A fault injection method for integrated electrical cabinet system

Abstract

The application discloses a fault injection method of an integrated electric cabinet system, and relates to the technical field of faults, which comprises the following steps: obtaining a fault injection instruction, and parsing a target storage unit identifier, a fault mode parameter and an application environment parameter of the electric cabinet system from the fault injection instruction; constructing extreme working condition simulation of the target storage unit based on the application environment parameter; generating a fault factor combination according to the fault mode parameter and the extreme working condition simulation; generating a dynamic adaptive fault atlas based on the fault factor combination; and performing a fault injection operation on a target storage unit indicated by the target storage unit identifier and collecting characteristic data of the electric cabinet system in operation based on the dynamic adaptive fault atlas. Through the cooperation of each step, the application effectively alleviates the problems in the prior art, such as disconnection between a fault injection scene and reality, and inability to reflect dynamic propagation of faults, and improves the authenticity and effectiveness of fault injection.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of faults, in particular to a fault injection method of an integrated electrical cabinet system. BACKGROUND

[0002] The integrated electrical cabinet system is widely used in key fields such as industrial control, data center, energy management, and the like. The storage unit as a core data carrier may cause serious consequences such as system paralysis and data loss once a fault occurs. In order to ensure the stable operation of the system in a complex environment, it is necessary to simulate various potential faults through fault injection technology to verify the fault tolerance capability and fault handling mechanism of the system, and therefore, accurate and efficient fault injection on the storage unit becomes a key link to improve the reliability of the electrical cabinet system.

[0003] At present, the fault injection for the storage unit of the electrical cabinet system mostly adopts a static injection method: fixed fault types (such as data bit flip, read / write timeout) and injection parameters (such as injection time, influence range) are preset, a single storage unit is independently simulated for fault in a laboratory environment, and system response data is recorded through a monitoring device. This scheme relies on manual presetting of fault scenarios, the injection process is isolated from the actual running environment of the electrical cabinet, and only the fault performance of the storage unit in an ideal state can be verified.

[0004] In the existing scheme, the application environment differences (such as extreme working conditions such as high temperature, high humidity, and electromagnetic interference) of the electrical cabinet system and the dynamic propagation characteristics of the fault among multiple units are not considered, the injected fault scenarios are inconsistent with the actual fault evolution law in the running, and the simulation results cannot truly reflect the fault tolerance capability of the system in complex working conditions, which makes it difficult to effectively guide the reliability design and optimization of the electrical cabinet system. SUMMARY

[0005] In order to solve the above technical problems, the present application provides a fault injection method of an integrated electrical cabinet system to at least alleviate the above technical problems.

[0006] The technical scheme provided by the embodiments of the present application is as follows:

[0007] A fault injection method of an integrated electrical cabinet system, the method comprising:

[0008] Step 1, obtaining a fault injection instruction to parse a target storage unit identifier, a fault mode parameter, and an application environment parameter of the electrical cabinet system therefrom;

[0009] Step 2, based on the application environment parameter, constructing an extreme working condition simulation of the target storage unit;

[0010] Step 3, generating a fault factor combination according to the fault mode parameter and the extreme working condition simulation;

[0011] Step 4, generating a dynamic adaptive fault atlas based on the fault factor combination;

[0012] Step 5, performing a fault injection operation on the target storage unit indicated by the target storage unit identifier based on the dynamic adaptive fault atlas and collecting feature data of the electrical cabinet system in operation.

[0013] In the technical solution of the present application, by obtaining the fault injection instruction and parsing the target storage unit identifier, fault mode parameters and application environment parameters, the problem that the existing scheme does not consider the application environment difference of the electrical cabinet system is solved. By clearly defining the application environment parameters, a basis is provided for subsequent construction of actual working condition simulation, so that the fault injection is no longer isolated from the actual operating environment. In addition, based on the application environment parameters, extreme working condition simulation is constructed, which directly addresses the defect that the existing scheme does not involve extreme working conditions. By simulating extreme environments such as high temperature and high humidity, the fault injection scenario is closer to the complex environment in the actual operation of the electrical cabinet, avoiding the distortion of the simulation results due to the lack of environmental factors. Furthermore, the fault factor combination is generated according to the fault mode parameters and the extreme working condition simulation, which combines the fault mode with the actual working condition, solving the problem that the fault scene in the existing scheme is single and disconnected with the actual situation. The fault injection is no longer isolated fault type simulation, but a comprehensive fault scene considering the influence of working conditions. Moreover, based on the fault factor combination, a dynamic adaptive fault atlas is generated, which addresses the problem that the existing scheme does not consider the dynamic propagation characteristics of faults among multiple units. The atlas can reflect the dynamic evolution and correlation of faults, so that the fault injection can reflect the propagation law of faults, rather than static single fault simulation. Finally, based on the dynamic adaptive fault atlas, the fault injection is performed and feature data is collected, ensuring that the fault injection operation is performed based on the consideration of dynamic fault propagation and actual working conditions. The collected feature data can truly reflect the response of the system under complex working conditions, solving the problem that the simulation results of the existing scheme cannot truly reflect the fault tolerance capability of the system, and providing an effective basis for the reliability design and optimization of the electrical cabinet system. BRIEF DESCRIPTION OF DRAWINGS

[0014] Figure 1 FIG. 1 is a flowchart of a fault injection method for an integrated electrical cabinet system according to an embodiment of the present application. DETAILED DESCRIPTION

[0015] As shown in FIG. 1, the present application provides a fault injection method for an integrated electrical cabinet system, which comprises: Figure 1

[0016] Step 1, obtaining a fault injection instruction to parse a target storage unit identifier, fault mode parameters and application environment parameters of an electrical cabinet system therefrom;

[0017] Step 2, based on the application environment parameters, constructing an extreme working condition simulation of the target storage unit; ​

[0018] Step 3, generating a fault factor combination according to the fault mode parameters and extreme working condition simulation;

[0019] Step 4, generating a dynamic adaptive fault atlas based on the fault factor combination;

[0020] Step 5, based on the dynamic adaptive fault atlas, performing a fault injection operation on the target storage unit indicated by the target storage unit identifier and collecting feature data of the electrical cabinet system during operation.

[0021] Optionally, the step 1 specifically comprises:

[0022] Step 11: analyzing the data packet structure of the fault injection instruction to obtain a multi-level nested data frame;

[0023] Step 12, tracing the storage unit field in the multi-level nested data frame to determine the target storage unit identifier;

[0024] Step 13, deconstructing the fault feature field of the multi-level nested data frame to determine the fault mode parameters;

[0025] Step 14: deduce the environment parameter field of the multi-level nested data frame to determine the application environment parameters of the electrical cabinet system.

[0026] Optionally, the step 11: analyzing the data packet structure of the fault injection instruction to obtain a multi-level nested data frame, specifically comprises the following steps:

[0027] Step 111: analyzing the data packet structure of the fault injection instruction to extract the data packet frame header identifier and frame tail identifier to determine the boundary range of the data frame;

[0028] Step 112: based on the boundary range of the data frame, performing nested level unpacking processing on the fault injection instruction to obtain a main frame structure and a sub-frame structure;

[0029] Step 113: performing field index anchoring processing on the main frame structure and the sub-frame structure according to the hierarchical relationship to form a multi-level nested data frame.

[0030] Specifically, in step 111, the accurate segmentation of data packets is realized by a dynamic boundary identification mechanism, breaking through the limitations of traditional fixed-length frame analysis. Specifically, first, a preset frame header identifier (such as a 16-bit synchronization code 0xAA55) is extracted from the binary data stream, but it is not simply dependent on the fixed code value, but rather the length of the identifier is automatically adapted in combination with the type field after the frame header (such as a 2-byte identifier for control frames and a 4-byte extended identifier for data frames). For the frame tail identifier, a dual verification mechanism of "CRC check + length redundancy check" is used: first, the check value of the data in the frame is calculated by the CRC16 algorithm, and the check result stored in the frame tail is compared; at the same time, the length field declared by the frame header is verified whether it is consistent with the actual number of bytes parsed (allowing ±1 byte error). This dual verification mechanism can effectively distinguish between the real frame tail and the pseudo identifier that occasionally appears in the data stream (such as the case where the data stream happens to contain 0xAA55), making the boundary identification accuracy above 99.9%, providing an absolutely reliable spatial boundary for subsequent hierarchical analysis. The core principle is to construct a unique "digital fingerprint" of the data frame through the coordinated verification of multiple dimensions (identifier code, length, check value), realizing anti-interference analysis in complex electromagnetic environments.

[0031] Specifically, in step 112, the intelligent splitting of nested data is realized based on an adaptive hierarchical unpacking algorithm, solving the problem of traditional fixed hierarchical analysis that cannot adapt to dynamic instruction structures. The technical essence of this algorithm is to simulate the "whole-part" decomposition logic of human cognition: first, when analyzing the main frame structure, the "level depth field" in the main frame header is used to predict the number of subframe nesting layers (such as 0x02 indicating the existence of 2 layers of subframes), and then the subframe offset table (storing the starting address and length of each subframe in the main frame) is used for positioning. For subframes with multiple layers of nesting (such as sub-subframes contained in the subframe), the algorithm uses a recursive call mechanism to automatically treat the current subframe as a new analysis object and repeat the above process until all hierarchical analysis is complete. Compared with traditional static analysis, the key improvement is the introduction of a "dynamic field mapping table" - through the field descriptor (Field Descriptor) in the main frame, the field type of the subframe (such as address segment, parameter segment, and check segment) is identified in real time, without the need for a preset analysis template to adapt to different manufacturers' instruction formats. For example, when analyzing the instructions of an unknown model electric cabinet, the algorithm can automatically analyze the parameter values through the data type identifier in the field descriptor (such as 0x03 indicating a 32-bit integer), making the compatibility improved by more than 80%, and the principle is similar to the dynamic type identification of general programming languages, realizing the adaptive adjustment of analysis logic through metadata driving.

[0032] Specifically, in step 113, a semantic network of hierarchical data is constructed by a field association engine to convert scattered field data into a "data organism" with logical association. The core principle of the engine is to establish an association relationship based on a "parent-child field mapping rule": the association between a master frame and a sub-frame is achieved by "parent frame ID + sub-frame type" (for example, the control instruction with master frame ID 0x01 must have a storage unit parameter with sub-frame type 0x0A); the association between fields at the same level is defined by "offset + dependency marker" (for example, the "voltage parameter" field depends on the calibration value of the "temperature parameter" field, so the temperature value needs to be obtained first before calculating the voltage correction amount). In specific implementation, the algorithm generates an "association relationship table" that records the "upstream dependent fields", "downstream affected fields", and "calculation rules" of each field (for example, when the temperature increases by 10°C, the voltage threshold decreases by 0.2V). Such association is not only a mapping of physical addresses, but also a binding of logical semantics - for example, the "storage unit address" field in the sub-frame will automatically associate with the "device type" field of the master frame through the association table. When the device type is "solid state storage", the address resolution automatically switches to the block-page addressing mode of the flash chip; when it is "mechanical hard disk", it switches to the cylinder-sector mode. The multi-level nested data frame formed finally is essentially a "live data structure" containing logical operation rules, providing plug-and-play semantic support for parameter extraction in subsequent steps, avoiding semantic ambiguity caused by field isolation in traditional parsing.

[0033] Optionally, in step 12, the storage unit field in the multi-level nested data frame is traced to determine the target storage unit identifier, specifically including the following steps:

[0034] Step 121: Call the cross-level field index table of the electrical cabinet system to obtain the index code of the storage unit field from the table;

[0035] Step 122: Based on the index code of the storage unit field, locate the storage unit field in the multi-level nested data frame and extract the distributed address information in the field, including the logical address segment, the board code segment, and the chip identifier segment;

[0036] Step 123: Perform address mapping conversion on the logical address segment to convert the logical address into a physical address base;

[0037] Step 124: Analyze the board code segment and the chip identifier segment to obtain the physical position code of the board and the array coordinates of the chip on the board;

[0038] Step 125: Concatenate the physical address base, the board physical position code, and the array coordinates to generate the target storage unit identifier.

[0039] Preferably, in step 121, a "navigation system" for establishing cross-level data retrieval is implemented, which quickly locates the position of the storage unit field in the multi-level nested data frame by calling the cross-level field index table pre-constructed by the cabinet system. This index table is not a simple address list, but a multi-dimensional mapping structure that integrates field type, hierarchical relationship, and data length. The index code of the storage unit field uses a composite coding method of "level identifier + field type ID" (e.g., "02-0A" represents the storage unit field of the 2nd level). Its core function is to solve the positioning problem caused by the scattered storage of fields in multi-level nested data, similar to the index mechanism in a database. Through the index code, the target field can be directly jumped to the sub-frame offset address, avoiding inefficient operations such as frame-by-frame traversal, and making the positioning efficiency reach O(1) level, providing accurate "data coordinates" for subsequent field extraction.

[0040] Preferably, in step 122, the accurate extraction and structured analysis of the storage unit field are realized based on the index code. Its innovation lies in converting continuous binary data into distributed address components with clear semantics. Based on the index code obtained in step 121, the starting byte of the storage unit field is located through field offset calculation, and then the "field separator + length identifier" parsing method is used to split out the logical address segment (such as 32-bit virtual address), board code segment (such as 16-bit geographic location code), and chip identification segment (such as 8-bit array number) from the continuous data stream. Here, the "structured data parsing algorithm" is used, which automatically identifies the boundaries of each segment through pre-set field format descriptors (such as bit width and coding format of each address segment). Even if the field length changes dynamically (such as different board coding bit numbers), the algorithm can still adaptively adjust the parsing range through length identification, ensuring accurate extraction of the core components constituting the storage unit address from complex nested data.

[0041] Preferably, in step 123, the mapping and conversion of logical address to physical address are realized to solve the correspondence problem between virtual address and actual hardware address. The core is to complete the address space conversion through the "address mapping algorithm". The essence of this algorithm is based on the memory management table of the cabinet system (containing the mapping relationship between logical address range and physical address block), which uses a segmented mapping mechanism: first, the logical address is split into segment selection sub and segment offset, then the corresponding physical segment base address is found in the memory management table through the segment selection sub, and finally the segment offset is added to the physical segment base address to get the physical address base. Unlike traditional linear mapping, this algorithm introduces a "dynamic relocation" mechanism. When the storage unit changes its physical location due to hot plug or fault switching, the corresponding relationship in the mapping table can be automatically updated, ensuring that the logical address always points to the correct physical storage unit, providing a reliable base address for subsequent physical location positioning.

[0042] The essence of step 124 is the decoding of hardware physical location information and the conversion of spatial coordinates, which converts abstract encoded data into locatable physical location parameters. When parsing the card encoding segment, the "geocoding decoding algorithm" is used, which converts the 16-bit card encoding into a three-dimensional physical location encoding of "cabinet-drawer-slot" (such as "03-02-05" representing the 5th slot of the 2nd drawer of the 3rd cabinet) through preset encoding rules (such as the first 8 bits representing the cabinet number and the last 8 bits representing the slot position). When parsing the chip identification segment, the "array coordinate conversion algorithm" is used to convert the chip identification (such as 8-bit binary code) into a two-dimensional array coordinate of the chip on the card according to the row-first or column-first rule (such as "04-07" representing the chip in the 7th column of the 4th row). The essence of these two parsing processes is to reversely convert the encoding information fixed during hardware production into a spatial coordinate system that can be used for physical positioning, realizing the mapping from data encoding to physical location.

[0043] Preferably, in step 125, the complete storage unit unique identifier is constructed to realize the full-chain association from logical address to physical location, and the core is to generate a globally unique identification code through the "multi-dimensional information fusion algorithm". This algorithm is not a simple string concatenation, but a structured combination according to the hierarchical relationship of "physical address base + physical location encoding + array coordinates", while introducing a check bit mechanism (such as performing CRC8 check on the three parts of information to obtain a 1-byte check code, which is appended at the end of the identifier). The generated target storage unit identifier (such as "0x100000-03-02-05-04-07-0x3A") not only contains the memory address of the storage unit, but also covers its physical installation location in the cabinet system, realizing the three-dimensional association of "logical address-physical location-chip array", so that the subsequent fault injection operation can not only accurately locate the data area in the memory, but also be associated with the physical location of the hardware device, providing full-dimensional identification support for accurate execution of fault injection and subsequent fault location.

[0044] Optionally, in step 13, the fault feature field of the multi-level nested data frame is deconstructed to determine the fault mode parameter, which includes the following steps:

[0045] Step 131: dynamically slice the fault feature field according to the preset feature granularity to generate several bit segments;

[0046] Step 132: based on the constructed fault mode feature map library and each bit segment, perform fault mode matching to filter out bit segments with a matching degree higher than a threshold value from the feature template of the fault type;

[0047] Step 133: reconstruct the parameters of the filtered bit segments to generate the fault mode parameter.

[0048] Preferably, in step 131, the refinement of the fault feature field is achieved by dynamically slicing the continuous binary data into independently analyzable feature units. The core of this is the "dynamic granularity adaptation" mechanism: the preset feature granularity is not a fixed value, but is automatically adjusted according to the fault type (such as 8-bit granularity for hardware faults and 16-bit granularity for protocol faults) and the field length. For example, when analyzing the "data check error" type of fault, the bit stream is divided according to the number of check code bits (such as 32-bit CRC) to ensure that each segment contains complete check features; for the "timing anomaly" type of fault, the slice length is dynamically adjusted according to the clock period (such as 10ns / segment). This dynamic slicing algorithm breaks through the limitations of fixed length splitting, similar to the idea of adaptive region division according to object edges in image segmentation, so that each bit segment carries meaningful fault feature information, laying the foundation for subsequent pattern matching.

[0049] Preferably, in step 132, the precise identification of the fault mode is achieved through feature matching, and the core is the collaborative mechanism of "fault feature atlas library" and "multi-dimensional matching algorithm". The fault mode feature atlas library is not a simple template set, but a knowledge graph containing fault type, feature vector, and correlation weight (such as the "voltage sag" corresponding feature vector is [0x01, 0x3C, 0x05], and the associated "capacitor aging" and "power fluctuation" sub-modes). In the matching process, a "layered weighted matching algorithm" is used: first, the core features of the bit segments (such as the first bit identifier of the error code) are compared for quick screening, then the Hamming distance is calculated to calculate the similarity between the segment and the template, and finally the correlation weight in the atlas library (such as the probability of a feature appearing in historical faults) is combined to calculate the comprehensive matching degree. When the matching degree exceeds the threshold (such as 85%), it is determined to be an effective segment. This algorithm is similar to the "feature point comparison + confidence weighting" logic in biometric identification, which not only ensures the matching speed, but also improves the recognition accuracy through multi-dimensional verification, solving the problem of single feature matching being easily disturbed by noise.

[0050] Preferably, in step 133, the discrete bit segments are reconstructed into structured fault mode parameters, realizing the transformation from feature data to executable parameters. The core is the "parameter mapping and verification mechanism": first, through segment association analysis (such as determining the logical sequence of "trigger condition segments" and "impact range segments" according to the time sequence), the screened bit segments are sorted according to the parameter structure of the fault mode (such as trigger condition, duration, impact level); then, using the "feature-parameter conversion algorithm", the binary segments are converted into physical quantity parameters (such as converting the 0x1011 bit segment into "3.3V±5%" according to the voltage coding rule); finally, through cross verification (such as verification of the physical reasonableness of the trigger condition parameters and the impact range parameters), the consistency of the reconstructed parameters is ensured. For example, when the "temperature trigger threshold" segment is converted to 85℃, and the "hardware tolerance temperature" parameter is 80℃, the algorithm will automatically correct the threshold to a reasonable range. This reconstruction logic is similar to the integration process from vocabulary to semantics in natural language processing. The final output of the fault mode parameters not only retains the accuracy of the original features, but also meets the physical constraints in engineering practice.

[0051] Optionally, step 14: deduce the environmental parameter field of the multi-level nested data frame to determine the application environment parameters of the electrical cabinet system, specifically including the following steps:

[0052] Step 141: multi-dimensional feature extraction is performed on the environmental parameter field to obtain a basic parameter vector, and principal component analysis is performed on the basic parameter vector to obtain an environmental parameter key vector;

[0053] Step 142: based on the environmental parameter key vector and the configured application scene feature fingerprint library, Bayesian inference is performed to obtain the posterior probability distribution of the scene, and Markov chain Monte Carlo sampling is performed on it to generate a matching probability vector of the application scene;

[0054] Step 143: based on the matching probability vector of the application scene and the scene configuration item, evidence fusion is performed to generate a fusion decision vector, and based on the fusion decision vector, the application environment parameters of the electrical cabinet system are retrieved from the scene environment parameter library.

[0055] Preferably, in step 141, the core features are extracted from the environmental parameter field, realizing data dimension reduction and key information focusing. First, multi-dimensional feature extraction is performed on the environmental parameter field, covering physical quantities such as temperature, humidity, vibration frequency, and electromagnetic interference strength. These discrete parameters are integrated into a basic parameter vector containing the distribution characteristics of the original data. Then, the principal component analysis (PCA) algorithm is used, which essentially converts the high-dimensional parameter vector that may have correlations into linearly independent low-dimensional key vectors through orthogonal transformation: the covariance matrix of the basic parameter vector is calculated, the eigenvalues and eigenvectors of the matrix are solved, the first few eigenvectors with cumulative contribution rate exceeding 85% are selected as principal components to form the environmental parameter key vector. This process is not simply data compression, but preserves the most discriminative information in the original environmental parameters for scene classification (such as the significant increase in the weight of the temperature parameter in a high-temperature environment), which simplifies the subsequent calculation complexity and ensures that the core environmental features are not lost, providing high-quality input data for scene matching.

[0056] Preferably, in step 142, precise matching of environmental features and application scenarios is achieved based on probabilistic reasoning, the core of which is to construct a scene probability model through Bayesian inference and Markov Chain Monte Carlo (MCMC) sampling. The technical essence of Bayesian inference is to update the prior probability of the application scenario using the environmental parameter key vector (new evidence): based on the application scenario feature fingerprint library (containing typical environmental feature distributions in different scenarios), the key vector is compared with the scene features in the fingerprint library, and the posterior probability distribution of each scene given the key vector (i.e. P(Scenario|Key Vector)) is calculated. Since the posterior probability distribution may have a complex multi-peak shape, direct solution is difficult, and the MCMC sampling algorithm constructs a Markov chain such that the stationary distribution of the chain is equal to the target posterior distribution. After a sufficient number of sampling steps (such as 10,000 iterations), the distribution of the sample approximates the posterior distribution, and the final matching probability vector is essentially a quantitative result of the similarity between each scene and the current environmental features, providing a probabilistic basis for scene decision-making.

[0057] Preferably, in step 143, the fusion of multi-source information generates the final decision, and the uncertainty of the single probability model is solved by evidence fusion. Based on the matching probability vector (the possibility of each scene) and the scene configuration item (such as the typical environmental parameter range of the scene and the equipment tolerance threshold), the evidence theory (Dempster-Shafer theory) is used for fusion: the probability vector is regarded as evidence from different sources, and the scene configuration item is regarded as a constraint condition. The trust degree and the likelihood of each scene are calculated by the evidence combination rule to generate a fusion decision vector. This vector not only contains the final matching probability of the scene, but also integrates the hard constraints of the configuration item (such as the trust degree of a scene is significantly reduced when the humidity threshold of the scene conflicts with the current key vector). Finally, the standard environmental parameters of the corresponding scene (such as the temperature fluctuation range of the industrial workshop and the electromagnetic compatibility level of the medical place) are retrieved from the scene environmental parameter library according to the fusion decision vector. The essence is to convert the probability reasoning result into deterministic environmental parameters that can be directly used for extreme working condition simulation, realizing the closed loop from data to decision.

[0058] Optionally, in step 2, based on the application environment parameters, the extreme working condition simulation of the target storage unit is constructed, specifically including the following steps:

[0059] Step 21: Based on the application environment parameters, the extreme working condition feature threshold of the target storage unit is determined.

[0060] Step 22: Based on the extreme working condition feature threshold of the target storage unit, the gradient evolution of the storage unit sensitive physical quantity is performed to generate a scene-based dynamic gradient sequence.

[0061] Step 23: According to the hardware tolerance limit of the target storage unit and the real-time state feedback, the scene-based dynamic gradient sequence is cooperatively coupled to generate an extreme working condition simulation.

[0062] Optionally, in step 21, based on the application environment parameters, the extreme working condition feature threshold of the target storage unit is determined, specifically including the following steps:

[0063] Step 211: The scene feature of the application environment parameters is deconstructed to determine the scene-specific constraint factor.

[0064] Step 212: Based on the hardware parameters of the target storage unit, the extreme working condition feature threshold of the target storage unit is determined according to the scene-specific constraint factor.

[0065] Preferably, in step 211, the scene-specific constraints on the storage unit are extracted from the application environment parameters, and the macro-environment parameters are converted into quantifiable constraint factors through scene feature decomposition. Specifically, using the "scene-parameter mapping decomposition algorithm", the application environment parameters are first classified according to physical properties (such as temperature, humidity, vibration, etc.), and then the influence weight of each parameter is analyzed for different scenes (such as industrial workshop, medical device room): for example, the weight of vibration frequency and electromagnetic interference in the industrial scene is significantly higher than that of other parameters, and the weight of temperature stability is higher in the medical scene. By calculating the deviation coefficient of each parameter in the scene (deviation rate from the standard value of the scene), the key parameters with a deviation coefficient exceeding 30% are selected as scene-specific constraint factors (such as "high-frequency vibration factor" and "strong electromagnetic interference factor" in the industrial scene), which are essentially the most stringent environmental stress indicators of the scene on the storage unit, providing a scene-based benchmark for subsequent threshold determination.

[0066] Preferably, in step 212, the hardware inherent properties and scene constraints are fused to construct the extreme working condition tolerance threshold of the storage unit in a specific scene, and the core is to realize parameter adaptation through the "hardware-scene collaborative threshold algorithm". This algorithm takes the hardware parameters of the target storage unit (such as the maximum working temperature of the chip 85℃, the voltage tolerance range 9-15V) as the basic framework, and takes the scene-specific constraint factors determined in step 211 as the correction coefficient: for example, in a high-temperature and high-humidity scene, the "humidity enhancement factor" will lower the temperature threshold (such as from 85℃ to 75℃), and the "vibration factor" will narrow the voltage stability range (such as from 9-15V to 10-14V). The "marginal tolerance model" is introduced in the correction process to calculate the safety margin of the hardware parameters under the scene constraints (such as leaving a 5℃ buffer space for the temperature threshold), and finally determine the extreme working condition characteristic threshold (such as temperature 75℃, voltage 10-14V) that meets the simulation needs of the scene for extreme environments and does not exceed the physical tolerance limit of the hardware, achieving a balance between the severity of the scene and the safety of the hardware.

[0067] Optionally, in step 22, based on the extreme working condition characteristic threshold of the target storage unit, the gradient evolution of the sensitive physical quantities of the storage unit is performed to generate a scene-based dynamic gradient sequence, which specifically includes the following steps:

[0068] Step 221: Extract the threshold boundary of the sensitive physical quantity from the extreme working condition characteristic threshold of the target storage unit, and establish a sensitive physical quantity-threshold mapping table;

[0069] Step 222: Based on the sensitive physical quantity-threshold mapping table, divide each sensitive physical quantity according to the set gradient step to generate a linear gradient sequence of the physical quantity;

[0070] Step 223: Time axis and amplitude parameter enhancement of the linear gradient sequence of physical quantities to generate a scenario dynamic gradient sequence.

[0071] Preferably, in step 221, the precise correspondence between sensitive physical quantities and threshold boundaries is established to provide a quantitative basis for subsequent gradient evolution. Through the "threshold boundary extraction algorithm", the sensitive physical quantities (such as core voltage, operating temperature, data transmission rate) that have a significant impact on the performance of the storage unit are selected from the extreme working condition characteristic thresholds, and the upper and lower limit thresholds of each physical quantity (such as voltage 10-14V, temperature -5℃-75℃) are extracted, and a sensitive physical quantity-threshold mapping table is constructed. This mapping table is not simply a list of parameters, but through the association of the units, accuracy levels and impact weights of physical quantities (such as voltage fluctuation weight 0.4, temperature fluctuation weight 0.3) on the storage unit, a structured threshold system is formed. In essence, it is to convert abstract threshold requirements into a "physical quantity-boundary" data dictionary that can be directly used for gradient calculation, ensuring that subsequent gradient division always focuses on the most sensitive parameters to the storage unit.

[0072] Preferably, in step 222, the threshold boundaries are converted into a continuous gradient change sequence, and linear division is used to achieve step-by-step simulation of extreme working conditions. Based on the sensitive physical quantity-threshold mapping table, the "dynamic step gradient division algorithm" is used to divide the intervals of each physical quantity: small steps (such as 0.1V / step) are used for parameters that are sensitive to changes (such as voltage), and large steps (such as 2% / step) are used for relatively stable parameters (such as humidity), and the step size is inversely proportional to the sensitivity of the physical quantity (sensitivity is calculated by the correlation between parameter changes and failure rates in historical failure data). For example, the voltage threshold 10-14V can generate 41 gradient nodes with a step size of 0.1V, forming a linear gradient sequence from 10V to 14V. This process essentially converts static threshold intervals into dynamic gradual change paths, simulating the continuous evolution process of physical quantities from normal to extreme states, providing a basic gradient skeleton for scenario simulation.

[0073] Preferably, in step 223, the dynamic characteristics of the linear gradient sequence are given scenario characteristics to make them more consistent with the actual evolution of the working conditions. Through the "spatiotemporal parameter enhancement algorithm", a duration parameter is added to each gradient node in the time dimension (for example, the duration of the extreme temperature node in the high temperature scenario is 3 times that of the normal temperature node), and the gradient change rate is set according to the scenario characteristics (for example, the temperature rise rate of the industrial scenario is 2 times that of the civil scenario); In the amplitude dimension, introduce scenario disturbance factors (such as superimposing random fluctuation amplitude of vibration parameters to simulate the mechanical vibration characteristics of industrial environment). For example, combine the temperature linear gradient sequence with the "high temperature sudden rise" scenario characteristics, so that the time interval of the gradient node gradually shortens as the temperature rises (from 5 minutes / step to 1 minute / step), while the amplitude is superimposed ±1℃ random fluctuation, The finally generated scenario dynamic gradient sequence is essentially a "dynamic working condition script" that combines time rhythm and scenario disturbance characteristics, making the extreme working condition simulation closer to the complex change law of physical quantities in the real environment.

[0074] Optionally, in step 23, according to the hardware tolerance limit of the target storage unit and the real-time state feedback data, the scenario dynamic gradient sequence is coupled to generate an extreme working condition simulation, which includes the following steps:

[0075] Step 231: Convert the hardware tolerance parameters of the target storage unit into an executable parameter safety boundary table, and dynamically correct the parameter safety boundary table in combination with the real-time collected storage unit aging indicators;

[0076] Step 232: Compare the parameter nodes in the scenario dynamic gradient sequence with the real-time state feedback data point by point to determine the characteristic coupling strategy of the application scenario;

[0077] Step 233: Based on the characteristic coupling strategy of the application scenario, the scenario dynamic gradient sequence is coupled according to the corrected safety boundary to generate a working condition execution table containing the time axis and parameter association relationship as an extreme working condition simulation.

[0078] Preferably, in step 231, a dynamically updated hardware security shield is constructed, combining the inherent tolerance of the storage unit with real-time conditions to form an executable security boundary. First, the tolerance parameters in the hardware manual (such as the maximum tolerance voltage 15V, which is determined by the semiconductor material breakdown voltage experiment during chip design, and set after leaving a 20% safety margin; the maximum temperature 85℃, based on the thermal stability test of the chip packaging material, taking the critical temperature value of long-term work) are converted into a parameter security boundary table containing parameter type, upper and lower limits, and alarm threshold. The key is to introduce real-time aging indicators (such as chip usage time, cumulative failure times) to dynamically modify the boundary table, using the "aging coefficient decay model": aging coefficient = 1 - (usage time / design life) x 0.3, where 0.3 is the aging influence weight coefficient, which is an empirical value obtained by analyzing 1000+ devices of the same type throughout their life cycle — statistics show that the impact of device aging on tolerance is about 30% of the design life. The coefficient is used to scale the original security boundary (such as 85℃ x aging coefficient 0.8 = 68℃, the aging coefficient of 0.8 corresponds to about 67% of the design life of the device, at this time, according to the aging decay law, the tolerance temperature needs to be adjusted to 80% of the original threshold), so that the security boundary gradually tightens with the aging of the device. This process essentially converts static hardware parameters into a dynamic security red line that changes with the state of the device, ensuring that subsequent condition simulation does not exceed the actual tolerance of the hardware.

[0079] Preferably, in step 232, the adaptation strategy of the scenario-based gradient sequence to the actual device state is determined through real-time data verification, each parameter node in the scenario-based dynamic gradient sequence (such as temperature 60℃, which is the medium extreme temperature point in this scenario; voltage 12V, which is 1.2 times the rated voltage of the device) is compared with real-time state feedback data (such as current actual temperature 55℃, voltage 11.8V) point by point, and the deviation value (such as temperature deviation 5℃, voltage deviation 0.2V) is calculated. According to the range of the deviation value (such as deviation <10% for slight, this threshold is based on the statistical range of normal fluctuations of the device, most devices do not fluctuate more than 10% when running stably; 10%-30% for moderate, this range is within the adjustment and correction range; >30% for significant, which exceeds the self-adjusting ability of the device and may cause failure), the preset characteristic coupling strategy is matched: slight deviation uses "progressive coupling" (advance according to the original gradient rate), moderate deviation uses "compensation coupling" (adjust the gradient step to offset the deviation), and significant deviation uses "pause coupling + alarm" (to prevent parameter mutation). This process essentially establishes a feedback loop between the gradient sequence and the real-time state of the device, providing targeted strategy basis for cooperative coupling.

[0080] Preferably, in step 233, the gradient sequence, the safety boundary and the scenario strategy are fused into an executable extreme working condition scheme, the parameter nodes in the scenario dynamic gradient sequence are rearranged on the time axis based on the characteristic coupling strategy, and the legality is verified by referring to the corrected safety boundary (such as eliminating the nodes exceeding the safety upper limit). For parameters with interaction influence (such as temperature and voltage), the “parameter correlation degree weighted coupling” is adopted: the correlation coefficient between parameters is calculated according to historical data (for example, the voltage needs to be reduced by 0.5V for every 10℃ rise in temperature, and this coefficient comes from the voltage stability test in high temperature environment — experiments show that the voltage tolerance range of the chip will narrow by 0.5V for every 10℃ rise in temperature), and the related parameter nodes are adjusted in linkage. The final generated working condition execution table contains the parameter combination of each time point, the duration and the safety verification result, and the essence is to condense the dispersed gradient data, safety rules and scenario strategy into a spatiotemporally coordinated working condition execution scheme, realizing the unity of safety and scenario reality of extreme working condition simulation.

[0081] Optionally, step 3, according to the fault mode parameters and the extreme working condition simulation, a fault factor combination is generated, specifically including the following steps:

[0082] Step 31: The fault mode parameters are analyzed by feature factorization, and the core features including the fault trigger condition, the fault influence range and the fault evolution rate are extracted, and the core features are converted into a quantifiable fault feature vector;

[0083] Step 32: The parameter sequence in the extreme working condition simulation is discretized to generate a working condition parameter matrix, and each element in the matrix is associated with the working condition intensity and the duration timestamp;

[0084] Step 33: The correlation degree of the fault feature vector and each element in the working condition parameter matrix is determined, and the fault feature vector and the working condition parameter matrix are time-coupled based on the correlation degree to generate a fault factor combination.

[0085] Optionally, step 31: The fault mode parameters are analyzed by feature factorization, and the core features including the fault trigger condition, the fault influence range and the fault evolution rate are extracted, and the core features are converted into a quantifiable fault feature vector, specifically including the following steps:

[0086] Step 311: The fault mode parameters are semantically analyzed to generate a calculable logical expression;

[0087] Step 312: The calculable logical expression is processed by spatiotemporal correlation mining based on historical fault data to obtain core features including the fault trigger condition, the fault influence range and the fault evolution rate in the form of triplets;

[0088] Step 313: Dimensional normalization and vector encoding of core features to generate quantifiable fault feature vectors.

[0089] Preferably, in step 311, the abstract fault mode parameters are converted into logical rules that can be directly involved in calculations. By constructing a fault domain vocabulary (containing professional terms such as "overvoltage" and "data loss" and corresponding logical symbols), the natural language description of the fault mode parameters is segmented and analyzed. For example, "trigger data error when voltage exceeds 12V and temperature is higher than 60℃" is parsed into the logical expression "(voltage>12V ∧ temperature>60℃)→data error". During the parsing process, a rule-based semantic mapping mechanism is used to convert fuzzy expressions (such as "long-time overvoltage") into precise quantitative conditions (such as "voltage>12V and duration≥5s"), where "5s" is a threshold value obtained from historical data statistics — experiments show that most storage units will not appear detectable faults until 5s of overvoltage, thus ensuring the calculability of the logical expression. This step essentially builds a bridge between natural language and machine-executable logic, providing a structured calculation basis for subsequent feature extraction.

[0090] Preferably, in step 312, the spatio-temporal association rules of fault features are mined from historical data. Based on historical fault data (including records of fault occurrence time, location, and impact range), the logical expressions generated in step 311 are analyzed in the time and space dimensions. In the time dimension, the time difference distribution between fault trigger conditions and fault occurrence is calculated to determine the typical trigger delay (e.g., 2s after overvoltage triggers a fault). In the spatial dimension, the association between fault impact range and device physical topology is calculated (e.g., a certain chip fault affects adjacent 3 storage units). Through frequent pattern mining (such as Apriori algorithm), association rules with support>20% are selected (20% is based on the minimum confidence requirement to ensure the statistical significance of the rules), and finally a triple "fault trigger condition - fault impact range - fault evolution rate" is formed, where the evolution rate is calculated by the expansion slope of the fault impact range over time (e.g., the impact range expands by 2 units per minute). This process essentially combines logical expressions with actual fault data to extract core fault features with spatio-temporal characteristics.

[0091] Preferably, in step 313, the heterogeneous fault features are converted into a unified vector form, realizing the quantization and standardization of the features, and the dimension normalization adopts min-max standardization: each feature value in the triple (such as trigger voltage 12V, influence range 3 units, evolution rate 2 units / minute) is mapped to the interval [0,1], the formula is (x-min) / (max-min), where max and min are taken from the extreme values in the historical data (such as voltage max=15V, min=9V, based on the voltage tolerance range of the storage unit), ensuring that features of different dimensions can be directly compared. The vector encoding adopts the combination of One-Hot encoding and numerical splicing: the discrete features (such as trigger condition type) are One-Hot encoded, and the continuous features (such as evolution rate) are directly retained as normalized numerical values, finally generating a fixed-length fault feature vector (such as a 10-length vector containing 3 trigger condition encoding bits, 4 influence range numerical bits, and 3 evolution rate numerical bits). The essence of this step is to compress the high-dimensional heterogeneous fault features into low-dimensional homogeneous quantized vectors, providing a unified mathematical representation for subsequent correlation degree calculation.

[0092] Optionally, step 32: discretize the working condition parameter sequence in the extreme working condition simulation to generate a working condition parameter matrix, each element in the matrix is associated with the working condition intensity and the duration timestamp, which includes the following steps:

[0093] Step 321: time-domain segmentation of continuous parameter sequence in extreme working condition simulation to generate space-time semantic units, and dynamically identify feature points of working condition parameter change from space-time semantic units using variable step-size sliding window, and divide the continuous parameter sequence into several semantic paragraphs according to the feature points, each semantic paragraph contains start time, end time and parameter change trend type;

[0094] Step 322: intensity quantization of each semantic paragraph, calculate the probability distribution of different working condition parameters in the semantic paragraph by kernel density estimation, construct a four-tuple array, and arrange all four-tuples in time sequence to form a working condition parameter matrix.

[0095] Preferably, in step 321, the continuous working condition parameter sequence is converted into discrete semantic units with time boundaries and change characteristics. The continuous parameter sequence in extreme working condition simulation (such as a continuous change curve of temperature from 25°C to 75°C) is first preliminarily divided in time domain. The initial step length of the sliding window is set to 1s, which is determined based on the sampling frequency (1Hz) of the electrical cabinet system sensor, ensuring that the window can capture the parameter change of each sampling point. When identifying feature points, the algorithm dynamically adjusts the window step length: when the parameter change rate ≤0.5°C / s (this threshold is derived from the statistics of stable working conditions, and the parameter change rate under most normal working conditions does not exceed 0.5°C / s), a large step length of 5s is used to reduce redundant segmentation; when the change rate >0.5°C / s, automatically switch to a small step length of 0.5s to accurately capture the mutation feature points (such as the starting time of temperature sudden rise). The continuous sequence is divided into semantic paragraphs through feature points (parameter change rate mutation points, extreme points), and each paragraph contains the starting / ending time and change trend type (such as linear rise, stepwise decline). The essence is to describe the dynamic evolution process of the parameter with discrete time segments, providing clear analysis units with time boundaries for subsequent intensity quantification.

[0096] Preferably, in step 322, the working condition intensity of the discrete semantic paragraph is quantitatively represented. The overall characteristics of the parameter in the paragraph are described by probability distribution. For each semantic paragraph, the kernel density estimation (KDE) algorithm is used to calculate the probability distribution of the working condition parameter: taking all parameter sampling values in the paragraph as samples, the probability density curve of the parameter in the paragraph is fitted by a Gaussian kernel function (the bandwidth is set to 0.8, which is determined based on the parameter measurement accuracy ±0.5°C, ensuring that the distribution curve is smooth and retains details), and the parameter value corresponding to the peak value of the curve is the typical intensity of the paragraph (such as the highest probability of temperature 50°C). Based on the distribution characteristics, a four-element array (starting time, ending time, typical intensity value, intensity fluctuation variance) is constructed, where the fluctuation variance reflects the parameter stability (such as variance <2°C², which is twice the parameter measurement error, ensuring that the true fluctuation is distinguished from measurement noise). The working condition parameter matrix formed by arranging all four-element groups in time order essentially quantifies the working condition state of each time segment by statistical characteristics, realizing the conversion of continuous parameters into structured matrices, and providing a calculable numerical basis for subsequent correlation analysis with fault features.

[0097] Optionally, step 33: determine the correlation degree of the fault feature vector and each element in the working condition parameter matrix, and time-couple the fault feature vector and the working condition parameter matrix based on the correlation degree to generate a fault factor combination, which specifically includes the following steps:

[0098] Step 331: compare each dimension of the fault feature vector with the elements in the working condition parameter matrix one by one, and calculate the linear correlation degree between the two by using the Pearson correlation coefficient;

[0099] Step 332: Screen out the fault feature and working condition parameter combination whose linear correlation degree exceeds the set correlation degree threshold;

[0100] Step 333, determine the timestamp in the working condition parameter matrix and take it as the reference axis;

[0101] Step 334, based on the reference axis, time sequence fusion is performed on the fault feature and working condition parameter in the fault feature and working condition parameter combination to generate the fault factor combination.

[0102] Preferably, in step 331, the linear correlation degree between the fault feature and the working condition parameter is quantified, and the covariance of each dimension of the fault feature vector (such as the fault triggering voltage) and the standard deviation ratio of the working condition parameter matrix element (such as the real-time voltage value) is calculated, and the formula is wherein, represents the standard deviation of the data of a certain dimension of the fault feature vector, and the physical meaning is the dispersion degree of the fault feature value (such as the fault triggering voltage, the fault evolution rate, etc.) in this dimension, which reflects the fluctuation range of the fault feature in multiple observations; represents the standard deviation of the data of the corresponding element in the working condition parameter matrix, and the physical meaning is the dispersion degree of the working condition parameter (such as the real-time voltage value, the temperature value, etc.), which reflects the change range of the working condition parameter in the extreme working condition simulation process. The product of the two as the denominator, divided by the covariance cov(X, Y) (reflecting the overall correlation degree of the fault feature and the working condition parameter) of the numerator, the correlation coefficient r obtained can standardize the linear correlation strength between the two. The value range of r is [-1, 1], and the closer the absolute value is to 1, the stronger the linear correlation. For example, when the fault feature is “overvoltage triggering”, and the “voltage intensity” dimension in the working condition parameter is calculated to get r=0.85, which indicates that there is a strong positive correlation between the two. In the calculation process, each fault feature dimension and all working condition parameter elements are compared, and the essence is to identify “which fault features will change with which working condition parameters” from the mathematical level, which provides a quantitative basis for subsequent screening of effective correlation.

[0103] Preferably, in step 332, the fault and working condition combinations with significant correlations are screened out, and weakly correlated noise is removed by setting a correlation threshold. The correlation threshold is usually set to 0.6, and the value is determined based on a large number of fault case statistics: when r≥0.6, the causal relationship between the occurrence of the fault and the change of the working condition parameter has a confidence of more than 90% (obtained by hypothesis testing, P value <0.05), and the correlation below this value is mostly random coincidence. For example, the correlation r=0.72 (>0.6) between “temperature fluctuation” and “data check error” is retained, and the correlation r=0.35 (<0.6) between “humidity change” and “bus fault” is removed. This step is essentially a statistical significance screening that focuses on the working condition parameters that actually affect the fault, reduces the redundancy of subsequent calculations, and ensures the effectiveness of the correlation combination.

[0104] Preferably, in step 333, a unified time reference framework is established to provide a reference for the time alignment of the fault features and the working condition parameters. The time stamps in the working condition parameter matrix (such as 16:00:00, 16:00:05) come from the time axis of the extreme working condition simulation. The reason for taking it as the reference axis is that the change of the working condition parameter has a clear time continuity, and the triggering and evolution of the fault feature also depend on the time dimension. By extracting the time stamps of all elements in the matrix and arranging them in ascending order, a complete time axis (such as 0-300s) from the beginning to the end of the fault injection is formed, and each time point corresponds to a unique working condition state. This step is essentially to construct a “time-parameter” mapping coordinate system to ensure that the correspondence between the fault and the working condition in the time dimension is clear and identifiable in the subsequent fusion process.

[0105] Preferably, in step 334, the screened correlation combinations are integrated along the time axis to form a spatiotemporal coupling sequence with both fault features and working condition parameters. Based on the time points of the reference axis, the fault features and the working condition parameters at each time are bound: for example, at 16:00:10, the “overvoltage fault feature (triggering voltage 12V)” and the “working condition voltage 12.5V, lasting 5s” are fused into a fault factor unit. During the fusion process, the correlation combinations with a time stamp deviation of ≤1s are aligned by interpolation (since the sampling synchronization error of the sensor and the fault monitoring is usually within 1s), ensuring consistent time granularity. The final generated fault factor combination is a sequence of correlation units arranged in chronological order, which essentially visualizes the process of “how the fault feature evolves under a specific working condition” to provide basic data containing three-dimensional information of time, fault, and working condition for subsequent generation of fault atlas.

[0106] Optionally, in step 4, a dynamic adaptive fault atlas is generated based on the fault factor combination, specifically including the following steps:

[0107] Step 41: from the fault factor combination, the fault type is parsed, the associated working condition parameters are parsed, each fault type is taken as a network node, the weight of the edge is taken as the association strength of the associated working condition parameters, and an initial node network is constructed in the time axis order, wherein the node attribute contains the fault characteristic value, and the edge attribute contains the coupling time stamp; Step 42: acquiring real-time data collected by the built-in monitoring module of the storage unit in the fault injection process, determining the dynamic deviation coefficient of the real-time data and the fault characteristic value, and dynamically adjusting the weight and edge in the initial node network based on the dynamic deviation coefficient to generate a dynamic adaptive fault graph.

[0108] Preferably, in step 41, the fault factor combination is converted into a visual network topology, and an initial correlation model of fault evolution is constructed. From the fault factor combination, discrete fault types (such as "data verification error" and "bus transmission interruption") and corresponding associated working condition parameters (such as "voltage 12.5V" and "temperature 60℃") are parsed. Each fault type is mapped to an independent node in the network, and the node attribute contains the fault characteristic value (such as "verification error rate threshold 1%", which is set based on the data integrity requirement of the storage unit. When the error rate exceeds 1%, it will affect the data reliability). The weight of the edge is determined by the association strength of the associated working condition parameters. The absolute value of the Pearson correlation coefficient r calculated in step 331 is used (for example, if r = 0.85, the weight is 0.85), because the correlation coefficient has quantified the linear correlation degree between the fault and the working condition, and can directly represent the connection strength of the edge. When constructing the network in the time axis order, 10ms is taken as the time granularity (matching the fault monitoring response time of the storage unit), and each time slice corresponds to a group of node connection relationships. The coupling time stamp in the edge attribute records the accurate time when the fault correlation occurs (for example, 16:00:10.005). The whole process is essentially to convert the time-series fault-working condition correlation data into a spatialized network topology. The node represents the fault entity, the edge represents the causal correlation between faults, and the weight and time stamp give the initial attributes of the dynamic evolution of the network, providing a structured basis for subsequent dynamic adjustment.

[0109] Optionally, in step 42, real-time data collected by the built-in monitoring module of the storage unit in the fault injection process is acquired, a dynamic deviation coefficient of the real-time data and the fault characteristic value is determined, and the weight and edge in the initial node network are dynamically adjusted based on the dynamic deviation coefficient to generate a dynamic adaptive fault graph, which specifically includes the following steps:

[0110] Step 421: from the real-time data, the key indicators associated with the fault characteristic values are filtered out, the key indicators are aligned with the corresponding fault characteristic values in the node attribute, and the dynamic deviation coefficient of the real-time data and the fault characteristic value is calculated;

[0111] Step 422: when the dynamic deviation coefficient is positive, the weight of the corresponding edge is amplified; when the dynamic deviation coefficient is negative, the weight of the edge is reduced;

[0112] Step 423, determine the trend of the dynamic deviation coefficient to determine the change of the fault correlation strength, if the increase amplitude of the fault correlation strength exceeds the set increase amplitude threshold, add a temporary edge, otherwise, delete the corresponding edge, to generate a dynamic adaptive fault graph.

[0113] Preferably, in step 421, the quantitative comparison relationship between real-time data and fault characteristic values is established to provide deviation basis for subsequent network adjustment. First, key indicators (such as real-time error rate, temperature drift value) are selected from real-time data. The selection basis is the correlation threshold (r≥0.6) determined in step 332 to ensure that the indicators are directly related to the fault characteristic values. After aligning the key indicators with the fault characteristic values (such as error rate threshold 1%, temperature critical value 60℃) in the node attributes by dimension, the deviation is calculated by the formula “dynamic deviation coefficient = (real-time indicator value - characteristic threshold) / characteristic threshold”. For example, the deviation coefficient of real-time error rate 1.5% and threshold 1% is 0.5. This process essentially normalizes and compares real-time running state with preset fault characteristics, and quantifies the deviation degree of the two with dimensionless coefficient, providing a directly applicable quantitative parameter for network weight and edge adjustment.

[0114] Preferably, in step 422, the weight of the network edge is dynamically adjusted according to the real-time deviation to realize real-time update of the fault correlation strength. When the dynamic deviation coefficient is positive (such as 0.5), it means that the real-time indicator exceeds the characteristic threshold, and the fault correlation strength should be enhanced. At this time, the weight is enlarged according to “weight = original weight × (1 + deviation coefficient)” (such as original weight 0.85 × 1.5 = 1.275). When the coefficient is negative (such as -0.3), the weight is reduced according to “weight = original weight × (1 + deviation coefficient)” (0.85 × 0.7 = 0.595), and the weight is kept at least 30% of the initial value (this threshold is set based on network stability requirements. Less than 30% will result in too weak correlation, which cannot reflect the actual fault propagation). This process essentially dynamically adjusts the closeness of fault correlation through the positive and negative and size of the deviation coefficient, so that the network weight reflects the actual strength of fault evolution in real time.

[0115] Preferably, in step 423, the edge structure of the network is dynamically maintained based on the deviation trend to realize adaptive updating of the fault correlation relationship. The first order difference of the dynamic deviation coefficient is calculated (for example, the coefficient increases from 0.2 to 0.5 in three consecutive periods, and the absolute value of the difference is 0.3) to judge the trend. When the trend shows that the increase amplitude of the fault correlation strength exceeds the set threshold (usually 50%, which is obtained by analyzing the fault propagation cases, and an increase amplitude exceeding 50% indicates the appearance of a new stable correlation), a temporary edge is added, and the initial weight is set to the average weight of the same type of historical correlation (for example, 0.6). If the increase amplitude continues to be lower than the threshold and the edge weight is less than or equal to 0.1 (which is the correlation failure threshold, and edges below this value have no practical significance), the corresponding edge is deleted. This process essentially allows the network edge structure to dynamically change with the generation and disappearance of fault correlations, ensuring that the fault graph can capture new fault propagation paths in real time and eliminate invalid correlations, ultimately forming a dynamic adaptive network that truly reflects fault evolution.

[0116] Optionally, in step 5, based on the dynamic adaptive fault graph, a fault injection operation is performed on the target storage unit indicated by the target storage unit identifier, and characteristic data of the electrical cabinet system during operation is collected. Specifically, the following steps are included:

[0117] Step 51: Analyze the topology structure of the dynamic adaptive fault graph to extract the node identifier of the corresponding target storage unit, the edge weight of the fault correlation strength, and the coupling timestamp. Filter out the core fault paths with a weight value greater than a set weight threshold, and sort the core fault paths according to the impact level "high / medium / low" to generate a fault injection execution list containing fault types, injection time, and target unit address. High-impact-level faults are marked as priority injection items.

[0118] Step 52: Generate a hierarchical fault injection mechanism based on the fault injection execution list. Start the hierarchical fault injection mechanism through a fault injection trigger signal. For high-impact-level faults, use a hardware-level direct injection method. For medium and low-impact-level faults, use software simulation injection. Compare the coupling timestamps in real time during the injection process.

[0119] Step 53: Use the fault injection trigger signal as the time reference point to collect characteristic data of the electrical cabinet system during operation, including electrical characteristics, data characteristics, and environmental characteristics. Electrical characteristics include input and output voltage ripple, working current dynamic change curve, and real-time power loss value. Data characteristics include read and write response time of the target storage unit, data verification error code, and bus transmission frame error rate. Environmental characteristics include target storage unit shell temperature, cabinet air flow speed, and humidity change.

[0120] Preferably, in step 51, key fault information is extracted from the dynamic adaptive fault graph and an executable injection plan is generated. First, the graph topology is parsed to extract the node identifier corresponding to the target storage unit, edge weight (reflecting the fault correlation strength), and coupling timestamp (the exact time when the fault correlation occurs). By setting a weight threshold (usually 0.7, which is based on fault impact significance analysis, and paths with a weight greater than 0.7 have a confidence level of more than 95% in affecting system stability), the core fault paths are filtered out. When sorting by impact level, the "weighted scoring method" is used: the weighted sum of edge weight (0.4), fault propagation speed (0.3), and the number of units involved (0.3) is ≥80 points for high impact level, 60-80 points for medium impact level, and <60 points for low impact level. An execution list containing fault type, injection timing (based on coupling timestamp, such as 16:00:10), and target address is generated. High-impact-level faults are marked as priority items. The essence is to convert the abstract correlation of the graph into specific execution tasks sorted by priority, ensuring that fault injection focuses on the path with the greatest impact on the system.

[0121] Preferably, in step 52, differentiated fault injection is implemented according to the fault impact level. Time synchronization mechanism is used to ensure the accuracy of injection. In the hierarchical fault injection mechanism, high-impact-level faults use hardware-level direct injection (such as writing error instructions through JTAG interface), as they need to simulate physical layer faults (such as chip pin short circuit), and the response delay of hardware injection is <1ms (satisfying real-time requirements); medium and low-level faults use software simulation injection (such as calling API to simulate data verification error), and the delay is allowed within 10ms. During the injection process, the coupling timestamp is compared with the system clock in real time, and if the deviation exceeds 5ms, calibration is triggered (this threshold is based on the time sensitivity of fault evolution, and a deviation of more than 5ms will cause the injection timing to be out of sync with the fault correlation). Time synchronization is achieved by adjusting the sending time of the injection trigger signal. This process essentially selects the appropriate injection method based on the fault impact level and ensures that the injection timing is consistent with the predicted evolution rhythm of the fault graph.

[0122] Preferably, in step 53, a multi-dimensional data acquisition system is constructed to comprehensively capture the system state after fault injection. The fault injection trigger signal is taken as the time reference point (denoted as t=0), the electrical characteristics are sampled at a rate of 1MHz (since the frequency of voltage ripple is usually in the hundreds of kHz, a high sampling rate can capture high-frequency fluctuations), the data characteristics are sampled at a rate matching the bus transmission rate (such as 100kHz, ensuring complete recording of frame errors), and the environmental characteristics are sampled at a rate of 10Hz (temperature and humidity change slowly, and do not need to be collected at a high frequency). The collected characteristic data covers three dimensions of electrical (reflecting the hardware load state), data (reflecting the information transmission quality), and environment (reflecting external influencing factors), forming a complete data chain of fault injection and system response. For example, voltage ripple collection needs to retain a measurement accuracy of ±0.1V (based on the voltage tolerance range of the storage unit, a small ripple may also cause data errors), and the essence is to provide original data support containing causal relationships for subsequent fault analysis through multi-dimensional, high-synchronization data collection.

[0123] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Those skilled in the art can make various modifications and changes to the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method of fault injection for an integrated electrical cabinet system, the method comprising: The method comprises: Step 1, obtaining a fault injection instruction to parse a target storage unit identifier, a fault mode parameter and an application environment parameter of the electric cabinet system therefrom; Step 2, comprising: Step 21: determining the extreme working condition characteristic threshold of the target storage unit based on the application environment parameter; Step 22: generating a scenario dynamic gradient sequence by gradient evolution of the storage unit sensitive physical quantity based on the extreme working condition characteristic threshold of the target storage unit; Step 23: generating an extreme working condition simulation by synergistic coupling of the scenario dynamic gradient sequence according to the hardware tolerance limit of the target storage unit and the real-time state feedback; Step 3, comprising: Step 31: feature factorization analysis is performed on the fault mode parameter to extract core features including fault trigger condition, fault influence range and fault evolution rate, and the core features are converted into a quantifiable fault feature vector; Step 32: discretizing the parameter sequence in the extreme working condition simulation to generate a working condition parameter matrix, each element in the matrix being associated with working condition intensity and duration timestamp; Step 33: determining the correlation degree of the fault feature vector and each element in the working condition parameter matrix to perform time sequence coupling of the fault feature vector and the working condition parameter matrix based on the correlation degree to generate a fault factor combination; Step 4, comprising: Step 41: parsing the fault type and associated working condition parameter from the fault factor combination, taking each fault type as a network node, taking the associated strength of the associated working condition parameter as the weight of the edge, and constructing an initial node network in time axis order, wherein the node attribute contains the fault feature value, and the edge attribute contains the coupling timestamp; Step 42: acquiring real-time data collected by the storage unit built-in monitoring module during the fault injection process, determining the dynamic deviation coefficient of the real-time data and the fault feature value, and dynamically adjusting the weight and edge in the initial node network based on the dynamic deviation coefficient to generate a dynamic adaptive fault graph; Step 5, based on the dynamic adaptive fault graph, performing a fault injection operation on the target storage unit indicated by the target storage unit identifier and collecting characteristic data of the electric cabinet system during operation.

2. The integrated cabinet system fault injection method of claim 1, wherein, The step 1 specifically comprises the following steps: Step 11: analyzing the data packet structure of the fault injection instruction to obtain a multi-level nested data frame; Step 12, tracing the storage unit field in the multi-level nested data frame to determine the target storage unit identifier; Step 13, deconstructing the fault feature field of the multi-level nested data frame to determine the fault mode parameter; Step 14: deduce the environment parameter field of the multi-level nested data frame to determine the application environment parameter of the electric cabinet system.

3. The method of claim 2, wherein, The step 11 specifically comprises the following steps: Step 111: analyzing the data packet structure of the fault injection instruction to extract the data packet frame header identifier and frame tail identifier therein to determine the boundary range of the data frame; Step 112: based on the boundary range of the data frame, performing nested level unpacking processing on the fault injection instruction to obtain a main frame structure and a sub-frame structure; Step 113: field index anchoring processing is performed on the main frame structure and the sub-frame structure according to the hierarchical relationship to form a multi-level nested data frame.

4. The method of claim 2, wherein, Step 13, specifically comprising the following steps: Step 131: slice the fault feature field into several bit segments according to the preset feature granularity; Step 132: based on the constructed fault mode feature map library and each bit segment, perform fault mode matching to filter out part of the bit segments with a matching degree higher than a threshold value from the fault type feature template; Step 133: reconstruct the parameters of the filtered bit segments to generate fault mode parameters.

5. The method of claim 1, wherein, Step 21, specifically comprising the following steps: Step 211: scene feature deconstruction is performed on the application environment parameters to determine the scene-specific constraint factors; Step 212: based on the hardware parameters of the target storage unit, the extreme working condition feature threshold of the target storage unit is determined according to the scene-specific constraint factors.

6. The method of claim 1, wherein, Step 33, specifically comprising the following steps: Step 331: compare each dimension of the fault feature vector with the elements in the working condition parameter matrix one by one to calculate the linear correlation degree between them using the Pearson correlation coefficient; Step 332: filter out the fault feature and working condition parameter combination with a linear correlation degree exceeding a set correlation degree threshold; Step 333, determine the timestamp in the working condition parameter matrix and take it as a reference axis; Step 334, based on the reference axis, time sequence fusion is performed on the fault feature and working condition parameter in the fault feature and working condition parameter combination to generate a fault factor combination.

7. The method of claim 1, wherein, Step 5, specifically comprising the following steps: Step 51: analyze the topology structure of the dynamic adaptive fault graph to filter out the core fault paths with a weight value greater than a set weight threshold, and sort the core fault paths to generate a fault injection execution list; Step 52: generate a hierarchical fault injection mechanism based on the fault injection execution list, and start the hierarchical fault injection mechanism through a fault injection trigger signal; Step 53: take the fault injection trigger signal as a time reference point to collect feature data during the operation of the electrical cabinet system, including electrical features, data features, and environmental features.

Images