A sample generation method for solid propellant failure identification

Abstract

The application discloses a kind of sample generation methods for solid propellant failure identification.The method includes: interpolation normalization to multi-strain rate tensile sample, extracts conditional characteristics and prior physical characteristics;Input latent distribution extraction variation auto-encoder, complete normal sampling or abnormal disturbance by double-track generation mechanism, decode reverse normalization output initial state curve;Combine initial state curve with prior physical characteristics, execute physical law constraint, mechanics correction and defect modeling in turn, output expanded sample set with physical meaning.The application adopts the series architecture of generation and physical time sequence decoupling, and the variation auto-encoder focuses on learning the diversity of data distribution, and the physical hard constraint is determined by the prior physical characteristics for step-by-step deterministic correction, and the generation is responsible for the diversity of the breadth, and the physics is responsible for the accuracy of mechanics, and the two are not compromised, realizes the mathematical two-way collaborative efficiency, guarantees the mechanics rationality and physical authenticity, fundamentally solves the problem of solid propellant failure sample extreme scarcity.

Description

Technical Field

[0001] This invention belongs to the field of mechanical behavior characterization, failure identification and data enhancement technology of polymer composite materials, and specifically relates to a sample generation method for failure identification of solid propellants. Background Technology

[0002] Solid propellants, as a typical high-fill-ratio composite material, are highly susceptible to failure under the complex temperature and pressure conditions inside engines, leading to dehydration, abnormal envelope, and tensile point shift due to internal matrix tearing or particle dehydration. Accurate identification of these failure modes is crucial for ensuring the storage reliability and operational safety of solid propellant engines. However, propellant failure identification relies on high-quality sample datasets covering normal conditions and multiple failure modes. Due to their high cost and risk, failure samples are extremely difficult to obtain in large quantities through physical testing, resulting in severe class imbalance and small sample size problems for failure identification models, thus limiting their generalization ability and robustness. Therefore, establishing a method for rapidly, massively, and faithfully generating solid propellant samples covering normal and failure conditions under multiple strain rates is a key prerequisite for overcoming the data bottleneck in failure identification and has significant engineering application value.

[0003] In existing technologies, sample generation mainly relies on purely data-driven generative deep learning models. For example, the document with publication number "CN109886388A" discloses "A Method for Expanding Training Sample Data Based on Variational Autoencoders." When processing high-dimensional data features, it mainly uses statistical probability distribution and reconstruction error as representation constraints for the latent space. However, the components of propellant have strict physicochemical coupling constraints. Existing pure data generation methods rely solely on mathematical probability fitting during the dimensionality reduction and reconstruction of the latent space, which cannot characterize the thermodynamic and kinetic relationships between the components. This results in distortion of the physical mechanism and low usability of the generated samples.

[0004] Regarding the introduction of physical constraints, existing deep learning methods that incorporate physical constraints, such as the "Method for Solving Fluid Dynamics Equations Based on Physical Information Neural Networks" disclosed in document CN119989887A, typically only introduce simple explicit linear penalty terms such as basic partial differential equations, mainly used to predict the steady-state characteristics of known materials under normal conditions. However, when generating samples for solid propellants, complex nonlinear physical boundaries such as abrupt changes in the critical burning rate exponent and thermodynamic collapse caused by compatibility conflicts are easily encountered, resulting in the problem that simple linear constraints cannot avoid the risk of nonlinear physical instability.

[0005] In summary, existing methods for generating solid propellant samples suffer from several drawbacks. At the purely data-driven level, they exhibit issues such as distortion of physical mechanisms and low usability of generated samples. At the level of introducing physical constraints, simple linear constraints cannot avoid the risk of nonlinear physical instability. Both of these shortcomings, combined, result in virtual samples that are even less able to accurately reproduce the nonlinear mechanical characteristics of the propellant failure process, thus failing to provide effective training data support for failure identification models. Summary of the Invention

[0006] This invention provides a sample generation method for solid propellant failure identification. By introducing prior physical feature extraction and hard constraints based on physical laws, it aims to solve the problem that existing technologies cannot realistically reproduce the nonlinear mechanical characteristics of the propellant failure process and cannot provide effective training data support for the failure identification model. This enables the generation of expanded solid propellant samples with physical meaning under multiple strain rates.

[0007] To achieve the above objectives, the technical solution provided by this invention is: a sample generation method for solid propellant failure identification, comprising the following steps:

[0008] Step 1, Feature Extraction and Data Preprocessing: Interpolate and normalize the real tensile data under multiple strain rates in sequence, and calculate the condition features; extract the elastic modulus, desiccation point strain, fracture point strain and fracture point stress as a priori physical features, which serve as the absolute judgment criteria for the physical hard constraint module in Step 3.

[0009] Step 2: Generation of the initial state curve of the latent space based on the variational autoencoder:

[0010] The interpolated and normalized data is input into a variational autoencoder to extract the latent distribution. Then, based on this distribution, different latent vectors are generated in the latent space through a dual-track generation mechanism: normal samples are sampled according to the distribution probability, abnormal samples are subjected to directional perturbation, and the latent vectors are concatenated with the conditional features to form a conditional latent vector. After decoding and inverse normalization, the initial state curve with upper and lower boundary envelopes is output.

[0011] Step 3: Post-processing module based on physical hard constraints: The initial state curve with upper and lower boundary envelopes generated in Step 2 is combined with the prior physical features extracted in Step 1. By sequentially executing basic physical law constraints, mechanical performance correction and physical defect feature modeling, the initial state curve is processed step by step, and finally an expanded sample dataset with physical meaning is output.

[0012] Furthermore, the specific implementation of step one above includes:

[0013] First, linear interpolation is performed on the real tensile data of multiple strain rates to resample the strain data of unequal lengths to a fixed length to meet the alignment requirements of the neural network tensors. Second, normalization is performed on the interpolated data to eliminate the dimensional differences between different physical quantities and multiple strain rates. Based on the normalized data, the mean and standard deviation sequences of stress for each working condition are calculated and combined into conditional features. Finally, prior physical features, including elastic modulus, desiccation point strain, fracture point strain, and fracture point stress, are extracted. The above conditional features and prior physical features are saved independently to provide standardized input data and absolute judgment criteria for subsequent processing modules.

[0014] Furthermore, the specific implementation of step two above includes:

[0015] First, the normalized real tensile data with multiple strain rates is input into the encoder network of the variational autoencoder. Through multi-layer feature mapping, the high-dimensional time series is reduced to a continuous low-dimensional latent space, and the potential distribution representing the macroscopic evolution law of the current working condition is extracted.

[0016] Secondly, based on the extracted latent distribution, a dual-track generation mechanism is initiated: for normal samples, standard Gaussian sampling is performed according to the distribution probability to obtain latent vectors; for abnormal samples, random Gaussian perturbation vectors are injected on the basis of the latent vectors of normal samples to complete the directional shift of features.

[0017] Finally, the generated latent vector is concatenated with the conditional features saved in step one along the feature dimension to form a conditional latent vector that integrates macroscopic constraints and microscopic randomness. This vector is then mapped and reconstructed by a decoder network to obtain stress components and standard deviation components. Finally, through inverse normalization calculation, an initial state curve with upper and lower boundary envelopes is generated.

[0018] Furthermore, the specific implementation of step three above includes:

[0019] First, the initial state curve generated by decoding is constrained by basic physical laws. By executing non-negative hard constraints, it is ensured that the material stress during the stretching process conforms to basic physical common sense.

[0020] Secondly, the prior physical features extracted in step one are introduced to correct the mechanical properties. The initial elastic modulus of the pure elastic data segment of the curve is corrected by calculating the truncation scaling factor, and the stress at the fracture point of the entire curve is corrected according to the tolerance range of the prior fracture point stress.

[0021] Next, based on the normal / abnormal sample labels input in step two, the samples are categorized into specific abnormal behavior judgment nodes for sorting: normal samples are skipped directly, while abnormal samples are based on prior physical characteristics and, according to the four preset defect labels including dehumidification point abnormality, tensile break point abnormality, envelope abnormality and composite abnormality, the corresponding empirical mechanical attenuation model or boundary condition constraints are introduced to perform physical defect feature modeling.

[0022] Finally, the upper and lower boundary envelopes are reconstructed using the established stress mean and standard deviation sequences, and the resulting augmented sample dataset with physical meaning is then integrated and output.

[0023] Compared with the prior art, the beneficial effects of the present invention are:

[0024] (1) Dual-track generation mechanism to solve the problem of scarce anomalous samples. Existing sample generation methods based on variational autoencoders only sample from the training set distribution and cannot generate anomalous samples outside the distribution, resulting in a continued scarcity of failed samples. In step two, this invention proposes a "normal sampling - anomalous perturbation" dual-track generation mechanism: normal samples are sampled along the standard of the latent manifold to maintain statistical characteristics, while anomalous samples are injected with random Gaussian perturbations to achieve directional shift of features. The continuity and structure of the latent manifold enable anomalous samples to have real evolutionary laws rather than random noise, fundamentally solving the problem of the extreme scarcity of solid propellant failed samples.

[0025] (2) Multi-level physical hard constraints ensure mechanical rationality. Existing soft constraint methods based on physical information neural networks embed the physical equation residuals into the loss function, causing the physical constraints and reconstruction loss to compete with each other within the same optimization objective, eventually converging to a compromise solution and failing to eliminate local mechanical anomalies. This invention designs the physical constraints as a post-processing hard constraint module independent of the generation process: it sequentially executes basic physical law constraints, mechanical performance correction based on prior physical features, and physical defect feature modeling. The three levels of hard constraints are determined independently and do not interfere with each other, ensuring the point-by-point mechanical rationality of all generated samples throughout the full stretching period.

[0026] (3) Decoupling generation from physical time to achieve synergistic effect between mathematics and physics. Existing methods that combine variational autoencoders with physical constraints adopt a parallel fusion paradigm. The smoothness of the latent distribution and the sharpness of the physical boundary conflict with each other within the same loss function, resulting in a trade-off between generation quality and physical rationality. This invention proposes a time-decoupled architecture: in step two, the variational autoencoder is not affected by physical constraints and focuses on learning the diversity of data distribution; in step three, the physical hard constraints use the prior physical features extracted in step one as the absolute judgment benchmark, are not affected by generation loss, and perform step-by-step deterministic correction of the initial state curve. Generation is responsible for the breadth of diversity, and physics is responsible for mechanical accuracy. The two are linked and do not compromise with each other, achieving a two-way synergistic effect between mathematics and physics that cannot be achieved by a pure parallel fusion paradigm.

[0027] (4) The sample generation idea based on the variational autoencoder dual-track generation mechanism proposed in this invention, combined with the post-processing correction method of prior physical features and hard constraints, is also applicable to generating sample data of other polymer composite materials with similar highly nonlinear mechanical responses and complex failure mechanisms. Attached Figure Description

[0028] Figure 1 This is a block diagram illustrating the implementation of the method of the present invention;

[0029] Figure 2 The above are comparison diagrams before and after the generation of the expanded sample of the present invention. (a) is the distribution diagram of the actual tensile data obtained. (b) is the expanded sample diagram of normal working conditions generated by the method of the present invention. (c), (d), (e), and (f) respectively show the expanded sample diagrams of solid propellants under four working conditions: envelope anomaly, tensile breakage point anomaly, dehydration point anomaly, and composite anomaly. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. Obviously, the described embodiments are only some embodiments of this invention and are used only to illustrate the invention, but are not intended to limit the scope of the invention.

[0031] like Figure 1 As shown, the basic idea of ​​this invention is as follows: First, preprocess the real tensile data with multi-strain rate to extract prior physical features characterizing the real mechanical properties, and extract conditional features to guide generation after interpolation and normalization. Second, learn the latent distribution of the data based on a variational autoencoder, and generate standard sampling of normal samples and random perturbation of abnormal samples in the latent space through a dual-track generation mechanism. Then, concatenate the generated latent vectors and conditional features into a conditional latent vector, and output the initial state curve after decoding, reconstruction, and inverse normalization. Finally, use a post-processing module based on physical hard constraints to sequentially perform basic physical law constraints, mechanical property correction, and specific physical defect modeling on the generated abnormal samples. Finally, integrate all processed stress mean and standard deviation envelopes to output an expanded sample dataset with physical meaning.

[0032] Based on the above design concept, this invention provides a sample generation method for solid propellant failure identification, specifically including the following steps:

[0033] Step 1: For real tensile data under multiple strain rates, perform conditional feature extraction and prior physical feature extraction sequentially:

[0034] First, conditional feature extraction is performed through linear interpolation and normalization: data interpolation is performed by resampling strain data of unequal lengths to a fixed length, resulting in interpolated data; then, the interpolated data is normalized to eliminate dimensional differences, and the corresponding mean stress and standard deviation sequences are calculated based on the normalized data, combined to form conditional features. The specific process is as follows:

[0035] To meet the strict alignment requirements of the input tensor dimension for subsequent neural networks, the acquired real tensile data under multiple strain rates are uniformly resampled to a fixed length using a linear interpolation algorithm. Then, to eliminate the absolute dimensional differences between different physical quantities and multiple strain rates, the stress mean sequence is standardized, and the standard deviation envelope is normalized based on the global maximum value, as shown in the following equation:

[0036] (1)

[0037] in, Represents the normalized i-th Stress values ​​at each node, This represents the interpolated physical quantity. and These represent the statistical mean and standard deviation of the global stress sequence under this working condition, respectively.

[0038] (2)

[0039] in, For the normalized first The standard deviation envelope value of each node This represents the interpolated physical quantity. This represents the global maximum value of the standard deviation envelope under this operating condition. To prevent extremely small positive numbers with a denominator of zero.

[0040] The normalized mean stress sequence and the standard deviation envelope sequence are concatenated to form a conditional feature that reflects the macroscopic mechanical evolution distribution of the material under a specific strain rate.

[0041] Then, prior physical feature extraction is performed: Simultaneously, macroscopic mechanical analysis is conducted on the acquired real tensile data under multiple strain rates to extract prior physical features reflecting the essential properties of the solid propellant material, including: elastic modulus, desiccation point strain, fracture point strain, and fracture point stress. The specific extraction method is as follows:

[0042] The first 10% of data points are extracted and subjected to least squares linear fitting to calculate the elastic modulus of the material. The mathematical definition of the elastic modulus is shown in the following equation:

[0043] (3)

[0044] in, For the extracted elastic modulus, This represents the set of valid data points that participate in the linear fitting. Indicates the first The true strain value of each data point The independent variable represents the search for the minimum sum of squared errors. .

[0045] By calculating the smoothed first and second derivatives of stress relative to strain, the local minimum point of the second derivative is found, thereby accurately locating the desiccation point strain. The mathematical definition of the desiccation point strain is shown in the following equation:

[0046] (4)

[0047] in, The strain at the desiccation point is the point of extraction. This represents the first derivative of stress with respect to tensile strain. This represents a one-dimensional smoothing operation. This means that taking the derivative of the outer layer function again results in the second derivative of the stress as a whole. To find the optimal dehumidification point within the effective window period, This indicates the search for the strain that minimizes the second derivative. .

[0048] The coordinates at which the curve reaches its peak value are recorded as the stress and strain at the fracture point. The mathematical definitions of fracture point stress and strain are shown in the following equations:

[0049] (5)

[0050] in, For the extracted fracture point stress, Represents the global stress sequence vector under the current working condition. The point with the largest absolute value in the middle. This represents the strain at the fracture point, corresponding to the elongation at which the peak stress occurs. Its value is equal to the stress value and the maximum stress. When, the corresponding strain The value of .

[0051] Through the processing in step one, normalized real tensile data with multiple strain rates, conditional features, and prior physical features are obtained.

[0052] Step 2: Generation of the initial state curve of the latent space based on the variational autoencoder:

[0053] A generative network centered on a variational autoencoder is constructed, consisting of an encoder network and a decoder network. The encoder maps the input normalized time-series data to a continuous low-dimensional latent space and extracts the latent distribution. The decoder reconstructs the conditional latent vector, which is the result of concatenating the latent vector with conditional features, into stress components and standard deviation components. This step uses a dual-track generation mechanism to perform standard sampling of normal samples and directional perturbation of abnormal samples in the latent space. The generated latent vector is then concatenated with the conditional features output from step one to form a conditional latent vector. After decoding and inverse normalization, an initial state curve with upper and lower boundary envelopes is output. The specific steps are as follows:

[0054] 2.1 The normalized real tensile data with multiple strain rates is input into the encoder network of the variational autoencoder. Through multi-layer feature mapping, the high-dimensional time series is reduced to a continuous low-dimensional latent space, and the latent distribution representing the macroscopic evolution law of the current working condition is extracted, including the latent mean vector and the latent log-variance vector. The specific definitions are as follows:

[0055] (6)

[0056] in, and To obtain the latent mean vector and latent log-variance vector, Representative parameters are encoder network, This is the normalized true tensile data for multiple strain rates.

[0057] 2.2 To achieve sample diversification and controllable expansion of defect features, a dual-track generation mechanism is executed based on the potential distribution of the macroscopic evolution law of the current input operating conditions to obtain the potential vectors required for the generation stage:

[0058] The dual-track generation mechanism is as follows: For the generation of normal samples, a reparameterization technique is introduced to directly sample randomly from the Gaussian distribution output by the encoder to obtain normal latent vectors with a physically reasonable fluctuation range; for the generation of abnormal samples, based on the latent vectors of normal samples, an additional random Gaussian perturbation vector is injected to simulate actual engineering defects such as premature dehumidification points and shifted breakage points, thereby completing the directional shift of features. The specific definitions are shown in the following equation:

[0059] (7)

[0060] in, This represents the latent vector of the generated normal sample. and This represents the latent distribution mean vector and latent log-variance vector output by the encoder. Indicates standard deviation recovery calculation This represents a random noise vector that follows a standard normal distribution.

[0061] (8)

[0062] in, This represents the latent vector of the generated anomalous samples. This represents the random perturbation direction vector drawn from the standard normal distribution. This represents the disturbance intensity coefficient.

[0063] 2.3: After obtaining the latent feature vectors (normal latent feature vectors and abnormal latent feature vectors), they are concatenated with the conditional features extracted in step one along the channel dimension to form a conditional latent vector carrying specific strain rate constraint information. Subsequently, the conditional latent vector is fed into the decoder network for decoding and mapping to obtain the stress components and standard deviation components. Then, through inverse normalization calculation, an initial state curve with upper and lower boundary envelopes is generated. The specific mathematical definitions are shown in the following equations:

[0064] (9)

[0065] in, This represents the concatenated conditional latent vector. and These represent the potential vector and the conditional vector, respectively.

[0066] (10)

[0067] in, Represents the stress component and the standard deviation component. Representative parameters are The decoder network.

[0068] (11)

[0069] in, The stress components generated after denormalization, and These represent the statistical mean and standard deviation of the global stress sequence, respectively.

[0070] (12)

[0071] in, The standard deviation components are generated after denormalization. This represents the global maximum value of the standard deviation envelope under this operating condition. To prevent extremely small positive numbers with a multiplier of zero.

[0072] 2.4: Training objective of variational autoencoder:

[0073] Before deploying the actual generation process, the variational autoencoder needs to be trained offline. During the training phase, normalized real tensile data is input into the encoder, undergoing latent distribution extraction, reparameterization sampling, and decoder reconstruction to complete forward propagation. A hybrid loss function of reconstruction loss and KL divergence is used as the optimization objective to enable the encoder to accurately learn the latent distribution of the nonlinear characteristics of material mechanics and ensure that the latent space has good generation capabilities. The specific definitions are shown in the following equation:

[0074] (13)

[0075] in, This represents the total loss function of the variational autoencoder, with the first term in the formula being the reconstruction loss: This indicates the number of samples in a batch during network training. Indicates the first input encoder A real stretching data, The output of the decoding network is represented by the first... Initial state curve data; the second term in the formula is the KL divergence: and They represent the first The mean and log-variance vectors corresponding to each sample in the latent space. To adjust the hyperparameters.

[0076] Step 3: Post-processing method based on physical hard constraints:

[0077] The initial-state curve with upper and lower boundary envelopes generated in step two is combined with the prior physical features extracted in step one. By sequentially executing fundamental physical law constraints, mechanical performance corrections, and physical defect feature modeling, the initial-state curve is processed step by step, ultimately outputting a physically meaningful expanded sample dataset. The specific steps are as follows:

[0078] 3.1 To address the potential non-physical domain values ​​in the neural network output, a strict non-negative hard constraint is first applied to the inverse-normalized initial stress value to ensure that the material stress during the tensile process conforms to basic physical principles. The calculation formula is as follows:

[0079] (14)

[0080] in, The stress value after constraint. The initial stress value after denormalization. It is the lower bound of the minimal positive number field.

[0081] 3.2 Mechanical property correction:

[0082] To ensure that the macroscopic mechanical properties of the generated curve do not deviate from the intrinsic material properties at that strain rate, mechanical property corrections are performed using prior physical characteristics, specifically elastic modulus correction and fracture point stress correction. Elastic modulus correction involves extracting data from the first 10% of the global strain sequence of the generated curve, calculating the initial slope of this segment, and adjusting the stress using a truncation scaling factor. To ensure a smooth transition of the entire curve after correction, full stiffness scaling is applied to the first 10% of the purely elastic data segment, while the remaining 90% of the data segment is scaled down by a specific damping ratio. Fracture point stress correction assesses the maximum stress of the entire curve; if the current maximum stress exceeds the set allowable error range, global scaling is performed. The specific definitions are shown in the following equation:

[0083] (15)

[0084] in, To cut off the scaling factor, Represents an interval-limited amplitude function. For the true prior elastic modulus, This represents the actual initial slope of the generated curve within the first 10% interval. It is the lower bound of the minimal positive number field. This is the maximum allowable slope adjustment range. Index of the boundary points for the first 10% of strain data segments. The stress value after being constrained by basic physical laws This is the stress value after correction for the elastic modulus.

[0085] (16)

[0086] (17)

[0087] in, This is the global scaling factor for strength. This represents the peak stress at the actual fracture point. The set allowable error range, To complete the stress value after mechanical correction.

[0088] 3.3 Perform specific physical defect modeling on anomalous samples:

[0089] After completing the mechanical property correction, the current sample path is determined based on the normal / abnormal sample labels input by the dual-track generation mechanism in step two. If it is set as a normal sample, the system will skip the artificial defect application step and directly inherit the corrected reference stress; if it is set as an abnormal sample, the model will, based on the initial physical curve, use the extracted prior physical features as a basis, and according to the four preset defect labels including desiccation point anomaly, tensile break point anomaly, envelope anomaly, and composite anomaly, introduce the corresponding empirical mechanical attenuation model or boundary condition constraints to perform physical defect feature modeling processing. The desiccation point anomaly aims to simulate the decrease in load-bearing capacity caused by particle desiccation. Using the variable desiccation point as a trigger switch, a hyperbolic tangent function is introduced to construct a dynamic softening coefficient, nonlinearly reducing the local stress increment after desiccation, resulting in a smooth, weakening, and unloading trend on the macroscopic curve. The tensile fracture point anomaly aims to simulate premature fracture caused by stress concentration or crack propagation. When the tensile strain exceeds a set abnormal fracture threshold, the system forcibly assigns all subsequent stress points to a minimum residual stress value, mathematically reproducing the brittle fracture characteristics of instantaneous material failure. The envelope anomaly aims to simulate data dispersion distortion caused by batch differences. A random extreme scaling factor is applied to the standard deviation sequence, and a numerical truncation function is used to constrain the statistical mean and maximum values ​​of the real data as upper and lower limits, generating a physically reasonable abnormal dispersion envelope. The composite defect aims to simulate complex and harsh conditions such as aging or severe manufacturing defects. While applying stress softening to the main curve, the standard deviation envelope is directly amplified by 1.5 times, intuitively reflecting the decrease in tensile strength and unstable mechanical performance of the material under severe defects. Specific definitions are shown in the following equations:

[0090] (18)

[0091] (19)

[0092] in, Represents the dehumidification and softening attenuation coefficient. For the first The tensile strain value at each node represents the current location being calculated. The abnormal desiccation point strain after disturbance. For the first The final stress value at each node The initial stress value of the reference correction curve, This represents the local stress increment between two adjacent points on the reference curve.

[0093] (20)

[0094] in, For the first The final stress value at each node The minimum residual stress value, This indicates the array index position corresponding to the abnormal breakpoint.

[0095] (twenty one)

[0096] in, Indicates the first The final standard deviation envelope value of each data point This represents the initial standard deviation. The random envelope scaling factor. Represents an interval-limited amplitude function. The global average standard deviation This represents the global maximum standard deviation.

[0097] (twenty two)

[0098] 3.4: Output a physically plausible expanded sample dataset:

[0099] After completing the physical modeling of specific anomalous behaviors, the finalized stress sequence is linearly superimposed with the constrained standard deviation sequence to reconstruct the upper and lower envelopes reflecting the dispersion of the experimental data. This data is then stored in a tabular file to record the complete evolutionary data structure, constructing a physically meaningful augmented sample dataset. The calculation process is shown below:

[0100] (twenty three)

[0101] (twenty four)

[0102] in, and These represent the expanded samples at... The upper and lower envelope values ​​of stress at the first strain node.

[0103] Figure 2 This is a feature comparison diagram before and after the expanded sample generation for this invention. For example... Figure 2 As shown, subfigure a illustrates the distribution of the actual tensile data obtained. Subfigures b, c, d, e, and f respectively demonstrate the expanded samples of solid propellant under five working conditions—normal, envelope anomaly, tensile failure point anomaly, dehydration point anomaly, and composite anomaly—generated using the method of this invention. The comparison shows that the mean curve generated by the method of this invention can accurately track the mechanical evolution trend of the real data, and the generated envelope interval reasonably and effectively covers the statistical dispersion range of the real data. The data fidelity and distribution consistency are high, verifying the effectiveness of introducing physical mechanisms into the data generation model to expand the high-reliability mechanical dataset.

[0104] The above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Any modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A sample generation method for solid propellant failure identification, characterized in that: Includes the following steps: Step 1: Feature Extraction and Data Preprocessing The real tensile data under multiple strain rates are interpolated and normalized sequentially to calculate the conditional features; the elastic modulus, desiccation point strain, fracture point strain and fracture point stress are extracted as a priori physical features and used as the absolute judgment criteria for the physical hard constraint module in step three. Step 2: Generation of the initial state curve of the latent space based on the variational autoencoder: The interpolated and normalized data is input into a variational autoencoder to extract the latent distribution. Then, based on this distribution, different latent vectors are generated in the latent space through a dual-track generation mechanism: normal samples are sampled according to the distribution probability, abnormal samples are subjected to directional perturbation, and the latent vectors are concatenated with the conditional features to form a conditional latent vector. After decoding and inverse normalization, the initial state curve with upper and lower boundary envelopes is output. Step 3: Post-processing module based on physical hard constraints: The initial state curve with upper and lower boundary envelopes generated in step two is combined with the prior physical features extracted in step one. By sequentially executing basic physical law constraints, mechanical performance corrections, and physical defect feature modeling, the initial state curve is processed step by step, and finally, an expanded sample dataset with physical meaning is output.

2. The sample generation method for solid propellant failure identification according to claim 1, characterized in that: The specific implementation of step one includes: First, linear interpolation is performed on the real tensile data with multiple strain rates to uniformly resample the strain data of unequal lengths to a fixed length in order to meet the alignment requirements of neural network tensors. Secondly, the interpolated data is normalized to eliminate the dimensional differences between different physical quantities and multiple strain rates. Based on the normalized data, the stress mean and standard deviation sequence for each working condition is calculated and combined into conditional features. Finally, the prior physical features are extracted, including elastic modulus, desiccation point strain, fracture point strain, and fracture point stress. The aforementioned conditional features and prior physical features are stored independently to provide standardized input data and absolute judgment criteria for subsequent processing modules.

3. The sample generation method for solid propellant failure identification according to claim 2, characterized in that: The specific implementation of step two includes: First, the normalized real tensile data with multiple strain rates is input into the encoder network of the variational autoencoder. Through multi-layer feature mapping, the high-dimensional time series is reduced to a continuous low-dimensional latent space, and the potential distribution representing the macroscopic evolution law of the current working condition is extracted. Secondly, based on the extracted latent distribution, a dual-track generation mechanism is initiated: for normal samples, standard Gaussian sampling is performed according to the distribution probability to obtain latent vectors; for abnormal samples, random Gaussian perturbation vectors are injected on the basis of the latent vectors of normal samples to complete the directional shift of features. Finally, the generated latent vector is concatenated with the conditional features saved in step one along the feature dimension to form a conditional latent vector that integrates macroscopic constraints and microscopic randomness. This vector is then mapped and reconstructed by a decoder network to obtain stress components and standard deviation components. Finally, through inverse normalization calculation, an initial state curve with upper and lower boundary envelopes is generated.

4. The sample generation method for solid propellant failure identification according to claim 3, characterized in that: The specific implementation of step three includes: First, the initial state curve generated by decoding is constrained by basic physical laws. By executing non-negative hard constraints, it is ensured that the material stress during the stretching process conforms to basic physical common sense. Secondly, the prior physical features extracted in step one are introduced to correct the mechanical properties. The initial elastic modulus of the pure elastic data segment of the curve is corrected by calculating the truncation scaling factor, and the stress at the fracture point of the entire curve is corrected according to the tolerance range of the prior fracture point stress. Next, based on the normal / abnormal sample labels input in step two, the samples are categorized into specific abnormal behavior judgment nodes for sorting: normal samples are skipped directly, while abnormal samples are based on prior physical characteristics and, according to the four preset defect labels including dehumidification point abnormality, tensile break point abnormality, envelope abnormality and composite abnormality, the corresponding empirical mechanical attenuation model or boundary condition constraints are introduced to perform physical defect feature modeling. Finally, the upper and lower boundary envelopes are reconstructed using the established stress mean and standard deviation sequences, and the resulting augmented sample dataset with physical meaning is then integrated and output.

Images