A Fault Early Warning Method and System for Embedded Storage Modules Based on Vertical Connectors

Abstract

This invention relates to the field of embedded storage device fault diagnosis technology, and discloses a method and system for early warning of embedded storage module faults based on a vertical connector. The method includes: acquiring historical low-level data and preprocessing it to obtain a state sequence; extracting the cooperative change relationship of multi-dimensional physical quantities based on the state sequence to obtain a comprehensive feature matrix and performing feature fusion to obtain a comprehensive feature set; constructing a health benchmark model based on the comprehensive feature set; acquiring real-time low-level data and performing spatial distance calculation with the health benchmark model to obtain a state deviation value; if the deviation value is greater than a preset deviation threshold, a cooperative distortion is determined to have occurred; performing multi-dimensional feature space outlier evaluation on the real-time data with cooperative distortion; if the outlier degree exceeds an adaptive alarm benchmark, a module fault early warning signal is generated. This invention solves the problem that single-threshold monitoring cannot identify early cooperative degradation signals through multi-dimensional cooperative relationship modeling and dual anomaly verification.

Description

Technical Field

[0001] This invention relates to the field of embedded storage device fault diagnosis technology, and in particular to an embedded storage module fault early warning method and system based on a vertical connector. Background Technology

[0002] Currently, in high-reliability application scenarios such as spacecraft, wind power generation, and industrial control, embedded storage modules based on vertical connector structures undertake the tasks of storing and transmitting critical data. The stability of their operating status is directly related to the safety and continuous operation capability of the entire system, making them an important research object in the field of fault prediction and health management.

[0003] In one existing technology, digital virtualization technology is used to simulate the communication process of a spaceborne embedded system on a host machine. This is achieved by software virtualization of the underlying hardware communication interface, and by overlaying a preset hardware fault response mode onto the communication data stream to complete fault injection and evaluation. This solution can simulate and verify known types of hardware faults during the R&D phase, which helps improve the system's fault resistance design level. However, this technology is essentially an offline testing and verification method. Its core lies in simulating the injection effect of known fault types and evaluating the system response, rather than sensing the health degradation state of the storage module itself in real-time in a real operating environment. In the long-term operation scenario of embedded storage modules with vertical connector structures, there are complex collaborative relationships among multiple physical quantities such as voltage fluctuations, temperature changes, read / write latency, and data throughput. When early degradation occurs inside the module, the collaborative change pattern of each physical quantity will subtly shift, but this shift is often submerged within the normal fluctuation range.

[0004] In summary, existing technologies cannot accurately identify potential faults that have already occurred but have not yet triggered a single-point threshold in the early stages. Summary of the Invention

[0005] This invention provides a method and system for early warning of embedded storage module faults based on a vertical connector, in order to solve the problem that existing technologies cannot accurately identify potential faults that have not yet triggered a single-point threshold but have already caused collaborative distortion in the early stages.

[0006] In a first aspect, to solve the above-mentioned technical problems, the present invention provides a fault early warning method for embedded storage modules based on a vertical connector, comprising: Obtain the underlying data of the embedded storage module during its historical operating cycle, and preprocess the underlying data to obtain the time series sequence of the reference physical quantity; Based on the time series sequence of the benchmark physical quantity, physical quantity sequences of multiple dimensions are separated. Based on the physical quantity sequences, a cooperative change relationship map is constructed, and the cooperative change relationship map is subjected to feature projection processing to obtain a comprehensive feature matrix characterizing the operating state. The comprehensive feature matrix is ​​subjected to feature fusion processing to obtain a comprehensive feature set; A high-dimensional feature space is obtained by mapping the comprehensive feature set to a high-dimensional space. The boundary of the normal operating state of the high-dimensional feature space is fitted to determine the health benchmark model. The real-time underlying data of the embedded storage module is obtained, and the spatial distance between the real-time underlying data and the health benchmark model is calculated to obtain the state deviation value. If the state deviation value is greater than the preset deviation threshold, it is determined that a cooperative distortion has occurred. The real-time underlying data that is determined to have undergone collaborative distortion is subjected to outlier evaluation in the multidimensional feature space to obtain an anomaly score value and a dynamic limit value. If the anomaly score value is greater than the dynamic limit value, a module fault warning signal is generated.

[0007] Secondly, the present invention provides an embedded storage module fault early warning system based on a vertical connector, comprising: a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor executes the program to implement the method described above.

[0008] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described above.

[0009] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention obtains a complete operating state sequence by performing time series alignment and missing value filling on the underlying data of the embedded storage module during historical operating cycles. Then, it separates the physical quantity sequences of multiple dimensions and performs frequency domain coupling analysis, extracts the coupling weights between dimensions, and constructs a collaborative change relationship map. This makes the multidimensional collaborative distortion information of the vertical connector, which was originally submerged in the normal fluctuation range, explicit and numerical. A comprehensive feature matrix is ​​obtained through dimensionality reduction mapping, and a comprehensive feature set is obtained through feature fusion. This overcomes the shortcomings of existing technologies that rely on preset fixed thresholds or single-dimensional anomaly judgment, resulting in insufficient information and inability to capture collaborative offsets between multidimensional physical quantities. This significantly improves the perceptibility of the health status of the embedded storage module.

[0010] (2) This invention obtains a high-dimensional feature space by performing kernel function mapping based on the comprehensive feature set, and determines the data that maximizes the classification interval through hyperplane construction and classification interval optimization, thereby dividing the normal data boundary and establishing a health benchmark model. This overcomes the shortcomings of existing technologies that cannot autonomously learn the normal state boundary from historical data and rely on manually setting fixed thresholds, resulting in poor adaptability, and realizes the adaptive construction and dynamic updating of the health benchmark model.

[0011] (3) This invention maps the spatial relationship between real-time data and health boundaries into a continuous algebraic deviation to quantify the degree of state deviation. Compared with the traditional discretization method that triggers an alarm when the threshold is exceeded, this continuous processing not only eliminates logical definition conflicts but also intuitively reflects the early slow degradation trajectory of the storage module, making up for the shortcomings of traditional rigid threshold judgment that cannot perceive the degradation process.

[0012] (4) Based on the preliminary screening of collaborative distortion, this invention introduces a multi-dimensional feature space outlier assessment and distribution variance adaptive verification mechanism. It quantifies the degree of outlier by calculating the local distribution density of sample points and dynamically sets the alarm benchmark according to the discrete characteristics of sample scores. A multi-level anomaly early warning closed loop from coarse screening to fine verification is constructed, which effectively eliminates false alarms in a single dimension and significantly improves the accuracy and reliability of fault early warning in complex storage environments. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the fault warning method for embedded storage modules based on vertical connectors provided in the first embodiment of the present invention. Detailed Implementation

[0014] 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.

[0015] Reference Figure 1 The first embodiment of the present invention provides a fault early warning method for embedded storage modules based on a vertical connector, including the following steps: S11, Obtain the underlying data of the embedded storage module in the historical operation cycle, and preprocess the underlying data to obtain the reference physical quantity time series; S12, Based on the time series sequence of the benchmark physical quantity, separate the physical quantity sequence of multiple dimensions, construct a cooperative change relationship map based on the physical quantity sequence, and perform feature projection processing on the cooperative change relationship map to obtain a comprehensive feature matrix characterizing the operating state; S13, Perform feature fusion processing on the comprehensive feature matrix to obtain a comprehensive feature set; S14, A high-dimensional feature space is obtained by mapping the comprehensive feature set to a high-dimensional space, and the boundary of the normal operating state of the high-dimensional feature space is fitted to determine the health benchmark model. S15, obtain the real-time underlying data of the embedded storage module, calculate the spatial distance between the real-time underlying data and the health benchmark model to obtain the state deviation value, and if the state deviation value is greater than the preset deviation threshold, determine that a cooperative distortion has occurred. S16, perform multi-dimensional feature space outlier evaluation on the real-time underlying data that is determined to have cooperative distortion to obtain an anomaly score value and a dynamic limit value. If the anomaly score value is greater than the dynamic limit value, generate a module fault warning signal.

[0016] In step S11, the underlying data of the embedded storage module during its historical operating cycle is obtained, and the underlying data is preprocessed to obtain a reference physical quantity time series, including: S111, Obtain the underlying data of the embedded storage module during its historical operating cycle, and extract the timestamp and sampling frequency from the underlying data; S112, perform sequence alignment processing on the underlying data according to the timestamp and the sampling frequency to obtain an initial state sequence with missing positions; S113, if the continuous span of the vacancy position is less than the maximum allowable continuous vacancy threshold, then interpolation calculation is performed on the vacancy position to determine the filling value; S114, the missing value is filled into the initial state sequence to perform missing value filling operation, and a reference physical quantity time series sequence is obtained.

[0017] In step S111, the underlying data of the embedded storage module within the historical operating cycle is obtained, and the timestamp and sampling frequency are extracted from the underlying data. The duration of the historical operating cycle is determined according to the equipment's maintenance interval. For example, in a wind turbine monitoring scenario, the routine maintenance cycle of the wind turbine is usually used as the length of the historical operating cycle, which is set to 30 days according to industry operation and maintenance standards. The underlying operating data includes at least voltage fluctuation data, temperature change data, read / write latency data, and data throughput data. The data of each physical quantity are collected by the sensors and controller logs integrated within the embedded storage module according to a predetermined sampling strategy and stored in the storage medium.

[0018] The system reads the timestamp corresponding to each data record from the underlying data as a timestamp; it also reads the sampling frequency parameter configured during data acquisition. The sampling frequency is determined based on the highest frequency component of the monitored physical signal and follows the Nyquist sampling theorem. For example, in the scenario of monitoring the vibration of the main shaft of a wind turbine generator, the fundamental frequency of the main shaft rotation is usually below 100 Hz. According to the Nyquist theorem, the sampling frequency is set to be no less than 200 Hz. The specific value is determined by the hardware configuration of the storage module, such as 50 Hz or 200 Hz.

[0019] In step S112, the underlying data is sequence aligned according to the timestamp and the sampling frequency to obtain an initial state sequence with missing positions.

[0020] Using a standard timeline as a reference, the underlying data undergoes sequence alignment processing based on the extracted timestamps and sampling frequency. Specifically, a standard time grid with equal intervals is constructed using the reciprocal of the sampling frequency as the theoretical sampling interval. For example, when the sampling frequency is 50 Hz, the theoretical sampling interval is 0.02 seconds, so the interval between adjacent nodes in the standard time grid is 0.02 seconds. The actual physical quantity data collected is mapped to the corresponding theoretical time nodes in the standard time grid according to their timestamps. If a corresponding actual collected data exists at a theoretical time node, the data is filled into that node; if no corresponding actual collected data exists at a theoretical time node, the node is marked as a vacant position. After traversing and mapping all the underlying data, an initial state sequence with actual data and vacant positions distributed on the standard time grid is obtained. This alignment processing is used to eliminate time misalignment problems caused by network latency or sensor transient failures, ensuring strict equidistance of the time series data.

[0021] In step S113, if the continuous span of the vacancy position is less than the maximum allowable continuous vacancy threshold, then interpolation calculation is performed on the vacancy position to determine the filling value.

[0022] The initial state sequence is traversed and detected for each missing position, identifying all consecutive missing intervals formed by consecutively arranged adjacent missing positions. The number of missing positions contained in each consecutive missing interval is counted, and the product of the number of missing positions and the sampling frequency is taken as the continuous span of the consecutive missing interval. The maximum allowable consecutive missing threshold is set to 0.1 seconds; it is dynamically calculated based on the sampling frequency. When the sampling frequency is 50 Hz, the threshold corresponds to 5 sampling points; when the sampling frequency is 200 Hz, the threshold corresponds to 20 sampling points. The conversion formula is sampling frequency multiplied by 0.1. Taking a sampling frequency of 50 Hz as an example, the time span corresponding to 5 consecutive sampling points is 0.1 seconds. When the sampling frequency changes, the corresponding number of sampling points is recalculated according to the actual sampling frequency. This span meets the technical requirements of data continuity for frequency domain analysis and ensures the effectiveness of interpolation operations within the local feature range.

[0023] The continuous span of the detected consecutive gap interval is compared with the maximum allowable consecutive gap threshold. If the continuous span is less than the maximum allowable consecutive gap threshold, the filling value is determined by interpolation using any one of the nearest neighbor interpolation algorithm, linear interpolation algorithm, or spline interpolation algorithm for each gap position within the consecutive gap interval. If the continuous span is greater than or equal to the maximum allowable consecutive gap threshold, no interpolation filling is performed on the consecutive gap interval to avoid introducing unreliable interpolated data.

[0024] When using the nearest neighbor interpolation algorithm, the nearest valid data value is searched forward and backward along the time axis, with the missing position as the center. This nearest valid data value is used as the filling value for the missing position. If the nearest valid data values ​​before and after the missing position are equidistant, the average of the two is taken as the filling value.

[0025] When using the linear interpolation algorithm, the nearest valid data value is searched forward on the time axis as the first reference value, and the nearest valid data value is searched backward as the second reference value. The line connecting the first reference value and the second reference value is used as the interpolation line, and the value on the interpolation line corresponding to the missing position on the time axis is used as the filling value.

[0026] When using the spline interpolation algorithm, multiple valid data values ​​are selected as control points on the time axis, both forward and backward, with the missing position as the center. The interpolation curve is obtained by performing cubic spline fitting on each control point, and the value on the interpolation curve corresponding to the missing position on the time axis is used as the filling value.

[0027] In one implementation, the continuous gap interval contains three gap positions with a continuous span of 0.06 seconds, which is less than the maximum allowable continuous gap threshold of 0.1 seconds. The filling value for each gap position is calculated using any one of the following algorithms: nearest neighbor interpolation, linear interpolation, or spline interpolation. Preferably, nearest neighbor interpolation is used to reduce computational overhead in scenarios with gradual data changes; linear interpolation is used in scenarios with linear data changes; and spline interpolation is used in scenarios with non-linear data changes and requiring high interpolation accuracy.

[0028] In step S114, the missing value is filled into the initial state sequence to perform a missing value filling operation, thereby obtaining a reference physical quantity time series sequence.

[0029] All interpolated values ​​are added to the initial state sequence according to their corresponding missing positions, overwriting the original missing markers, thus completing the missing value filling operation. After processing all missing positions, the initial state sequence is transformed into a reference physical quantity time series sequence where all nodes on the standard time grid have valid data values. This reference physical quantity time series sequence has a complete time series structure, with each data node evenly distributed at equal intervals on the time axis, making it suitable for subsequent frequency domain analysis and multidimensional feature extraction operations.

[0030] Furthermore, before performing interpolation calculations on the missing positions, the process includes outlier detection and removal steps for the initial state sequence. Specifically, the local mean and local standard deviation of each valid data value in the initial state sequence are calculated. If a valid data value deviates from the local mean by more than a preset multiple of the local standard deviation, such as four times, the valid data value is marked as an outlier and treated as a missing position. The preset multiple is determined statistically based on the normal fluctuation range in historical data. This outlier preprocessing avoids interpolation reference shifts caused by occasional spike pulses from the sensor, improving the reliability of the filled data.

[0031] In step S12, multiple-dimensional physical quantity sequences are separated based on the time series sequence of the baseline physical quantities. A cooperative change relationship map is constructed based on the physical quantity sequences, and feature projection processing is performed on the cooperative change relationship map to obtain a comprehensive feature matrix characterizing the operating state, including: S121, Based on the reference physical quantity time series, separate multiple independent dimensions of initial physical quantity sequences, and perform frequency domain transformation processing on the initial physical quantity sequences of each dimension to obtain the frequency domain fluctuation amplitude set of each dimension. S122, calculate the coupling weights between different dimensions based on the frequency domain fluctuation amplitude set, and generate an initial correlation matrix based on the coupling weights; S123, establish connection edges between the dimensions corresponding to the elements in the initial association matrix that are greater than the preset collaboration threshold, and construct a collaboration change relationship graph based on the selected nodes and connection edges; S124, Perform spatial dimensionality reduction mapping on the cooperative change relationship map to obtain a comprehensive feature matrix.

[0032] In step S121, multiple independent-dimensional initial physical quantity sequences are separated from the reference physical quantity time series, and frequency domain transformation processing is performed on the initial physical quantity sequences of each dimension to obtain the frequency domain fluctuation amplitude set of each dimension.

[0033] The benchmark physical quantity time series contains synchronous observations of multidimensional physical quantities along the time axis recorded by an embedded storage module, with each physical quantity corresponding to a different data dimension. Based on the data channel identifier of each physical quantity, initial physical quantity sequences for each independent dimension are separated from the benchmark physical quantity time series. In one embodiment, for a wind turbine generator main shaft vibration monitoring scenario, the benchmark physical quantity time series contains spatial composite vibration information. Through coordinate system decoupling, the composite vibration information is separated into initial physical quantity sequences for three independent dimensions: a transverse vibration sequence, a longitudinal vibration sequence, and an axial vibration sequence.

[0034] The initial physical quantity sequences for each dimension are subjected to frequency domain transformation, converting each sequence from a time domain representation to a frequency domain representation. Specifically, a Fast Fourier Transform (FFT) is performed on the initial physical quantity sequences for each dimension to obtain the amplitude spectrum of each sequence in the frequency domain. Frequency domain fluctuation amplitude values ​​within a preset core frequency band are extracted from the amplitude spectrum to form a set of frequency domain fluctuation amplitudes corresponding to each dimension. The core frequency band is determined based on the physical characteristics of the monitored object. For example, for the vibration signal of the main shaft of a wind turbine generator, the fundamental frequency of the main shaft rotation and its harmonic energy are mainly concentrated in the 10 Hz to 500 Hz frequency band, so this frequency band is determined as the core frequency band.

[0035] In step S122, the coupling weights between different dimensions are calculated based on the set of frequency domain fluctuation amplitudes, and an initial correlation matrix is ​​generated based on the coupling weights.

[0036] The set of frequency domain fluctuation amplitudes corresponding to each dimension is used as the frequency domain feature vector of that dimension. The correlation coefficient between the frequency domain feature vectors of different dimensions is calculated, and the correlation coefficient is used as the coupling weight between the corresponding dimensions. Specifically, the Pearson correlation coefficient algorithm is used to calculate the correlation value of the frequency domain fluctuation amplitude between each pair of dimensions, and the coupling weight representing the coupling strength between each dimension is obtained. The coupling weights of all the calculated dimension pairs are arranged according to the dimension correspondence to generate an initial correlation matrix. The initial correlation matrix is ​​a symmetric matrix, where the rows and columns of the matrix correspond to each physical quantity dimension, and the element values ​​in the matrix represent the coupling weight between the corresponding row dimension and column dimension. In one embodiment, for the three-dimensional scenario of spindle vibration monitoring, the coupling weights between the horizontal and vertical, the vertical and axial, and the horizontal and axial are calculated respectively. The three sets of coupling weights are arranged according to the dimension correspondence to generate a 3x3 initial correlation matrix.

[0037] In step S123, connecting edges are established between the dimensions corresponding to the elements in the initial association matrix that are greater than the preset collaboration threshold, and a collaboration change relationship graph is constructed based on the selected nodes and connecting edges.

[0038] The preset collaboration threshold is determined based on the statistical significance of the Pearson correlation coefficient at a specified sample size. The system calculates the statistical critical value of the correlation coefficient based on the set confidence level and the current sample size of the monitoring window, and uses this critical value as a benchmark. To ensure the robustness of collaboration determination in engineering applications, the system introduces a preset safety offset to correct the critical value, ensuring that the final collaboration threshold is higher than the statistical critical value.

[0039] In one implementation, the system monitoring window is set with a sample size of 30 and a confidence level of 95%. The corresponding statistical significance threshold is 0.36. The system adds a correction offset of 0.24 to this threshold, with the offset ranging from 20% to 40% of the current threshold, determined by engineering experience within this range, thereby setting the collaboration threshold to 0.6. If the sample size in the monitoring window changes, the system automatically recalculates the statistical significance threshold using the standard correlation coefficient distribution function and adds the same safety offset, achieving dynamic adjustment of the collaboration threshold.

[0040] Iterate through the elements in the initial correlation matrix and identify dimension pairs with element values ​​greater than a preset collaboration threshold as having significant collaboration relationships. Using each physical quantity dimension as a node, establish connection edges between dimension pairs with significant collaboration relationships. The weight of the connection edges between dimensions is determined according to the corresponding coupling weight.

[0041] Based on all the selected nodes and connecting edges, a cooperative change relationship graph is constructed. The cooperative change relationship graph is an undirected weighted graph, where each node corresponds to a physical quantity dimension, each connecting edge represents a significant cooperative change relationship between the corresponding two dimensions, and the weight of the connecting edge represents the strength of the cooperative change.

[0042] In step S124, the cooperative change relationship graph is subjected to spatial dimensionality reduction mapping to obtain a comprehensive feature matrix. Spatial distribution features are extracted from the cooperative change relationship graph to obtain the spatial topological distribution features of each node in the graph. The spatial distribution feature extraction is implemented using a graph convolutional neural network.

[0043] The structure of the graph convolutional neural network is as follows: the input layer receives the adjacency matrix and node feature matrix of the cooperative change relationship graph; the hidden layer uses two graph convolutional layers, with the output dimension of the first graph convolutional layer being 64 and the output dimension of the second graph convolutional layer being 32; each graph convolutional layer is followed by a modified linear unit activation layer to introduce nonlinear expression; the output layer outputs the feature vector corresponding to each node; the output dimension is determined according to the number of nodes in the graph, with 64 being about twice the number of nodes and 32 being equivalent to the number of nodes. This setting enables the encoder to achieve layer-by-layer feature compression while preserving topological information.

[0044] The graph convolutional neural network is pre-trained in an unsupervised manner, with the training objective being to ensure that the output feature vector retains the topological structure and node attribute information of the input graph. Low-level data from multiple historical periods are collected from the embedded storage module under normal operating conditions and processed according to steps S111 to S123 to obtain multiple sample co-variation relationship graphs. The adjacency matrix and node feature matrix of each sample graph are sequentially input into the graph convolutional neural network encoder. The encoder contains two graph convolutional layers: the first layer has an output dimension of 64, and the second layer has an output dimension of 32. Each graph convolutional layer is followed by a modified linear unit activation layer, and the output feature vector corresponding to each node is calculated through forward propagation. The output feature vector is then input into the decoder network, which consists of two graph convolutional layers and is used to reconstruct the feature vector into an adjacency matrix and a node feature matrix. The mean squared error is used to calculate the errors between the reconstructed adjacency matrix and the original adjacency matrix, and between the reconstructed node feature matrix and the original node feature matrix. The two types of errors are summed to obtain the reconstruction loss. The Adam optimizer is used to jointly iteratively update the parameters of the graph convolutional encoder and decoder networks with a learning rate of 0.001 until the reconstruction loss converges to below 0.01. After pre-training, the decoder network is discarded, and only the graph convolutional encoder is retained for subsequent feature extraction.

[0045] The set of frequency domain fluctuation amplitudes obtained in S121 is used as the initial feature vector of the corresponding node, and concatenated in node order to form a node feature matrix. The cooperative change relationship graph is input into the trained graph convolutional neural network, and the feature vectors corresponding to each node output by the hidden layer are extracted through forward propagation. The feature vectors of each node are arranged in node order to form the initial feature matrix.

[0046] The initial feature matrix undergoes dimensionality reduction mapping. Specifically, principal component analysis (PCA) can be used to perform dimensionality reduction mapping on the initial feature matrix. The covariance matrix of the initial feature matrix is ​​calculated, and eigenvalue decomposition is performed on the covariance matrix to obtain each eigenvalue and its corresponding eigenvector. The eigenvalues ​​are arranged in descending order, and the cumulative contribution rate of each eigenvalue is calculated. The cumulative contribution rate is the proportion of the sum of the first k eigenvalues ​​to the sum of all eigenvalues. The smallest k value that makes the cumulative contribution rate exceed a preset contribution rate threshold is selected as the target dimension, and the corresponding k eigenvectors are used to construct a projection matrix. The preset contribution rate threshold is determined based on the requirements for the degree of feature information retention in subsequent modeling, for example, set to 85%. The initial feature matrix is ​​multiplied by the projection matrix to obtain the dimensionality-reduced comprehensive feature matrix.

[0047] In step S13, the comprehensive feature matrix is ​​subjected to feature fusion processing to obtain a comprehensive feature set, including: S131, Perform orthogonal transformation on the comprehensive feature matrix to obtain the projection component sequence; S132, perform information entropy calculation on the projection component sequence to obtain the fusion weight value of each projection component; S133, determine whether the fusion weight value of each projection component is greater than a preset weight threshold, and retain the projection components that are greater than the preset weight threshold and their corresponding fusion weight values. S134, perform weighted summation on the corresponding projection components according to the retained fusion weight values ​​to obtain the fusion feature vector; S135, Perform feature reconstruction processing on the fused feature vector to determine the comprehensive feature set.

[0048] In step S131, the comprehensive feature matrix is ​​subjected to orthogonal transformation to obtain the projection component sequence.

[0049] The row vectors of the composite feature matrix correspond to features in different dimensions, and there is information redundancy between these dimensions. To eliminate the correlation between dimensions and extract the independent components of each feature in mutually orthogonal directions, the composite feature matrix undergoes orthogonal transformation. Specifically, the Schmitt orthogonalization algorithm is used to process the row vectors of the composite feature matrix. The row vectors of the composite feature matrix are used as an initial vector group, and a set of mutually orthogonal unit basis vectors is constructed using the Schmitt orthogonalization algorithm. The number of orthogonal basis vectors is the same as the number of rows in the composite feature matrix. Each row vector in the composite feature matrix is ​​then multiplied by its corresponding orthogonal basis vector, and the projection length of each row vector in each orthogonal basis direction is calculated. The projection lengths of each row vector in the same orthogonal basis direction are arranged in row order to form the projection component sequence corresponding to that orthogonal basis direction. All projection component sequences corresponding to orthogonal basis directions together form a projection component sequence group. The values ​​in the projection component sequence reflect the energy distribution of the original feature vectors in mutually orthogonal spatial dimensions.

[0050] In step S132, the information entropy of the projection component sequence is calculated to obtain the fusion weight value of each projection component.

[0051] Specifically, for each projected component sequence, the frequency of occurrence of each value in the sequence is statistically analyzed, and the probability of each value's occurrence is calculated. These probabilities are then substituted into the Shannon information entropy formula to calculate the information entropy value of the projected component sequence. A higher information entropy value indicates a more dispersed numerical distribution in the projected component sequence and contains more effective information. The information entropy values ​​of each projected component are arranged in descending order and normalized. The normalized information entropy values ​​are mapped to the range of 0 to 100 to obtain the fusion weight values.

[0052] In step S133, it is determined whether the fusion weight value of each projection component is greater than a preset weight threshold. Projection components with a weight greater than the preset weight threshold and their corresponding fusion weight values ​​are retained. The preset weight threshold is used to filter effective features. During the training phase, the system calculates the fusion weight values ​​of all projection components and selects the sum of the median of the value distribution and 0.5 times the interquartile range as the threshold. This calculation rule ensures that the threshold falls at the critical point between the distribution of noise-dominant components and effective physical information components. By eliminating projection components with weights lower than this threshold, the system retains effective physical features with high information content, ensuring the integrity of the physical information of the comprehensive feature set.

[0053] In step S134, the corresponding projection components are weighted and summed according to the retained fusion weight values ​​to obtain a fusion feature vector. Each retained projection component sequence is multiplied by its corresponding fusion weight value to obtain a weighted feature sequence for each projection component. All weighted feature sequences are then superimposed element-wise according to the order of the projection components to obtain a one-dimensional fusion feature vector. This fusion feature vector integrates the information of each effective projection component, and the contribution of each component is adjusted by a weight value determined by information entropy quantification, ensuring that components with greater information content occupy a higher proportion in the fusion result.

[0054] In step S135, the fused feature vector undergoes feature reconstruction processing to determine the comprehensive feature set. Nonlinear activation processing is then applied to the fused feature vector to enhance its nonlinear expressive power. Specifically, each element in the fused feature vector is input into a modified linear unit activation function, which evaluates the input elements. If the element value is greater than zero, it remains unchanged; otherwise, it is set to zero.

[0055] After the linear unit activation process is corrected, negative elements in the fused feature vector are set to zero, while positive elements are retained, enhancing the discriminative ability of the features while maintaining feature sparsity. The activated fused feature vector is then concatenated with the historical baseline feature vector. The historical baseline feature vector is a historical fused feature vector obtained by the embedded storage module collecting underlying data during the initial 72-hour standard health period after factory delivery, truncating it through a 30-day standard historical period window, using sampling parameters completely consistent with real-time data, and following the complete processing flow from S11 to S13. The two feature dimensions are naturally aligned and are pre-stored in the storage medium. The activated fused feature vector and the historical baseline feature vector are concatenated end-to-end to form the concatenated vector. The sampling parameters and processing steps of the historical baseline feature vector and the current fused feature vector are completely consistent, ensuring dimensional consistency and satisfying the preconditions for dimensional concatenation.

[0056] In step S14, a high-dimensional feature space is obtained by mapping the comprehensive feature set to a high-dimensional space. The boundary of the normal operating state of the high-dimensional feature space is fitted to determine the health benchmark model, including: S141, perform kernel function mapping based on the comprehensive feature set to obtain a high-dimensional feature space; S142, construct a hyperplane in the high-dimensional feature space to obtain initial classification interval data; S143, determine whether the initial classification interval data is less than the preset classification interval threshold. If so, introduce a penalty factor to optimize the initial classification interval data to obtain the maximized classification interval data. S144, Determine the normal data boundary based on the maximized classification interval data, and determine the health benchmark model.

[0057] In step S141, kernel function mapping is performed based on the comprehensive feature set to obtain a high-dimensional feature space.

[0058] The feature vectors in the comprehensive feature set may exhibit a non-linear distribution in the original space. A kernel function is used to map the comprehensive feature set from the original low-dimensional space to a high-dimensional feature space. The Gaussian radial basis function (RBF) kernel is selected as the mapping kernel function. Each feature vector in the comprehensive feature set is taken as input, and the kernel function calculates the inner product similarity between each feature vector, mapping the feature vectors in the original space to a high-dimensional Hilbert space, outputting the high-dimensional feature space. The width parameter of the Gaussian RBF kernel is determined based on the average distance between each feature vector in the comprehensive feature set. The average distance between all pairwise feature vectors in the comprehensive feature set is calculated as the scale parameter, and the specific value of the width parameter is set according to this scale parameter, so that the kernel function has an appropriate local response range in densely populated regions of data points.

[0059] In step S142, a hyperplane is constructed in the high-dimensional feature space to obtain initial classification interval data.

[0060] In a high-dimensional feature space, a hypersphere capable of enclosing all normal-state samples is sought, with the surface of this hypersphere serving as the boundary of the normal data distribution. A support vector data description algorithm is used to construct a hyperplane in the high-dimensional feature space. The mapping points of each feature vector in the comprehensive feature set in the high-dimensional feature space are used as training samples. An optimization problem minimizing the hypersphere's volume is solved to obtain the hypersphere's center position and radius. The distance from each training sample to the center is calculated, and the maximum value is taken as the initial classification interval data, expressed as a dimensionless relative distance. In one embodiment, the comprehensive feature set contains multiple samples under normal operating conditions; after calculation by the support vector data description algorithm, the initial classification interval data is 15.

[0061] In step S143, it is determined whether the initial classification interval data is less than the preset classification interval threshold. If so, a penalty factor is introduced to optimize the initial classification interval data to obtain the maximized classification interval data.

[0062] A preset classification interval threshold is used to determine whether the currently constructed hyperplane can adequately cover the distribution range of the normal operating state. The preset classification interval threshold is determined by continuously collecting 72 hours of underlying operating data during the initial stage of operation after the embedded storage module passes factory inspection. This data is considered as standard healthy state samples. Multiple sets of comprehensive features are obtained through steps S11 to S13. The initial classification interval data corresponding to each set of comprehensive features is calculated, and the average of all calculation results is taken as the baseline interval value. The baseline interval value is multiplied by a preset amplification factor to obtain the classification interval threshold. The preset amplification factor is determined by the compactness of the distribution of normal operating state samples in the feature space. The system calculates the distance distribution from all normal samples in the health baseline training set to the center of the hypersphere, and sets the ratio of the 95th quantile of this distance distribution to the baseline interval value as the preset amplification factor. This factor characterizes the dispersion of the normal operating state samples after boundary fitting. The more dispersed the distribution of normal samples, the larger the amplification factor. This dynamic adjustment mechanism ensures that the health baseline model maintains a consistent boundary relaxation and anomaly detection sensitivity in different application scenarios.

[0063] The calculated initial classification margin data is compared with the preset classification margin threshold. If the initial classification margin data is greater than or equal to the preset classification margin threshold, it indicates that the current hyperplane can sufficiently cover the normal state distribution range, and no optimization adjustment is needed; the initial classification margin data is directly used as the maximum classification margin data. If the initial classification margin data is less than the preset classification margin threshold, it indicates that the current hyperplane is too compact, and a penalty factor needs to be introduced for relaxation optimization.

[0064] The penalty factor is introduced by adding a slack variable to the optimization objective of the support vector data description algorithm, allowing some training samples to lie outside the hypersphere. A penalty term is added to the objective function to constrain the number of samples outside the hypersphere. The penalty factor is a weight coefficient of the penalty term, used to balance the weight relationship between minimizing the hypersphere volume and minimizing the number of samples outside the boundary. The penalty factor is the reciprocal of the total number of training samples; that is, the larger the number of training samples, the smaller the influence of each sample on the hypersphere boundary, and the smaller the penalty factor; the smaller the number of training samples, the greater the influence of each sample, and the larger the penalty factor. The constrained optimization problem is transformed into an unconstrained dual problem using the Lagrange multiplier method. The penalty factor is substituted into the objective function of the dual problem, and iterative optimization is performed using a quadratic programming algorithm. While tolerating a small amount of edge noise data, the center position and radius of the hypersphere are recalculated, and the optimized radius value is used as the data to maximize the classification margin.

[0065] In step S144, the normal data boundary is divided according to the maximized classification interval data, and the health benchmark model is determined.

[0066] Using the optimized hypersphere's center as the center and the data with the maximized classification margin as the radius, the hypersphere's surface is determined as the normal data boundary in the high-dimensional feature space. The region inside the hypersphere is defined as the normal operating state space, and the region outside the hypersphere is defined as the abnormal state space. The hypersphere's center position and radius value together constitute the health baseline model.

[0067] In step S15, real-time underlying data of the embedded storage module is obtained, and spatial distance calculation is performed between the real-time underlying data and the health benchmark model to obtain a state deviation value. If the state deviation value is greater than a preset deviation threshold, it is determined that a cooperative distortion has occurred, including: S151, Obtain the real-time underlying data of the embedded storage module, perform time-frequency domain conversion on the real-time underlying data, and obtain the real-time feature vector; S152, The real-time feature vector is input into the health benchmark model for spatial projection to determine the mapping coordinates of the real-time feature vector in high-dimensional space; S153, calculate the algebraic deviation of the mapped coordinates from the normal data boundary in the health baseline model; S154, Perform normalization mapping quantization processing on the algebraic deviation to obtain the state deviation value representing the degree of deviation; S155, determine whether the state deviation value is greater than a preset deviation threshold; if so, determine that the real-time underlying data has undergone collaborative distortion.

[0068] In step S151, the real-time underlying data of the embedded storage module is obtained, and the real-time underlying data is converted into a time-frequency domain to obtain a real-time feature vector.

[0069] The system acquires real-time low-level operational data from the embedded storage module. This real-time low-level data shares the same physical quantity type and sampling frequency as the historical low-level data. A continuous time-domain data segment of a preset length is extracted from the real-time low-level data as an analysis window. The length of the analysis window is determined based on the sampling frequency and the frequency resolution required for subsequent frequency-domain analysis. A time-frequency domain transformation is performed on the continuous time-domain data segment to convert the time-domain signal into a frequency-domain signal. Specifically, a Fast Fourier Transform is performed on the continuous time-domain data segment to obtain the amplitude spectrum of the signal in the frequency domain. Frequency-domain feature values ​​within a preset frequency band are extracted from the amplitude spectrum and arranged in frequency order to form a real-time feature vector.

[0070] In step S152, the real-time feature vector is input into the health benchmark model for spatial projection to determine the mapping coordinates of the real-time feature vector in high-dimensional space.

[0071] During training, the health baseline model establishes a projection mapping relationship from the original feature space to the high-dimensional feature space. This projection mapping relationship is determined by the parameters of the kernel function and the center position of the hypersphere. Using the real-time feature vector as input, and following the same kernel function mapping method as in step S141, the real-time feature vector is mapped from the original low-dimensional space to the high-dimensional feature space. The spatial coordinates of the real-time feature vector in the high-dimensional feature space are determined, and these spatial coordinates are used as the mapped coordinates. These mapped coordinates represent the current operating state's position in the high-dimensional baseline feature space.

[0072] In step S153, the algebraic deviation of the mapped coordinates from the normal data boundary in the health baseline model is calculated.

[0073] The health baseline model includes normal data boundaries represented by a hypersphere, which has a defined center position and radius. The system calculates the geometric distance of the mapped coordinates relative to the center of the hypersphere and defines the difference between the geometric distance and the hypersphere radius as the algebraic deviation.

[0074] The algebraic deviation is used to quantitatively assess the degree to which the operating state deviates from the normal boundary. When the mapped coordinates are inside the hypersphere, the algebraic deviation is negative or zero, indicating that the operating state is within the normal data boundary; when the mapped coordinates are outside the hypersphere, the algebraic deviation is positive, indicating that the operating state has exceeded the normal data boundary. The algebraic deviation serves as input data for subsequent numerical calculations of the state deviation.

[0075] In step S154, normalization mapping quantization processing is performed on the algebraic deviation to obtain a state deviation value that characterizes the degree of deviation.

[0076] Multiple sets of real-time data samples of historical normal operation status are acquired, the algebraic deviation of each sample is calculated and statistically distributed, the mean of the distribution is set as the distance benchmark, and the sum of the distance benchmark and three times the standard deviation of the algebraic deviation distribution of normal operation status is used as the preset distance threshold.

[0077] During quantization, the statistical extreme value in the algebraic deviation distribution of normal operating conditions is selected as the scaling benchmark, which is determined to be the difference between the statistical extreme value and a preset distance threshold. During calculation, if the actual algebraic deviation is greater than the preset distance threshold, the excess amount between the algebraic deviation and the preset distance threshold is calculated, and the ratio of the excess amount to the scaling benchmark is linearly scaled to a numerical range of 0 to 100 to obtain the state deviation value; if the actual algebraic deviation is less than or equal to the preset distance threshold, the state deviation value is set to zero.

[0078] In step S155, it is determined whether the state deviation value is greater than a preset deviation threshold. If so, it is determined that the real-time underlying data has undergone collaborative distortion.

[0079] The preset deviation threshold is used to determine whether the current operating state meets the triggering conditions for collaborative distortion. The preset deviation threshold is a value within a percentage scoring range. According to the industry-standard health scoring system, a score below 60 is usually considered abnormal, so the preset deviation threshold is set to 50.

[0080] A time-series fluctuation sequence of real-time underlying data within a preset time window is extracted. The length of the time-series fluctuation sequence is determined based on the dynamic response characteristics of the monitored object. In one embodiment, a time-series fluctuation sequence containing 2048 sampling points is extracted. The time-series fluctuation sequence is reconstructed into a phase space, embedding the one-dimensional time-series sequence into a high-dimensional phase space to obtain a distortion feature matrix containing multi-dimensional state vectors. Based on the distortion feature matrix, the local discrete distribution density of each state vector in the multi-dimensional space is calculated. A local anomaly factor algorithm is used to calculate the local anomaly factor of each state vector. The threshold of the local anomaly factor is set based on the distribution of local anomaly factors of normally operating samples during the training phase, and the 95th quantile of this distribution is selected as the preset factor threshold. The neighborhood size parameter in the local anomaly factor algorithm is set to the logarithm of the total number of samples in the current time window, rounded to the nearest integer. If the calculated local anomaly factor is greater than the preset factor threshold, a secondary confirmation is performed to determine that cooperative distortion has occurred, and the process proceeds to step S16 to perform a tertiary screening.

[0081] In step S16, a multi-dimensional feature space outlier assessment is performed on the real-time underlying data where cooperative distortion is determined to have occurred to obtain an anomaly score and a dynamic limit value. If the anomaly score is greater than the dynamic limit value, a module fault warning signal is generated, including: S161, calculate the local distribution density of the real-time underlying data that has undergone cooperative distortion in the feature space, and calculate the outlier degree of each data sample based on the local distribution density and the feature distribution benchmark under normal operation. S162, based on the outlier degree, deduce the abnormal score value of each data sample, determine whether the abnormal score value is greater than the preset high-risk threshold, and if so, extract the corresponding data sample as a high-risk sample set. S163, calculate the dynamic limit value based on the distribution variance of the abnormal score values ​​of the samples in the high-risk sample set, determine whether the abnormal score value is greater than the dynamic limit value, and if so, generate a module fault warning signal.

[0082] In step S161, the local distribution density of the real-time underlying data that has been determined to have undergone cooperative distortion in the feature space is calculated, and the outlier degree of each data sample is calculated based on the local distribution density and the feature distribution benchmark under normal operation.

[0083] Data sample points identified as having co-distorted characteristics are mapped to a high-dimensional feature space. A neighborhood search algorithm based on sample points is used to calculate the local neighborhood density of each data sample point in the feature space. The feature distribution benchmark under normal operating conditions is constructed from the sample density distribution of the healthy benchmark model during the training phase. The system quantifies the outlier degree of a sample point deviating from the normal region by comparing the local neighborhood density of real-time data sample points with the feature distribution benchmark. The larger the outlier degree value, the sparser the distribution of the sample in the feature space and the more it deviates from the dense region. Specifically, the feature distribution benchmark consists of the mean center and covariance matrix of all normal samples in the feature space during the training phase. Together, they constitute a multidimensional Gaussian distribution model, serving as a comparison benchmark for judging the degree of deviation of real-time samples.

[0084] In step S162, the outlier score of each data sample is derived based on the outlier degree, and it is determined whether the outlier score is greater than a preset high-risk threshold. If so, the corresponding data sample is extracted into a high-risk sample set.

[0085] The outlier score of each sample point is linearly normalized to distribute its value within the range of 0 to 100, resulting in an anomaly score. The upper bound of the normalization is three times the maximum outlier score of normal samples during the training phase, and the lower bound is the minimum outlier score of normal samples during the training phase, ensuring that the mapped numerical distribution reasonably covers the normal and anomaly ranges. The system compares the anomaly score of each data sample with a preset high-risk threshold. If the anomaly score is greater than the preset high-risk threshold, the sample is classified into the high-risk sample set.

[0086] A preset high-risk threshold is used to further filter samples with more prominent anomalies from those identified as having cooperative distortion. The preset high-risk threshold is determined based on the distribution characteristics of anomaly scores within the range of 0 to 100. In one embodiment, referencing industry-standard anomaly detection scoring systems, 75 points is typically used as the distinction between significant and general anomalies, and the preset high-risk threshold is set to 75. The anomaly score of each data sample is compared with the preset high-risk threshold. If the anomaly score is greater than the preset high-risk threshold, the data sample is included in the high-risk sample set; if the anomaly score is less than or equal to the preset high-risk threshold, it is excluded.

[0087] In step S163, a dynamic limit value is calculated based on the distribution variance of the abnormal score values ​​of the samples in the high-risk sample set, and it is determined whether the abnormal score value is greater than the dynamic limit value. If so, a module fault warning signal is generated.

[0088] The system calculates the variance of the abnormal scores for all samples in the high-risk sample set, reflecting the discrete nature of the anomaly severity. The system adds this variance to the mean of the abnormal scores calculated relative to the characteristic distribution benchmark for all normal samples during the training phase, obtaining a dynamic threshold value. This dynamic threshold value adaptively adjusts according to the distribution characteristics of the high-risk sample set. The more dispersed the sample scores, the higher the dynamic threshold value becomes to suppress false alarms caused by data fluctuations; conversely, the more concentrated the sample scores, the lower the dynamic threshold value becomes to improve warning sensitivity. If the abnormal score value of a sample exceeds the dynamic threshold value, the sample is identified as a significant outlier, and a module fault warning signal is generated.

[0089] Furthermore, before generating the module fault warning signal, a step of physical-level confirmation of high-risk samples is included. Specifically, the storage medium status parameters of the embedded storage module corresponding to the real-time underlying data are extracted. These storage medium status parameters include read / write latency characteristics and erase / write balance status. The read / write latency characteristics are quantified by the average time taken for the storage chip to execute page programming instructions. The rated value of this average time is given by the storage chip datasheet. In one embodiment, the rated value is 200 microseconds. An actual average time exceeding 50% of the rated value is considered an abnormal read / write latency. The erase / write balance status is quantified by the difference between the maximum and minimum erase counts in the storage block. The normal range of this difference is determined according to the industry-standard wear leveling algorithm for storage media. In one embodiment, a difference of less than 100 counts is considered a normal balance status, and a difference of 100 counts or more is considered an abnormal balance status. If at least one of the read / write latency characteristics and erase / write balance status is determined to be abnormal, a module fault warning signal is generated. If both are normal, a module fault warning signal is not generated temporarily to avoid false alarms based on a single dimension of data distribution characteristics.

[0090] In summary, this invention discloses a fault early warning method for embedded storage modules based on a vertical connector, comprising: acquiring and preprocessing historical underlying data to construct a health baseline model; calculating the algebraic deviation between real-time underlying data and the model, and determining collaborative distortion based on the state deviation value; and performing fault early warning judgment by utilizing a multi-dimensional feature space outlier evaluation mechanism combined with local distribution density and variance adaptive verification. This invention, through multi-dimensional collaborative relationship modeling and a dynamic adaptive distribution feature mechanism, achieves accurate capture and effective identification of early collaborative degradation signals in complex embedded storage modules, significantly improving the sensitivity and reliability of fault early warning.

[0091] It should be noted that the embedded storage module fault early warning system based on a vertical connector provided in this embodiment of the invention is used to execute all the process steps of the embedded storage module fault early warning method based on a vertical connector in the above embodiment. The working principle and beneficial effect of the two are one-to-one, so they will not be described again.

[0092] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0093] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A fault early warning method for embedded storage modules based on a vertical connector, characterized in that, include: Obtain the underlying data of the embedded storage module during its historical operating cycle, and preprocess the underlying data to obtain the time series sequence of the reference physical quantity; Based on the time series sequence of the benchmark physical quantity, physical quantity sequences of multiple dimensions are separated. Based on the physical quantity sequences, a cooperative change relationship map is constructed, and the cooperative change relationship map is subjected to feature projection processing to obtain a comprehensive feature matrix characterizing the operating state. The comprehensive feature matrix is ​​subjected to feature fusion processing to obtain a comprehensive feature set; A high-dimensional feature space is obtained by mapping the comprehensive feature set to a high-dimensional space. The boundary of the normal operating state of the high-dimensional feature space is fitted to determine the health benchmark model. The real-time underlying data of the embedded storage module is obtained, and the spatial distance between the real-time underlying data and the health benchmark model is calculated to obtain the state deviation value. If the state deviation value is greater than the preset deviation threshold, it is determined that a cooperative distortion has occurred. The real-time underlying data that is determined to have undergone collaborative distortion is subjected to outlier evaluation in the multidimensional feature space to obtain an anomaly score value and a dynamic limit value. If the anomaly score value is greater than the dynamic limit value, a module fault warning signal is generated.

2. The method for early warning of embedded storage module faults based on a vertical connector according to claim 1, characterized in that, The process of acquiring the underlying data of the embedded storage module during its historical operating cycle and preprocessing the underlying data to obtain the time series sequence of the reference physical quantity includes: Obtain the underlying data of the embedded storage module during its historical operating cycle, and extract the timestamp and sampling frequency from the underlying data; The underlying data is sequence aligned based on the timestamp and the sampling frequency to obtain an initial state sequence with missing positions. If the continuous span of the vacancy position is less than the maximum allowable continuous vacancy threshold, then interpolation calculation is performed on the vacancy position to determine the filling value; The missing value is filled into the initial state sequence to perform a missing value filling operation, thereby obtaining the time series sequence of the reference physical quantity.

3. The method for early warning of embedded storage module faults based on a vertical connector according to claim 2, characterized in that, The step of interpolating and calculating the filling value for the vacancy includes: The missing positions are calculated using a nearest neighbor interpolation algorithm, a linear interpolation algorithm, or a spline interpolation algorithm to determine the filling value.

4. The method for early warning of embedded storage module faults based on a vertical connector according to claim 1, characterized in that, The process involves separating multiple dimensions of physical quantity sequences based on the time-series sequence of the baseline physical quantities, constructing a cooperative change relationship map based on the physical quantity sequences, and performing feature projection processing on the cooperative change relationship map to obtain a comprehensive feature matrix characterizing the operating state, including: Based on the time series sequence of the reference physical quantity, multiple independent dimensions of the initial physical quantity sequence are separated. The initial physical quantity sequence of each dimension is then subjected to frequency domain transformation to obtain the set of frequency domain fluctuation amplitudes for each dimension. Calculate the coupling weights between different dimensions based on the frequency domain fluctuation amplitude set, and generate an initial correlation matrix based on the coupling weights; Establish connection edges between the dimensions corresponding to the elements in the initial association matrix that are greater than the preset collaboration threshold, and construct a collaboration change relationship graph based on the selected nodes and connection edges; The collaborative change relationship map is subjected to spatial dimensionality reduction mapping to obtain a comprehensive feature matrix.

5. The method for early warning of embedded storage module faults based on a vertical connector according to claim 1, characterized in that, The feature fusion processing of the comprehensive feature matrix yields a comprehensive feature set, including: The comprehensive feature matrix is ​​subjected to orthogonal transformation to obtain the projection component sequence; The information entropy of the projection component sequence is calculated to obtain the fusion weight value of each projection component; Determine whether the fusion weight value of each projection component is greater than a preset weight threshold, and retain the projection components that are greater than the preset weight threshold and their corresponding fusion weight values; The corresponding projection components are weighted and summed according to the retained fusion weight values ​​to obtain the fusion feature vector; The fused feature vector is subjected to feature reconstruction processing to determine the comprehensive feature set.

6. The method for early warning of embedded storage module faults based on a vertical connector according to claim 1, characterized in that, The process of mapping a high-dimensional space based on the comprehensive feature set to obtain a high-dimensional feature space, fitting the boundary of the normal operating state of the high-dimensional feature space, and determining the health benchmark model includes: Based on the comprehensive feature set, a kernel function mapping is performed to obtain a high-dimensional feature space; A hyperplane is constructed in the high-dimensional feature space to obtain initial classification interval data; Determine whether the initial classification interval data is less than a preset classification interval threshold. If so, introduce a penalty factor to optimize the initial classification interval data and obtain the maximum classification interval data. The normal data boundary is defined based on the data with the maximum classification interval, and the health benchmark model is determined.

7. The method for early warning of embedded storage module faults based on a vertical connector according to claim 1, characterized in that, The process involves acquiring real-time low-level data from the embedded storage module, calculating the spatial distance between the real-time low-level data and the health benchmark model to obtain a state deviation value, and determining that a cooperative distortion has occurred if the state deviation value is greater than a preset deviation threshold. This includes: The real-time underlying data of the embedded storage module is obtained, and the real-time underlying data is converted into the time-frequency domain to obtain the real-time feature vector. The real-time feature vector is input into the health benchmark model for spatial projection to determine the mapping coordinates of the real-time feature vector in high-dimensional space. Calculate the algebraic deviation of the mapped coordinates from the normal data boundary in the health baseline model; The algebraic deviation is subjected to normalized mapping quantization to obtain a state deviation value that represents the degree of deviation. Determine whether the state deviation value is greater than a preset deviation threshold; if so, determine that the real-time underlying data has undergone collaborative distortion.

8. The method for early warning of embedded storage module faults based on a vertical connector according to claim 1, characterized in that, The real-time underlying data that is determined to have undergone cooperative distortion is subjected to multi-dimensional feature space outlier evaluation to obtain an anomaly score value and a dynamic limit value. If the anomaly score value is greater than the dynamic limit value, a module fault warning signal is generated, including: Calculate the local distribution density of the real-time underlying data that has been determined to have undergone cooperative distortion in the feature space, and calculate the outlier degree of each data sample based on the local distribution density and the feature distribution benchmark under normal operation. Based on the outlier degree, the abnormal score value of each data sample is derived, and it is determined whether the abnormal score value is greater than the preset high-risk threshold. If so, the corresponding data sample is extracted as a high-risk sample set. Calculate the dynamic limit value based on the distribution variance of the abnormal score values ​​of the samples in the high-risk sample set, and determine whether the abnormal score value is greater than the dynamic limit value. If so, generate a module fault warning signal.

9. A fault early warning system for an embedded storage module based on a vertical connector, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the method described in any one of claims 1 to 8.

Images