Carbon fiber composite material damage prediction method and device based on machine learning

Abstract

The application discloses a carbon fiber composite material damage prediction method and device based on machine learning, and relates to the field of machine learning.The method comprises the following steps: acquiring target material production timing parameters, combining material models and expected service life to search for a matching sample set, determining possible damage types and occurrence rates, screening potential damage types with occurrence rates exceeding a threshold, calculating damage detection factors, selecting a network integrated target detector from a damage detection network pool based on the factors, calculating damage probabilities through multi-network voting, and identifying the carbon fiber composite material as high-risk material if any potential damage probability exceeds a preset threshold.The application introduces machine learning technology, realizes accurate prediction of potential damage probabilities of carbon fiber composite materials, guides optimal allocation of detection resources based on the prediction, and significantly improves overall detection efficiency.

Description

Technical Field

[0001] This invention relates to the field of machine learning, and more specifically to a method and apparatus for predicting damage in carbon fiber composite materials based on machine learning. Background Technology

[0002] Carbon fiber composites are widely used in aerospace, automobile manufacturing and other fields due to their advantages such as high strength and lightweight. However, they are prone to various damages such as fiber breakage, matrix cracking, interface debonding and delamination during the production process due to process fluctuations.

[0003] Currently, the testing of carbon fiber composite materials mainly relies on traditional non-destructive testing methods such as ultrasonic and X-ray testing. Each material sample needs to be tested individually using hardware. This method has problems such as long testing cycle, low efficiency, and high testing cost in mass production. Summary of the Invention

[0004] This invention addresses the technical problem of low detection efficiency in mass production due to the need for hardware testing of each material sample in existing technologies, and provides a damage prediction method and device for carbon fiber composite materials based on machine learning.

[0005] The technical solution of the present invention to solve the above-mentioned technical problems is as follows:

[0006] In a first aspect, the present invention provides a damage prediction method for carbon fiber composite materials based on machine learning, comprising:

[0007] Obtain the production time sequence parameters of the target carbon fiber composite material, and determine multiple possible damage types of the target carbon fiber composite material and the damage incidence rate of each possible damage type based on the production time sequence parameters;

[0008] Select possible damage types with a damage incidence rate greater than the corresponding incidence rate threshold, determine multiple potential damage types, and obtain the damage detection factor for each potential damage type.

[0009] Based on the damage detection factors for each of the potential damage types, the damage probability of multiple potential damage types is determined.

[0010] When the probability of any of the multiple potential damage types is greater than the corresponding preset probability threshold, the target carbon fiber composite material is identified as a high-risk material.

[0011] Secondly, the present invention provides a machine learning-based damage prediction device for carbon fiber composite materials, comprising:

[0012] The beneficial effects of this invention are:

[0013] This invention, by introducing machine learning technology, achieves accurate prediction of the probability of potential damage to carbon fiber composite materials, and uses this to guide the optimal allocation of testing resources, significantly improving overall testing efficiency. First, by analyzing production time-series parameters and integrating historical sample data, the probability of various potential damages is accurately identified and quantified. Second, based on damage detection factors, the configuration scale of the testing network is dynamically adjusted, enabling a resource allocation strategy that prioritizes the detection of high-risk damage types and rapidly screens low-risk types, effectively avoiding resource redundancy and waste. Third, by leveraging multi-network integration and voting mechanisms, damage assessment is completed efficiently and stably, reducing reliance on traditional manual or hardware testing, shortening the testing cycle, and better adapting to the needs of mass production scenarios. Finally, based on preset probability thresholds, fully automated quality judgment and labeling are achieved, ensuring the objectivity and consistency of the assessment process and significantly improving the accuracy and reliability of product quality control. Attached Figure Description

[0014] Figure 1 A flowchart illustrating the machine learning-based damage prediction method for carbon fiber composite materials provided by this invention.

[0015] Figure 2 A schematic diagram of the structure of the machine learning-based carbon fiber composite material damage prediction device provided by the present invention.

[0016] In the attached diagram, the components represented by each number are as follows:

[0017] Module 11 for parameter acquisition and initial damage type judgment, module 12 for potential damage screening and factor acquisition, module 13 for damage probability calculation, and module 14 for quality judgment and identification. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0020] In the description of this invention, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this invention is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed herein.

[0021] Example 1, as Figure 1 As shown, embodiments of the present invention provide a damage prediction method for carbon fiber composite materials based on machine learning, including:

[0022] S10: Obtain the production time sequence parameters of the target carbon fiber composite material, and determine multiple possible damage types of the target carbon fiber composite material and the damage incidence rate of each possible damage type based on the production time sequence parameters;

[0023] First, when the damage probability of multiple potential damage types is less than or equal to the corresponding preset probability threshold, the target carbon fiber composite material is identified as a low-risk material.

[0024] Potential damage types refer to high-risk damage categories obtained by screening possible damage types with a damage incidence rate greater than a corresponding incidence rate threshold, such as fiber breakage, matrix cracking, interface debonding, and delamination damage. These are the damage forms that need to be verified in the testing process. Each potential damage type corresponds to an independent damage probability, which is used to quantify the likelihood of damage.

[0025] In addition, the preset probability threshold is a quantitative critical value set in advance to measure whether a potential damage type constitutes a quality risk, based on the application scenario, safety level requirements, industry standards and expected service requirements of the target carbon fiber composite material in order to achieve material quality judgment. Each potential damage type has an independent preset probability threshold. For example, the delamination damage probability threshold for key structural components in the aerospace field can be set to ≤5%, and the matrix cracking probability threshold for ordinary industrial composite materials can be set to ≤12%.

[0026] In the non-destructive testing (NDT) process for carbon fiber composites, the ultimate goal of NDT is to determine the material's quality status, and the quality labeling includes two results: "low risk" and "high risk." When the probability of a certain potential damage type is greater than its corresponding preset probability threshold, the carbon fiber composite is labeled as a high-risk material; when the probability of all potential damage types is less than or equal to their respective preset probability thresholds, the carbon fiber composite is labeled as a low-risk material.

[0027] Based on damage prediction results, differentiated detection strategies can be configured. For materials predicted to be high-risk, enhanced detection measures are adopted: increasing the frequency of detection to achieve 100% full inspection; using a combination of multiple detection methods, such as ultrasonic testing, X-ray testing, and infrared thermal imaging, to ensure comprehensive coverage of all potential damage; and equipping the materials with experienced testing personnel and high-precision testing equipment, with priority given to scheduling testing times.

[0028] For materials predicted to be low-risk, an optimized testing strategy is adopted: sampling testing can be carried out at a certain proportion, such as a sampling rate of 10%-20%; the most suitable single testing method can be selected to avoid the waste of resources from multiple tests; key stress-bearing parts should be tested, and the testing process should be appropriately simplified while ensuring quality; testing can be scheduled during the idle time of the testing equipment to improve equipment utilization.

[0029] This differentiated testing strategy based on prediction results ensures the quality control requirements of high-risk materials while avoiding over-testing of low-risk materials. It optimizes the allocation of testing resources, significantly improves overall testing efficiency, and reduces testing costs, making it particularly suitable for quality control needs in mass production environments.

[0030] In this embodiment of the application, non-destructive testing of carbon fiber composite materials first requires obtaining the production timeline parameters of the target carbon fiber composite material. Based on the production timeline parameters, multiple possible damage types of the target carbon fiber composite material and the damage incidence rate of each possible damage type are determined, including:

[0031] Based on the production timing parameters of the target carbon fiber composite material, a production parameter timing feature is generated, and the material type and expected service life of the target carbon fiber composite material are obtained.

[0032] Based on the time-series characteristics of the production parameters, the material type, and the expected service life, a sample set of target carbon fiber composite materials is retrieved.

[0033] The damage record data of each historical carbon fiber composite material sample in the target carbon fiber composite material sample set are statistically analyzed to identify the damage types that occur and determine multiple possible damage types.

[0034] The damage incidence rate of each of the possible damage types is obtained by calculating the ratio of the number of occurrences of each possible damage type in the target carbon fiber composite material sample set to the total number of samples in the target carbon fiber composite material sample set.

[0035] Specifically, production sequence parameters refer to a set of parameters for the target carbon fiber composite material throughout the entire production process, dynamically changing with the production progress and directly related to key material forming processes. These parameters are core foundational data reflecting process fluctuations and tracing the causes of potential material damage. They include key process parameters such as fiber layup angle sequence, resin curing temperature curve, curing pressure variation curve, curing time, and environmental humidity variation. Specifically, the fiber layup angle sequence is a sequence of angle values ​​and time points used to record the layup angle at different time points in the fiber layup process; the resin curing temperature curve is a continuous curve of temperature and time, reflecting the temperature rise, holding, and cooling processes in the resin curing process; the curing pressure variation curve is a continuous curve of pressure values ​​and time, recording the pressure fluctuations during the curing process; the curing time is quantified duration data, including the total curing time and the specific duration of each stage, such as the holding and pressure holding stages; and the environmental humidity variation is continuous data corresponding to humidity values ​​on the time axis, covering the dynamic changes in environmental humidity throughout the entire production process, especially during the fiber layup and resin curing stages.

[0036] Production parameter time series features are quantitative features extracted after preprocessing the original production time series parameters, such as removing outliers and performing segmented statistics. These features include parameter mean, standard deviation of fluctuation, peak / valley values, number of outliers, etc. Their purpose is to simplify high-dimensional original time series data, retain key information on process fluctuations, and provide quantifiable comparison basis for subsequent sample matching.

[0037] Data processing was performed on the acquired production time-series parameters of the target carbon fiber composite material to generate production parameter time-series characteristics. Simultaneously, two key attribute information types were collected: material type and expected service life. Material type refers to the standardized identifier of the target carbon fiber composite material, including core attributes such as fiber type (e.g., T700, T800), matrix type (e.g., epoxy resin, phenolic resin), and layup structure (e.g., alternating [0° / 90°] layups). Differences in composition and structure between different material types directly lead to the specificity of damage types and causes, serving as the basic screening condition for sample retrieval. Expected service life refers to the designed service time of the target carbon fiber composite material in actual application scenarios, measured in years. For example, aerospace components are expected to have a service life of 10 years, and automotive components 5 years. Damage incidence is highly correlated with service life; materials with long service life are more prone to aging, delamination, and other damage. This parameter ensures the timeliness of historical sample damage data for reference regarding the target material.

[0038] Further, based on the time-series characteristics of the production parameters, the material type, and the expected service life, a sample set of target carbon fiber composite materials is retrieved, including:

[0039] Based on the material type, historical carbon fiber composite material samples with the same material type are retrieved to obtain the first carbon fiber composite material sample set;

[0040] Based on the expected service life, the first carbon fiber composite material sample set is screened, and historical carbon fiber composite material samples with a service life less than or equal to the expected service life are selected to obtain the second carbon fiber composite material sample set.

[0041] Calculate the similarity between the time-series characteristics of the production parameters and the time-series characteristics of the production parameters of each historical carbon fiber composite material sample in the second carbon fiber composite material sample set. Select historical carbon fiber composite material samples with similarity greater than a preset similarity threshold to obtain a third carbon fiber composite material sample set, which is used as the target carbon fiber composite material sample set.

[0042] Specifically, historical carbon fiber composite material samples refer to past production instances of carbon fiber composite materials stored in a sample database, possessing complete production information and damage records. Screening samples from these historical samples that are highly relevant to the target material provides a reliable data foundation for subsequent statistical analysis of damage types and calculation of damage incidence rates. To ensure a high degree of matching between the final sample set and the target material in terms of core properties, service conditions, and production process characteristics, the target carbon fiber composite material sample set is retrieved based on the temporal characteristics of production parameters, material type, and expected service life. A three-level progressive screening process achieves precise focusing on historical samples.

[0043] First, a primary screening based on material type is conducted to form the first carbon fiber composite material sample set. The material type, as the core identifier of carbon fiber composite materials, directly relates to the fiber type, matrix composition, and layup structure, and is a key factor determining the type and cause of material damage. For example, materials made of high-modulus fibers are more prone to fiber fracture, and brittle matrix materials are more prone to matrix cracking. Therefore, retrieving historical samples that are completely consistent with the target material type ensures from the outset that all samples in the first carbon fiber composite material sample set are consistent with the target material in terms of basic properties such as composition and structure, eliminating damage type mismatches caused by differences in the inherent properties of the materials themselves.

[0044] Secondly, a secondary screening based on the expected service life was conducted to obtain a second carbon fiber composite material sample set from the first sample set. The expected service life is the designed service time of the target material in actual application scenarios. The damage incidence rate of carbon fiber composite materials is positively correlated with the service life. Materials with long-term service are more prone to delamination and accelerated interface debonding due to environmental aging, fatigue load accumulation, and other factors, and their damage types differ significantly from those of short-term service materials. Therefore, by using a service life less than or equal to the expected service life as a screening criterion, a secondary filtering of the first carbon fiber composite material sample set was performed. This excluded historical samples whose damage conditions exceeded the actual risk range of the target material due to excessively long service times, ensuring that the sample damage data in the second carbon fiber composite material sample set accurately reflects the potential damage risks that the target material may face within its expected service life.

[0045] Finally, based on the similarity of the time-series characteristics of production parameters, a third carbon fiber composite material sample set was selected from the second carbon fiber composite material sample set as the target sample set. The time-series characteristics of production parameters are key feature quantities extracted after preprocessing the production time-series parameters of the target material. They directly reflect the process fluctuations during the material production process, and these fluctuations are the core cause of material damage. For example, abnormal curing temperature rise rates may trigger matrix cracking, and excessive ply tension fluctuations may lead to interlayer delamination. By calculating the similarity between the time-series characteristics of the target material's production parameters and the time-series characteristics of each historical sample in the second carbon fiber composite material sample set, and selecting samples with similarity greater than a preset similarity threshold (set according to detection accuracy requirements, such as 80%), it is ensured that the historical samples in the third carbon fiber composite material sample set highly match the production process fluctuation characteristics of the target material. This ensures the damage type and damage pattern of the samples, and can be directly used as a reference for the potential damage analysis of the target material. The third carbon fiber composite material sample set is the target carbon fiber composite material sample set.

[0046] Furthermore, damage inspection reports and quality records for each historical sample in the target carbon fiber composite material sample set were extracted. Damage forms appearing in all samples were statistically analyzed, damage categories were identified and classified, and multiple possible damage types for the target material were ultimately determined. Since materials are prone to similar damage under similar production processes and service conditions, statistical analysis of historical sample damage records can objectively pinpoint the potential damage risks to the target material, avoiding unfounded guesses about damage types. For example, if 80% of the historical samples in the sample set exhibit delamination damage and fiber breakage, then the possible damage types of the target material include delamination damage and fiber breakage, providing a clear direction for subsequent targeted testing.

[0047] Furthermore, the damage incidence rate of each possible damage type is calculated. The damage incidence rate is the ratio of the number of historical samples of a specific damage type to the total number of samples in the sample set. It is the core indicator for quantifying the probability of various potential damage risks of the target material. Specifically, the damage incidence rate = (number of times a certain possible damage type appears in the sample set ÷ total number of samples in the target carbon fiber composite material sample set) × 100%.

[0048] For example, suppose the target carbon fiber composite material sample set contains 100 historical samples. After statistics, it is found that delamination damage occurred in 35 samples and fiber breakage occurred in 18 samples. Then: the occurrence rate of delamination damage = (35 ÷ 100) × 100% = 35%, and the occurrence rate of fiber breakage = (18 ÷ 100) × 100% = 18%.

[0049] Damage incidence rate directly reflects the potential risk probability of different damage types in target materials, ensuring that detection resources are focused on high-risk damage and improving detection efficiency and targeting.

[0050] S20: Select possible damage types with damage incidence rates greater than the corresponding incidence rate threshold, determine multiple potential damage types, and obtain the damage detection factor for each potential damage type.

[0051] Specifically, potential damage types with an incidence rate greater than a corresponding incidence rate threshold are selected to determine multiple potential damage types, and damage detection factors for each potential damage type are obtained, including:

[0052] A first possible damage type is determined from the possible damage types, and a first damage occurrence rate of the first possible damage type is extracted to obtain the corresponding first occurrence rate threshold.

[0053] When the first injury occurrence rate is greater than or equal to the first occurrence rate threshold, the first possible injury type is determined as the first potential injury type;

[0054] Based on the first damage incidence rate and the first incidence rate threshold, obtain the first damage detection factor for the first potential damage type;

[0055] The remaining possible damage types are processed to obtain multiple potential damage types and damage detection factors for each potential damage type.

[0056] First, the first possible damage type is determined from the potential damage types, its first damage incidence rate is extracted, and the corresponding first incidence rate threshold is obtained. The first possible damage type refers to the first damage type to be judged selected from multiple possible damage types obtained through previous statistical analysis of the target sample set, in any order, such as from high to low damage incidence rate. The first damage incidence rate is the statistical result of this damage type in the target sample set, directly reflecting the historical probability of this damage type occurring under similar materials and similar process conditions. The first incidence rate threshold is a pre-set risk assessment threshold for this damage type, and its setting is related to the safety requirements of the material application scenario and industry quality standards. For example, in the aerospace field, the incidence rate threshold for damage such as "fiber fracture," which directly affects the structural load-bearing capacity, might be set at 5%, while in ordinary industrial fields it can be relaxed to 10%. Each possible damage type corresponds to an independent incidence rate threshold; not all damage types share the same standard, ensuring that the threshold matches the risk level of the damage type.

[0057] Secondly, when the incidence rate of the first damage is greater than or equal to the first incidence rate threshold, the first possible damage type is identified as the first potential damage type. Specifically, only when the incidence rate of a certain possible damage type reaches or exceeds a preset incidence rate threshold is it determined to be a "potential damage type" requiring focused attention. For example, if the incidence rate of the first damage of delamination damage is 15%, and the corresponding first incidence rate threshold is 10%, then 15% ≥ 10%, and this damage type becomes the first potential damage type. If the incidence rate of the first damage of fiber breakage is 3%, which is lower than the first incidence rate threshold of 5%, it indicates that its historical occurrence probability is low, the risk to the target material is controllable, and it does not need to be included in the subsequent key detection scope. By comparison, low-risk types can be quickly identified from many possible damage types, and detection resources can be concentrated on high-risk potential damage types, improving the targeting and efficiency of subsequent detection.

[0058] Next, based on the first injury incidence rate and the first incidence rate threshold, the first injury detection factor for the first potential injury type is obtained. The injury detection factor is a core parameter used to adjust the injury detection strategy, calculated based on the quantitative relationship between the injury incidence rate and the corresponding threshold. Essentially, it reflects the risk level of the potential injury type in numerical form, providing a weighting basis for selecting the injury detection network and calculating the injury probability. When the injury detection factor is 1.0, it indicates that the injury incidence rate has just reached the incidence rate threshold, and the risk is at a critical state. When the injury detection factor is greater than 1.0, the larger the value, the higher the injury risk, and more detection networks need to be configured to improve detection accuracy. Injury detection factor = Injury incidence rate ÷ Incidence rate threshold.

[0059] For example, if the incidence rate of layered damage is 15% and the incidence rate threshold is 10%, then its damage detection factor is 1.5. The higher the value, the higher the risk level of this type of damage, and more detection networks need to be selected for evaluation to ensure the detection accuracy of high-risk damage; the lower the value, but not lower than 1.0, the lower the risk level, since the damage incidence rate is ≥ the incidence rate threshold, and the detection resource allocation can be adjusted appropriately to balance detection efficiency while ensuring accuracy.

[0060] Finally, following the same method used to process the first possible damage type, the remaining possible damage types are processed, resulting in multiple potential damage types and damage detection factors for each potential damage type. This process continues until all possible damage types have been processed, completing the screening and parameter configuration for all high-risk damage types. This lays a complete data foundation for the subsequent construction of a target damage detector based on the damage detection factors and the calculation of damage probabilities.

[0061] S30: Based on the damage detection factors of each potential damage type, determine the damage probability of multiple potential damage types;

[0062] Specifically, based on the damage detection factors for each of the potential damage types, the damage probability of multiple potential damage types is determined, including:

[0063] Traverse multiple potential damage types, obtain the target potential damage type, and extract the corresponding damage detection factors to obtain the target damage detection factor;

[0064] The corresponding damage detection network pool is retrieved according to the target potential damage type, and the damage detection network is selected in the damage detection network pool based on the target damage detection factor, and integrated to obtain the target damage detector.

[0065] Based on the target damage detector, damage assessment is performed on the production time series parameters to obtain the damage probability of the target potential damage type;

[0066] Based on the method of obtaining the damage probability of the target potential damage type, the damage probabilities of other potential damage types are obtained, resulting in multiple potential damage probabilities.

[0067] First, multiple potential damage types are traversed to obtain the target potential damage type and extract the corresponding target damage detection factor. Traversal refers to processing each of the multiple potential damage types obtained through threshold screening in a preset order, such as from high to low damage risk, focusing on only one target potential damage type at a time. After determining the target type, the target damage detection factor corresponding to that type needs to be extracted. This factor is calculated in step S20, and its value directly reflects the intensity of the damage type's demand for detection resources. That is, the larger the target damage detection factor value, the higher the damage risk, and the denser the detection network needs to be configured.

[0068] Secondly, the corresponding damage detection network pool is retrieved according to the target potential damage type, and the network is selected and integrated based on the target damage detection factors to obtain the target damage detector. The damage detection network pool is a pre-built resource library containing multiple detection models. These models are trained with different training data and network architectures, enabling damage identification from different dimensions and ensuring comprehensive detection.

[0069] Specifically, the steps for constructing the damage detection network pool include:

[0070] Multiple sample production time-series parameters were collected to construct a sample production time-series parameter set. Then, based on the expected service duration, the potential damage types of the multiple sample production time-series parameters were labeled to construct a sample damage label set.

[0071] Construct multiple damage detection network architectures;

[0072] Multiple damage detection network architectures are trained based on the sample production time series parameter set and the sample damage label set to obtain multiple damage detection networks.

[0073] The multiple damage detection networks are stored to obtain the damage detection network pool.

[0074] First, production time-series parameters for multiple carbon fiber composite samples were collected and a sample production time-series parameter set was constructed. These parameters include: fiber layup angle sequence (fiber layup angle data at different time points during the layup process, in degrees Celsius; angle changes must be recorded sequentially to reflect dynamic adjustments in layup direction); resin curing temperature curve (in degrees Celsius; recording the complete temperature change process from heating, holding, to cooling during the curing process, including peak temperature, heating rate (degree Celsius / min), and holding time); curing pressure variation curve (in MPa; recording pressure fluctuations during the curing process, reflecting the dynamic impact of pressure on material forming); curing time (total curing time in minutes and the specific duration of each stage, such as heating, holding, and cooling, reflecting the time dimension of the curing process); and environmental humidity variation (recording the change in environmental humidity over time throughout the entire production process, in %RH, capturing the potential impact of humidity on the material's interfacial bonding performance). Production time series parameters are summarized by sample dimension to ensure that the time series parameters of each sample fully cover the key production links, and finally form a sample production time series parameter set containing data from multiple batches and multiple process scenarios.

[0075] Secondly, based on the expected service life of each sample, the potential damage types of the materials corresponding to the production timeline parameters are labeled to construct a sample damage label set. For each sample in the sample production timeline parameter set, its designed expected service life is first determined, such as 5 years, 10 years, etc. Then, physical testing methods, such as ultrasonic testing, X-ray testing, and infrared thermal imaging testing, are used to conduct actual damage testing on the sample to identify and determine whether the sample has any potential damage types. For example, if a sample has an expected service life of 8 years, and its fiber layup angle sequence has local angle deviations and the resin curing temperature curve has significant fluctuations, physical testing confirms that the sample has "delamination damage," then the "delamination damage" label is marked next to the sample's production timeline parameter data. If another sample has "porosity defects" due to insufficient curing pressure, then the "porosity defect" label is marked. Finally, a sample damage label set is formed with a one-to-one correspondence between production timeline parameters, expected service life, and potential damage types, providing a clear "input and output" mapping relationship for the subsequent training of the damage detection network architecture.

[0076] Furthermore, multiple damage detection network architectures are constructed. To adapt to the diversity of production time-series parameters and damage types, reduce the prediction bias of a single network, improve the reliability of damage judgment, and support adaptive network configuration based on damage detection factors, multiple damage detection network architectures need to be constructed and form a damage detection network pool.

[0077] First, based on the core attributes of the target production time-series parameters, they are categorized by parameter type and data characteristics as follows: Continuous time-series parameters, such as resin curing temperature curves and curing pressure change curves, exhibit continuous fluctuation trends and require capturing long-term time-series patterns; the corresponding architecture must possess long-term time-series feature memory capabilities. Discrete time-series parameters, such as fiber laying angle sequences and environmental humidity jump data, contain abrupt change nodes and require identifying local anomalies; the corresponding architecture must possess local feature extraction capabilities. Multi-parameter coupled scenarios, such as synchronous fluctuations in temperature and pressure and correlation changes between laying angle and tension, require analyzing the synergistic effects between parameters; the corresponding architecture must possess multi-input correlation analysis capabilities.

[0078] Meanwhile, since the identification criteria for different potential damage types are different, for example, delamination damage is often caused by abnormal trends in the curing temperature / pressure time sequence, such as too slow heating or insufficient pressure holding, the architecture needs to be able to capture the long-term trend of continuous parameters and match the time sequence memory architecture; fiber fracture is often related to a sudden drop in local layup tension and abrupt change in layup angle, the architecture needs to be able to identify local anomalies of discrete parameters and match the local feature architecture; matrix cracking may be caused by a sudden rise or fall in temperature and fluctuations in environmental humidity, the architecture needs to be able to correlate and analyze multiple types of parameters and match the multi-parameter fusion architecture.

[0079] Therefore, based on the aforementioned adaptation directions, various network types with different core capabilities are selected as the basic framework for the architecture design. For example, a Long Short-Term Memory (LSTM) network is chosen to capture the long-term evolution trends of production time-series parameters, such as slow deviations in curing temperature and insufficient holding time, adapting to the detection of delamination damage and aging damage. LSTM is a special type of recurrent neural network used to process long-sequence data, particularly adept at capturing long-term dependencies in time series. The core components of LSTM include memory cells and gating units. Memory cells refer to the information channels running through the network, used to store key long-term information in the time-series data. As information flows through the cells, it is selectively modified only by the gating units. The gating units include: a forget gate, used to determine which historical information in the "memory cell" needs to be discarded; an input gate, used to determine which new information needs to be stored in the "memory cell"; and an output gate, used to determine which information to output to the next time step based on the current state of the "memory cell".

[0080] In the specific construction, the input layer is used to receive single-class continuous time-series parameters, such as the time-series data of temperature curves, and the input dimension is divided according to the time step (e.g., 1 min / step); the hidden layer is used to set 2-3 layers of LSTM units, each layer containing 64-128 neurons, and uses gating mechanisms (input gate, forget gate, output gate) to memorize long-term time-series features and suppress irrelevant short-term fluctuations; the output layer uses the Sigmoid activation function to output the judgment result of "damage or no damage", which corresponds to the time-series correlation features of potential damage types, such as the risk of stratification if the temperature is lower than the standard value for a long time.

[0081] Convolutional Neural Networks (CNNs) are used to identify local abrupt changes or abnormal peaks in production time-series parameters, such as sudden drops in layup tension or sudden increases in resin injection rate, adapting to the detection of porosity defects and fiber fractures. CNNs are deep learning algorithms specifically designed for processing mesh-structured data. Through a combination of convolutional layers, pooling layers, and fully connected layers, they achieve automatic extraction and learning from pixel-level features to high-level semantic features. Their core utilizes convolution operations, namely, extracting local features and sharing parameters through a sliding window. The same convolutional kernel is reused at different locations in the image to reduce computation and capture local spatial correlations. Specifically, the input layer transforms the time-series parameters into a two-dimensional matrix corresponding to time and parameter values ​​(e.g., time on the horizontal axis and temperature on the vertical axis), simulating image-like input. The convolutional layers use 2-3 kernels, such as 3×3 or 5×5, to extract local abnormal features through a sliding window. The pooling layer uses max pooling to retain key abnormal information in the local features, reducing data dimensionality. The output layer, through a fully connected layer and a softmax activation function, outputs the damage type judgment result corresponding to the local abnormality.

[0082] A CNN-LSTM fusion architecture is selected to analyze the coupling features of multiple production time-series parameters, such as interface debonding caused by simultaneous temperature rises and pressure drops, adapting to complex damage detection caused by multi-parameter collaboration. Specifically, the input layer receives parallel data of multiple time-series parameters (such as temperature, pressure, and humidity) and aligns them by time step; the convolutional submodule performs a single convolutional operation on each parameter to extract local features; the concatenation layer concatenates the local features of multiple parameters by time step to form a multi-dimensional feature matrix; the LSTM submodule sets up 1-2 LSTM units to learn the temporal correlations between multi-parameter features; and the output layer outputs the damage judgment results under multi-parameter coupling scenarios.

[0083] Constructing multiple damage detection network architectures can cover different dimensions of damage identification and achieve functional complementarity.

[0084] Furthermore, multiple damage detection network architectures are trained based on the sample production time-series parameter set and the sample damage annotation set. First, the sample production time-series parameter set and the sample damage annotation set need to be standardized. Various parameters in the sample production time-series parameter set are normalized according to a unified rule, such as mapping to the 0-1 interval, to avoid the excessive influence of certain parameters on the training results due to differences in parameter dimensions. Second, the target potential damage types in the sample damage annotation set, such as delamination and fiber breakage, are converted into binary labels, i.e., "delamination damage" is labeled as [1, 0, 0], and "fiber breakage" is labeled as [0, 1, 0], ensuring that the output layers of all architectures correspond to the same format of labels, avoiding inconsistent training objectives due to differences in annotation formats. Simultaneously, the parameter data and damage annotations are aligned by time to ensure that parameter changes in each time segment correspond to a clear damage state label. The preprocessed sample production time series parameter set and sample damage annotation set are divided into training set and validation set in a ratio, such as 7:3. All architectures share the same partitioning result, that is, the samples contained in the training set and the samples contained in the validation set are completely consistent. This ensures that the training effect of each architecture can be compared under the same data benchmark and avoids the bias of training results caused by the difference in sample partitioning.

[0085] Specifically, training parameters are adjusted according to the core characteristics of different architectures. For example, the batch size for LSTM is set to 64, the adaptive moment estimation (Adam) optimizer is selected, the learning rate is set to 0.001, and the loss function is binary cross-entropy. For CNN, the batch size is set to 32, the stochastic gradient descent (SGD) optimizer is selected, the momentum parameter is 0.9, the learning rate is 0.0005, and the loss function is also cross-entropy. For the CNN-LSTM multi-parameter fusion architecture, the batch size is set to 96, the optimizer is AdamW (Adam with weight decay), the learning rate is set to 0.0008, the loss function is designed for complex damage types caused by multi-parameter coupling, and attention weight parameters are added to the fusion layer to ensure that the parameter correlation features that significantly affect the damage are focused on during training.

[0086] Furthermore, the production time-series parameters from the unified training set are input into the LSTM in batches. Feature extraction and data processing are performed through the input and hidden layers of the architecture, such as the gating units of the LSTM. Finally, the output layer outputs the predicted damage type. All architectures process the same batch of training samples in this manner to ensure the consistency of the input data. Next, the predicted results output by the architecture are compared with the corresponding damage labels on the unified training set, and the prediction error is calculated using a preset loss function. Based on the error value, backpropagation algorithms, such as the BPTT algorithm for LSTM and the backward gradient descent algorithm for CNN, are used to adjust the weights and bias parameters of each layer of the architecture, such as the forget gate weights of LSTM and the convolutional kernel parameters of CNN, to minimize the prediction error. Finally, the above steps are iterated until the validation set error stabilizes and reaches 90% accuracy, indicating convergence. Training of the architecture is then stopped to avoid overfitting. All architectures use the same convergence criterion to ensure the comparability of training termination timing.

[0087] After all architectures have been trained, multiple damage detection networks are obtained. These networks are then stored to form a damage detection network pool. To ensure that a sufficient number of networks with suitable types can be selected for different detection factor requirements, and to improve the reliability and stability of the multi-network voting mechanism and cover more diverse production process fluctuations and damage scenarios, an adaptive network configuration mechanism based on damage detection factors is needed to expand the network pool size. This requires data partitioning and network architecture combination.

[0088] Specifically, the sample production time-series parameter set and sample damage annotation set are divided N times, with random sampling with replacement used during the division, resulting in N sets of training data. Simultaneously, M different damage detection network architectures are constructed, and each of the M architectures is trained using the N sets of training data, ultimately yielding N×M damage detection networks. This forms a damage detection network pool containing N×M networks.

[0089] For example, suppose the original sample set contains 1000 historical samples, covering different production batches and process fluctuation scenarios. The number of partitions is set to N=5. Each time, the same number of samples as the original set are drawn, i.e., 1000 samples are drawn each time, and repeated sampling is allowed, resulting in 5 sets of training data. M=3 complementary architectures are designed to ensure that each architecture can adapt to the recognition needs of specific damage types, including LSTM networks, CNN networks, and CNN-LSTM fusion architectures. M network architectures are trained on N sets of training data respectively. When N=5 and M=3, 15 networks can be obtained. The weights and bias parameters of each network vary depending on the "training data distribution" and "architectural characteristics". This combination method not only increases the number of networks but also improves network diversity, which is beneficial for dynamically selecting the optimal network combination for damage probability assessment based on damage detection factors, thereby improving the accuracy and reliability of detection.

[0090] Further, based on the target damage detection factor, damage detection networks are selected from the damage detection network pool and integrated to obtain a target damage detector, including:

[0091] Obtain network integration baseline data;

[0092] The target network configuration quantity is obtained by adjusting the network integration baseline number using the target damage detection factor.

[0093] Based on the target network configuration quantity, damage detection networks are randomly selected from the damage detection network pool to obtain multiple selected detection networks;

[0094] The target damage detector is obtained by integrating multiple selected detection networks through a network output processing layer.

[0095] The network integration baseline number serves as the basic reference value for subsequent adjustments to the number of target networks. Its setting needs to be comprehensively determined based on the size of the damage detection network pool, common detection requirements for potential damage types, and computational resource limitations. For example, if the damage detection network pool contains 15 differentiated networks, the network integration baseline number can be set to 5. This avoids both insufficient reliability of multi-network voting due to an excessively small baseline number and waste of computational resources due to an excessively large baseline number.

[0096] The target network integration baseline is adjusted using a target damage detection factor. The target network configuration quantity = network integration baseline quantity × target damage detection factor. For example, if the baseline quantity is 5 and the target damage detection factor is 1.8 (high risk), then the target network configuration quantity = 5 × 1.8 = 9, rounded up. If the target damage detection factor is 1.2 (low risk), then the target network configuration quantity = 5 × 1.2 = 6. This adjustment logic ensures that the target network configuration quantity is positively correlated with the damage risk, facilitating adaptive configuration with more resources for high-risk networks and fewer resources for low-risk networks.

[0097] Furthermore, based on the calculated number of target network configurations, multiple selected detection networks are randomly chosen from a pre-constructed damage detection network pool. These selected detection networks are then integrated through a network output processing layer. The network output processing layer is the core module for multi-network collaboration, and its functions include unifying input and output formats, summarizing detection results, and ensuring the stability of the integrated detector. Specifically, the network output processing layer first unifies the input interfaces of the selected detection networks, synchronously inputting the production time-series parameters of the target carbon fiber composite material into all selected networks in the same format, ensuring that each network can perform independent detection based on consistent input data. Secondly, the processing layer receives the binary detection results of "damaged or undamaged" from each selected network. Finally, the processing layer integrates multiple selected detection networks into a unified functional module, namely the target damage detector. This detector is designed only for the current target's potential damage type, directly receiving production time-series parameters and outputting multi-network detection results, providing direct support for calculating damage probabilities.

[0098] Further, based on the target damage detector, damage assessment is performed on the production time-series parameters to obtain the damage probability of the target potential damage type, including:

[0099] The production time series parameters are input into the target damage detector, and the target damage detector performs damage detection on the production time series parameters through multiple selected detection networks to obtain multiple damage detection results.

[0100] The number of detection networks identified as having damage from the multiple damage detection results is counted to obtain the number of damage detection networks. Based on the number of damage detection networks and the total number of detection networks, the damage probability of the target potential damage type is obtained.

[0101] First, the production time-series parameters of the target carbon fiber composite material are input into the target damage detector. Multiple selected detection networks perform damage detection in parallel and output the results. The multiple selected detection networks inside the target damage detector will independently detect the input production time-series parameters based on their respective training backgrounds and architectural characteristics. Each network will output a binary detection result of "damaged" or "no damaged" according to its own damage identification logic, ultimately forming multiple independent sets of detection results.

[0102] Secondly, the number of networks that are identified as "damaged" among multiple damage detection results is counted, and the damage probability of the potential damage type of the target is calculated based on the total number of detection networks. In practice, the output results of all selected detection networks are first categorized and statistically analyzed. For example, if 10 detection networks are selected, and 6 of them are identified as "damaged" and 4 are identified as "no damage", then the number of damage detection networks is 6, the total number of detection networks is 10, and the damage probability = number of damage detection networks ÷ total number of detection networks × 100%, the damage probability is (6 ÷ 10) × 100% = 60%.

[0103] This probability calculation method based on "multi-network voting" can effectively offset the prediction bias caused by the limitations of training data or architectural characteristics of a single network. For example, a network may misjudge a certain type of process fluctuation scenario because it does not cover it, while the correct judgment of other networks can balance this bias, significantly improving the stability and reliability of damage probability assessment. At the same time, high-risk damage types, due to the configuration of more detection networks, have stronger statistical significance in their voting results and higher probability calculation accuracy, ultimately providing a precise quantitative basis for the subsequent determination of "high-risk / low-risk material identification".

[0104] Finally, in the same way as obtaining the damage probability of the target potential damage type, the damage probabilities of the other potential damage types are obtained. The other potential damage types are traversed, and the other potential damage types and their corresponding damage detection factors are determined in turn. For each other potential damage type, the steps of calling the network pool and configuring the target damage detector are repeated to obtain the damage probability of each other potential damage type until the damage probability calculation of each potential damage type is completed. Finally, the damage probabilities of each of the multiple potential damage types are obtained.

[0105] S40: When the probability of any of the multiple potential damage types is greater than the corresponding preset probability threshold, the target carbon fiber composite material is identified as a high-risk material.

[0106] Specifically, the probability of each potential damage type is compared with its corresponding preset threshold. If the probability of any potential damage type exceeds its threshold, a "high-risk material identification" is triggered. For example, the potential damage types of a carbon fiber composite sample include: delamination damage (15% probability, preset threshold 12%), fiber breakage (8% probability, preset threshold 10%), and matrix cracking (20% probability, preset threshold 22%). Since the 15% probability of delamination damage is greater than 12%, even if the probabilities of the other two damage types do not exceed the preset thresholds, the sample still needs to be identified as a high-risk material. This veto-judgment method can minimize the risk of material failure due to a single high-risk damage.

[0107] In summary, the embodiments of this application have at least the following technical effects:

[0108] This invention effectively predicts the potential damage probability of carbon fiber composite materials by introducing machine learning technology, achieving intelligent optimization of detection resources and significantly improving detection efficiency. First, by analyzing production time-series parameters and combining them with historical sample data, the probability of various potential damages is accurately identified and quantified, providing a reliable basis for material quality assessment. Second, based on damage detection factors, the number of detection networks is dynamically adjusted, implementing a resource allocation strategy that prioritizes the detection of high-risk damage types and moderately detects low-risk types, avoiding resource waste. Third, through multi-network integration and a voting mechanism, efficient and stable damage assessment is achieved, reducing reliance on traditional manual or hardware detection, shortening the detection cycle, and adapting to the needs of mass production. Finally, automated quality labeling based on preset probability thresholds ensures the objectivity and consistency of the judgment process, effectively improving the accuracy and reliability of product quality control.

[0109] Example 2, as Figure 2 As shown, based on the same inventive concept as the machine learning-based carbon fiber composite material damage prediction method provided in Embodiment 1, this embodiment of the invention also provides a machine learning-based carbon fiber composite material damage prediction device, including:

[0110] The parameter acquisition and damage type preliminary judgment module 11 is used to acquire the production time sequence parameters of the target carbon fiber composite material, and determine multiple possible damage types of the target carbon fiber composite material and the damage occurrence rate of each possible damage type based on the production time sequence parameters.

[0111] The potential damage screening and factor acquisition module 12 is used to select possible damage types with a damage occurrence rate greater than the corresponding occurrence rate threshold, determine multiple potential damage types, and acquire the damage detection factor for each potential damage type.

[0112] Damage probability calculation module 13 is used to determine the damage probability of multiple potential damage types based on the damage detection factors of each potential damage type.

[0113] The quality judgment and identification module 14 is used to identify the target carbon fiber composite material as a high-risk material when the probability of any of the multiple potential damage types is greater than the corresponding preset probability threshold.

[0114] Specifically, the parameter acquisition and damage type preliminary judgment module 11 is used for:

[0115] When the damage probability of multiple potential damage types is less than or equal to the corresponding preset probability threshold, the target carbon fiber composite material is identified as a low-risk material.

[0116] Further, production timeline parameters of the target carbon fiber composite material are obtained, and multiple possible damage types of the target carbon fiber composite material and the damage incidence rate of each possible damage type are determined based on the production timeline parameters, including:

[0117] Based on the production timing parameters of the target carbon fiber composite material, a production parameter timing feature is generated, and the material type and expected service life of the target carbon fiber composite material are obtained.

[0118] Based on the time-series characteristics of the production parameters, the material type, and the expected service life, a sample set of target carbon fiber composite materials is retrieved.

[0119] The damage record data of each historical carbon fiber composite material sample in the target carbon fiber composite material sample set are statistically analyzed to identify the damage types that occur and determine multiple possible damage types.

[0120] The damage incidence rate of each of the possible damage types is obtained by calculating the ratio of the number of occurrences of each possible damage type in the target carbon fiber composite material sample set to the total number of samples in the target carbon fiber composite material sample set.

[0121] Further, based on the time-series characteristics of the production parameters, the material type, and the expected service life, a sample set of target carbon fiber composite materials is retrieved, including:

[0122] Based on the material type, historical carbon fiber composite material samples with the same material type are retrieved to obtain the first carbon fiber composite material sample set;

[0123] Based on the expected service life, the first carbon fiber composite material sample set is screened, and historical carbon fiber composite material samples with a service life less than or equal to the expected service life are selected to obtain the second carbon fiber composite material sample set.

[0124] Calculate the similarity between the time-series characteristics of the production parameters and the time-series characteristics of the production parameters of each historical carbon fiber composite material sample in the second carbon fiber composite material sample set. Select historical carbon fiber composite material samples with similarity greater than a preset similarity threshold to obtain a third carbon fiber composite material sample set, which is used as the target carbon fiber composite material sample set.

[0125] Specifically, the potential damage screening and factor acquisition module 12 is used for:

[0126] Possible damage types with an incidence rate greater than a corresponding incidence rate threshold are selected to determine multiple potential damage types, and damage detection factors for each potential damage type are obtained, including:

[0127] A first possible damage type is determined from the possible damage types, and a first damage occurrence rate of the first possible damage type is extracted to obtain the corresponding first occurrence rate threshold.

[0128] When the first injury occurrence rate is greater than or equal to the first occurrence rate threshold, the first possible injury type is determined as the first potential injury type;

[0129] Based on the first damage incidence rate and the first incidence rate threshold, obtain the first damage detection factor for the first potential damage type;

[0130] The remaining possible damage types are processed to obtain multiple potential damage types and damage detection factors for each potential damage type.

[0131] The damage probability calculation module 13 is specifically used for:

[0132] Based on the damage detection factors for each of the aforementioned potential damage types, the damage probabilities of multiple potential damage types are determined, including:

[0133] Traverse multiple potential damage types, obtain the target potential damage type, and extract the corresponding damage detection factors to obtain the target damage detection factor;

[0134] The corresponding damage detection network pool is retrieved according to the target potential damage type, and the damage detection network is selected in the damage detection network pool based on the target damage detection factor, and integrated to obtain the target damage detector.

[0135] Based on the target damage detector, damage assessment is performed on the production time series parameters to obtain the damage probability of the target potential damage type;

[0136] Based on the method of obtaining the damage probability of the target potential damage type, the damage probabilities of other potential damage types are obtained, resulting in multiple potential damage probabilities.

[0137] The steps for constructing the damage detection network pool include:

[0138] Multiple sample production time-series parameters were collected to construct a sample production time-series parameter set. Then, based on the expected service duration, the potential damage types of the multiple sample production time-series parameters were labeled to construct a sample damage label set.

[0139] Construct multiple damage detection network architectures;

[0140] Multiple damage detection network architectures are trained based on the sample production time series parameter set and the sample damage label set to obtain multiple damage detection networks.

[0141] The multiple damage detection networks are stored to obtain the damage detection network pool.

[0142] Further, based on the target damage detection factor, damage detection networks are selected from the damage detection network pool and integrated to obtain a target damage detector, including:

[0143] Obtain network integration baseline data;

[0144] The target network configuration quantity is obtained by adjusting the network integration baseline number using the target damage detection factor.

[0145] Based on the target network configuration quantity, damage detection networks are randomly selected from the damage detection network pool to obtain multiple selected detection networks;

[0146] The target damage detector is obtained by integrating multiple selected detection networks through a network output processing layer.

[0147] Further, based on the target damage detector, damage assessment is performed on the production time-series parameters to obtain the damage probability of the target potential damage type, including:

[0148] The production time series parameters are input into the target damage detector, and the target damage detector performs damage detection on the production time series parameters through multiple selected detection networks to obtain multiple damage detection results.

[0149] The number of detection networks identified as having damage from the multiple damage detection results is counted to obtain the number of damage detection networks. Based on the number of damage detection networks and the total number of detection networks, the damage probability of the target potential damage type is obtained.

[0150] Specifically, the quality determination and identification module 14 is used for:

[0151] In the non-destructive testing (NDT) process for carbon fiber composites, the ultimate goal of NDT is to determine the material's quality status, and the quality labeling includes two results: "high risk" and "low risk." When the probability of a certain potential damage type is greater than its corresponding preset probability threshold, the carbon fiber composite is labeled as a high-risk material; when the probability of all potential damage types is less than or equal to their respective preset probability thresholds, the carbon fiber composite is labeled as a low-risk material.

[0152] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0153] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0154] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. A damage prediction method for carbon fiber composite materials based on machine learning, characterized in that, The method includes: Obtain the production time sequence parameters of the target carbon fiber composite material, and determine multiple possible damage types of the target carbon fiber composite material and the damage incidence rate of each possible damage type based on the production time sequence parameters; Possible damage types with an incidence rate greater than a corresponding incidence rate threshold are selected to determine multiple potential damage types, and damage detection factors for each potential damage type are obtained, including: A first possible damage type is determined from the possible damage types, and a first damage occurrence rate of the first possible damage type is extracted to obtain the corresponding first occurrence rate threshold. When the first injury occurrence rate is greater than or equal to the first occurrence rate threshold, the first possible injury type is determined as the first potential injury type; Based on the first damage incidence rate and the first incidence rate threshold, obtain the first damage detection factor for the first potential damage type; Continue processing the remaining possible damage types to obtain multiple potential damage types and damage detection factors for each potential damage type; Based on the damage detection factors for each of the aforementioned potential damage types, the damage probabilities of multiple potential damage types are determined, including: Traverse multiple potential damage types, obtain the target potential damage type, and extract the corresponding damage detection factors to obtain the target damage detection factor; The corresponding damage detection network pool is retrieved according to the target potential damage type, and the damage detection network is selected in the damage detection network pool based on the target damage detection factor, and integrated to obtain the target damage detector. Based on the target damage detector, damage assessment is performed on the production time series parameters to obtain the damage probability of the target potential damage type; Based on the method of obtaining the damage probability of the target potential damage type, the damage probabilities of other potential damage types are obtained, resulting in multiple potential damage probabilities. The steps for constructing the damage detection network pool include: Multiple sample production time-series parameters were collected to construct a sample production time-series parameter set. Then, based on the expected service duration, the potential damage types of the multiple sample production time-series parameters were labeled to construct a sample damage label set. Construct multiple damage detection network architectures; Multiple damage detection network architectures are trained based on the sample production time series parameter set and the sample damage label set to obtain multiple damage detection networks. The multiple damage detection networks are stored to obtain the damage detection network pool; Specifically, based on the target damage detection factor, damage detection networks are selected from the damage detection network pool and integrated to obtain a target damage detector, including: Obtain network integration baseline data; The target network configuration quantity is obtained by adjusting the network integration baseline number using the target damage detection factor. Based on the target network configuration quantity, damage detection networks are randomly selected from the damage detection network pool to obtain multiple selected detection networks; The target damage detector is obtained by integrating multiple selected detection networks through a network output processing layer; Specifically, based on the target damage detector, damage assessment is performed on the production time-series parameters to obtain the damage probability of the target potential damage type, including: The production time series parameters are input into the target damage detector, and the target damage detector performs damage detection on the production time series parameters through multiple selected detection networks to obtain multiple damage detection results. The number of detection networks that are identified as having damage in the multiple damage detection results is counted to obtain the number of damage detection networks. Based on the number of damage detection networks and the total number of detection networks, the damage probability of the target potential damage type is obtained. When the probability of any of the multiple potential damage types is greater than the corresponding preset probability threshold, the target carbon fiber composite material is identified as a high-risk material.

2. The method according to claim 1, characterized in that, When the damage probability of multiple potential damage types is less than or equal to the corresponding preset probability threshold, the target carbon fiber composite material is identified as a low-risk material.

3. The method according to claim 1, characterized in that, Obtain production timeline parameters of the target carbon fiber composite material, and determine multiple possible damage types of the target carbon fiber composite material and the damage incidence rate of each possible damage type based on the production timeline parameters, including: Based on the production timing parameters of the target carbon fiber composite material, a production parameter timing feature is generated, and the material type and expected service life of the target carbon fiber composite material are obtained. Based on the time-series characteristics of the production parameters, the material type, and the expected service life, a sample set of target carbon fiber composite materials is retrieved. The damage record data of each historical carbon fiber composite material sample in the target carbon fiber composite material sample set are statistically analyzed to identify the damage types that occur and determine multiple possible damage types. The damage incidence rate of each of the possible damage types is obtained by calculating the ratio of the number of occurrences of each possible damage type in the target carbon fiber composite material sample set to the total number of samples in the target carbon fiber composite material sample set.

4. The method according to claim 3, characterized in that, Based on the time-series characteristics of the production parameters, the material type, and the expected service life, a sample set of target carbon fiber composite materials is retrieved, including: Based on the material type, historical carbon fiber composite material samples with the same material type are retrieved to obtain the first carbon fiber composite material sample set; Based on the expected service life, the first carbon fiber composite material sample set is screened, and historical carbon fiber composite material samples with a service life less than or equal to the expected service life are selected to obtain the second carbon fiber composite material sample set. Calculate the similarity between the time-series characteristics of the production parameters and the time-series characteristics of the production parameters of each historical carbon fiber composite material sample in the second carbon fiber composite material sample set. Select historical carbon fiber composite material samples with similarity greater than a preset similarity threshold to obtain a third carbon fiber composite material sample set, which is used as the target carbon fiber composite material sample set.

5. A machine learning-based damage prediction device for carbon fiber composite materials, characterized in that, For performing the method according to any one of claims 1-4, comprising: The parameter acquisition and damage type preliminary judgment module is used to acquire the production time sequence parameters of the target carbon fiber composite material, and determine multiple possible damage types of the target carbon fiber composite material and the damage occurrence rate of each possible damage type based on the production time sequence parameters. The potential damage screening and factor acquisition module is used to select possible damage types with a damage incidence rate greater than a corresponding incidence rate threshold, determine multiple potential damage types, and acquire damage detection factors for each potential damage type, including: A first possible damage type is determined from the possible damage types, and a first damage occurrence rate of the first possible damage type is extracted to obtain the corresponding first occurrence rate threshold. When the first injury occurrence rate is greater than or equal to the first occurrence rate threshold, the first possible injury type is determined as the first potential injury type; Based on the first damage incidence rate and the first incidence rate threshold, obtain the first damage detection factor for the first potential damage type; Continue processing the remaining possible damage types to obtain multiple potential damage types and damage detection factors for each potential damage type; The damage probability calculation module is used to determine the damage probability of multiple potential damage types based on the damage detection factors of each potential damage type, including: Traverse multiple potential damage types, obtain the target potential damage type, and extract the corresponding damage detection factors to obtain the target damage detection factor; The corresponding damage detection network pool is retrieved according to the target potential damage type, and the damage detection network is selected in the damage detection network pool based on the target damage detection factor, and integrated to obtain the target damage detector. Based on the target damage detector, damage assessment is performed on the production time series parameters to obtain the damage probability of the target potential damage type; Based on the method of obtaining the damage probability of the target potential damage type, the damage probabilities of other potential damage types are obtained, resulting in multiple potential damage probabilities. The steps for constructing the damage detection network pool include: Multiple sample production time-series parameters were collected to construct a sample production time-series parameter set. Then, based on the expected service duration, the potential damage types of the multiple sample production time-series parameters were labeled to construct a sample damage label set. Construct multiple damage detection network architectures; Multiple damage detection network architectures are trained based on the sample production time series parameter set and the sample damage label set to obtain multiple damage detection networks. The multiple damage detection networks are stored to obtain the damage detection network pool; Specifically, based on the target damage detection factor, damage detection networks are selected from the damage detection network pool and integrated to obtain a target damage detector, including: Obtain network integration baseline data; The target network configuration quantity is obtained by adjusting the network integration baseline number using the target damage detection factor. Based on the target network configuration quantity, damage detection networks are randomly selected from the damage detection network pool to obtain multiple selected detection networks; The target damage detector is obtained by integrating multiple selected detection networks through a network output processing layer; Specifically, based on the target damage detector, damage assessment is performed on the production time-series parameters to obtain the damage probability of the target potential damage type, including: The production time series parameters are input into the target damage detector, and the target damage detector performs damage detection on the production time series parameters through multiple selected detection networks to obtain multiple damage detection results. The number of detection networks that are identified as having damage in the multiple damage detection results is counted to obtain the number of damage detection networks. Based on the number of damage detection networks and the total number of detection networks, the damage probability of the target potential damage type is obtained. The quality assessment and identification module is used to identify the target carbon fiber composite material as a high-risk material when the probability of any of the multiple potential damage types is greater than the corresponding preset probability threshold.

Images