wavelength channels. Each wavelength channel is separated by an adjustable filter and then enters the corresponding photodetector, TIA, PGA, and... ADC channel.

[0158] In this embodiment, the spacing between adjacent wavelengths can be set according to the filter passband width and crosstalk requirements, for example, 0.8 nm or a corresponding frequency spacing. During the rapid coarse calibration stage, the control module adjusts the adjustable optical attenuators of each channel to preferably control the optical power deviation of each wavelength channel within ±0.3 dB; simultaneously, it adjusts the PGA gain and channel bias to preferably reduce the difference in the root mean square amplitude of the output of each channel to less than 1% to 2%. Since resolution is prioritized in this operation, the OSR can be set to 128 or 256, coupled with a relatively large decimation ratio. Suppress in-band quantization noise.

[0159] After entering steady-state operation, if the performance status frame If a decrease in optical power occurs in a certain wavelength channel, accompanied by a decrease in the output amplitude of that channel and an increase in the channel gain residual, the deep learning module classifies this condition as either optical power imbalance or channel response attenuation. At this point, the candidate control vector... This includes improving the transmittance of the corresponding channel optical attenuator, correcting the PGA gain, or updating the digital gain normalization factor.

[0160] If performance status frame If the center wavelength reading deviates from the target wavelength, accompanied by changes in channel amplitude and phase response or increased in-band spurious emissions, the deep learning module classifies this condition as a center wavelength shift. At this point, the candidate control vector... This includes fine-tuning the center frequency of the tunable filter, updating wavelength channel labels, and digital merging weights.

[0161] If performance status frame The modulator bias monitoring value deviates from the current operating configuration. Furthermore, the reconstructed output spectrum shows an increase in second- or third-order spurious signals, which the deep learning module identifies as modulator bias drift or enhanced optoelectronic nonlinearity. At this point, the candidate control vector... This includes small-step modulator bias point callback, reducing front-end drive amplitude, or updating digital nonlinear compensation coefficients.

[0162] Before execution, the safety protection unit checks the upper and lower limits of optical power, PGA / TIA saturation flags, OSR and decimation ratio matching, and modulator bias adjustment. For optical power, PGA gain, and modulator bias, the control module uses small-step soft updates; for digital compensation coefficients, merging weights, and normalization factors, a double-buffered approach is used for switching at frame boundaries. After each update, the control module retests ENOB, SFDR, noise floor, in-band spurious emissions, and channel equalization status. If the retest results show an improvement in SFDR and no increase in stability risk, the update is confirmed as effective; if the retest shows no improvement or spurious emissions further increase, the configuration reverts to the pre-update configuration.

[0163] Through this embodiment, the system can invoke wavelength multiplexing mode under high-precision operating conditions, and distinguish the source of error based on the coupling relationship between optical power, center wavelength, modulator bias and spectral spurious signals, thereby taking optical power equalization, wavelength correction or bias compensation actions respectively, which helps to maintain high ENOB and high SFDR output.

[0164] Example 3 In this embodiment, the system is applied to scenarios involving broadband spectrum sensing, software-defined radio, or multi-band signal monitoring. Under these conditions, the total input bandwidth is relatively wide, but the effective signal is typically concentrated within a few sub-bands; therefore, the system preferably uses spectrum slicing mode.

[0165] When the target input bandwidth Located at the second bandwidth threshold With the first bandwidth threshold Between, for example Alternatively, when the input signal exhibits a clear multi-subband distribution, the control module selects the spectrum slicing mode. Taking an input monitoring bandwidth of 2GHz as an example, the control module can divide this bandwidth into four subbands of approximately 500MHz and set transition regions between adjacent subbands to facilitate frequency domain splicing and smooth fusion.

[0166] After rapid coarse calibration, the control module records the center frequency of each sub-band, sub-band bandwidth, adjustable filter parameters, optical power of each channel, PGA gain, OSR, decimation ratio, and frequency domain stitching weight as the current operating configuration. Under this operating condition, Used to mark the current spectrum slice boundaries and splicing relationships, enabling the deep learning module to determine, based on the current configuration, whether subband anomalies are caused by boundary settings, phase compensation, gain normalization, or local nonlinearity.

[0167] During steady-state operation, if the performance status frame The display shows a stable step in the power spectral density at the boundary of adjacent sub-bands, and this step persists across multiple steady-state windows. The deep learning module identifies this condition as a spectral splicing error. At this point, the candidate control vector... This includes adjusting the splicing weights of overlapping areas in adjacent sub-bands, correcting the sub-band phase compensation amount, and fine-tuning the center frequency and bandwidth of the adjustable filter when necessary.

[0168] If performance status frame The display shows inconsistent phase slopes between adjacent sub-bands, leading the deep learning module to determine this condition as insufficient group delay compensation. At this point, the candidate control vector... This mainly includes updating the digital phase compensation amount and the frequency domain stitching weights. (If the performance status frame...) The display shows that the energy of a certain sub-band is significantly higher than that of other sub-bands, and the bitstream density of this sub-band deviates from the normal range or the intra-band intermodulation component is increased. The deep learning module judges this condition as local sub-band overload or local nonlinear enhancement. At this time, the candidate control vector... This includes reducing the PGA gain of the sub-band channel, reducing the corresponding optical power, or increasing the OSR of the channel.

[0169] When the security protection unit reviews the candidate control vector, it focuses on checking three constraints: whether each sub-band still continuously covers the target frequency band after adjustment; whether the dimensions of the channel mask, splicing weight, and phase compensation are consistent; and whether the decimation ratio, digital filter bandwidth, and output data rate match. After the review is passed, the control module first updates the digital splicing weight and phase compensation, and then adjusts the adjustable filter parameters in groups. After each group update, the control module collects short-window data to retest the sub-band boundary continuity, noise floor, SFDR, and target frequency band coverage. If the retest shows that the sub-band boundary step has decreased and the target frequency band is still continuously covered, the update is confirmed to be effective; if the retest shows that the frequency band coverage gap or splicing spurious has increased, it reverts to the configuration before the update.

[0170] Through this embodiment, the system can invoke the spectrum slicing mode under broadband monitoring conditions, and dynamically adjust the slice boundary, phase compensation and splicing weight according to the amplitude, phase and boundary continuity relationship between sub-bands, which helps to reduce sub-band splicing errors and local overload risks in broadband spectrum reconstruction.

[0171] Example 4 In this embodiment, the system is applied to portable measurement devices, edge spectrum monitoring systems, or long-term monitored broadband sensing devices. Under these conditions, the system needs to reduce power consumption when the signal is sparse, and also needs to compensate for slow-varying errors caused by temperature drift and device aging during long-term operation.

[0172] During low-power operation, the system preferentially operates in spectrum slicing mode. The control module statistically analyzes the energy distribution, output bandwidth, ENOB, SFDR, channel effective mask, and estimated power consumption of each channel within a continuous steady-state window. If a certain sub-band is in continuous steady-state mode... If the energy in each steady-state window is below a preset energy threshold, and the target frequency band coverage still meets the requirements after closing the sub-band, then the control module marks the sub-band as an idle candidate sub-band. Preferably, The value can be 3 to 10, and the energy threshold can be 3 to 6 dB above the system noise level.

[0173] The deep learning module is based on performance status frames. Determine whether the subband can be shut down. If the performance prediction results indicate that ENOB, SFDR, and target frequency band coverage still meet preset requirements after shutting down the subband, the deep learning module generates a low-power candidate control vector. This includes shutting down the corresponding optical channel and ADC channel, reducing the optical power of the channel, reducing the local clock frequency, shutting down the corresponding digital filter calculation unit, updating the channel start / stop mask, and reconstructing the weights.

[0174] Before execution, the security protection unit confirms that the remaining subband still covers the target frequency band after the channel is shut down, and confirms that the channel mask, phase compensation amount, and splicing weight dimensions are consistent. After approval, the control module first updates the channel mask and reconstructs the weights in the digital domain, then gradually reduces the corresponding optical power and local clock frequency, and finally shuts down the corresponding ADC channel. After shutdown, the control module collects short-window data to retest the target frequency band coverage, subband boundary continuity, ENOB, and SFDR. If the retest passes, the low-power configuration is recorded as the new most recently stable configuration. If a subsequent window displays that the previously closed sub-band has regained valid energy, the control module will restore that sub-band channel in reverse order.

[0175] During long-term operation, if the rate of temperature change increases continuously across multiple steady-state windows while ENOB or SFDR decreases slowly, the control module records this state as a drift state to be analyzed. Preferably, the continuous... The trigger condition is that the cumulative decrease in ENOB bits in a window exceeds 0.3 bits. A value of 5 can be used; alternatively, the center wavelength offset exceeding 5% of the filter passband width can be used as the trigger condition.

[0176] The deep learning module does not rely solely on temperature values ​​for judgment. Instead, it identifies drift sources by combining optical power, center wavelength, modulator bias, channel amplitude, spectral spurious signals, and historical performance variations. When the primary error source is slowly varying optical power, candidate control vectors include rebalancing the optical power and correcting the digital gain normalization factor. When the primary error source is modulator bias drift, candidate control vectors include adjusting the modulator bias point with small steps and updating the nonlinear compensation coefficients. When the primary error source is accumulated channel delay error, candidate control vectors include fine-tuning the optical delay and adjusting the digital phase compensation.

[0177] For the aforementioned slow-varying compensation actions, the control module employs small-step soft updates. The adjustment amount for optical power and bias parameters is preferably no more than 1% to 5% of the corresponding adjustable range each time; digital weights and phase compensation amounts are switched at frame boundaries using a double-buffered approach. After each adjustment, the control module retests ENOB, SFDR, channel consistency, and stability risks. If the retest results show performance improvement, compensation continues in that direction; if the retest shows no improvement or stability risks increase, updates in that direction are stopped and the module reverts to the previous state. .

[0178] During long-term operation, when the control module detects a continuously increasing deviation between the performance prediction results and actual retest results of the deep learning module, it can trigger an online model update. The control module extracts performance state frames, candidate control vectors, security audit results, and post-execution retest results from the historical state cache to form incremental samples. After incremental training, the candidate model runs in parallel as a shadow model, only outputting prediction results and not directly controlling the hardware. The current model is only allowed to replace the current model when the shadow model outperforms the current model in multiple consecutive verification windows and does not generate high-risk candidate actions; if short-window retests show performance degradation or abnormal risk assessment after replacement, the previous model version is restored.

[0179] Through this embodiment, the system can invoke low-power control actions when the signal is sparse to reduce the power consumption of invalid channels; and invoke drift compensation and controlled model update mechanisms during long-term operation to reduce performance degradation caused by temperature drift and device aging, thereby improving the long-term operating efficiency and stability of the system.

[0180] Example 5 In this embodiment, the system is applied to sudden large input fluctuations, abnormal optical power, abnormal channel synchronization, or... The loop is at critical instability. Under this condition, continuing aggressive control actions aimed at increasing bandwidth or resolution may lead to front-end saturation, quantizer overload, or... The loop diverges, so the system prioritizes calling the security guard and the three-level rollback mechanism.

[0181] During normal operation, when the amplitude of the input signal suddenly increases within a short period of time, the steady-state window... Quantizer input peak Integrator amplitude and overflow rate It will rise rapidly, at the same time The bitstream may exhibit consecutive increases in length, deviations in bitstream density, and overloaded bit patterns. These characteristics collectively indicate that the modulator loop may be approaching saturation or instability.

[0182] In this embodiment, the safety trigger conditions can be set as follows: the quantizer input peak value exceeds 80% of the maximum stable input amplitude (MSA), or the overflow rate exceeds a preset threshold for two consecutive short windows, or the OBG estimate exceeds a preset upper limit. When any of these conditions are met, the performance status frame... It carries a stability risk flag when entering the deep learning module.

[0183] After the deep learning module determines that the system is in a critically unstable state, candidate control vectors are generated. Prioritize actions such as reducing PGA gain, reducing optical power, increasing OSR, increasing decimation ratio, reducing modulator order, reducing loop gain, and freezing mode switching. If candidate control vectors simultaneously contain actions that conflict with stability recovery, such as increasing gain, reducing OSR, expanding the number of channels, or fast mode switching, the safety protection unit will block these actions and retain only conservative recovery actions.

[0184] The control module then performs a recovery operation via a soft update. Preferably, it first reduces the PGA gain or optical power to lower the quantizer input peak value; secondly, it increases the OSR or decimation ratio to increase the modulator stability margin; and finally, it determines whether to restore the original output data rate based on the retest results. After each update step, the control module acquires short-window data retest quantizer input peak value, integrator amplitude, overflow rate, ENOB, and SFDR.

[0185] If the retest results show that the quantizer input peak value drops below 70% of the MSA and the overflow rate recovers to the preset range, the control module confirms successful recovery and records the recovered configuration as the new most recently stable configuration. If the retest results show that the risk has not been eliminated, the control module will stop subsequent updates and execute a three-level rollback.

[0186] During a Level 3 rollback, the control module first uses a reverse soft update method to restore the most recent stable configuration. If the system still cannot meet the stability requirements, it will enter the fast self-calibration configuration. The system re-performs local channel calibration to restore the channel amplitude, bias, delay, and optical power references. If local calibration fails, the number of effective channels falls below the minimum channel count threshold, or serious anomalies such as sampling clock lockout or critical channel failure persist, the control module reverts to the default safe configuration. Disable aggressive parameter optimization and mode switching, and only maintain the lowest risk data output. If the system is still abnormal under these conditions, it will directly enter the alarm state.

[0187] Through this embodiment, the system can prioritize invoking the safety protection mechanism under sudden overload or critical instability conditions, blocking control actions that may exacerbate instability, and in accordance with... , , The order is reversed step by step, thereby reducing Risks include loop divergence, front-end saturation, and channel reconstruction failure.

Claims

1. A control method for a reconfigurable photon sampling analog-to-digital converter, characterized in that, Includes the following steps: S1. Obtain the startup security configuration and perform an initialization self-test. If the initialization self-test results meet the preset security startup conditions, the initialization self-test is deemed to have passed, and S2 is executed. S2. After the initial self-test passes, determine the target analysis frequency band based on the target working range; Based on the target application scenario, select the working mode from time interleaving mode, wavelength multiplexing mode and spectrum slicing mode; The initial sampling configuration is determined based on the startup security configuration, working mode and target analysis frequency band. Under the initial sampling configuration, calibration samples corresponding to each active channel are collected from the target analysis frequency band. Based on the calibration samples, the channel error vector corresponding to each channel is constructed. When the corresponding error component meets the preset threshold requirement, the coarse calibration configuration is obtained. Based on the calibration samples, the channel error vector corresponding to each channel is constructed as follows: ， in, Indicates the first The relative amplitude error of each channel, Indicates the first DC bias error of each channel Indicates the first The relative delay error of each channel, Indicates the first Optical power error of each channel; S3. Under coarse calibration configuration, the analog signal to be converted is input into the reconfigurable multi-channel photonic sampling module. The optical domain is sampled according to the selected working mode and distributed to M channels. After photoelectric conversion, M electrical signals are obtained. The M electrical signals are then subjected to transimpedance amplification, programmable gain and bias adjustment, ΣΔ oversampling noise shaping and quantization, and digital filtering and decimation downsampling to obtain M digital signals. Digital signals and intermediate monitoring data from each channel are collected and stitched together to form a performance status frame. The performance status frame is then input into a pre-trained deep learning error estimation model, which analyzes and outputs the error estimation results for each channel. S4. Generate candidate control vectors based on error estimation results, review the candidate control vectors for safety constraints, and use the approved candidate control vectors as system control vectors. Perform soft updates on the target parameters in the system control vectors. The candidate control vector is generated based on the error estimation results. Specifically, based on the error estimation results, combined with the current target application mode, hardware constraints and stability margin, initial parameters are generated, and the initial parameters are converted into candidate control vectors. Candidate control vectors undergo safety constraint review. Those that pass the review are used as system control vectors. Specifically, based on risk indicators, hard fault flags, and model confidence levels, candidate control vectors are marked as allowed to execute, restricted to execute, or prohibited from execution: when all risk indicators are below the first safety threshold, the model confidence level meets the standard, and there are no hard faults, they are marked as allowed to execute and the review is passed; when risk indicators exceed the second safety threshold, the model confidence level is below the standard lower limit, or a hard fault occurs, they are marked as prohibited from execution; otherwise, they are marked as restricted to execute. Only candidate control vectors that pass the review are used as system control vectors. Based on the maximum single-step change in the candidate control vector, the target parameter is decomposed into multiple small-step updates. After each update step is completed, short-window data is collected, and the overflow rate, quantizer input peak value, integrator amplitude, channel consistency index, and output performance index are recalculated. When the retest result meets the safety conditions, the next update is executed until all data is updated. When the retest result does not meet the safety conditions, subsequent updates are stopped and the rollback process is initiated. Maintain at least three levels of rollback configurations, namely the most recently stable configuration. Quick coarse calibration configuration and Startup Security Configuration ,in, Configured for stable operation in the most recent update, and used as a priority rollback in case of a failed soft update; The configuration saved after a successful rapid coarse calibration is used for... If system stability cannot be restored, re-establish the channel amplitude, offset, and delay references; This is a low-risk configuration loaded during system power-up to maintain a minimum-risk operating state in the event of severe anomalies. If the following occurs during a soft update: configuration readback failure, performance retest failure, continuously rising overflow rate, quantizer input peak value approaching the maximum stable input amplitude (MSA), OBG exceeding limits, sampling clock lockout, critical channel failure, channel synchronization anomaly, or abnormal model confidence, immediately stop subsequent updates and execute a rollback. During the rollback, a reverse soft update method should be used to restore to the previous state. If restored to If the system still does not meet the stability condition, then proceed to... Then, re-perform channel calibration; if calibration fails, the number of valid channels is lower than the minimum channel number threshold, or the system still has serious stability risks, then roll back to... Disable aggressive parameter optimization and mode switching, and only maintain the lowest risk data output; if the system still does not meet the stability requirements under this configuration, it will enter a safety alarm state and wait for maintenance.

2. The control method for a reconfigurable photon sampling analog-to-digital converter according to claim 1, characterized in that, In S2, the target analysis frequency band is determined based on the target working interval, specifically as follows: When the target working interval is represented by a frequency range [lowest frequency, highest frequency], this frequency range is used as the initial analysis frequency band; When the target working interval is at the center frequency and target bandwidth When indicating, will As the initial analysis frequency band; The target analysis frequency band is obtained by taking the intersection of the initial analysis frequency band and the frequency band that the system hardware can cover.

3. The control method for a reconfigurable photon sampling analog-to-digital converter according to claim 1, characterized in that, Before generating the performance status frame in S3, the following steps are also included: Before collecting digital signals and intermediate monitoring data from each channel, a stabilization waiting time is set, and sampling points generated during the stabilization waiting period are discarded.

4. A reconfigurable photon sampling analog-to-digital converter control system, used to implement the reconfigurable photon sampling analog-to-digital converter control method according to any one of claims 1-3, characterized in that, include: The initialization module obtains the startup security configuration and performs an initialization self-test. If the initialization self-test results meet the preset security startup conditions, the initialization self-test is deemed to have passed. The coarse registration module determines the target analysis frequency band based on the target working interval after the initial self-test passes. Based on the target application scenario, select the working mode from time interleaving mode, wavelength multiplexing mode and spectrum slicing mode; The initial sampling configuration is determined based on the startup security configuration, working mode and target analysis frequency band. Under the initial sampling configuration, calibration samples corresponding to each active channel are collected from the target analysis frequency band. Based on the calibration samples, the channel error vector corresponding to each channel is constructed. When the corresponding error component meets the preset threshold requirement, the coarse calibration configuration is obtained. The error estimation module, under coarse calibration configuration, inputs the analog signal to be converted into the reconfigurable multi-channel photonic sampling module, performs optical domain sampling according to the selected working mode and distributes it to M channels. After photoelectric conversion, M electrical signals are obtained. The M electrical signals are then subjected to transimpedance amplification, programmable gain and bias adjustment, ΣΔ oversampling noise shaping and quantization, and digital filtering and decimation downsampling to obtain M digital signals. Collect digital signals and intermediate monitoring data from each channel, and stitch them together to form a performance status frame; The performance status frame is input into a pre-trained deep learning error estimation model, which analyzes and outputs the error estimation results for each channel. The update module generates candidate control vectors based on the error estimation results, performs safety constraint review on the candidate control vectors, uses the approved candidate control vectors as the system control vectors, and performs soft update of the target parameters in the system control vectors. Candidate control vectors are generated based on the error estimation results. Specifically, based on the error estimation results, combined with the current target application mode, hardware constraints, and stability margin, initial parameters are generated, and the initial parameters are converted into candidate control vectors.

5. A reconfigurable photon sampling analog-to-digital converter control system according to claim 4, characterized in that, The system is applied to a reconfigurable multi-channel photonic sampling analog-to-digital converter. The front end of the reconfigurable multi-channel photonic sampling analog-to-digital converter is configured to support time-interleaving mode, wavelength multiplexing mode, and spectrum slicing mode through its internal optical switch matrix, tunable optical delay unit, and tunable optical filter. It can also dynamically switch between the three modes during operation according to the mode switching instructions in the candidate control vector.

6. A control device for a reconfigurable photon sampling analog-to-digital converter, characterized in that, It includes a processor and a memory, wherein the processor executes a computer program stored in the memory to implement a reconfigurable photon sampling analog-to-digital converter control method as described in any one of claims 1-3.

Images