Dynamic pruning and fine-tuning method and system for large aviation models based on hierarchical stability analysis

Abstract

The present invention relates to the technical field of large aviation model optimization, and provides a method and system for dynamic pruning and fine-tuning of large aviation models based on hierarchical stability analysis, in order to solve the problem that the existing technology relies on static thresholds to judge the importance of hierarchies, resulting in large fluctuations in model performance and slow convergence after pruning. The method of the present invention comprises: constructing a fault mode coupling map based on historical fault data of aircraft engines, generating a hierarchical association topology of the large aviation model; calculating a sensitivity energy evaluation value, and determining the hierarchy below the sensitivity energy evaluation value as an energy stable layer; when the energy cumulative change rate of the energy stable layer meets the set vibration energy level critical threshold, marking it as a removal layer; adding the temporal feature code of the removed layer to the adjacent retained layer, performing fine-tuning compensation processing, and obtaining the large aviation model after dynamic pruning and fine-tuning. The present invention solves the core defects of the existing technology that static threshold decisions lead to the accidental deletion of key layers, and progressive fine-tuning causes performance fluctuations and slow convergence.

Description

Technical Field

[0001] The present invention relates to the technical field of aviation large model optimization, and in particular to a method and system for dynamic pruning and fine-tuning of aviation large models based on hierarchical stability analysis. Background Art

[0002] As the intelligence level of aviation equipment continues to increase, fault diagnosis and prediction for key components such as aircraft engines rely on high-precision, high-efficiency, large-scale deep learning models. In actual deployment, due to the complex and ever-changing flight mission scenarios, limited edge computing resources, and the high demand for timely model responses, an intelligent fine-tuning mechanism is urgently needed that can dynamically adjust the model structure and balance inference efficiency and diagnostic performance.

[0003] Current research has proposed a model lightweighting method based on dynamic freezing of attention weights. This solution introduces a learnable attention module during the training phase to quantitatively evaluate the importance of each model layer, and automatically freezes low-contribution network layers based on a preset threshold, allowing them to skip calculations during the inference phase, thereby achieving model compression and acceleration. A progressive fine-tuning strategy is adopted to perform local parameter updates on the remaining layers after pruning to alleviate the problem of model performance degradation. However, existing solutions rely on static thresholds to judge the importance of layers, which can easily lead to mispruning or underpruning. Their fine-tuning compensation mechanism only acts on the adjacent layers of the frozen layers, and lacks effective modeling of changes in the overall model topology. This results in large performance fluctuations and slow convergence in the pruned model, making it difficult to meet the stringent requirements of the aviation field for model robustness and real-time performance. Summary of the Invention

[0004] The present invention provides a dynamic pruning and fine-tuning method and system for large aviation models based on hierarchical stability analysis, which is used to solve the problems in the existing technology that rely on static thresholds to judge the importance of hierarchies, which easily leads to mispruning or underpruning; the fine-tuning compensation mechanism lacks effective modeling of changes in the overall model topology structure, and the pruned model has large performance fluctuations and slow convergence speed.

[0005] In a first aspect, the present invention provides a method for dynamic pruning and fine-tuning of a large aviation model based on hierarchical stability analysis, comprising:

[0006] Based on historical failure data of aircraft engines, a fault mode coupling map is constructed to generate a hierarchical correlation topology of the aviation large model;

[0007] Calculating a sensitivity energy evaluation value according to the hierarchical association topology, and determining a layer in the aviation large model within a preset training period that is lower than the sensitivity energy evaluation value as an energy stable layer;

[0008] Setting a critical threshold value of a vibration energy level, and when the energy cumulative change rate of the energy stabilization layer meets the critical threshold value of the vibration energy level, marking the energy stabilization layer in the hierarchical association topology of the large aviation model as a removal layer;

[0009] The temporal feature code of the removed layer is added to the adjacent retained layer to obtain the added retained layer, and fine-tuning compensation processing is performed on the added retained layer to obtain the large aviation model after dynamic pruning and fine-tuning.

[0010] Optionally, based on historical aircraft engine failure data, a fault mode coupling map is constructed to generate a hierarchical association topology of the aviation large model, including:

[0011] Analyze the waveform characteristics of historical aircraft engine failure data, identify periodically recurring characteristic segments, and define the corresponding failure mode for each characteristic segment;

[0012] Detect multiple failure modes that occur simultaneously in the same time interval to calculate the interaction strength value between the modes;

[0013] Creating a node relationship table according to the inter-mode interaction strength value, and establishing a bidirectional connection relationship for the fault mode pairs whose inter-mode interaction strength values in the node relationship table exceed a preset connection threshold, so as to construct a fault mode coupling map;

[0014] Each fault mode in the fault modal coupling graph is associated with a specific level of the aviation model to obtain an associated level, and the associated levels are connected according to the bidirectional connection relationship to generate a hierarchical association topology of the aviation model.

[0015] Optionally, a node relationship table is created based on the inter-mode interaction strength value, and a bidirectional connection relationship is established for the fault mode pairs whose inter-mode interaction strength values in the node relationship table exceed a preset connection threshold, so as to construct a fault mode coupling graph, including:

[0016] Creating a two-dimensional table containing all failure modes, wherein the row headers and column headers of the two-dimensional table correspond to the failure modes, and filling each cell where the rows and columns intersect with the interaction strength value between the modes;

[0017] When the inter-mode interaction strength value corresponding to the non-diagonal cell in the two-dimensional table is greater than a preset connection threshold, an undirected connection line with a weight is established between the fault mode corresponding to the row header and the fault mode corresponding to the column header, where the weight is the inter-mode interaction strength value;

[0018] The fault modes are used as nodes, the weighted undirected connecting lines are used as edges, and the nodes and edges are combined to generate a fault mode coupling graph.

[0019] Optionally, calculating a sensitivity energy evaluation value according to the hierarchical association topology, and determining a layer in the aviation large model within a preset training period that is lower than the sensitivity energy evaluation value as an energy stable layer includes:

[0020] Calculating the parameter offset amplitude of each associated layer in the aviation large model within a preset training period during the training process, so as to calculate the inter-layer influence factor according to the parameter offset amplitude;

[0021] determining, according to the hierarchical association topology, a target post-association level to which each post-association level is connected, to generate an association level group;

[0022] The parameter offset amplitude of each associated level is superimposed with the inter-level influencing factor to obtain multiple associated level response values, and the distribution median of all associated level response values is used as the sensitivity energy evaluation value;

[0023] The post-association layer whose response value of the post-association layer is lower than the sensitivity energy evaluation value is determined to be an energy stable layer.

[0024] Optionally, a vibration energy level critical threshold is set, and when the energy cumulative change rate of the energy stabilization layer meets the vibration energy level critical threshold, the energy stabilization layer in the hierarchical association topology of the large aviation model is marked as a removal layer, including:

[0025] Query the preset material property database to obtain the vibration propagation attenuation coefficient corresponding to the aircraft engine model;

[0026] Recording the weight parameters corresponding to the start and end times of the energy stabilization layer at a preset training cycle, respectively, to calculate the total amount of parameter change, and generating the energy cumulative change rate based on the total number of training cycles;

[0027] Extracting the safety vibration energy corresponding to the aircraft engine model from a preset safety vibration energy table;

[0028] The vibration propagation attenuation coefficient is multiplied by the safe vibration energy to obtain a critical threshold value of the vibration energy level. When the energy cumulative change rate is less than the critical threshold value of the vibration energy level, the energy stable layer in the hierarchical association topology of the large aviation model is marked as a removal layer.

[0029] Optionally, the temporal feature code of the removed layer is added to an adjacent retained layer to obtain an added retained layer, and fine-tuning and compensating processing is performed on the added retained layer to obtain a large aerial model after dynamic pruning and fine-tuning, including:

[0030] Recombining waveform units of three consecutive training cycles in the waveform segment of the removed layer into a temporal feature code;

[0031] Determining, based on the bidirectional connection relationship of the layer association topology, a layer connected to the removed layer as an adjacent retained layer;

[0032] Acquire first position information of the removed layer at a data input moment within a preset training cycle, and locate second position information identical to the first position information in an input sequence of the adjacent retained layer;

[0033] Inserting the temporal feature code into the second position information to obtain an adjacent retained layer after insertion;

[0034] According to the periodic fluctuation reference value of the historical fault data of the aircraft engine, the waveform time axis parameters of the adjacent retained layer after insertion are adjusted to perform fine-tuning compensation processing on the retained layer after addition, so as to obtain the large aircraft model after dynamic pruning and fine-tuning.

[0035] Optionally, according to a periodic fluctuation reference value of historical fault data of an aircraft engine, a waveform time axis parameter of the adjacent retained layer after insertion is adjusted to perform fine-tuning compensation processing on the retained layer after addition, thereby obtaining a large aviation model after dynamic pruning and fine-tuning, including:

[0036] Selecting fault waveform samples with the same fault mode as the removed layer from the historical fault data, extracting an average time value of a preset peak interval from the fault waveform samples, and using the average time value as a period fluctuation reference value;

[0037] Obtaining a waveform sequence of adjacent retained layers after insertion under the action of a standard vibration test signal, detecting a time difference between adjacent main peaks in the waveform sequence to calculate an actual period value, and calculating a period proportional factor in combination with the period fluctuation reference value;

[0038] According to the period scale factor, the waveform time axis parameters of the adjacent retained layer after insertion are scaled to recalculate the actual period value after scaling. When the absolute difference between the actual period value after scaling and the period fluctuation reference value does not exceed the preset difference, it is confirmed that the fine-tuning compensation processing is completed, and the large aviation model after dynamic pruning and fine-tuning is obtained.

[0039] In a second aspect, the present invention provides a dynamic pruning and fine-tuning system for large aviation models based on hierarchical stability analysis, comprising:

[0040] A construction module is used to construct a fault mode coupling map based on historical fault data of aircraft engines to generate a hierarchical correlation topology of the aviation large model;

[0041] a calculation module, configured to calculate a sensitivity energy evaluation value according to the hierarchical association topology, and determine a layer in the aviation large model within a preset training period that is lower than the sensitivity energy evaluation value as an energy stable layer;

[0042] a marking module, configured to set a critical threshold value of a vibration energy level, and mark the energy stabilization layer in the hierarchical association topology of the large aviation model as a removal layer when the energy cumulative change rate of the energy stabilization layer meets the critical threshold value of the vibration energy level;

[0043] The fine-tuning module is used to add the temporal feature code of the removed layer to the adjacent retained layer to obtain the added retained layer, perform fine-tuning compensation processing on the added retained layer, and obtain the aerial large model after dynamic pruning and fine-tuning.

[0044] In a third aspect, an embodiment of the present invention provides a computing device comprising a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement a dynamic pruning and fine-tuning method for large aviation models based on hierarchical stability analysis as described in the first aspect above.

[0045] In a fourth aspect, an embodiment of the present invention provides a computer storage medium storing a computer program. When the computer program is executed by a computer, it implements a dynamic pruning and fine-tuning method for a large aviation model based on hierarchical stability analysis as described in the first aspect.

[0046] In the present invention, a fault mode coupling map is constructed based on historical fault data of aircraft engines to generate a hierarchical association topology of an aviation large model; a sensitivity energy evaluation value is calculated based on the hierarchical association topology, and the layers in the aviation large model within a preset training cycle that are lower than the sensitivity energy evaluation value are determined to be energy stable layers; a vibration energy level critical threshold is set, and when the energy cumulative change rate of the energy stable layer meets the vibration energy level critical threshold, the energy stable layer in the hierarchical association topology of the aviation large model is marked as a removal layer; the temporal feature code of the removal layer is added to the adjacent retained layer to obtain an added retained layer, and fine-tuning compensation processing is performed on the added retained layer to obtain the aviation large model after dynamic pruning and fine-tuning.

[0047] Beneficial effects of the present invention:

[0048] The technical solution provided by the present invention solves the problem of ignoring the physical coupling relationship between faults in existing solutions. It converts the concurrent characteristics of multiple faults of aircraft engines (such as vibration conduction paths and thermal coupling effects) into the hierarchical association topology of large aviation models, ensuring the ability of the pruned model to retain the multi-fault coupling characteristics from the source, avoiding the feature decoupling distortion caused by static pruning; breaking through the limitations of static threshold judgment, quantifying the impact of fault conduction (such as the transmission intensity of blade cracks on bearing vibration) through hierarchical association topology, and generating aviation scenario-adaptive sensitivity energy evaluation values combined with dynamic parameter changes in a preset training cycle, significantly reducing the false shearing rate and under-shearing rate; integrating the physical vibration characteristics of the engine (such as the vibration propagation attenuation coefficient, etc.), replacing the traditional artificial threshold with a dynamic decision-making mechanism driven by vibration energy levels, ensuring that the removal layer selection conforms to the actual failure law of aviation components, and solving the problem of general pruning methods destroying the fault propagation path; through the position information matching and periodic characteristic calibration of the fault waveform, the fault timing characteristics are completely retained while the model scale is compressed, eliminating the defects of slow convergence speed and large performance fluctuations caused by the existing progressive fine-tuning. Furthermore, based on the physical characteristics of aircraft engine fault waveforms, waveform segments with clear periodic patterns in the removed layers are reorganized into timing codes, and the adjacent retained layers are located through the bidirectional connection relationship of the hierarchical topology. Feature codes are injected at the precise position of the training data input moment, and the waveform time axis parameters are calibrated and scaled according to the periodic fluctuation benchmark value of the historical fault waveform library to achieve spatiotemporal consistency migration and compensation of fault features. This overcomes the fundamental defect of the existing fine-tuning mechanism that only updates parameters locally and ignores overall topological changes. Specifically, through fault waveform reorganization and spatiotemporal alignment injection, the fragmentation of timing features caused by pruning is avoided, and the integrity of the multi-fault coupled propagation link is maintained. Parameter calibration driven by the periodic fluctuation benchmark value eliminates waveform phase distortion during the fine-tuning process, accelerates model convergence and suppresses performance fluctuations.

[0049] These and other aspects of the present invention will become more readily apparent from the following description of the embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following is a brief introduction to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0051] Figure 1 A flowchart of a dynamic pruning and fine-tuning method for a large aviation model based on hierarchical stability analysis provided by the present invention is shown;

[0052] Figure 2The present invention shows a structural diagram of a large aviation model dynamic pruning and fine-tuning system based on hierarchical stability analysis;

[0053] Figure 3 A schematic structural diagram of a computing device provided by the present invention is shown. DETAILED DESCRIPTION

[0054] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0055] In some of the processes described in the specification and claims of the present invention and the above-mentioned figures, multiple operations that appear in a specific order are included, but it should be clearly understood that these operations may not be executed in the order in which they appear in this article or may be executed in parallel. The serial numbers of the operations, such as 101, 102, etc., are only used to distinguish between different operations, and the serial numbers themselves do not represent any execution order. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions of "first", "second", etc. in this article are used to distinguish different messages, devices, modules, etc., and do not represent the order of precedence, nor do they limit "first" and "second" to be different types.

[0056] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative efforts shall fall within the scope of protection of the present invention.

[0057] To address the challenges of balancing inference efficiency and diagnostic performance faced by large-scale deep learning model deployment in aircraft engine fault diagnosis, existing technologies achieve model lightweighting through dynamic freezing of attention weights. However, these technologies rely on static thresholds to determine layer importance, which can easily lead to mispruning or underpruning. Furthermore, their fine-tuning compensation mechanism only acts on local adjacent layers, making it difficult to effectively model changes in the overall topology. This results in large fluctuations in model performance and slow convergence after pruning, making it difficult to meet the aviation industry's requirements for high real-time performance and robustness. To this end, the present invention first constructs a fault mode coupling map based on historical fault data to generate a hierarchical topology for the large aviation model, characterizing the functional dependencies between layers at a structural level. Secondly, through sensitivity energy evaluation, it dynamically identifies energy-stable layers with low energy contributions within a preset training cycle. Thirdly, by setting a critical threshold for vibration energy levels, it further screens out layers that change slightly during dynamic processes and can be removed, avoiding the misjudgment problem caused by traditional fixed thresholds. Finally, after pruning, the temporal feature encoding of the removed layers is fused to the adjacent retained layers, and compensation fine-tuning is performed to restore the model's expressive power from a global perspective, improving the stability and diagnostic accuracy of the pruned layers. Figure 1 The present invention provides a flowchart of a method for dynamic pruning and fine-tuning of a large aviation model based on hierarchical stability analysis, as shown in FIG. Figure 1 As shown, the method includes:

[0058] Step 101: construct a fault mode coupling map based on historical fault data of the aircraft engine to generate a hierarchical correlation topology of the aviation large model;

[0059] In this step, the fault mode coupling graph refers to a topological network composed of nodes (failure modes) and weighted edges (bidirectional connections) based on historical aircraft engine failure data, with edge weights greater than 0.35. The aviation macro model refers to a deep learning model used for engine fault prediction, whose hierarchical structure includes an input layer, hidden layers, and an output layer. The hierarchical association topology refers to the hierarchical connection rules formed by mapping the fault mode coupling graph to the aviation macro model, including [failure mode, model-specific level] and [edges in the fault mode coupling graph, bidirectional data paths between levels].

[0060] In an embodiment of the present invention, based on historical fault data of aircraft engines, periodically recurring feature segments are identified through waveform feature analysis, and a corresponding fault mode is defined for each feature segment; the frequency of multiple fault modes occurring simultaneously in the same time interval is detected as the inter-mode interaction strength value; a node relationship table is created based on the inter-mode interaction strength value, and a bidirectional connection relationship is established for fault mode pairs that exceed a preset connection threshold to form a fault modal coupling map; finally, the fault modes in the map are associated with specific levels of the aircraft large model, and the levels are connected according to the bidirectional connection relationship to generate a hierarchical association topology.

[0061] Step 102: Calculate a sensitivity energy evaluation value based on the hierarchical association topology, and determine the layers in the aviation large model within a preset training period that are lower than the sensitivity energy evaluation value as energy stable layers;

[0062] In this step, the sensitivity energy evaluation value refers to the distribution median of all associated hierarchical response values generated by superimposing the arithmetic mean of the parameter offset amplitude in the associated hierarchical group parameters and the parameter offset amplitude, which is used to determine the energy stability layer.

[0063] In an embodiment of the present invention, according to the hierarchical association topology, the parameter offset amplitude of each associated layer during the training process is calculated; the target associated layer directly connected to each associated layer is determined to form an associated layer group; the mean parameter offset amplitude of all layers in the associated layer group is calculated as the inter-layer influence factor; the parameter offset amplitude of each layer and its inter-layer influence factor are superimposed to generate a post-associated layer response value; the median of the distribution of all response values is taken as the sensitivity energy evaluation value; and finally, the layer with a response value lower than the evaluation value is determined as the energy stable layer.

[0064] Step 103: Setting a critical threshold value of the vibration energy level. When the energy cumulative change rate of the energy stabilization layer meets the critical threshold value of the vibration energy level, marking the energy stabilization layer in the hierarchical association topology of the large aviation model as a removal layer.

[0065] In this step, the critical vibration energy level threshold refers to the product of the aircraft engine's vibration propagation attenuation coefficient and the safe vibration energy, which is used to determine the removal layer. Temporal feature encoding refers to the reconstructed waveform segments of three consecutive training cycles in the removal layer.

[0066] In an embodiment of the present invention, a material property database is queried to obtain a vibration propagation attenuation coefficient corresponding to an aircraft engine model; parameter values of an energy stabilization layer at the start and end times of a training cycle are recorded, and the absolute value of the parameter change is calculated and divided by the total number of cycles to obtain an energy cumulative change rate; safe vibration energy corresponding to the same model is extracted from a safe vibration energy table; the attenuation coefficient is multiplied by a reference value to obtain a critical threshold value of the vibration energy level; and when the energy cumulative change rate is less than the threshold value, the energy stabilization layer is marked as a removal layer in the hierarchical association topology.

[0067] Step 104: adding the temporal feature code of the removed layer to the adjacent retained layer to obtain the added retained layer, performing fine-tuning compensation processing on the added retained layer to obtain the aerial model after dynamic pruning and fine-tuning;

[0068] In an embodiment of the present invention, waveform segments of three consecutive training cycles in the removed layer are reorganized into timing feature codes; the adjacent retained layer is located according to the bidirectional connection relationship of the hierarchical association topology; the first position information of the removed layer data input moment is obtained, and the second position information of the same moment is matched in the adjacent retained layer input sequence; the timing feature code is inserted at the second position to form the adjacent retained layer after insertion; the waveform time axis parameters of the inserted layer are adjusted according to the periodic fluctuation reference value of the historical fault data until the absolute difference between the scaled actual period value calculated according to the periodic fluctuation reference value and the periodic fluctuation reference value does not exceed the preset difference, thereby completing fine-tuning compensation.

[0069] The fault modal coupling map of the embodiment of the present invention achieves deep binding between the physical characteristics of aviation faults and the model structure, breaking through the technical barriers of general pruning methods. Specifically, the hierarchical association topology inherits the engine vibration conduction path to avoid feature decoupling caused by pruning; the critical threshold of the vibration energy level integrates the material attenuation characteristics and the safety vibration energy to solve the problem of false pruning or under-pruning caused by the static threshold; the position alignment and period calibration of the timing feature encoding eliminate slow fine-tuning convergence and performance fluctuations.

[0070] The present invention provides a specific embodiment, step 102, constructing a fault mode coupling map based on historical aircraft engine fault data to generate a hierarchical association topology of an aircraft large model, specifically comprising the following steps:

[0071] Step 201: performing waveform feature analysis on historical fault data of aircraft engines, identifying periodically recurring feature segments, and defining a corresponding fault mode for each feature segment;

[0072] In this step, a characteristic segment refers to the smallest identifiable unit of periodically repetitive waveforms in historical fault data, including amplitude peak-valley sequences (e.g., three consecutive peaks separated by 8.3ms) and energy density characteristics (i.e., the integral value of vibration energy per unit time). Fault modes refer to typical aircraft engine failure types defined based on characteristic segments, including mechanical failures (e.g., blade cracks, bearing pitting) and thermal failures (e.g., combustion chamber oscillation, turbine overheating).

[0073] In an embodiment of the present invention, waveform feature analysis is performed on historical fault data of an aircraft engine (such as vibration sensor time series records). Specifically, the waveform data is scanned through a sliding window to identify characteristic segments whose amplitude changes conform to a periodic repetition pattern, such as a sequence of three consecutive amplitude peaks and valleys. Based on the aircraft engine fault feature library, a corresponding fault mode is defined for each identified characteristic segment. For example, a characteristic segment with an amplitude peak interval of 8.3ms is defined as a blade crack failure mode.

[0074] Step 202: Detecting multiple fault modes occurring simultaneously within the same time interval to calculate the interaction strength value between the modes;

[0075] In this step, the inter-mode interaction strength value refers to the parameter that quantifies the degree of fault coupling. The inter-mode interaction strength value = the number of co-occurrence windows of fault modes A and B ÷ the total number of observation windows.

[0076] In this embodiment of the present invention, multiple fault modes that occur simultaneously within the same time interval (such as a 10-second standard diagnostic window) are detected, and the number of co-occurrences of each pair of fault modes in the historical data is counted. The number of co-occurrences is divided by the total number of time windows to obtain the inter-mode interaction strength value. For example, if blade cracks and bearing wear co-occur 38 times in 100 windows, the inter-mode interaction strength value is 0.38.

[0077] Step 203: creating a node relationship table based on the inter-mode interaction strength value, and establishing a bidirectional connection relationship for the fault mode pairs whose inter-mode interaction strength values in the node relationship table exceed a preset connection threshold, so as to construct a fault mode coupling graph;

[0078] In this step, the node relationship table is a two-dimensional table whose row and column indices are fault modes. Each cell stores the inter-mode interaction strength value, and the diagonal is forced to zero (to exclude self-connections). The preset connection threshold refers to the critical value for vibration energy transmission in aircraft engines (default 0.35), which is derived from the minimum vibration transmission energy of titanium alloy blades. A fault mode pair refers to two possible coupled fault combinations (e.g., blade cracks and bearing wear) that must satisfy the inter-mode interaction strength value greater than the preset connection threshold. A bidirectional connection relationship refers to the physical transmission path between fault modes (e.g., vibration transmission from compressor blades to bearings). The edge weight equals the inter-mode interaction strength value.

[0079] In an embodiment of the present invention, a node relationship table is created. Specifically, a two-dimensional table is constructed in which both row and column headers are fault modes, and the inter-mode interaction strength values are filled into corresponding cells. Non-diagonal cells are scanned, and when the inter-mode interaction strength value exceeds a preset connection threshold (such as the lower limit of vibration transmission of aviation materials of 0.35), a weighted bidirectional connection relationship is established between the fault mode pairs corresponding to the row and column headers (such as blade cracks and bearing wear). Finally, all nodes and edges are collected to generate a fault mode coupling map.

[0080] Step 204: Associating each fault mode in the fault mode coupling graph with a specific level of the aviation model to obtain an associated level, and connecting the associated levels according to the bidirectional connection relationship to generate a hierarchical association topology of the aviation model;

[0081] In this step, the specific layer refers to the network layer in the large aviation model that matches the fault function (for example, the vibration feature extraction layer corresponds to the blade crack pattern). The associated layer refers to the model layer that completes the fault mode mapping (for example, the blade crack pattern corresponds to the third layer), retaining the original layer parameter structure.

[0082] In an embodiment of the present invention, the fault modes in the fault modal coupling map are associated with specific levels of the aviation model, such as adding the blade crack mode to the third convolutional layer; based on the bidirectional connection relationship of the fault modal coupling map, such as the connecting line between blade cracks and bearing wear, data connection channels with the same weight are established between the associated levels, such as the bidirectional connection between the third layer and the fifth layer, to generate a hierarchical association topology.

[0083] The embodiment of the present invention converts the coupling relationship of engine physical faults into a computable topological structure through waveform feature analysis and quantification of the interaction intensity between modes, avoiding pruning that destroys the multi-fault correlation characteristics; the strict binding of fault modes and model levels ensures that the pruned model retains the fault response mechanism unique to aviation scenarios.

[0084] For example, in the deployment of a predictive maintenance system for a commercial turbofan engine, historical fault data was analyzed to identify a characteristic segment with an amplitude peak interval of 8.3ms±5%, which was defined as a blade crack fault mode; a characteristic segment with a temperature rise followed by a slow drop was identified and defined as a combustion chamber oscillation fault mode; within 10,000 10-second time windows, the number of co-occurrences of blade cracks and combustion chamber oscillations was counted (1,200 times); the interaction strength value between patterns was calculated (1200 / 10000 = 0.12); and a node relationship table was created, where The row title is Blade Crack, the column title is Combustion Chamber Oscillation, and the cell value is 0.12. Because 0.12 is less than the preset connection threshold of 0.35, no connection line is established. A bidirectional connection with a weight of 0.42 is established for blade cracks and bearing wear (the interaction strength value between modes is 0.42>0.35). Finally, the blade crack fault mode is associated with the third layer of the aviation large model, and the bearing wear fault mode is associated with the fifth layer. Based on the atlas connection, a bidirectional data channel is established between the third and fifth layers, and the hierarchical association topology is output to guide subsequent dynamic pruning.

[0085] The present invention provides a specific embodiment, step 203, creating a node relationship table based on the inter-mode interaction strength value, establishing a bidirectional connection relationship for the fault mode pairs in the node relationship table whose inter-mode interaction strength value exceeds a preset connection threshold, so as to construct a fault mode coupling map, specifically comprising the following steps:

[0086] Step 211: creating a two-dimensional table containing all failure modes, wherein the row and column headers of the two-dimensional table correspond to failure modes, and each cell where a row and a column intersect is filled with an interaction strength value between modes;

[0087] In this step, the two-dimensional table refers to a matrix data container whose row and column structures are both fault modes. The row-column intersection cell refers to the data unit at the intersection of a row and a column in the two-dimensional table.

[0088] In an embodiment of the present invention, a two-dimensional table containing all fault modes is created. Specifically, an N×N table (N is the total number of fault modes) is constructed based on the fault mode list; both row and column headers are set to the fault mode name, such as the row header is blade crack and the column header is bearing wear; for each row and column intersection cell, such as the intersection of the blade crack row and the bearing wear column, the interaction strength value between the modes (such as 0.42) is filled; in particular, diagonal cells with the same row or column (such as the blade crack row and the blade crack column) are forcibly set to zero to eliminate self-connection of the fault mode.

[0089] Step 212: When the inter-mode interaction strength value corresponding to the non-diagonal cell in the two-dimensional table is greater than a preset connection threshold, establish an undirected connection line with a weight between the fault mode corresponding to the row header and the fault mode corresponding to the column header, where the weight is the inter-mode interaction strength value;

[0090] In an embodiment of the present invention, when the inter-mode interaction strength value in a two-dimensional table is greater than a preset connection threshold, an undirected connection line is established. Specifically, all non-diagonal cells (such as blade crack rows and bearing wear column cells) are scanned; it is determined whether the corresponding inter-mode interaction strength value is greater than a preset connection threshold, which defaults to 0.35; when the condition of being greater than the preset connection threshold is met, a weighted undirected connection line is established between the row header fault mode and the column header fault mode, where the line width = inter-mode interaction strength value × 10 pixels, and the weight value directly adopts the inter-mode interaction strength value in the cell.

[0091] Step 213: using the fault modes as nodes and the weighted undirected links as edges, and combining the nodes and edges to generate a fault mode coupling graph;

[0092] In this step, the undirected connection lines are visualization carriers of the bidirectional conduction relationships between fault modes.

[0093] In an embodiment of the present invention, nodes and edges are combined to generate a fault modal coupling graph. Specifically, each fault mode (such as blade cracks and bearing wear) is converted into a graph node, where the node diameter corresponds to the historical occurrence rate of the fault mode; all weighted undirected connecting lines are aggregated as edges, such as the weight of blade cracks and bearing wear edges is 0.42; the nodes are partitioned and arranged according to the type of engine component, such as the compressor fault mode group is in the upper part and the turbine fault group is in the lower part. Finally, a fault modal coupling graph containing node positions, edge connection relationships and weights is output.

[0094] The embodiment of the present invention uses a matrix representation of a two-dimensional table to convert implicit fault interactions (such as blade cracks causing bearing wear) into visual weighted connection lines, solving the problem of ambiguous expression of coupling relationships in the existing technology. The preset connection threshold is bound to the vibration conduction characteristics of the engine material to ensure that the map only retains valid connections that comply with the laws of aviation physics.

[0095] The present invention provides a specific embodiment, step 102, calculating a sensitivity energy evaluation value based on the hierarchical association topology, and determining a layer in the aviation large model within a preset training period that is lower than the sensitivity energy evaluation value as an energy stable layer, specifically includes the following steps:

[0096] Step 201: Calculating the parameter offset amplitude of each associated layer in the aviation large model during the training process within a preset training period, and calculating the inter-layer influence factor according to the parameter offset amplitude;

[0097] In this step, the post-association layer refers to associating the fault mode to a specific layer of the aviation model, preserving the original layer parameter structure and carrying the physical characteristics of the fault. The parameter offset amplitude refers to the average absolute value of the change in layer weights during the training process. The inter-layer impact factor is the physical quantification of the average parameter offset amplitude, which represents the average propagation strength of the fault between topologically connected layers.

[0098] In an embodiment of the present invention, the parameter offset amplitude of each associated layer in the aviation large model during the training process within a preset training cycle is calculated. Specifically, a weight parameter snapshot of each associated layer in continuous training iterations is first obtained, and the absolute difference of the parameters between adjacent iterations is calculated. After accumulating all absolute differences within a preset cycle (such as 100 iterations), the average value is taken and used as the parameter offset amplitude of the layer, such as the offset amplitude of the third layer = 0.02; the mean of all parameter offset amplitudes is calculated to obtain the inter-level influence factor. Specifically, the parameter offset amplitude values of all associated layers in the associated layer group are extracted, such as the second layer = 0.03, the fifth layer = 0.04), and the arithmetic mean is calculated, such as (0.03+0.04)÷2=0.035, and the arithmetic mean is used as the inter-level influence factor.

[0099] Step 202: Determine the target post-association level to which each post-association level is connected according to the hierarchical association topology to generate an association level group;

[0100] In this step, the target post-association level refers to the directly adjacent levels identified by the bidirectional connection relationship in the level association topology, reflecting the fault transmission path. The associated level group refers to the set of all target post-association levels centered on the current level and is used to quantify the fault transmission impact.

[0101] In an embodiment of the present invention, by parsing the bidirectional connection relationship in the hierarchical association topology, such as the direct connection between the 3rd layer and the 5th layer, all layers that have a direct physical connection with the current associated layer are included in the associated hierarchical group. For example, the connection target of the 3rd layer is the 2nd and 5th layers, forming a fault conduction impact area, that is, generating an associated hierarchical group, such as the 3rd layer associated hierarchical group = [layer 2, layer 5].

[0102] Step 203: superimposing the parameter offset amplitude of each associated level with the inter-level influencing factor to obtain multiple associated level response values, and using the distribution median of all associated level response values as the sensitivity energy evaluation value;

[0103] In this step, the post-correlation hierarchical response value refers to the comprehensive sensitivity index generated by the superposition formula. The distribution median refers to the middle value (for odd-numbered hierarchical levels) or the average of the two middle values (for even-numbered hierarchical levels) after sorting all post-correlation hierarchical response values by size, which is used as the global sensitivity benchmark.

[0104] In an embodiment of the present invention, the parameter offset amplitude of each associated level (such as 0.02 for the third level) and its inter-level influence factor (such as 0.035) are weighted and superimposed in a fixed ratio (such as 0.7:0.3) to generate an associated level response value, such as 0.02×0.7+0.035×0.3=0.023; all the associated level response values are sorted, and the middle value (when it is an odd-numbered level) or the average of the two middle values (when it is an even-numbered level) is taken as the sensitivity energy evaluation value, such as the sixth response value of 0.031 among the 11 levels.

[0105] Step 204: determining the associated layer whose response value of the associated layer is lower than the sensitivity energy evaluation value as an energy stable layer;

[0106] In this embodiment of the present invention, the response value of each associated layer is compared with the sensitivity energy evaluation value (such as 0.031). When the response value of the associated layer is lower than the sensitivity energy evaluation value (such as 0.023<0.031 for the third layer), the layer is determined to be an energy stable layer, that is, a layer with low contribution to fault conduction.

[0107] The embodiments of the present invention quantify the propagation intensity of faults in the topological path by associating hierarchical groups, thereby solving the problem that traditional pruning destroys the fault propagation chain; and uses the median of the distribution to generate a scenario-adaptive sensitivity energy evaluation value to avoid false pruning (retaining the critical fault layer) or under-pruning (not removing the redundant layer) caused by static thresholds.

[0108] The present invention provides a specific embodiment, step 103, setting a critical threshold value of a vibration energy level. When the cumulative energy change rate of the energy stabilization layer meets the critical threshold value of the vibration energy level, the energy stabilization layer in the hierarchical association topology of the large aviation model is marked as a removal layer, specifically comprising the following steps:

[0109] Step 301: querying a preset material property database to obtain a vibration propagation attenuation coefficient corresponding to the aircraft engine model;

[0110] In this step, the pre-set material property database refers to a structured dataset that stores the physical properties of aircraft engine materials. This includes fields such as model identification, component material composition, and vibration propagation attenuation coefficient. This data originates from airworthiness certification reports and can be directly mapped to the material attenuation coefficient through the model identification. The vibration propagation attenuation coefficient refers to the energy loss rate of vibration waves when propagating through specific engine materials, reflecting the natural attenuation of vibration energy in components such as blades.

[0111] In an embodiment of the present invention, a material property database storing physical properties of aviation materials is retrieved according to the aircraft engine model identification to extract the vibration propagation attenuation coefficient corresponding to the material of the core component of the model (such as the compressor blade).

[0112] Step 302: Record the weight parameters corresponding to the start and end times of the energy stabilization layer at the preset training cycle to calculate the total parameter change, and generate the energy cumulative change rate based on the total number of training cycles;

[0113] In this step, the total parameter change refers to the cumulative sum of the absolute values of the layer weight changes within the preset training cycle. It is calculated by accumulating the absolute differences between the weight parameters of adjacent iterations. The cumulative energy change rate refers to the average parameter fluctuation intensity per training cycle. It is calculated by dividing the total parameter change by the total number of training cycles.

[0114] In an embodiment of the present invention, a set of weight parameters of the energy stabilization layer is obtained at the beginning and end of a preset training cycle (such as 100 iterations), and the absolute difference of the weight parameters between adjacent iterations (i.e., the absolute value of the weight difference between the kth and k-1th iterations) is calculated; the absolute differences in all training cycles are accumulated to obtain the total parameter change; the total parameter change is divided by the total number of training cycles to obtain the energy cumulative change rate.

[0115] Step 303: extracting the safety vibration energy corresponding to the aircraft engine model from a preset safety vibration energy table;

[0116] In this step, the pre-set safe vibration energy table maps aircraft engine models to safe vibration energy. It contains three columns: model identification, sensor type, and reference value. The reference value is derived from airworthiness standards. Safe vibration energy refers to the upper limit of the engine's permissible vibration energy at maximum thrust. It is calibrated through airworthiness certification testing and is used to ensure flight safety.

[0117] In an embodiment of the present invention, according to the aircraft engine model identification, the corresponding model row is located in the safety vibration energy table, where the fields in the safety vibration energy table are [model, sensor type, reference value]; and the safety vibration energy under rated operating conditions is extracted.

[0118] Step 304: Multiplying the vibration propagation attenuation coefficient by the safe vibration energy to obtain a critical vibration energy level threshold. When the energy cumulative change rate is less than the critical vibration energy level threshold, marking the energy stable layer in the hierarchical association topology of the large aviation model as a removal layer.

[0119] In this embodiment of the present invention, the vibration propagation attenuation coefficient is multiplied by the safe vibration energy to obtain a critical vibration energy level threshold. When the cumulative energy change rate is less than the critical vibration energy level threshold, the energy-stable layer in the hierarchical topology of the large aviation model is marked as a removal layer. For example, if the vibration propagation attenuation coefficient is 0.35, the safe vibration energy is 120, and the critical vibration energy level threshold = 0.35 × 120 = 42, which exceeds the cumulative energy change rate of 38, the energy-stable layer is marked as a removal layer.

[0120] The embodiment of the present invention replaces the artificial experience threshold with the critical threshold of the vibration energy level driven by the physical characteristics of the engine through dynamic calculation of the material attenuation coefficient and the safe vibration energy, thereby solving the problem that the general pruning method is divorced from aviation reality; based on the comparison mechanism of parameter fluctuation intensity and physical threshold, it ensures that only the layers where the vibration energy conduction is lower than the safety limit are removed, avoiding the accidental deletion of critical fault layers.

[0121] The present invention provides a specific embodiment, step 104, adding the temporal feature code of the removed layer to the adjacent retained layer to obtain the added retained layer, performing fine-tuning compensation processing on the added retained layer to obtain a large aerial model after dynamic pruning and fine-tuning, specifically including the following steps:

[0122] Step 401: reorganizing waveform units of three consecutive training cycles in the waveform segment of the removed layer into a temporal feature code;

[0123] In this step, the waveform segment refers to a continuous segment with complete periodic characteristics in the removal layer weight parameter change curve, and the duration is equal to the typical fault vibration period of the engine.

[0124] In an embodiment of the present invention, three consecutive complete cycles of waveform units (such as an amplitude peak-valley sequence) are extracted from the waveform segment of the removed layer; the three waveform units are smoothly connected end to end with the peak position as the zero phase point to form a standardized time series feature code.

[0125] Step 402: Determine, based on the bidirectional connection relationship of the layer association topology, the layer connected to the removed layer as an adjacent retained layer;

[0126] In an embodiment of the present invention, the layer identifier that has a direct bidirectional connection with the removed layer in the hierarchical association topology is read; if the removed layer is connected to both the upstream layer and the downstream layer in the hierarchical association topology, both are treated as adjacent retained layers, such as [layer 2, layer 4].

[0127] Step 403: Acquire first position information of the removed layer at the data input moment within a preset training cycle, and locate second position information identical to the first position information in the input sequence of the adjacent retained layer;

[0128] In this step, the data input time refers to the iteration index (e.g., iteration 1500) at which the data batch enters the model layer during training, recorded in the timestamp field of the training log. The first position information specifically refers to the index value of the time in the training sequence when the removed layer receives input data, which is used as a spatiotemporal alignment benchmark. The input sequence refers to the time-ordered data stream received by the adjacent retained layer during training, with the structure [iteration 1 data, iteration 2 data, ..., iteration N data]. The second position information refers to the time index position in the input sequence of the adjacent retained layer corresponding to the first position information.

[0129] In an embodiment of the present invention, the time index at which the removed layer receives input data within a preset training cycle (such as 1500 iterations) is extracted from the training log to obtain first position information; the same time index position is searched for in the input sequence of the adjacent retained layer, such as the 1500th position of the second layer input sequence, and finally the second position information that is the same as the first position information is located.

[0130] Step 404: inserting the temporal feature code into the second position information to obtain an adjacent retained layer after insertion;

[0131] In this step, the adjacent retained layer after insertion refers to the hierarchical state after the injection of the temporal feature code is completed, wherein 60% of the original parameters are retained, 40% of the injected temporal feature code is injected, and the time axis parameters are to be calibrated.

[0132] In the embodiment of the present invention, a temporal feature code is inserted at the second position information of the input sequence of the adjacent reserved layer, wherein the original parameters of the reserved layer and the inserted temporal feature code are weighted and superimposed in a ratio of 0.6:0.4 to form the adjacent reserved layer after insertion.

[0133] Step 405: adjusting the waveform time axis parameters of the adjacent retained layer after insertion according to the periodic fluctuation reference value of the historical fault data of the aircraft engine, so as to perform fine-tuning compensation processing on the retained layer after insertion, thereby obtaining a large aircraft model after dynamic pruning and fine-tuning;

[0134] In this step, the periodic fluctuation baseline refers to the standard period length of a specific aircraft engine failure mode in historical failure data. Waveform timeline parameters include time delay parameters and frequency response parameters. The time delay parameter controls the characteristic response timing (such as the convolution layer delay coefficient), while the frequency response parameter determines the feature extraction frequency band (such as the filter cutoff frequency).

[0135] In an embodiment of the present invention, a periodic fluctuation reference value of the same fault mode is extracted from historical fault data; the waveform sequence of the inserted retained layer under the action of a standard vibration test signal is detected to obtain an actual periodic value, and the periodic scaling factor is calculated in combination with the periodic fluctuation reference value to scale the waveform time axis parameters of the adjacent retained layer after insertion until the absolute difference between the actual periodic value and the periodic fluctuation reference value after scaling does not exceed a preset difference, thereby generating a large aviation model after dynamic pruning and fine-tuning.

[0136] The embodiments of the present invention use waveform reorganization and timing code injection to accurately migrate the fault features of the removed layer to the adjacent layer, solving the feature fragmentation problem caused by traditional pruning. Time axis parameter calibration based on aviation physics eliminates feature phase offset and speeds up fine-tuning convergence. Adjacent layer positioning driven by bidirectional connection relationships maintains the integrity of the fault conduction path.

[0137] The present invention provides a specific embodiment, step 405, adjusting the waveform time axis parameters of the adjacent retained layer after insertion based on the periodic fluctuation reference value of the historical fault data of the aircraft engine to perform fine-tuning compensation processing on the retained layer after insertion, thereby obtaining a large aircraft model after dynamic pruning and fine-tuning, specifically comprising the following steps:

[0138] Step 411: selecting fault waveform samples with the same fault mode as the removed layer from the historical fault data, extracting an average time value of a preset peak interval from the fault waveform samples, and using the average time value as a periodic fluctuation reference value;

[0139] In this step, the fault waveform sample refers to the waveform segment in the historical fault data that strictly matches the fault mode of the removed layer and contains ≥3 complete cycles. The preset peak interval refers to the set number of peaks used when calculating the cycle fluctuation baseline value (default is 10), which is based on the minimum sample size standard for aviation fault diagnosis.

[0140] In an embodiment of the present invention, fault waveform samples with the same fault mode as that associated with the removed layer are selected from historical fault data; a preset number (e.g., 10) of peak interval values are extracted; and the arithmetic mean of these peak interval values is calculated to generate an average time value as a reference value for periodic fluctuations.

[0141] Step 412: Obtain a waveform sequence of adjacent retained layers after insertion under the action of a standard vibration test signal, detect the time difference between adjacent main peaks in the waveform sequence to calculate an actual period value, and calculate a period proportional factor in combination with the period fluctuation reference value;

[0142] In this step, the standard vibration test signal refers to an aviation excitation signal that meets standard specifications. The waveform sequence refers to the time-ordered output data stream of the adjacent retained layers after insertion under the action of the standard vibration test signal, with a sampling rate ≥ 2 times the fault frequency. Adjacent main peaks refer to extreme points in the waveform sequence that meet two conditions: amplitude > sequence maximum × 0.7 and greater than the amplitude values of the previous and next points. The time difference refers to the time coordinate difference between adjacent main peaks. The actual period value refers to the arithmetic mean of N consecutive time differences (N ≥ 5), reflecting the average period length of the output waveform. The period scale factor refers to the ratio of the period fluctuation reference value to the actual period value, retained to three decimal places to meet the accuracy requirements of the aviation controller.

[0143] In an embodiment of the present invention, a standard vibration test signal is input into the adjacent retained layer after insertion, and the output waveform sequence is recorded, the positions of adjacent main peaks in the waveform sequence are detected, and the average value of the time difference of 5 consecutive main peaks is taken as the actual period value; the period fluctuation reference value is divided by the actual period value, and the period scale factor is output with three decimal places retained.

[0144] Step 413: Scaling the waveform time axis parameters of the adjacent retained layer after insertion according to the period scale factor to recalculate the actual period value after scaling. When the absolute difference between the actual period value after scaling and the period fluctuation reference value does not exceed a preset difference, the fine-tuning compensation process is confirmed to be completed, thereby obtaining the large aviation model after dynamic pruning and fine-tuning.

[0145] In this step, the actual period value after scaling refers to the remeasured period length after parameter adjustment. It must satisfy the condition |actual period value after scaling - period fluctuation reference value| ≤ the preset difference. The preset difference refers to the upper limit of period deviation permitted by aviation airworthiness standards and is derived from the material thermal expansion coefficient and sensor accuracy limits.

[0146] In an embodiment of the present invention, the specific scaling process includes: multiplying the time delay parameter of the waveform time axis parameter by the period scale factor, and dividing the frequency response parameter by the period scale factor to obtain the inserted adjacent retained layer after adjusting the waveform time axis parameter; re-inputting the same standard vibration test signal into the inserted adjacent retained layer, and measuring the actual period value after scaling; calculating the absolute difference between the actual period value after scaling and the reference value, and when the absolute difference is ≤ a preset difference (such as 0.4ms is allowed for a blade crack mode), completing the fine-tuning compensation process to generate a large aviation model after dynamic pruning and fine-tuning.

[0147] The embodiment of the present invention extracts periodic fluctuation reference values through same-mode samples, eliminates cross-mode interference, and solves the problem of feature confusion and distortion in traditional fine-tuning; the parameter classification and scaling driven by the periodic scale factor accurately matches the engine vibration response mechanism, reducing the number of fine-tuning iterations; the preset difference is bound to the aviation airworthiness tolerance standard to ensure that the model output after calibration meets flight safety specifications.

[0148] Figure 2 The present invention provides a structural diagram of a large-scale aviation model dynamic pruning and fine-tuning system based on hierarchical stability analysis, as shown in FIG. Figure 2 As shown, the system includes:

[0149] A construction module 21 is used to construct a fault mode coupling map based on historical fault data of the aircraft engine to generate a hierarchical correlation topology of the aviation large model;

[0150] A calculation module 22 is configured to calculate a sensitivity energy evaluation value based on the hierarchical association topology, and determine a layer in the aviation large model within a preset training period that is lower than the sensitivity energy evaluation value as an energy stable layer;

[0151] a marking module 23 for setting a critical threshold value of a vibration energy level, and marking the energy stabilization layer in the hierarchical association topology of the large aviation model as a removal layer when the energy cumulative change rate of the energy stabilization layer meets the critical threshold value of the vibration energy level;

[0152] The fine-tuning module 24 is configured to add the temporal feature code of the removed layer to an adjacent retained layer to obtain an added retained layer, and perform fine-tuning compensation processing on the added retained layer to obtain a large aerial model after dynamic pruning and fine-tuning.

[0153] Figure 2 The dynamic pruning and fine-tuning system for large aviation models based on hierarchical stability analysis can be executed Figure 1The implementation principles and technical effects of the illustrated embodiment of a dynamic pruning and fine-tuning method for large aircraft models based on hierarchical stability analysis are not further elaborated. The specific manner in which each module and unit performs operations in the aforementioned embodiment of a dynamic pruning and fine-tuning system for large aircraft models based on hierarchical stability analysis has been described in detail in the related embodiments and will not be further elaborated here.

[0154] In one possible design, Figure 2 The embodiment shown is a large-scale aviation model dynamic pruning and fine-tuning system based on hierarchical stability analysis, which can be implemented as a computing device, such as Figure 3 As shown, the computing device may include a storage component 31 and a processing component 32;

[0155] The storage component 31 stores one or more computer instructions, wherein the one or more computer instructions are called and executed by the processing component 32 .

[0156] The processing component 32 is used for the above Figure 1 The embodiment provides a dynamic pruning and fine-tuning method for a large aviation model based on hierarchical stability analysis.

[0157] The processing component 32 may include one or more processors to execute computer instructions to complete all or part of the steps in the above method. Of course, the processing component may also be implemented as one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the above method.

[0158] The storage component 31 is configured to store various types of data to support operations at the terminal. The storage component can be implemented by any type of volatile or non-volatile memory device, or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk, or optical disk.

[0159] Of course, a computing device may also include other components, such as input / output interfaces, display components, communication components, etc.

[0160] The input / output interface provides an interface between the processing component and the peripheral interface module, which can be an output device, an input device, etc.

[0161] The communication component is configured to facilitate, among other things, wired or wireless communications between the computing device and other devices.

[0162] Among them, the computing device can be a physical device or an elastic computing host provided by a cloud computing platform, etc. In this case, the computing device can refer to a cloud server, and the above-mentioned processing components, storage components, etc. can be basic server resources rented or purchased from the cloud computing platform.

[0163] The embodiment of the present invention further provides a computer storage medium storing a computer program, which can achieve the above-mentioned Figure 1 The illustrated embodiment provides a dynamic pruning and fine-tuning method for a large aviation model based on hierarchical stability analysis.

[0164] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0165] The device 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, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0166] Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented using software plus a necessary general-purpose hardware platform, or of course, hardware. Based on this understanding, the essence of the above technical solution, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for causing a computer device (such as a personal computer, server, or network device) to execute the methods described in each embodiment or certain portions of the embodiments.

[0167] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A dynamic pruning and fine-tuning method for large aviation models based on hierarchical stability analysis, characterized in that: include: Based on the historical fault data of aircraft engines, a fault mode coupling map is constructed to generate the hierarchical correlation topology of the aviation large model; Calculating a sensitivity energy evaluation value according to the hierarchical association topology, and determining a layer in the aviation large model within a preset training period that is lower than the sensitivity energy evaluation value as an energy stable layer; Setting a critical threshold value of a vibration energy level, and when the energy cumulative change rate of the energy stabilization layer meets the critical threshold value of the vibration energy level, marking the energy stabilization layer in the hierarchical association topology of the large aviation model as a removal layer; The temporal feature code of the removed layer is added to the adjacent retained layer to obtain the added retained layer, and fine-tuning compensation processing is performed on the added retained layer to obtain the large aviation model after dynamic pruning and fine-tuning.

2. The method according to claim 1, characterized in that Based on historical aircraft engine failure data, a fault modal coupling map is constructed to generate a hierarchical association topology for the large aviation model, including: Analyze the waveform characteristics of historical aircraft engine failure data, identify periodically recurring characteristic segments, and define the corresponding failure mode for each characteristic segment; Detect multiple fault modes that occur simultaneously in the same time interval and calculate the interaction strength value between the modes; Creating a node relationship table according to the inter-mode interaction strength value, establishing a bidirectional connection relationship for the fault mode pairs whose inter-mode interaction strength values in the node relationship table exceed a preset connection threshold, and constructing a fault mode coupling map; Each fault mode in the fault modal coupling graph is associated with a specific level of the aviation model to obtain an associated level, and the associated levels are connected according to the bidirectional connection relationship to generate a hierarchical association topology of the aviation model.

3. The method according to claim 2, characterized in that According to the inter-mode interaction strength value, a node relationship table is created, and for the fault mode pairs whose inter-mode interaction strength values in the node relationship table exceed a preset connection threshold, a bidirectional connection relationship is established to construct a fault mode coupling map, including: Creating a two-dimensional table containing all failure modes, wherein the row headers and column headers of the two-dimensional table correspond to the failure modes, and filling each cell where the rows and columns intersect with the interaction strength value between the modes; When the inter-mode interaction strength value corresponding to the non-diagonal cell in the two-dimensional table is greater than a preset connection threshold, an undirected connection line with a weight is established between the fault mode corresponding to the row header and the fault mode corresponding to the column header, where the weight is the inter-mode interaction strength value; The fault modes are used as nodes, the weighted undirected connecting lines are used as edges, and the nodes and edges are combined to generate a fault mode coupling graph.

4. The method according to claim 1, wherein Calculating a sensitivity energy evaluation value according to the hierarchical association topology, and determining a layer in the aviation large model within a preset training period that is lower than the sensitivity energy evaluation value as an energy stable layer, including: Calculating the parameter offset amplitude of each associated layer in the aviation large model during the training process within a preset training period, and calculating the inter-layer influence factor based on the parameter offset amplitude; Determining, according to the hierarchical association topology, a target post-association level to which each post-association level is connected, and generating an association level group; The parameter offset amplitude of each associated level is superimposed with the inter-level influencing factor to obtain multiple associated level response values, and the distribution median of all associated level response values is used as the sensitivity energy evaluation value; The post-association layer whose response value of the post-association layer is lower than the sensitivity energy evaluation value is determined to be an energy stable layer.

5. The method according to claim 1, characterized in that Setting a critical threshold value of a vibration energy level, and when the energy cumulative change rate of the energy stabilization layer meets the critical threshold value of the vibration energy level, marking the energy stabilization layer in the hierarchical association topology of the large aviation model as a removal layer, including: Query the preset material property database to obtain the vibration propagation attenuation coefficient corresponding to the aircraft engine model; Recording the weight parameters corresponding to the start and end times of the energy stabilization layer at a preset training cycle, calculating the total amount of parameter change, and generating an energy cumulative change rate based on the total number of training cycles; Extracting the safety vibration energy corresponding to the aircraft engine model from a preset safety vibration energy table; The vibration propagation attenuation coefficient is multiplied by the safe vibration energy to obtain a critical threshold value of the vibration energy level. When the energy cumulative change rate is less than the critical threshold value of the vibration energy level, the energy stable layer in the hierarchical association topology of the large aviation model is marked as a removal layer.

6. The method according to claim 1, characterized in that Adding the temporal feature code of the removed layer to an adjacent retained layer to obtain an added retained layer, performing fine-tuning compensation processing on the added retained layer to obtain a large aerial model after dynamic pruning and fine-tuning, including: Recombining waveform units of three consecutive training cycles in the waveform segment of the removed layer into a temporal feature code; Determining, based on the bidirectional connection relationship of the layer association topology, a layer connected to the removed layer as an adjacent retained layer; Acquire first position information of the removed layer at a data input moment within a preset training cycle, and locate second position information identical to the first position information in an input sequence of the adjacent retained layer; Inserting the temporal feature code into the second position information to obtain an adjacent retained layer after insertion; According to the periodic fluctuation reference value of the historical fault data of the aircraft engine, the waveform time axis parameters of the adjacent retained layer after insertion are adjusted, and fine-tuning compensation processing is performed on the retained layer after addition to obtain the large aircraft model after dynamic pruning and fine-tuning.

7. The method according to claim 6, characterized in that According to the periodic fluctuation reference value of the historical fault data of the aircraft engine, the waveform time axis parameters of the adjacent retained layer after insertion are adjusted, and fine-tuning compensation processing is performed on the retained layer after addition to obtain the large aircraft model after dynamic pruning and fine-tuning, including: Selecting a fault waveform sample with the same fault mode as the removed layer from the historical fault data, extracting an average time value of a preset peak interval from the fault waveform sample, and using the average time value as a period fluctuation reference value; Obtaining a waveform sequence of adjacent retained layers after insertion under the action of a standard vibration test signal, detecting a time difference between adjacent main peaks in the waveform sequence, calculating an actual period value, and calculating a period proportional factor in combination with the period fluctuation reference value; According to the period scale factor, the waveform time axis parameters of the adjacent retained layer after insertion are scaled to recalculate the actual period value after scaling. When the absolute difference between the actual period value after scaling and the period fluctuation reference value does not exceed the preset difference, it is confirmed that the fine-tuning compensation processing is completed, and the large aviation model after dynamic pruning and fine-tuning is obtained.

8. A dynamic pruning and fine-tuning system for large aviation models based on hierarchical stability analysis, characterized by: include: A construction module is used to construct a fault mode coupling map based on historical fault data of aircraft engines to generate a hierarchical correlation topology of the aviation large model; a calculation module, configured to calculate a sensitivity energy evaluation value according to the hierarchical association topology, and determine a layer in the aviation large model within a preset training period that is lower than the sensitivity energy evaluation value as an energy stable layer; a marking module, configured to set a critical threshold value of a vibration energy level, and mark the energy stabilization layer in the hierarchical association topology of the large aviation model as a removal layer when the energy cumulative change rate of the energy stabilization layer meets the critical threshold value of the vibration energy level; The fine-tuning module is used to add the temporal feature code of the removed layer to the adjacent retained layer to obtain the added retained layer, perform fine-tuning compensation processing on the added retained layer, and obtain the aerial large model after dynamic pruning and fine-tuning.

9. A computing device, characterized in that The invention comprises a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement a dynamic pruning and fine-tuning method for a large aviation model based on hierarchical stability analysis as described in any one of claims 1 to 7.

10. A computer storage medium, characterized in that A computer program is stored, and when the computer program is executed by a computer, the method for dynamic pruning and fine-tuning of a large aviation model based on hierarchical stability analysis as described in any one of claims 1 to 7 is implemented.

Images