High-overload-resistant data analysis method and system based on multi-dimensional buffer protection

Abstract

The application discloses a high-overload-resistant data analysis method and system based on multi-dimensional buffer protection, and relates to the technical field of data acquisition and storage. The system is composed of a plurality of function modules, including: a transient impedance module, which acquires stress wave propagation parameters and material strain data collected by a recorder node in real time, constructs a dynamic digital twin model of a buffer structure, and outputs an impact feature vector and a stress field distribution matrix; a signal reconstruction module, which dynamically configures a sensing signal based on the impact feature vector, including gain, bandwidth and filtering parameters of an analog front end, carries out conditioning and analog-digital conversion on the sensing signal, and carries out sparse reconstruction on abnormal data by using a compression sensing and dictionary learning algorithm, to generate a high-fidelity data stream and a storage data integrity identifier of each storage block; and a federal storage module, which inputs the stress field distribution matrix and the storage data integrity identifier into a side reinforcement learning decision engine together, to calculate failure risk probabilities of the storage blocks.

Description

Technical Field

[0001] This invention relates to the field of data acquisition and storage technology, specifically to a high-overload-resistant data analysis method and system based on multi-dimensional buffer protection. Background Technology

[0002] Data recording is performed in high-overload environments with peak accelerations exceeding 10,000 times the Earth's gravitational acceleration, with an impact time of less than one-hundredth of a second. This technology has key applications, particularly in high-speed collision tests and spacecraft reentry. The technology has undergone three development stages: the first stage was pure physical protection, using a multi-level buffer structure with a metal shell, potting material, and energy-absorbing layer to reduce the impact load transmitted to core components to a tolerable level; the second stage was redundant storage, introducing storage redundancy mechanisms on top of physical protection, including parallel writing of multiple flash memory chips and data striping based on disk array concepts; the third stage is embedded intelligence, where lightweight machine learning models are being deployed at the firmware level to dynamically adjust data protection strategies based on acceleration sensor data.

[0003] In existing technologies, recorders are all designed as isolated nodes. When the impact intensity exceeds the limit of the buffer structure and causes the storage chip to shatter, the data cannot be recovered, and there is a lack of cross-node collaborative backup mechanism. The storage management strategy cannot be continuously evolved. The edge decision model is trained on offline data with fixed parameters, which cannot adapt to the high uncertainty of the actual impact environment. The fixed threshold strategy causes significant fluctuations in data retention rate under different operating conditions. The buffer structure will suffer cumulative damage after multiple impacts, and the storage unit will also suffer from implicit degradation. Existing technologies cannot remotely obtain this health information, and maintenance depends on post-event disassembly and inspection. Existing technologies treat the buffer structure design, signal conditioning and acquisition, and data storage management as independent links. Storage management cannot use the stress field distribution information inside the buffer structure for accurate data migration. The signal acquisition quality cannot be fed back to the data migration strategy, and the data migration decision cannot guide the adjustment of signal conditioning parameters, resulting in the system failing to achieve the optimal collaborative state. Summary of the Invention

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] A high-overload resistant data analysis system based on multi-dimensional buffer protection includes:

[0006] The transient impedance module acquires stress wave propagation parameters and material strain data collected in real time by the recorder node, constructs a dynamic digital twin model of the buffer structure, and outputs the impact feature vector and stress field distribution matrix.

[0007] The signal reconstruction module dynamically configures the sensing signal based on the impact feature vector, including the gain, bandwidth and filtering parameters of the analog front end, conditions and converts the sensing signal from analog to digital, and uses compressed sensing and dictionary learning algorithms to sparsely reconstruct abnormal data, generating a high-fidelity data stream and storage data integrity identifiers for each storage block.

[0008] The federated storage module inputs the stress field distribution matrix and the storage data integrity identifier into the edge-side reinforcement learning decision engine to calculate the failure risk probability of each storage block, execute the active migration action of storage data, and at the same time perform redundant backup of high-fidelity data stream through the communication window. It also aggregates the data through the federated learning model to generate a cross-node backup mapping table and updated edge-side decision model parameters.

[0009] The twin maintenance module constructs a digital twin health record based on a cross-node backup mapping table and updated end-side decision model parameters, integrating failure risk probabilities. It uses a fatigue accumulation model and Weibull life distribution to predict the cumulative damage degree of the buffer structure and the remaining life of the storage unit in the digital twin health record, calculates the maintenance urgency index, and outputs predictive maintenance instructions.

[0010] Furthermore, the process of outputting the impact eigenvector and stress field distribution matrix is ​​as follows:

[0011] By acquiring real-time pressure time-series signals and multi-point strain values, and performing spatiotemporal alignment, a sparse measured strain field is generated using radial basis function interpolation. Based on Kalman filtering, the residual between the measured strain field and the simulated strain field of the benchmark model is used as the observation update quantity to dynamically adjust the material's effective elastic modulus, wave impedance, and interface bonding stiffness, and a dynamic digital twin model is obtained through iteration. Peak acceleration, impact pulse width, stress wave dominant frequency, impact direction angle, and impact energy density are extracted from the dynamic digital twin model to form an impact feature vector.

[0012] The internal space of the buffer structure is discretized into a regular grid, and the stress components of each grid node are calculated and arranged into a three-dimensional stress field distribution matrix.

[0013] Furthermore, the process of conditioning and analog-to-digital conversion of the sensor signal is as follows:

[0014] The system analyzes the peak acceleration, stress wave frequency, and impact pulse width in the impact characteristic vector; dynamically configures the gain level of the programmable gain amplifier based on the peak acceleration; sets the cutoff frequency of the anti-aliasing filter based on the stress wave frequency; configures the parameters of the programmable bandpass filter and notch filter based on the impact direction angle and energy density; conditions the signal in the order of impedance transformation, gain adjustment, filtering, and sample-and-hold; and dynamically selects the sampling rate of the analog-to-digital converter based on the impact pulse width and stress wave frequency.

[0015] Furthermore, the process of generating a high-fidelity data stream and storage data integrity identifiers for each storage block is as follows:

[0016] The system performs amplitude, continuity, and spectral checks on the stored data, classifying it as normal, slightly distorted, or severely damaged. A complete sparse dictionary is constructed offline using K-singular value decomposition, and a pre-defined online update interface is provided. For abnormal stored data, undamaged sampling points are extracted to form an observation vector. An atomic subset is pre-selected based on impulse characteristics, and the signal is reconstructed using an orthogonal matching pursuit algorithm, increasing the iteration count for severely damaged data. Confidence levels are calculated across three dimensions: reconstruction error, sparse representation, and spectral consistency, and a weighted integrity identifier is synthesized. After the task is completed, high-quality reconstructed segments are used as new samples.

[0017] Furthermore, the process of calculating the failure risk probability of each storage block is as follows:

[0018] The storage blocks are mapped to a stress field distribution matrix, and an equivalent average stress vector is obtained by using volume-weighted averaging. The peak stress, rise rate, and pulse width are extracted. The integrity identifier and its historical trend are analyzed. The physical baseline probability is calculated based on the offline calibrated stress failure probability surface. The integrity identifier is mapped to a risk correction factor through an exponential correction function. The physical baseline probability and correction factor are fused using an adaptive weighting method, where the baseline probability weight is increased when the physical stress level is low and the integrity identifier weight is increased when it is high. The failure risk probability is output after smoothing and filtering, and time series prediction is performed based on the stress wave propagation speed.

[0019] Furthermore, the process of generating the cross-node backup mapping table and the updated edge-side decision model parameters is as follows:

[0020] Establish a communication window and synchronize with the edge nodes; select high-value data blocks based on data integrity identifiers, failure risk probabilities, and importance levels, and encode them using erasure coding; request the allocation of target backup nodes from the edge nodes, transmit data slices and parity slices, and generate a cross-node backup mapping table that records the original block identifiers and storage locations; quantize and compress the parameters of the edge reinforcement learning model and upload them; the edge nodes perform federated average aggregation and distribute the global model; update the model locally and verify it using historical data, and roll back if the accuracy drops beyond a threshold.

[0021] Furthermore, the process of constructing a digital twin health record is as follows:

[0022] The system aggregates cross-node backup mapping tables, updated end-side decision model parameters, and failure risk probability time series for integrity verification; it establishes a hierarchical health profile for each node, including a metadata layer, an impact history layer, a storage health layer, a decision model layer, and a backup mapping layer; it aligns the failure risk probability with the stress field distribution matrix in time and space, and uses an offline-calibrated damage function to infer the cumulative damage contribution.

[0023] The backup mapping tables of all nodes are aggregated to build a global data distribution view and assess redundancy health; the model version tree is maintained, and model performance is evaluated offline using data from subsequent tasks; the archives are dynamically updated and a visualization interface is generated, and the data is persistently stored on the cloud platform.

[0024] Furthermore, the process of predicting the cumulative damage to the buffer structure and the remaining lifetime of the storage unit in the digital twin health record is as follows:

[0025] The stress field distribution matrix and impact feature vector are extracted, and the stress cycle is decomposed using the rainflow counting method. The material fatigue life curve is used for two-dimensional interpolation, and the damage contribution of each cycle is linearly accumulated based on Miner's law. Peak acceleration and energy density correction factors are introduced to dynamically update the cumulative damage degree. Historical health indicators of the memory block are extracted, and the failure time parameters are fitted using the Weibull distribution to estimate the remaining lifetime point and interval. A stress and temperature acceleration model is introduced to convert mechanical stress and thermal effects into equivalent programming erase number increments to correct lifetime prediction.

[0026] Furthermore, the process of outputting predictive maintenance instructions is as follows:

[0027] The comprehensive maintenance urgency index is synthesized by weighting three components: the cumulative damage degree of the buffer structure, the normalized reciprocal of the minimum remaining lifetime of the storage unit, and the proportion of insufficient cross-node replica redundancy; historical degradation trends are introduced to correct each component, and the threshold or bias is dynamically adjusted according to the mission criticality and the estimated impact intensity.

[0028] The comprehensive index is compared with multi-level thresholds to determine the urgency of maintenance; instructions for replacing buffer structures, calibrating memory chips, or re-executing cross-node redundancy are generated based on the dominant components; feedback data is collected after execution to compare with the prediction results and correct fatigue life curves, Weibull distributions, and acceleration model parameters.

[0029] The high-overload resistance data analysis method based on multidimensional buffer protection includes the following steps:

[0030] Step 1: Obtain stress wave propagation parameters and material strain data collected in real time by the recorder nodes, construct a dynamic digital twin model of the buffer structure, and output the impact feature vector and stress field distribution matrix.

[0031] Step 2: Dynamically configure the sensing signal based on the impact feature vector, including the gain, bandwidth and filtering parameters of the analog front end, condition and convert the sensing signal to analog-to-digital, and use compressed sensing and dictionary learning algorithms to sparsely reconstruct the abnormal data to generate a high-fidelity data stream and storage data integrity identifiers for each storage block.

[0032] Step 3: Input the stress field distribution matrix and the storage data integrity identifier into the edge-side reinforcement learning decision engine, calculate the failure risk probability of each storage block, execute the active migration action of storage data, and at the same time perform redundant backup of high-fidelity data stream through the communication window, and generate cross-node backup mapping table and updated edge-side decision model parameters through federated learning model aggregation.

[0033] Step 4: Based on the cross-node backup mapping table and the updated end-side decision model parameters, construct a digital twin health record by integrating failure risk probabilities. Use the fatigue accumulation model and Weibull life distribution to predict the cumulative damage degree of the buffer structure and the remaining life of the storage unit in the digital twin health record, calculate the maintenance urgency index, and output predictive maintenance instructions.

[0034] The high-overload-resistant data analysis method and system based on multi-dimensional buffer protection provided by this invention have the following beneficial effects:

[0035] (1) This invention establishes a communication window through the federated storage module, filters high-value data blocks for erasure coding, and uses side nodes to allocate target backup nodes to generate a cross-node backup mapping table. When a single recorder node is completely destroyed by extreme impact, the original data can be recovered through the data fragments and verification fragments stored in other surviving nodes, so that data survival no longer depends on the physical integrity of a single node, fundamentally solving the problem of permanent data loss caused by the breakage of the recorder body in the existing isolated node design.

[0036] (2) This invention calculates the failure risk probability of each storage block in real time and performs active migration through the edge reinforcement learning decision engine. At the same time, the edge model parameters are uploaded to the edge nodes. The edge nodes use the federated average algorithm to aggregate the experience data of multiple nodes and then distribute the global model. This mechanism enables the edge decision strategy to continuously absorb real impact experience in multiple task executions, realize the continuous optimization and generalization ability of the model, and thus adapt to the diverse impact waveforms generated by different target media such as concrete, soil, and metal, avoiding the problem of high decision error rate of fixed threshold strategy under different working conditions.

[0037] (3) The present invention drives the analog front-end dynamic configuration of the signal reconstruction module through the stress field distribution matrix output by the transient impedance module. The data integrity identifier generated by the signal reconstruction module and the stress field data are input together into the failure risk probability calculation of the federated storage module. The data migration instruction of the federated storage module controls the link parameter switching of the signal reconstruction module in reverse. The predictive maintenance instruction of the twin maintenance module is issued by the side and then adjusts the end-side migration strategy threshold. This closed-loop collaborative mechanism enables the physical protection of the buffer structure, the high-fidelity acquisition of the signal and the intelligent storage of data to perceive, restrict and optimize each other, realizing the optimal balance of the overall system performance, rather than the local independent optimization of each link. Attached Figure Description

[0038] Figure 1 This is a schematic diagram of the system of the present invention;

[0039] Figure 2 This is a schematic diagram of the overall method of the present invention. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] Example 1

[0042] Please see Figure 1 Embodiment 1 of this application provides a high-overload resistant data analysis system based on multi-dimensional buffer protection. The system includes:

[0043] The transient impedance module acquires stress wave propagation parameters and material strain data collected in real time by the recorder node, constructs a dynamic digital twin model of the buffer structure, and outputs the impact feature vector and stress field distribution matrix.

[0044] The process of real-time acquisition of stress wave propagation parameters and material strain data is as follows:

[0045] A distributed sensing network is pre-embedded at the multi-layer material interfaces and key stress concentration points within the buffer structure. This network consists of two types of sensors: a flexible piezoelectric thin-film sensor, which responds to the dynamic pressure signal generated during stress wave propagation and outputs a charge signal proportional to the rate of pressure change. This charge signal is converted into a voltage signal by a charge amplifier and then digitized by a high-speed analog-to-digital converter at a sampling rate of no less than 1 MHz; and a miniaturized fiber Bragg grating sensor, which is deployed at different depths and azimuths of the buffer structure. The sensor detects the Bragg wavelength shift in real time using a demodulator and calculates the local principal strain value and temperature compensation value at each measuring point.

[0046] All sensors are triggered by a synchronous clock and acquire raw time-series data in parallel during the impact process. The acquired raw data is preprocessed by the edge computing unit: first, low-frequency vibration interference and high-frequency circuit noise are removed by a bandpass filter, then the effective stress wave waveform is extracted by a wavelet threshold denoising algorithm, and then the stress wave propagation parameters, including wavefront arrival time, peak pressure, rise time, and full width at half maximum (FWHM) pulse width, are calculated from the waveform. At the same time, based on the strain measurements of multiple fiber optic gratings, the three-dimensional strain field distribution inside the material is reconstructed by a strain tensor inversion algorithm to obtain the material strain data, including the principal strain direction and amplitude at each measuring point.

[0047] The process of outputting the impact eigenvector and stress field distribution matrix is ​​as follows:

[0048] By establishing a multiphysics offline benchmark model of the buffer structure, a three-dimensional solid model containing geometric dimensions, material constitutive properties, and contact boundary conditions is constructed using finite element simulation software for the same type of buffer structure before execution. The material constitutive properties include elastoplasticity, viscoelasticity, and damage initiation and evolution criteria. By pre-setting various typical impact conditions in the model, i.e. different peak accelerations, pulse widths, and incident directions, the propagation law of stress waves in multilayer media and the spatiotemporal response of stress and strain inside the material are calculated. The simulation results are used as the initial library for parameterization of the benchmark model.

[0049] The real-time measurement data of the distributed sensor network is spatiotemporally aligned with the benchmark model. After receiving the pressure timing signal output by the flexible piezoelectric thin film sensor and the multi-point strain value output by the fiber optic grating sensor, time synchronization is performed. Taking the synchronization clock trigger time as the origin, the data of the two types of sensors are aligned along the same time axis and spatial interpolation is performed. The radial basis function interpolation method is used to map the strain measurement values ​​of a limited number of measurement points onto the three-dimensional mesh nodes of the buffer structure to form a sparse measured strain field.

[0050] By performing dynamic correction of the digital twin model, starting with an offline benchmark model as the initial state, the measured strain field is compared with the simulated strain field calculated by the benchmark model at the same spatial location to obtain the residual field. A real-time data assimilation algorithm based on Kalman filtering is used to dynamically adjust key parameters in the benchmark model, including the equivalent elastic modulus, wave impedance, and interface bonding stiffness of each layer of material, using the residual field as the observation update quantity. After several iterations, usually completed within hundreds of microseconds of the impact process, the root mean square error between the strain field output by the corrected model and the measured strain field is reduced to below a preset threshold. The model represents the current true state of the buffer structure, thus obtaining a dynamic digital twin model.

[0051] The impact feature vector is extracted from the dynamic digital twin model. This vector contains five components: peak acceleration, which is the maximum acceleration value along the projectile axis when the stress wave output by the model propagates to the device mounting surface; impact pulse width, which is the time taken for the acceleration waveform to travel from the 10% amplitude point of the rising edge to the 10% amplitude point of the falling edge; stress wave dominant frequency, which is the frequency component with the largest amplitude after performing a fast Fourier transform on the acceleration waveform; impact direction angle, which is the angle between the main impact direction and the projectile axis obtained by using the arctangent function based on the amplitude of the stress wave components in the three orthogonal directions calculated by the model; and impact energy density, which is the energy absorbed per unit volume obtained by integrating the stress field tensor output by the model within the component volume and dividing by the volume.

[0052] The stress field distribution matrix is ​​extracted from the dynamic digital twin model, and the internal space of the buffer structure is discretized into a regular grid, for example, along the radial and circumferential directions. The number of layers depends on the number of layers in the buffer structure. Three principal stress components and the maximum shear stress are calculated on each grid node. The values ​​are arranged into a three-dimensional matrix, with the dimensions of the matrix corresponding to the grid coordinates. Each element is a vector containing six independent stress components. This stress field distribution matrix provides the instantaneous stress value at each moment and also provides the stress pulse sequence of each grid point through time accumulation, which is used by subsequent modules to assess the mechanical damage risk of the storage unit location.

[0053] By outputting the impact feature vector and stress field distribution matrix, the five impact feature components calculated above are encapsulated into a feature vector, and the three-dimensional stress mesh data is encapsulated into a stress field distribution matrix. The data is then sent to the signal reconstruction module and the federated storage module via a high-speed internal bus in a preset data format. At the same time, the sensor data of the next moment is received and updated continuously until the impact process ends.

[0054] The signal reconstruction module dynamically configures the sensing signal based on the impact feature vector, including the gain, bandwidth and filtering parameters of the analog front end, conditions and converts the sensing signal from analog to digital, and uses compressed sensing and dictionary learning algorithms to sparsely reconstruct abnormal data, generating a high-fidelity data stream and storage data integrity identifiers for each storage block.

[0055] The process of conditioning and analog-to-digital conversion of the sensor signal is as follows:

[0056] The impact feature vector received from the transient impedance module contains five components: peak acceleration, impact pulse width, stress wave dominant frequency, impact direction angle, and impact energy density. The internal configuration controller extracts three parameters directly related to signal conditioning: peak acceleration, used to determine the maximum dynamic range of the signal; stress wave dominant frequency, used to determine the frequency distribution range of the effective signal; and impact pulse width, used to evaluate the time resolution requirements of the signal.

[0057] The gain of the analog front-end is dynamically configured. The analog front-end consists of a programmable gain amplifier and a programmable transconductance amplifier array. The configuration controller automatically sets the gain level according to the predicted peak acceleration value and the following principles: if the predicted peak acceleration is extremely high, for example, more than 50,000 times the acceleration due to gravity of the earth, a low gain level is selected to prevent signal saturation and clipping; if the predicted peak acceleration is moderate, a medium gain level is selected to balance the signal-to-noise ratio and no distortion; if the predicted peak acceleration is low, a high gain level is selected to amplify weak signals. The gain switching is completed in microseconds through a digitally controlled analog switch to ensure that the gain is stable before the arrival of the main impulse pulse.

[0058] The bandwidth of the analog front-end is dynamically configured. The configuration controller sets the cutoff frequency of the analog front-end anti-aliasing filter based on the stress wave's dominant frequency. Specifically, the cutoff frequency is set to 3 to 5 times the dominant frequency of the stress wave. This preserves the main frequency components of the signal while filtering out high-frequency noise and spurious interference above the Nyquist frequency. When the dominant frequency of the stress wave is high, the cutoff frequency is automatically increased to maintain the steepness of the signal's rising edge. When the dominant frequency is low, the cutoff frequency is appropriately decreased to suppress high-frequency noise. The bandwidth adjustment is achieved through an adjustable capacitor array or a switched capacitor filter, with an adjustment time of less than ten microseconds.

[0059] The dynamic configuration of filtering parameters includes an anti-aliasing filter, a programmable bandpass filter, and a notch filter in the analog front end. The configuration controller identifies potential interference sources based on the impact direction angle and energy density in the impact characteristic vector. For example, when the impact direction angle deviates significantly from the axial direction, the mechanical noise frequency generated by structural resonance is low. In this case, the lower limit frequency of the bandpass filter is adjusted upward to filter out the resonance noise. When the energy density is extremely high, the sensor itself may generate nonlinear saturation spikes. In this case, the notch filter is activated to suppress the characteristic frequency corresponding to the spikes. The center frequency, quality factor, and gain coefficient of all filters are written to the corresponding control registers through a digital interface.

[0060] The sensor signal is conditioned and processed by the dynamically configured analog front end in the following order: First, the charge or voltage signal output by the sensor enters the charge amplifier and voltage follower for impedance transformation; second, the signal is adjusted to the optimal input range of the analog-to-digital converter by the programmable gain amplifier; then the signal passes through the anti-aliasing low-pass filter, the programmable band-pass filter and the notch filter in sequence to remove out-of-band noise and specific interference; finally, the signal enters the sample-and-hold circuit, which holds the instantaneous value under precise timing control, waiting for analog-to-digital conversion.

[0061] The analog-to-digital conversion is performed, sending the conditioned analog signal to a high-speed analog-to-digital converter. The sampling rate of the analog-to-digital converter is not fixed, but dynamically selected by the configuration controller based on the impact pulse width and the dominant frequency of the stress wave: for narrow pulse width and high-frequency impacts, the highest sampling rate, such as above 10 MHz, is selected to capture details; for wide pulse width and low-frequency impacts, a moderate sampling rate is selected to reduce the amount of data. The converted digital signal is output in the form of a parallel data stream, and a data validity flag is also output.

[0062] The process of generating a high-fidelity data stream and storage data integrity identifiers for each storage block is as follows:

[0063] Through data anomaly detection and marking, the digital signal output by the analog-to-digital converter is sent to the anomaly detection unit in units of data blocks. Each data block corresponds to a continuous sampling point within a time window. The unit also receives status flags from the analog front end, including gain switching flags, filter saturation flags, and sampling rate change flags. For each data block, the anomaly detection unit performs three checks: amplitude check, which calculates the maximum and minimum values ​​of the sampled values ​​within the data block. If the values ​​exceed the preset normal range, such as being close to or above 95% of the full scale of the analog-to-digital converter or below 1%, it is marked as a saturation anomaly; continuity check, which calculates the difference between adjacent sampling points. If multiple consecutive difference values ​​exceed three times the expected rate of change, it is marked as a jump anomaly; and spectrum check, which performs a fast Fourier transform on the data block. If abnormal high-frequency components or power frequency interference characteristics appear in the spectrum, it is marked as an interference anomaly. Based on the check results, each data block is assigned an anomaly flag bit, which is divided into three levels: level 0 indicates normal data, level 1 indicates minor, repairable distortion, and level 2 indicates severely damaged data requiring reconstruction.

[0064] The dictionary learning algorithm is completed offline before task execution and continuously updated during operation through federated learning. During dictionary construction, a large number of high signal-to-noise ratio (SNR) clean signal samples under typical impact conditions are first collected. These samples come from laboratory calibration tests or high-quality data from previously successful recovery tasks. Then, these samples are segmented into fixed-length signal segments, typically 64 or 128 times the length of the sampling points. Next, the K-singular value decomposition algorithm is used to train these signal segments. The algorithm iteratively searches for a set of atoms such that each signal segment can be approximated by a linear combination of as few atoms as possible. Finally, a set of atoms, called the overcomplete dictionary, is obtained, with the number of atoms in the dictionary being much greater than the length of the signal segments. This dictionary is stored in the non-volatile memory of the signal reconstruction module and an interface is reserved for online updates to receive model parameter updates from the federated storage module.

[0065] When a data block is marked as a Level 1 or Level 2 anomaly, the compressed sensing reconstruction process is initiated. First, the locations and values ​​of undamaged sampling points in the data block are extracted, and these valid sampling points constitute the observation vector. The subset of atoms that best matches the current impact feature is selected from the dictionary. Based on the dominant frequency and pulse width parameters in the impact feature vector, atoms with similar frequency features in the dictionary are pre-screened. Then, the orthogonal matching pursuit algorithm is used to solve for the sparse coefficients: the algorithm iteratively selects the atom with the strongest correlation to the residual from the subset of atoms and updates the coefficient vector until the residual energy is lower than a preset threshold or the number of iterations reaches the upper limit. The linear combination of the obtained sparse coefficient vector and the corresponding atom is the reconstructed complete signal segment. For Level 2 anomalies, i.e., severe damage, the algorithm increases the number of iterations and the number of atoms to obtain more accurate reconstruction at a higher computational cost. For Level 1 anomalies, a fast greedy algorithm is used to complete the reconstruction in fewer iterations. The reconstructed signal segment replaces the original damaged data block, forming a high-fidelity data stream.

[0066] The integrity identifier is a value between zero and one, used to quantify the reliability of data in each storage block. The signal reconstruction module calculates confidence scores in three dimensions for each output data block and synthesizes the final integrity identifier. The first dimension is the reconstruction error confidence score, calculated by comparing the values ​​corresponding to the original undamaged sampling points in the reconstructed signal segment with the actual sampled values ​​and calculating the root mean square error; the smaller the error, the higher the confidence score. The second dimension is the sparse representation confidence score, calculated by statistically analyzing the number of atoms used in the reconstruction process and the total number of iterations; fewer atoms indicate that the signal can be sparsely represented, indicating better reconstruction quality and higher confidence. The third dimension is the spectral consistency confidence score, calculated by comparing the reconstructed signal with the spectral features obtained in the anomaly detection stage; if the main peak of the reconstructed signal's spectrum matches the main peak of the original valid sampling points' spectrum, the confidence score is high. The three confidence scores are weighted and averaged with weights of 0.5, 0.3, and 0.2 respectively to synthesize a comprehensive score, which is the integrity identifier of the data block. For data blocks without anomalies, the integrity identifier is directly set to the maximum value of 1.0.

[0067] The high-fidelity data stream and integrity identifier are encapsulated and output. The signal reconstruction module concatenates the reconstructed signal segments with normal data blocks that have not experienced anomalies in chronological order to form a continuous high-fidelity data stream. Simultaneously, for each storage block, typically corresponding to a fixed-length data stream segment (e.g., one kilobyte per block), a corresponding integrity identifier is generated, and the identifier value is quantized into an eight-bit binary number between 0.00 and 1.00, with a step size of approximately 0.04. The high-fidelity data stream and integrity identifier sequence are sent in parallel to the federated storage module via a high-speed internal bus for data migration decisions and cross-node backup.

[0068] After each impact task, the signal reconstruction module temporarily stores the signal segments with good reconstruction quality, integrity indicators greater than 0.9, and original anomalies marked as level one as new training samples. When the number of accumulated samples reaches a threshold, such as 1000 segments, the module starts incremental dictionary learning: based on the original dictionary, a sparse dictionary update method is used to adjust some atoms with new samples, retaining atoms with high usage frequency and replacing atoms with low usage frequency. The updated dictionary parameters are sent to the federated storage module through the communication interface, and then the federated storage module aggregates them on the edge, ultimately realizing the collaborative evolution of the dictionary among multiple nodes.

[0069] The federated storage module inputs the stress field distribution matrix and the storage data integrity identifier into the edge-side reinforcement learning decision engine to calculate the failure risk probability of each storage block, execute the active migration action of storage data, and at the same time perform redundant backup of high-fidelity data stream through the communication window. It also aggregates the data through the federated learning model to generate a cross-node backup mapping table and updated edge-side decision model parameters.

[0070] Calculate the failure probability of each storage block:

[0071] The stress field distribution matrix is ​​a three-dimensional grid data that records the time-varying sequence of six stress components—three normal stresses and three shear stresses—on each grid node within the buffer structure. The end-side reinforcement learning decision engine first acquires the stress field distribution matrix at the current moment and maps each storage block to the corresponding grid cell in the stress field matrix based on the physical coordinates of the storage block inside the recorder. Since the size of the storage block may cover multiple grids, the engine uses a volume-weighted average method: using the volume occupied by the storage block as the weight, the stress components of all covered grid nodes are weighted and averaged to obtain the equivalent average stress vector borne by the storage block at the current moment. At the same time, the engine also extracts the stress peak value, stress rise rate, and stress pulse width at the grid location during the impact process as additional dynamic stress features.

[0072] Each storage block corresponds to an integrity identifier, which is generated by the signal reconstruction module and is a value between 0 and 1. After receiving the integrity identifier sequence, the engine first checks the validity of the identifier: if the identifier is 0, it means that the data in the storage block is completely unusable, and the failure risk probability of the storage block is directly set to one, without further calculation; if the identifier is greater than zero, it is normalized to the range of zero to one. The engine also extracts the historical trend of the integrity identifier, that is, the difference value of the integrity identifier in the most recent sampling, to determine whether the data quality is deteriorating or tending to stabilize. The integrity identifier and the stress field data are aligned in time through a synchronous clock to ensure that they reflect the physical and signal states at the same moment.

[0073] For each storage block, the decision engine merges the stress field-related features and integrity identifier-related features mentioned above to form a multi-dimensional comprehensive feature vector. The stress field-related features include: equivalent average stress amplitude, peak stress, stress rise rate, stress pulse width, and the time delay of the stress wave arriving at the storage block. The integrity identifier-related features include: the current integrity identifier value, the historical mean of the integrity identifier, the slope of the integrity identifier's trend, and the variance of the integrity identifier's fluctuation. These features collectively describe the current mechanical environment and data quality status of the storage block.

[0074] The decision engine internally maintains a baseline risk model based on physical failure mechanisms. This model determines the instantaneous failure probability of a storage block under different stress levels, considering only stress effects and not data quality, based on offline calibration tests. Specifically, through numerous destructive tests, the proportion of physical damage to the storage chip, such as bond wire breakage, solder ball cracking, and chip cracking, under different combinations of stress amplitude and stress pulse width is recorded, and a stress failure probability surface is fitted. For the equivalent average stress amplitude and stress pulse width of the current storage block, the baseline failure probability caused by physical stress is obtained by interpolation on this surface.

[0075] Data integrity indicators reflect the degree of distortion at the signal level, and signal distortion is often correlated with microscopic damage to memory chips, such as threshold voltage drift and read interference. The decision engine uses an empirical correction function to map the integrity indicator to a risk correction factor: when the integrity indicator is close to one, the correction factor is close to zero, indicating good data quality, and even with greater physical stress, the actual failure risk does not increase significantly; when the integrity indicator decreases, the correction factor increases exponentially, indicating that data quality deterioration will significantly increase the failure risk; the specific shape of the correction factor is derived from historical data fitting, reflecting the statistical relationship between the integrity indicator and the actual damage of memory cells.

[0076] The decision engine fuses the physical stress baseline failure probability with the integrity label correction factor to obtain the final failure risk probability for each storage block. The fusion method adopts adaptive weighting: when the physical stress level is low, the weight of the physical stress baseline probability is higher, and its impact on the integrity label is smaller; when the physical stress level is high, the weight of the integrity label increases, because the data integrity label under high stress is better able to reflect the impending hardware damage. The final probability is passed through a smoothing filter to avoid decision jitter caused by instantaneous fluctuations. After the calculation is completed, the engine outputs the failure risk probability of each storage block, with a value ranging from 0 to 1, where 0 represents no risk and 1 represents inevitable failure.

[0077] In addition to the risk probability at the current moment, the decision engine also predicts the risk probability evolution within a short time window, such as one millisecond, based on the time change trend of the stress field distribution matrix. The prediction method uses linear extrapolation combined with stress wave propagation speed: based on the location and propagation direction of the stress wave, the peak stress and arrival time of the storage blocks that have not yet arrived are estimated, and the risk probability of these storage blocks at future moments is calculated in advance. The prediction results are used to guide the early execution of proactive data migration actions, that is, to start migration before the risk probability rises to the threshold, thereby gaining valuable migration time windows.

[0078] The decision engine outputs migration decision instructions based on the failure risk probability of each storage block, combined with preset risk thresholds: high-risk threshold and medium-risk threshold. When the failure risk probability of a storage block exceeds the high-risk threshold, a forced migration action is triggered: all data in the storage block is read out, erasure coding is performed to generate data fragments and parity fragments, which are then distributed and written to several spare storage blocks with the lowest risk probability within the node, and the address mapping table is updated. When the risk probability is between medium-risk and high-risk, a dual backup action is triggered: while the data remains in its original location, an additional copy is written to the spare area. The migration process uses a high-speed DMA channel to ensure that data transmission is completed before the shock stress wave reaches the storage block. After the migration is completed, the original storage block is marked as migrated or awaiting reclamation, and subsequent read and write operations are automatically redirected to the new address.

[0079] The process of generating the cross-node backup mapping table and the updated edge-side decision model parameters is as follows:

[0080] The federated storage module continuously monitors the communication link status. When an available communication window is detected, such as the self-test phase before missile launch, the telemetry window during flight, or the beacon window after landing, it performs a handshake synchronization with the side-side node onboard computer, the escort node, or the ground relay. The handshake process includes: exchanging node identifiers, confirming the communication protocol version, synchronizing the clock reference, and evaluating link quality parameters such as signal-to-noise ratio, bit error rate, and available bandwidth. The communication window has a limited duration, typically tens of milliseconds to several seconds, and the module needs to efficiently complete backup and aggregation tasks within this window.

[0081] The federated storage module selects data blocks from local storage that require cross-node backup. The selection criteria include three dimensions: a data integrity index greater than 0.9, indicating good data quality and worth backing up; a storage block failure risk probability greater than 0.7, indicating that the data is about to face physical damage risk; and the importance level of the data, such as sampled data at the peak of the impact, trigger signals, calibration parameters, etc., are marked as the highest priority. For each selected data block, the module performs erasure coding: dividing the data block into multiple data pieces of fixed size, such as 1000 bytes each, and generating a corresponding number of parity pieces. The encoding parameters and the ratio of the number of data pieces to the number of parity pieces are dynamically adjusted according to the link quality. The better the link quality, the lower the proportion of parity pieces to save bandwidth.

[0082] The module sends a backup request to the edge node, which includes the total amount of data to be backed up and the desired redundancy, such as wanting the data to be stored on three different nodes. The edge node queries the health status of other recorder nodes, remaining storage capacity, battery level, whether it has completed its own data acquisition, and the current communication link quality according to the global node registry. It then assigns a target backup node to each data segment and returns the allocation list to the requesting node. According to the allocation list, the module sends the data segments and check segments one by one through the wireless communication link. After sending each segment, it waits for an acknowledgment signal from the receiving node. If no acknowledgment is received within a specified time, it automatically retransmits. If multiple retransmissions fail, it requests a change of target node from the edge node. After all data segments have been transmitted, the module records the storage location of each data segment, including the target node identifier, storage address offset, and the original identifier of the corresponding data block.

[0083] Based on the aforementioned transmission records, the federated storage module generates a cross-node backup mapping table. The mapping table adopts a key-value pair structure, where the key is the original data block identifier, such as a logical block address, and the value is a record containing the following fields: encoding group identifier, total number of data fragments and total number of check fragments, target node identifier and storage address for each data fragment, and checksum hash value. This mapping table is stored in the module's non-volatile memory and a copy is simultaneously synchronized to the edge nodes to prevent the local mapping table from being lost due to node damage. The mapping table is used to quickly locate the backup copy during subsequent data recovery.

[0084] Within the communication window, in addition to data backup, the federated storage module also participates in federated learning model aggregation. When a sufficient amount of new experience data has been accumulated locally, such as the impact feature sequence, integrity identifier sequence, action sequence, and reward sequence collected in this task, the parameters of the current edge reinforcement learning decision model are quantized and compressed to reduce the amount of data transmitted. This data, along with the experience data, is packaged and uploaded to the edge node. The model parameters include the weight matrix and bias vector of each layer of the deep Q network. The compression uses eight-bit integer quantization, converting the thirty-two-bit floating-point number into an eight-bit integer, achieving a compression ratio of four to one.

[0085] After collecting model parameters and experience data uploaded by multiple recorder nodes, the side nodes perform a federated averaging algorithm: First, they count the amount of experience data uploaded by each node and calculate the proportion of each node's data to the total data; then, using this proportion as the weight, they perform a weighted average of the model parameters of each node to obtain the global model parameters; the side nodes can also use the aggregated experience data to perform an additional round of training on the global model to further improve model performance; after aggregation, the side nodes send the updated global model parameters back to each recorder node through the communication window.

[0086] After receiving the global model parameters, the federated storage module performs a local model update: replacing the parameters of the current edge decision model with the received global parameters, while retaining the old version as a backup; after the update, the module uses some historical data stored locally, such as data not used in the training of the most recent task, for quick verification: calculating the decision accuracy and average reward of the new model on the dataset, and comparing it with the performance of the old model; if the accuracy of the new model drops more than a preset threshold, such as a drop of more than 5%, it automatically rolls back to the old version and reports the anomaly to the edge node through the communication window; if the verification passes, the new model is officially enabled, and the updated model parameters are saved to non-volatile memory as the initial values ​​for the next federated learning.

[0087] When the communication window is closed, such as during the intrusion and penetration phase of a projectile or when the wireless link is interrupted due to an impact, the federated storage module stops all backup and aggregation activities. The generated cross-node backup mapping table and the updated edge-side decision model parameters are securely stored for later use. In data recovery scenarios, if the node is completely destroyed, the mapping table can be queried through edge nodes or other surviving nodes to collect data fragments and parity fragments distributed on each node, and erasure coding decoding can be performed to recover the original data.

[0088] The twin maintenance module, based on the cross-node backup mapping table and the updated end-side decision model parameters, integrates the failure risk probability to construct a digital twin health record. It uses the fatigue accumulation model and Weibull life distribution to predict the cumulative damage degree of the buffer structure and the remaining life of the storage unit in the digital twin health record, calculates the maintenance urgency index, and outputs predictive maintenance instructions.

[0089] The process of building a digital twin health record is as follows:

[0090] The twin maintenance module is deployed on the cloud platform. It receives data uploaded from various recorder nodes via wireless communication links, either through a wired interface or a wireless network within the telemetry window after data is recovered. The received data is divided into three categories: a cross-node backup mapping table, generated and uploaded by the federated storage module, recording the cross-node distribution location, erasure coding parameters, and verification information of each original data block; updated edge-side decision model parameters, including the neural network weight matrix, bias vector, and model version number of each node's reinforcement learning decision engine; and a time series of failure risk probabilities recorded by each storage block during the task. This series is calculated by the edge-side decision engine at each sampling time and bound to the corresponding timestamp and storage block physical address. The cloud platform first performs integrity verification on these three types of data: it calculates a cyclic redundancy check code for each data packet and compares it with the check value attached to the sender; if they do not match, it requests retransmission; simultaneously, it checks the continuity of timestamps and removes obviously abnormal or duplicate data points.

[0091] For each recorder node, the cloud platform maintains a unique digital twin health record. If the node is uploading data for the first time, a new record is created, a globally unique identifier is assigned, and static information such as its model, batch, manufacturing date, and initial buffer structure parameters are recorded. If the node already has historical records, the new data from the current task is appended to the record. The health record adopts a hierarchical storage structure: the first layer is the metadata layer, which stores the node identifier, task number, task time, and task type, such as overview information like concrete penetration, armor penetration, and high-speed collision; the second layer is the impact history layer, which stores each task in chronological order. The first layer is the impact feature vector sequence of the task, including peak acceleration, pulse width, main frequency, orientation angle and energy density, as well as the corresponding stress field distribution matrix summary; the second layer is the storage health layer, which stores the failure risk probability time series, integrity identification history and indicators such as raw bit error rate, number of programming erases and threshold voltage drift fed back by the self-test circuit for each storage block; the third layer is the decision model layer, which stores the end-side decision model parameters and their version number and verification accuracy obtained after each task; the fourth layer is the backup mapping layer, which stores a copy of the cross-node backup mapping table for tracing data distribution and recovery path.

[0092] The twin maintenance module spatiotemporally aligns the received failure risk probability time series with the corresponding positions in the stress field distribution matrix. Specifically, for each storage block, it extracts its failure risk probability curve throughout the entire impact process, and simultaneously extracts physical parameters such as the peak stress, stress pulse width, and cumulative energy density of the grid point where the storage block is located. Using the offline calibrated damage function, the risk probability is inversely converted into the cumulative damage contribution. For example, when the risk probability reaches 0.8, it is considered that the impact has caused damage to the storage block equivalent to one fatigue cycle. When the risk probability exceeds 0.9, the damage contribution is doubled. The cumulative damage contributions from all tasks are summed to obtain the total cumulative damage value for each storage block, which is used for subsequent lifetime prediction.

[0093] The twin maintenance module aggregates cross-node backup mapping tables uploaded by all nodes to construct a global data distribution view. This view uses the original data block as an index to record the node identifiers and storage addresses of all replicas of that data block, as well as the relationships between replicas, such as which replicas belong to the same erasure coding group. By analyzing the global view, the module can assess the redundancy health of the system: if the number of valid replicas of a data block is lower than a preset redundancy threshold, such as less than three, it is marked as insufficient redundancy, triggering an early warning. It can also detect nodes that are frequently selected as backup targets, as these nodes may be aging faster due to excessive write operations, thus reminding the system to balance the backup load in maintenance recommendations.

[0094] The twin maintenance module stores the edge-side decision model parameters after each task update into the decision model layer of the health record; it records the parent version of each model, i.e., the previous generation model, the child version updated based on this model, and the amount of federated learning aggregated data used for each update; for each version of the model, the module uses the actual data collected in subsequent tasks of that node, including impact features, integrity identifiers, executed actions, and the real rewards obtained through data recovery verification afterward, to perform offline evaluation and calculate performance indicators such as generalization error, decision accuracy, and mistransfer rate of the model; these evaluation results are stored together with the model parameters to guide future model selection: for example, when a node reports a decline in model performance, it can roll back to a previously better-performing version.

[0095] The twin maintenance module updates the digital twin health record in real time via streaming. Each time a new data packet is received, the module immediately performs the aforementioned aggregation, correlation, and fusion operations, updates the statistical summary in the record, and generates a visualization interface for maintenance personnel: displaying the impact history of each node in a timeline format, using color depth to represent the cumulative damage level; displaying the stress field distribution and storage block health status inside the buffer structure in a 3D rendering; displaying the trend of failure risk probability changing with the impact process in a graph; and listing the global view of the cross-node backup mapping table and redundancy health status in a table format. Maintenance personnel can intuitively grasp the full life cycle status of each recorder through this interface and promptly identify potential problems.

[0096] All digital twin health records are stored in a distributed database on the cloud platform, employing a multi-copy redundancy mechanism to ensure no data loss; the record data is encrypted and accessible only to authorized maintenance personnel; after each task is completed, the cloud platform automatically archives and backs up the newly added health record data, and associates it with previous historical records to form a complete lifecycle record; when a recorder node cannot be recovered due to physical damage, its last uploaded health record is still retained on the cloud platform, which can be used to analyze the cause of failure and improve subsequent designs.

[0097] Predicting the cumulative damage to the buffer structure and the remaining lifetime of the storage units:

[0098] The twin maintenance module extracts the full lifecycle data of the buffer structure of each recorder node from the digital twin health record, including: the time series of stress field distribution matrix, impact characteristic vector, peak acceleration, pulse width, dominant frequency, orientation angle and energy density output by the transient impedance module in each task, as well as the cumulative damage contribution obtained by back-calculation through failure risk probability; the cumulative damage degree of the buffer structure is defined as: the degree of irreversible degradation of the buffer material due to repeated high overload impacts from the first use to the current moment, represented by a value from 0 to 1, where 0 represents no damage and 1 represents that the critical failure threshold has been reached.

[0099] For the stress field distribution matrix recorded in each task, the module first extracts the stress time history at each material integration point inside the buffer structure, showing the curves of normal stress and shear stress changing with time. Since the impact process usually includes multiple stress peaks and troughs, such as multiple reflections and oscillation attenuation, the rainflow counting method is used to cyclically decompose the stress time history. The stress time history curve is rotated 90 degrees to simulate the path of rainwater flowing along the curve, and each process of stress amplitude rising from a trough to the next peak and then back to the trough is defined as a half-cycle. Two half-cycles are paired to form a complete stress cycle. Each stress cycle is characterized by two parameters: stress amplitude and half the difference between the peak and trough values, and the average of the average peak and trough values. After rainflow counting, all stress cycles generated in each task are extracted and grouped and statistically analyzed according to the magnitude of the stress amplitude.

[0100] The module pre-stores the fatigue life curves of each layer of materials in the buffer structure, such as aluminum foam, epoxy resin, and rubber pads. These curves represent the number of cycles a standard specimen undergoes before failing under different stress amplitudes. The curves are obtained through numerous standard fatigue tests and stored in a tabular format in the material database of the cloud platform. For each stress cycle obtained from rainflow counting, the module performs two-dimensional interpolation on the fatigue life curve of the corresponding material based on its stress amplitude and average stress to obtain the fatigue life of the material under that cycle alone, i.e., the allowable number of cycles. If the average stress is not zero, the Goodman correction formula is used for equivalent transformation to convert the asymmetric cycle into a symmetric cycle under the equivalent stress amplitude, and then the equivalent fatigue life is obtained by looking up the table.

[0101] For each material integration point, the damage contribution of all stress cycles generated in all tasks is linearly accumulated; the damage caused by each stress cycle is equal to the actual occurrence of that cycle divided by the fatigue life corresponding to that cycle; the cumulative damage is the sum of the damage of all stress cycles; for example, if an integration point experiences ten cycles with a certain stress amplitude in multiple tasks, and the fatigue life corresponding to that amplitude is one thousand cycles, then these ten cycles contribute 0.01 damage; if it also experiences two cycles with higher amplitudes, and the fatigue life is one hundred cycles, then it contributes 0.02 damage, and the total cumulative damage is 0.03; the cumulative damage distribution cloud map inside the entire buffer structure is obtained by calculating independently for each integration point.

[0102] The above linear cumulative calculation assumes no interaction between stress cycles, but in reality, high overload impacts may produce material strengthening or weakening effects. By introducing a correction factor, the cumulative damage is adjusted according to the peak acceleration and impact energy density in the impact characteristic vector: when the energy density is extremely high, the material may experience accelerated microcrack propagation, and the correction factor is greater than one; when the number of impacts is small and the stress amplitude is low, the material may experience cyclic hardening, and the correction factor is less than one. The correction factor is obtained by inverting historical failure data and is updated online as more task data accumulates. The corrected cumulative damage is stored in a digital twin health record and recalculated after each new task to achieve dynamic tracking.

[0103] The lifetime prediction of memory cells mainly targets non-volatile memory chips, such as NAND flash memory. Historical health indicators for each memory block are extracted from its health record, including: the cumulative number of program erase cycles, the time series of the raw bit error rate, the trajectory of threshold voltage drift, and the number of error correction codes used during read / write operations. Simultaneously, the temperature cycling data experienced by the memory block is extracted, as high overload shocks are often accompanied by rapid temperature changes and the mechanical stress history at corresponding locations in the stress field distribution matrix.

[0104] The Weibull distribution is used to describe the failure time distribution of memory blocks. The Weibull distribution contains two parameters: a scale parameter and a shape parameter. For each memory block, the module uses its historical health index time series and dynamically fits the Weibull parameters in the current state using the maximum likelihood estimation method. Specifically, the number of program erases is regarded as the equivalent time axis, since flash memory wear is mainly related to the number of program erases. The raw bit error rate after each task is regarded as the state observation value. When the raw bit error rate exceeds a preset threshold, such as 0.3%, the memory block is considered to have entered a soft failure state. When it exceeds a higher threshold, such as 1%, and the error correction code cannot correct it, it is considered to be completely failed. The failure ratio data of memory blocks in the same batch under different number of program erases are collected to fit the shape parameter of the Weibull distribution, which reflects whether the failure rate is increasing, constant, or decreasing, and the scale parameter reflects the characteristic lifetime. For typical flash memory with a shape parameter greater than one, the failure rate accelerates with the increase of the number of uses.

[0105] Based on the fitted Weibull distribution parameters, the module calculates the remaining lifetime of each storage block. The remaining lifetime is defined as the number of program erases or tasks remaining from the current moment until the probability of block failure reaches a preset critical value, such as 50%. The point estimate is obtained by solving the equation that the failure probability function equals the critical value. At the same time, a confidence interval is given, such as the lifetime range at a 90% confidence level, reflecting the uncertainty of the prediction. For storage blocks that have not yet shown obvious signs of degradation, a lower limit estimate is given based on historical data from the same batch.

[0106] Mechanical stress and temperature accelerate flash memory aging. By introducing an acceleration factor, the peak stress field and temperature changes experienced by the memory block are converted into an equivalent increment of program erase cycles. For example, if the memory block is subjected to high stress during an impact, it is equivalent to consuming several additional program erase cycles. The specific value of the acceleration factor is calibrated through a pre-tested accelerated life test: the life of flash memory samples is tested at different stress levels and temperatures, and the coefficients of the Arrhenius or inverse power law model are fitted. The peak stress, stress pulse width, and highest temperature of the memory block in each task are substituted into the acceleration model to calculate the equivalent wear caused by the impact, which is then added to the number of program erase cycles. The accumulated equivalent number of cycles is then used for Weibull lifetime prediction.

[0107] By outputting the predicted remaining lifetime of each storage block, expressed as the remaining number of programmable erases or remaining number of tasks, and the cumulative damage degree of each material integration point of the buffer structure, the prediction results are written back to the digital twin health profile and compared with the actual subsequent failure data to correct the model parameters. For buffer structures with a cumulative damage degree close to the critical value, such as greater than 0.8, or storage blocks with a predicted remaining lifetime of less than one task, an early warning flag is generated and passed to the subsequent maintenance instruction calculation process. All prediction processes retain complete calculation logs for easy post-event traceability and model verification.

[0108] The process of calculating maintenance urgency indicators and outputting predictive maintenance instructions is as follows:

[0109] The twin maintenance module defines the maintenance urgency index as a weighted composite of three components, each reflecting the health status of the buffer structure, storage units, and system redundancy. The buffer structure damage urgency is directly calculated using the cumulative damage level of the buffer structure predicted in the previous step, with a value ranging from zero to one. The storage unit lifetime urgency is defined as the minimum predicted remaining lifetime among all storage blocks, expressed as the normalized reciprocal of the number of remaining tasks; the fewer the remaining tasks, the closer this component is to one. The redundancy health urgency is calculated based on the redundancy of global data replicas using a cross-node backup mapping table: the percentage of data blocks with fewer than three replicas (those with fewer than a safe threshold) is used to determine the redundancy health urgency. The initial weights of the three components are dynamically set according to the system usage stage: in the early stages of equipment operation, the weight of buffer structure damage is lower, and the weight of storage unit lifetime is higher; in the later stages of equipment operation, the weight of buffer structure damage gradually increases.

[0110] Simply looking at the current state might mask the accelerating trend of degradation. To address this, historical degradation data for each storage block and buffer material integration point is extracted, and the incremental rate of change of cumulative damage over the last three tasks is calculated. If the damage increment shows an increasing trend—for example, the increment in the third task is 1.5 times that of the second, and the second is 1.3 times that of the first—an amplification factor is applied to the corresponding urgency component. This amplification factor is proportional to the product of the increment rate. Conversely, if the degradation rate tends to level off, a reduction factor is applied. The corrected component better reflects the rising rate of failure risk in the near future, avoiding maintenance delays due to currently acceptable values.

[0111] Different tasks have different requirements for data reliability. The criticality level of the next planned task is read from the task configuration file, which is divided into three levels: normal, important, and critical, as well as the expected impact intensity level, which is estimated by the target medium type, incident velocity, etc. For critical tasks, the overall threshold of the maintenance urgency index is lowered by 20%, making it easier to trigger maintenance warnings. For tasks with extremely high expected impact intensity, an additional bias term is added to the buffer structure damage urgency component, and the bias magnitude is proportional to the estimated impact energy density. In addition, if the node is about to be deployed to a non-recoverable environment, such as a single-use weapon, the output logic of the maintenance urgency index is reversed: it no longer recommends early maintenance, but instead outputs a special instruction that allows use with defects, but the number of backup nodes needs to be increased.

[0112] The module sums the three modified and weighted components to obtain a comprehensive maintenance urgency index, with a value range of 0 to 1. This index is then compared with three preset thresholds: a green threshold (e.g., below 0.2) indicates that the equipment is in good condition and requires no maintenance; a yellow threshold (0.3 to 0.6) indicates that attention is needed and it is recommended to check it before the next task; an orange threshold (0.6 to 0.8) indicates that maintenance is needed as soon as possible; and a red threshold (above 0.8) indicates that the equipment must be immediately shut down and replaced. The comparison result determines whether a maintenance command is triggered and the urgency level of the command.

[0113] Based on the components of the urgency index and the reasons for exceeding the threshold, the module generates different types of maintenance instructions. If the urgency contribution of buffer structure damage is the largest and exceeds the orange threshold, a buffer structure replacement instruction is output, along with a suggested replacement location, based on the most severely damaged area in the cumulative damage cloud map, such as the outer aluminum foam near the projectile's head. If the urgency contribution of storage unit lifespan is the largest, a storage chip calibration or replacement instruction is output, listing the physical addresses of the five storage blocks with the shortest predicted remaining lifespan. If the urgency contribution of redundancy health is the largest, a cross-node data redundancy re-execution instruction is output, requiring the backup of insufficiently replicated data blocks to other nodes within the next communication window. If multiple components exceed the threshold simultaneously, a comprehensive overhaul instruction is output, including the aforementioned sub-tasks.

[0114] When multiple nodes report maintenance requests simultaneously, and maintenance resources are limited, the module sorts all nodes to be maintained from highest to lowest according to the comprehensive maintenance urgency index and outputs a priority list. For nodes ranked higher, it is recommended to schedule maintenance immediately; for nodes ranked lower, it is recommended to postpone to the next maintenance cycle. Based on the specific instruction type of each node, the module estimates the time and consumables required for maintenance, such as the type and quantity of buffer materials to be replaced and the type of memory chip to be calibrated, and outputs a resource requirement list for reference in the maintenance plan formulation.

[0115] The generated predictive maintenance instructions are encapsulated in the form of structured data packets, including node identifiers, urgency index values, instruction types, suggested operating steps, reference data such as cumulative damage distribution maps, remaining life prediction curves, and timestamps. After being encrypted, the data packets are sent to the maintenance management information system through the cloud platform's output interface, and are also presented in a readable text format on the maintenance personnel's workstation interface. For emergency instructions with a red threshold, an alert is sent to the designated maintenance manager via SMS or instant messaging tools.

[0116] After the maintenance personnel execute the instructions, they enter the actual operation results, such as which layer of buffer material was replaced and the comparison test data before and after the replacement, into the system. The twin maintenance module reads this feedback data and compares it with the previously predicted damage level and remaining life: if the damage level found during the actual replacement is basically consistent with the prediction, the accuracy of the prediction model is verified; if there is a large deviation, the model correction process is triggered to adjust the fatigue life curve interpolation coefficient, Weibull distribution parameters, or acceleration model coefficients; the corrected model will be used for the prediction of subsequent nodes to achieve adaptive optimization of predictive maintenance.

[0117] Example 2

[0118] Please see Figure 2 Based on Example 1, Example 2 of this application also provides a high-overload data analysis method based on multi-dimensional buffer protection, including the following specific steps:

[0119] Step 1: Obtain stress wave propagation parameters and material strain data collected in real time by the recorder nodes, construct a dynamic digital twin model of the buffer structure, and output the impact feature vector and stress field distribution matrix.

[0120] Step 2: Dynamically configure the sensing signal based on the impact feature vector, including the gain, bandwidth and filtering parameters of the analog front end, condition and convert the sensing signal to analog-to-digital, and use compressed sensing and dictionary learning algorithms to sparsely reconstruct the abnormal data to generate a high-fidelity data stream and storage data integrity identifiers for each storage block.

[0121] Step 3: Input the stress field distribution matrix and the storage data integrity identifier into the edge-side reinforcement learning decision engine, calculate the failure risk probability of each storage block, execute the active migration action of storage data, and at the same time perform redundant backup of high-fidelity data stream through the communication window, and generate cross-node backup mapping table and updated edge-side decision model parameters through federated learning model aggregation.

[0122] Step 4: Based on the cross-node backup mapping table and the updated end-side decision model parameters, construct a digital twin health record by integrating failure risk probabilities. Use the fatigue accumulation model and Weibull life distribution to predict the cumulative damage degree of the buffer structure and the remaining life of the storage unit in the digital twin health record, calculate the maintenance urgency index, and output predictive maintenance instructions.

[0123] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0124] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0125] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A high-overload resistant data analysis system based on multi-dimensional buffer protection, characterized in that, The system includes: a transient impedance module, which acquires stress wave propagation parameters and material strain data collected in real time by the recorder nodes, constructs a dynamic digital twin model of the buffer structure, and outputs impact feature vectors and stress field distribution matrices; The signal reconstruction module dynamically configures the sensing signal based on the impact feature vector, including the gain, bandwidth and filtering parameters of the analog front end, conditions and converts the sensing signal from analog to digital, and uses compressed sensing and dictionary learning algorithms to sparsely reconstruct abnormal data, generating a high-fidelity data stream and storage data integrity identifiers for each storage block. The federated storage module inputs the stress field distribution matrix and the storage data integrity identifier into the edge-side reinforcement learning decision engine to calculate the failure risk probability of each storage block, execute the active migration action of storage data, and at the same time perform redundant backup of high-fidelity data stream through the communication window. It also aggregates the data through the federated learning model to generate a cross-node backup mapping table and updated edge-side decision model parameters. The process of calculating the failure risk probability of each storage block is as follows: The storage blocks are mapped to a stress field distribution matrix, and an equivalent average stress vector is obtained by using volume-weighted averaging. The peak stress, rise rate, and pulse width are extracted. The integrity identifier and its historical trend are analyzed. The physical reference probability is calculated based on the offline calibrated stress failure probability surface. The integrity identifier is mapped to a risk correction factor through an exponential correction function. The physical reference probability and correction factor are fused using adaptive weighting. The failure risk probability is output after smoothing and filtering, and time series prediction is performed based on the stress wave propagation speed. The integrity indicator reflects the degree of distortion at the signal level, and signal distortion is correlated with microscopic damage to the memory chip and read interference. An empirical correction function is used to map the integrity indicator to a risk correction factor: when the integrity indicator is close to one, the risk correction factor is close to zero, indicating good data quality. When the integrity indicator decreases, the risk correction factor increases exponentially, indicating that data quality deterioration will significantly increase the risk of failure. The specific shape of the risk correction factor is obtained by fitting historical data. The twin maintenance module constructs a digital twin health record based on a cross-node backup mapping table and updated end-side decision model parameters, integrating failure risk probabilities. It uses a fatigue accumulation model and Weibull life distribution to predict the cumulative damage degree of the buffer structure and the remaining life of the storage unit in the digital twin health record, calculates the maintenance urgency index, and outputs predictive maintenance instructions.

2. The high-overload resistant data analysis system based on multi-dimensional buffer protection according to claim 1, characterized in that, The process of outputting the impact eigenvector and stress field distribution matrix is ​​as follows: By acquiring real-time pressure time-series signals and multi-point strain values, and performing spatiotemporal alignment, a sparse measured strain field is generated using radial basis function interpolation. Based on Kalman filtering, the residual between the measured strain field and the simulated strain field of the benchmark model is used as the observation update quantity to dynamically adjust the material's effective elastic modulus, wave impedance, and interface bonding stiffness, and a dynamic digital twin model is obtained through iteration. Peak acceleration, impact pulse width, stress wave dominant frequency, impact direction angle, and impact energy density are extracted from the dynamic digital twin model to form an impact feature vector. The internal space of the buffer structure is discretized into a regular grid, and the stress components of each grid node are calculated and arranged into a three-dimensional stress field distribution matrix.

3. The high-overload resistant data analysis system based on multi-dimensional buffer protection according to claim 2, characterized in that, The process of conditioning and analog-to-digital conversion of sensor signals is as follows: The system analyzes the peak acceleration, stress wave frequency, and impact pulse width in the impact characteristic vector; dynamically configures the gain level of the programmable gain amplifier based on the peak acceleration; sets the cutoff frequency of the anti-aliasing filter based on the stress wave frequency; configures the parameters of the programmable bandpass filter and notch filter based on the impact direction angle and energy density; conditions the signal in the order of impedance transformation, gain adjustment, filtering, and sample-and-hold; and dynamically selects the sampling rate of the analog-to-digital converter based on the impact pulse width and stress wave frequency.

4. The high-overload resistant data analysis system based on multi-dimensional buffer protection according to claim 3, characterized in that, The process of generating a high-fidelity data stream and storage data integrity identifiers for each storage block is as follows: The stored data is subjected to amplitude, continuity and spectrum checks, and marked as normal, slightly distorted or severely damaged. A complete sparse dictionary is constructed offline using K-singular value decomposition and reserved in the online update interface. For abnormal stored data, undamaged sampling points are extracted to form observation vectors, atomic subsets are pre-selected based on impact characteristics, and signals are reconstructed through orthogonal matching pursuit algorithm. The number of iterations is increased when the data is severely damaged. Calculate the confidence scores of three dimensions: reconstruction error, sparse representation, and spectral consistency, and then synthesize a weighted integrity identifier. After the task is completed, the high-quality reconstructed fragments are used as new samples.

5. The high-overload resistant data analysis system based on multi-dimensional buffer protection according to claim 1, characterized in that, The process of generating the cross-node backup mapping table and the updated edge-side decision model parameters is as follows: Establish a communication window and synchronize with the edge nodes; select high-value data blocks based on data integrity identifiers, failure risk probability, and importance level, and encode them with erasure coding; request the allocation of target backup nodes from the edge nodes, transmit data fragments and parity fragments, and generate a cross-node backup mapping table that records the original block identifiers and storage locations; Upload the parameters of the edge reinforcement learning model after quantization and compression. Side nodes perform federated average aggregation and distribute the global model; they update the model locally and validate it using historical data, and roll back if the accuracy drops below a threshold.

6. The high-overload resistant data analysis system based on multi-dimensional buffer protection according to claim 1, characterized in that, The process of building a digital twin health record is as follows: The system aggregates cross-node backup mapping tables, updated edge-side decision model parameters, and failure risk probability time series for integrity verification; it also establishes a hierarchical health profile for each node, including a metadata layer, impact history layer, storage health layer, decision model layer, and backup mapping layer. The failure risk probability is spatiotemporally aligned with the stress field distribution matrix, and the cumulative damage contribution is inversely calculated through the offline calibrated damage function. The backup mapping tables of all nodes are aggregated to build a global data distribution view and assess redundancy health; the model version tree is maintained, and model performance is evaluated offline using data from subsequent tasks; the archives are dynamically updated and a visualization interface is generated and stored on the cloud platform.

7. The high-overload resistant data analysis system based on multi-dimensional buffer protection according to claim 6, characterized in that, The process of predicting the cumulative damage to the buffer structure and the remaining lifetime of the storage unit in a digital twin health record is as follows: The stress field distribution matrix and impact characteristic vector were extracted, and the stress cycle was decomposed using the rainflow counting method. Two-dimensional interpolation is performed using the material fatigue life curve. The damage contribution of each cycle is linearly accumulated based on Miner's law, and peak acceleration and energy density correction factors are introduced to dynamically update the cumulative damage degree. Historical health indicators of the storage block are extracted, and the failure time parameters are fitted using Weibull distribution to estimate the remaining lifetime point and interval. A stress and temperature acceleration model is introduced to convert mechanical stress and thermal effects into equivalent programming erase increments to correct lifetime prediction.

8. The high-overload resistant data analysis system based on multi-dimensional buffer protection according to claim 7, characterized in that, The process of outputting predictive maintenance instructions is as follows: The comprehensive maintenance urgency index is synthesized by weighting three components: the cumulative damage degree of the buffer structure, the normalized reciprocal of the minimum remaining lifetime of the storage unit, and the proportion of insufficient cross-node replica redundancy; historical degradation trends are introduced to correct each component, and the threshold or bias is dynamically adjusted according to the mission criticality and the estimated impact intensity. The overall index is compared with multi-level thresholds to determine the urgency of maintenance; Based on the dominant component, instructions for replacing the buffer structure, calibrating the memory chip, or re-executing cross-node redundancy are generated; after execution, feedback data is collected to compare with the prediction results, and fatigue life curves, Weibull distributions, and acceleration model parameters are corrected.

9. A high-overload resistance data analysis method based on multi-dimensional buffer protection, characterized in that, Includes the following steps: Step 1: Obtain stress wave propagation parameters and material strain data collected in real time by the recorder nodes, construct a dynamic digital twin model of the buffer structure, and output the impact feature vector and stress field distribution matrix. Step 2: Dynamically configure the sensing signal based on the impact feature vector, including the gain, bandwidth and filtering parameters of the analog front end, condition and convert the sensing signal to analog-to-digital, and use compressed sensing and dictionary learning algorithms to sparsely reconstruct the abnormal data to generate a high-fidelity data stream and storage data integrity identifiers for each storage block. Step 3: Input the stress field distribution matrix and the stored data integrity identifier into the edge-side reinforcement learning decision engine to calculate the failure risk probability of each storage block, execute the active migration of stored data, and simultaneously perform redundant backup of the high-fidelity data stream through the communication window. Then, aggregate the data through a federated learning model to generate a cross-node backup mapping table and updated edge-side decision model parameters. The process of calculating the failure risk probability of each storage block involves: mapping the storage block to the stress field distribution matrix, using volume-weighted averaging to obtain the equivalent average stress vector, and extracting the stress peak value, rise rate, and pulse width; analyzing the integrity identifier and its historical trend; calculating the physical baseline probability based on the offline calibrated stress failure probability surface; and using the index... A numerical correction function maps the integrity flag to a risk correction factor; an adaptive weighted fusion of the physical baseline probability and the correction factor is employed; the output is a smoothed and filtered failure risk probability, and time series prediction is performed based on the stress wave propagation speed; the integrity flag reflects the degree of distortion at the signal level, and signal distortion is correlated with microscopic damage to the memory chip and read interference; an empirical correction function is used to map the integrity flag to the risk correction factor: when the integrity flag is close to one, the risk correction factor is close to zero, indicating good data quality; when the integrity flag decreases, the risk correction factor increases exponentially, indicating that data quality deterioration will significantly increase the failure risk; the specific shape of the risk correction factor is derived by fitting historical data; Step 4: Based on the cross-node backup mapping table and the updated end-side decision model parameters, construct a digital twin health record by integrating failure risk probabilities. Use the fatigue accumulation model and Weibull life distribution to predict the cumulative damage degree of the buffer structure and the remaining life of the storage unit in the digital twin health record, calculate the maintenance urgency index, and output predictive maintenance instructions.

Images