A carbon fiber broken wire intelligent detection method and system

Abstract

The present application provides a method and system for intelligent detection of broken carbon fiber, which relates to the technical field of intelligent detection. The detection method includes the steps of determining an area, collecting data, first modeling, intelligent detection, and result judgment; the detection system includes an area determination module, a data collection module, a first modeling module, an intelligent detection module, and a result judgment module. The present application detects whether the carbon fiber is broken based on environmental data, thereby improving the accuracy of carbon fiber broken wire detection.

Description

Technical Field

[0001] The present application relates to the technical field of intelligent detection, and in particular to a method and system for intelligent detection of broken carbon fiber filaments. Background Art

[0002] In the field of modern industrial manufacturing, carbon fiber, as a lightweight, high-strength, corrosion-resistant advanced material, is widely used in a variety of industries such as aerospace, automobile manufacturing, sporting goods, and building structures. However, with the widespread application of carbon fiber composite materials in these fields, their safety and reliability issues are becoming increasingly prominent. During the use of carbon fiber, due to changes in environmental conditions (such as extreme temperature, humidity fluctuations, long-term stress and vibration, etc.) or factors such as aging and fatigue of the material itself, the fiber may break. Fiber breakage will not only reduce the overall performance of the material, but may also cause serious safety accidents.

[0003] The Chinese invention patent with the application publication date of May 12, 2023 and publication number CN116109642A provides a method, device and storage medium for detecting broken wire defects in carbon fiber. The patent performs threshold segmentation on the image data, extracts multiple target areas, obtains the minimum circumscribed rectangle of the target area, extracts its characteristic parameters, and compares the characteristic parameters with the preset threshold value, and determines whether there is a broken wire defect based on the comparison result. The patent mainly analyzes the image data of carbon fiber, extracts its characteristic parameters, and compares the characteristic parameters with the preset threshold value to obtain its judgment result. However, image processing can only detect broken carbon fiber, and it is impossible to know whether environmental factors will affect whether the carbon fiber will break in the future. Summary of the invention

[0004] In order to improve detection efficiency and accuracy and predict whether environmental factors will cause carbon fiber breakage, the present application provides a carbon fiber breakage intelligent detection method and system.

[0005] In the first aspect, the present application provides a method for intelligently detecting broken carbon fiber filaments, which adopts the following technical solution:

[0006] A method for intelligently detecting broken carbon fiber filaments comprises the following steps:

[0007] Determine the area: Determine the detection area;

[0008] Data collection: including the first collection and the second collection;

[0009] First acquisition: extracting the data of carbon fiber in the detection area, which is recorded as the first data;

[0010] Second collection: collecting environmental data;

[0011] First modeling: establishing a detection model, wherein the detection model includes: a convolutional neural network sub-model, a BI-LSTM sub-model and a decision tree sub-model;

[0012] Intelligent detection: including the first detection, the second detection and the third detection;

[0013] First detection: inputting the first data into the convolutional neural network sub-model to obtain a first output result;

[0014] Second detection: input the environmental data into the BI-LSTM sub-model to obtain the second output result;

[0015] Third test: input the first output result and the second output result into the decision tree sub-model, and output the third output result;

[0016] Result judgment: judge whether the third output result meets expectations. If so, output a detection result indicating no broken wires; if not, output a detection result indicating broken wires.

[0017] By adopting the above technical solution, through the combination of the first acquisition and the second acquisition, we not only pay attention to the characteristics of the carbon fiber itself (first data), but also consider the influence of external environmental factors (temperature, humidity, stress and vibration) on the state of the carbon fiber. This comprehensive data acquisition method helps to more accurately reflect the actual state of the carbon fiber. The detection model of the present application includes a convolutional network sub-model, a BI-LSTM sub-model and a decision tree sub-model. The convolutional network sub-model can automatically extract features and identify subtle changes in carbon fiber by processing carbon fiber data. The BI-LSTM sub-model is suitable for processing time series data, such as time-varying sequences such as temperature and humidity in environmental data, and can capture the temporal dependencies between environmental data. The decision tree sub-model, as the final decision layer, can integrate the outputs of the first two sub-models, perform logical judgments, and output the final detection results. The present application uses intelligent detection steps to perform detection in stages (first detection, second detection and third detection). The final output result combines the advantages of different models and improves the accuracy of the detection results. In particular, the third detection step uses the decision tree sub-model to make a comprehensive judgment on the first two detection results, reducing the possible misjudgment of a single model. Then, by judging whether the third output result meets expectations, it is judged whether the carbon fiber has the risk of broken wires.

[0018] Optionally, after executing the step of data collection and before executing the step of first modeling, the method further includes:

[0019] Data fusion: including network construction, second modeling, initial value determination, optimal value determination and update;

[0020] Network construction: build wireless sensor network;

[0021] Second modeling: establish BP neural network model;

[0022] Determine the initial value: Based on the sparrow search algorithm, determine the initial weight and initial threshold of the BP neural network model;

[0023] Determine the optimal value: adjust the weights and thresholds of the BP neural network model through the back propagation algorithm, obtain the optimal weights and thresholds of the BP neural network model, and transmit the optimal weights and thresholds of the BP neural network model to the wireless sensor network;

[0024] Update: The environmental data is fused using a wireless sensor network, and the fused environmental data is used as new environmental data.

[0025] By adopting the above technical solutions, building a wireless sensor network (WSN) can achieve real-time monitoring of the detection area environment, collect more comprehensive and detailed environmental data, and provide more valuable information for subsequent modeling. Using the sparrow search algorithm to determine the initial weights and initial thresholds of the BP neural network can reduce the risk of local optimal problems caused by traditional random initialization methods and improve convergence speed and performance. The parameters of the BP neural network model are adjusted by the back propagation algorithm to obtain the optimal weights and optimal thresholds. This process optimizes the prediction ability of the BP neural network model, enabling it to more accurately reflect the impact of environmental data on the state of carbon fiber. After the wireless sensor network fuses the environmental data, the fused data is used as the new environmental data, which helps to reduce noise and redundant information, improve the representativeness and reliability of environmental data, and provide more powerful support for subsequent intelligent detection. The BP neural network model determines the initial weights and initial thresholds through the sparrow search algorithm, and uses the back propagation algorithm to optimize its own parameters, which can better adapt to and fit the complexity and nonlinear relationship of environmental data, thereby improving detection accuracy.

[0026] Optionally, after performing the step of data fusion and before performing the step of first modeling, the method further includes:

[0027] Initial dimensionality reduction: including relevance analysis, relevance judgment, retention, elimination and iteration;

[0028] Correlation analysis: Based on the single factor analysis algorithm, the correlation between the i-th environmental data and the second output result is calculated and recorded as the second data;

[0029] Relevance judgment: judging whether the second data is greater than a preset relevance threshold, if so, executing the retaining step; if not, executing the eliminating step;

[0030] Keep: keep the i-th environment data;

[0031] Eliminate: delete the i-th environment data;

[0032] Iteration: Take the i+1th environmental data as the new ith environmental data and perform the correlation analysis step.

[0033] By adopting the above technical solution, by eliminating environmental data that is not related to the second output result (i.e., the output of the BI-LSTM sub-model), the amount of data that needs to be processed in the subsequent process can be reduced, thereby reducing the computational cost and improving the processing speed. Retaining environmental data that is highly related to the second output result helps the detection model to capture key features more accurately and improve the accuracy of prediction and classification. According to the results of the relevance judgment, flexibly choosing to retain or eliminate environmental data helps to remove noise and redundant information and retain key data that is useful for detection. The environmental data after the preliminary dimensionality reduction process is more refined and efficient. Using it as input data can significantly improve the training speed and prediction accuracy of the detection model. Since irrelevant data is removed, the detection model can focus more on learning key features during the training process, thereby improving the generalization ability and robustness of the detection model.

[0034] Optionally, after performing the step of preliminary dimensionality reduction and before performing the step of first modeling, the method further includes:

[0035] Deep dimensionality reduction: including integration, distance calculation, centering, inner product matrix construction, eigendecomposition, feature selection and obtaining results;

[0036] Integration: All environmental data in the retained steps are integrated into an environmental data set, and one environmental data is recorded as a sample in the environmental data set;

[0037] Calculate distance: Calculate the distance between all sample pairs and integrate all distances into the first distance matrix;

[0038] Centralization: Centralize the first distance matrix to obtain the second distance matrix;

[0039] Construct the inner product matrix: Based on the second distance matrix, calculate the inner product matrix G. The calculation model is as follows:

[0040] ;

[0041] ;

[0042] in, is the i-th vector in the second distance matrix and the jth vector The inner product of is a diagonal matrix constructed according to the second distance matrix; is a diagonal matrix The element at row i and column i; is a diagonal matrix The element at row j and column j; is the distance between the i-th sample and the j-th sample in the environmental data set;

[0043] Eigendecomposition: Perform eigendecomposition on the inner product matrix to obtain eigenvalues ​​and eigenvectors;

[0044] Feature selection: select the first n eigenvalues ​​and the eigenvectors corresponding to the eigenvalues;

[0045] Obtain results: Map the samples to the corresponding feature vectors to obtain the reduced-dimensional environmental data. The calculation model is as follows:

[0046] ;

[0047] Among them, X is the environmental data after dimension reduction; is an eigenvector matrix composed of n eigenvectors; is a diagonal matrix consisting of n eigenvalues.

[0048] By adopting the above technical solution, the present application calculates the distance between all sample pairs, reveals the similarities and differences between samples, and helps identify the key patterns and structures in the environmental data set. In the process of dimensionality reduction, centralization helps to reduce noise and redundant information in environmental data and improve the dimensionality reduction effect. Based on the first distance matrix after centralization, the inner product matrix is ​​constructed, and the distance information between samples is used to construct the similarity matrix between samples. The inner product matrix reflects the intrinsic relationship between samples and provides a basis for subsequent eigendecomposition. The inner product matrix is ​​subjected to eigendecomposition to obtain eigenvalues ​​and eigenvectors. These eigenvalues ​​and eigenvectors represent the main direction and degree of change in the sample. In the process of dimensionality reduction, by selecting important eigenvalues ​​and corresponding eigenvectors, the key information in the sample can be retained while removing redundancy and noise. By selecting the first n eigenvalues ​​and their corresponding eigenvectors, the dimension of the environmental data can be further reduced while retaining important information. The sample is mapped with the selected eigenvector to obtain the environmental data after dimensionality reduction, and the original environmental data is projected into a low-dimensional space while retaining the key information of the original environmental data as much as possible.

[0049] Optionally, after performing the step of deep dimensionality reduction and before performing the step of first modeling, the method further includes:

[0050] Expansion: Generate adversarial networks to learn the distribution characteristics of the reduced-dimensional environmental data and generate new environmental data. The reduced-dimensional environmental data and the generated environmental data are used as new environmental data.

[0051] By adopting the above technical solutions, the environmental data generated by the generative adversarial network (GAN) can effectively increase the amount of data, especially for sparse or unbalanced data sets. This method can significantly improve the generalization ability of model training. The expanded data set contains more types of environmental conditions, enabling the detection model to make predictions or decisions more stably when facing new or unknown environments. The application of GAN on dimensionality reduction data further demonstrates the potential of the detection model in complex data processing and enhancement. In addition, through the steps of preliminary dimensionality reduction and deep dimensionality reduction, the dimensionality reduction processing of the fused environmental data is achieved. At this time, expanding it will help increase the amount of data and thus fully train the detection model.

[0052] Optionally, after executing the step of data collection and before executing the step of first modeling, the method further includes:

[0053] Data processing: including preprocessing and time alignment;

[0054] Preprocessing: preprocessing the environmental data, and recording the preprocessed environmental data as the third data;

[0055] Time alignment: extract the third data at the same time and use the extracted third data as new environment data.

[0056] By adopting the above technical solutions, data processing can remove noise and outliers in environmental data, fill missing values, etc., thereby improving the overall quality of environmental data and providing a reliable data foundation for subsequent modeling. In environmental data collected by multiple data sources or sensors, due to various reasons (such as device delays, synchronization problems, etc.), the data may not be recorded completely at the same time. Time alignment is to adjust these environmental data to the same time frame to ensure that their observations at the same time can be extracted simultaneously. Time-aligned environmental data can more accurately reflect changes in environmental status, thereby improving the prediction or classification accuracy of the model. Especially in scenarios where multi-source data fusion is required, time alignment is the key to ensuring the effectiveness of the model. Preprocessed and time-aligned data can reduce the uncertainty of the detection model during the training process, reduce the sensitivity of the detection model to noise, and thus enhance its robustness. In applications that require multi-source environmental data fusion, preprocessing and time alignment help to fully utilize the advantages of various data sources and improve the performance of the detection model.

[0057] Optionally, after performing the time alignment step and before performing the first modeling step, the method further includes:

[0058] Extracting time interval: extracting the time interval of the third data in the step of obtaining time alignment;

[0059] Determine the time interval: determine whether the time interval is less than a preset time interval threshold, if so, execute the first modeling step; if not, execute the adjustment step;

[0060] Adjustment: Reduce the time interval for collecting environmental data.

[0061] By adopting the above technical solution, the update frequency of the third data can be quantified by extracting the time interval of the third data (i.e., the environmental data after preprocessing and time alignment) in the step of obtaining time alignment. By judging whether the time interval is less than the preset time interval threshold, it can be ensured that the environmental data used for modeling has sufficient timeliness. When the time interval is greater than the preset threshold, it means that the environmental data update speed is slow and cannot reflect the changes in the environmental state in time. By reducing the time interval for collecting environmental data, the timeliness and richness of environmental data can be increased, and the quality of environmental data can be improved. Using more frequently updated environmental data for modeling can enable the detection model to more accurately capture the changing trend of the environmental state, thereby improving the prediction accuracy of the detection model.

[0062] Optionally, the data collection step further includes:

[0063] The third collection: collect historical data of carbon fiber;

[0064] Fourth collection: Collect historical data of environmental data.

[0065] By adopting the above technical solution, by collecting historical data of carbon fiber and environmental data (such as temperature, humidity, stress, vibration), multiple key factors affecting the performance or state of carbon fiber can be fully covered. This multi-dimensional data collection helps to more accurately understand the overall status of carbon fiber and its environment. The collection of historical data means that the data has a time span and can reflect the changing trend of carbon fiber and its environment over time. The present application also uses the historical data of carbon fiber to train the convolutional neural network sub-model, so that the convolutional neural network sub-model has the ability to predict and identify changes in carbon fiber performance. By using the historical data of carbon fiber at different time points, the convolutional neural network sub-model can learn the evolution of carbon fiber performance over time, including the mode and rate of performance degradation and the correlation with specific environmental factors. During the training process, the convolutional neural network sub-model will automatically extract features from the historical data of carbon fiber, such as microstructural changes, fluctuations in mechanical performance parameters, etc., and construct a complex mapping relationship between these features and the current or future performance state of carbon fiber. Using historical data of environmental data (including temperature, humidity, stress, vibration) to train the BI-LSTM sub-model can fully explore and utilize the information and laws contained in these time series data. During the training process, the historical data of environmental data is input into the BI-LSTM sub-model, which builds a complex mapping relationship between environmental data and changes in carbon fiber performance by learning the time dependency and correlation in these environmental data. In other words, the BI-LSTM sub-model can capture the impact of long-term and short-term changes in environmental factors on carbon fiber performance, and how these impacts accumulate and evolve over time.

[0066] Optionally, after executing the step of data collection and before executing the step of first modeling, the method further includes:

[0067] Quantity judgment: judging whether the quantity of various historical data is greater than a preset quantity threshold, if so, executing the first acquisition step; if not, executing the second acquisition step;

[0068] First acquisition: acquiring historical data whose quantity is less than a preset quantity threshold, recorded as fourth data;

[0069] Increasing the number: using a random oversampling algorithm to increase the fourth data until the number of the fourth data is not less than a preset number threshold;

[0070] Second acquisition: acquiring historical data whose quantity is greater than a preset quantity threshold, and recording it as fifth data;

[0071] Reduce the number: Use a random undersampling algorithm to delete the fifth data until the number of the fifth data is no more than a preset number threshold.

[0072] By adopting the above technical solution, by judging whether the amount of historical data meets the preset quantity threshold, and performing the corresponding operation of increasing or decreasing the quantity accordingly, the model bias caused by too much or too little data in certain categories or time periods can be effectively reduced. Data balance is one of the important factors affecting the generalization ability of the model. By ensuring that the amount of data in each category or time period is relatively balanced, the prediction accuracy and stability of the detection model in different situations can be improved. For the case where the amount of historical data is insufficient, a random oversampling algorithm is used to increase the diversity of the data, which helps the detection model learn more features and improve the generalization ability of the detection model. For the case where the amount of historical data is too much, a random undersampling algorithm is used to reduce the redundancy of the data and reduce the computational complexity of the model. At the same time, by retaining representative samples of the data, the detection model can reduce the amount of data without losing key information.

[0073] In a second aspect, the present invention provides an intelligent detection system for broken carbon fibers, which adopts the following technical solution:

[0074] A carbon fiber broken wire intelligent detection system, comprising:

[0075] A region determination module is used to determine a detection region;

[0076] A data acquisition module, which is in communication with the area determination module and the intelligent detection module, and the data acquisition module includes a first acquisition unit and a second acquisition unit;

[0077] A first acquisition unit is used to extract data of carbon fibers in a detection area, which is recorded as first data;

[0078] A second collection unit, used for collecting environmental data;

[0079] A first modeling module is communicatively connected with the intelligent detection module and is used to establish a detection model, wherein the detection model includes: a convolutional neural network sub-model, a BI-LSTM sub-model and a decision tree sub-model;

[0080] An intelligent detection module, comprising a first detection unit, a second detection unit and a third detection unit;

[0081] A first detection unit, used to input first data into a convolutional neural network sub-model to obtain a first output result;

[0082] A second detection unit is used to input the environmental data into the BI-LSTM sub-model to obtain a second output result;

[0083] A third detection unit, used to input the first output result and the second output result into the decision tree sub-model, and output a third output result;

[0084] The result judgment module is used to judge whether the third output result meets expectations.

[0085] By adopting the above technical solution, the present application first clarifies the detection range by determining the regional module, and then the first acquisition unit and the second acquisition unit in the data acquisition module respectively focus on the accurate extraction of carbon fiber data and the comprehensive collection of environmental data (temperature, humidity, stress and vibration), ensuring the richness and accuracy of the data source. In the data processing stage, the first modeling module integrates the in-depth analysis of carbon fiber data by the convolutional neural network sub-model, the dynamic capture of time series environmental data by the BI-LSTM sub-model, and the intelligent decision-making of multi-dimensional detection results by the decision tree sub-model, which significantly improves the advancement and comprehensiveness of the detection model. The intelligent detection module further utilizes the output results of these sub-models, and realizes the efficiency of the detection process and the accuracy of the output results through parallel processing and decision fusion strategies. Finally, the result judgment module judges the third output result according to the preset standard, providing users with an intuitive and reliable carbon fiber status assessment, and providing solid data support for subsequent decision-making and action planning.

[0086] In summary, the present application includes at least one of the following beneficial technical effects:

[0087] 1. By combining the first acquisition and the second acquisition, the characteristics of the carbon fiber itself (first data) are taken into account, and the influence of external environmental factors (temperature, humidity, stress and vibration) on the state of the carbon fiber is considered. This comprehensive data acquisition method helps to more accurately reflect the actual state of the carbon fiber. The detection model of the present application includes a convolutional network sub-model, a BI-LSTM sub-model and a decision tree sub-model. The convolutional network sub-model can automatically extract features and identify subtle changes in carbon fiber by processing carbon fiber data. The BI-LSTM sub-model is suitable for processing time series data, such as time-varying sequences such as temperature and humidity in environmental data, and can capture the temporal dependencies between environmental data. The decision tree sub-model, as the final decision layer, can integrate the outputs of the first two sub-models, perform logical judgments, and output the final detection results. The present application uses intelligent detection steps to perform detection in stages (first detection, second detection and third detection). The final output results combine the advantages of different models and improve the accuracy of the detection results. In particular, the third detection step uses the decision tree sub-model to make a comprehensive judgment on the first two detection results, reducing the misjudgment that may be caused by a single model. Then, by judging whether the third output result meets expectations, it is judged whether the carbon fiber has the risk of broken wires.

[0088] 2. By eliminating environmental data that is not related to the second output result (i.e., the output of the BI-LSTM sub-model), the amount of data that needs to be processed in the subsequent process can be reduced, thereby reducing the computational cost and increasing the processing speed. Retaining environmental data that is highly related to the second output result helps the detection model capture key features more accurately and improves the accuracy of prediction and classification. According to the results of the relevance judgment, flexibly choosing to retain or eliminate environmental data helps to remove noise and redundant information and retain key data that is useful for detection. The environmental data after the initial dimensionality reduction process is more refined and efficient. Using it as input data can significantly improve the training speed and prediction accuracy of the detection model. Since irrelevant data is removed, the detection model can focus more on learning key features during the training process, thereby improving the generalization ability and robustness of the detection model.

[0089] 3. This application calculates the distance between all sample pairs, reveals the similarities and differences between samples, and helps identify key patterns and structures in environmental data sets. In the process of dimensionality reduction, centralization helps to reduce noise and redundant information in environmental data and improve the effect of dimensionality reduction. The inner product matrix is ​​constructed based on the first distance matrix after centralization, and the distance information between samples is used to construct the similarity matrix between samples. The inner product matrix reflects the intrinsic relationship between samples and provides a basis for subsequent eigendecomposition. The inner product matrix is ​​subjected to eigendecomposition to obtain eigenvalues ​​and eigenvectors. These eigenvalues ​​and eigenvectors represent the main direction and degree of change in the sample. In the process of dimensionality reduction, selecting important eigenvalues ​​and corresponding eigenvectors can retain the key information in the sample while removing redundancy and noise. By selecting the first n eigenvalues ​​and their corresponding eigenvectors, the dimension of the environmental data can be further reduced while retaining important information. The sample is mapped with the selected eigenvector to obtain the environmental data after dimensionality reduction, which realizes the projection of the original environmental data into a low-dimensional space while retaining the key information of the original environmental data as much as possible. BRIEF DESCRIPTION OF THE DRAWINGS

[0090] Figure 1 is a flow chart of Example 1 of the present application;

[0091] Figure 2 is a flow chart of S2 data collection in Example 1 of the present application;

[0092] Figure 3 This is a flow chart from S2 data collection to S3 first modeling in Example 1 of the present application;

[0093] Figure 4 is a flow chart of Example 2 of the present application;

[0094] Figure 5 is a flow chart of S6 data fusion of Example 2 of the present application;

[0095] Figure 6 This is a flowchart of the preliminary dimensionality reduction in S7 of Example 2 of the present application;

[0096] Figure 7 It is a flowchart of S8 deep dimensionality reduction in Example 2 of the present application;

[0097] Figure 8 This is a flow chart of Example 3 of the present application. DETAILED DESCRIPTION

[0098] The following combination Figures 1 to 8 This application is described in further detail.

[0099] Embodiment 1: This embodiment discloses a method for intelligently detecting broken carbon fiber filaments. Figure 1 As shown, the detection method includes: S1 determining the area, S2 data collection, S3 first modeling, S4 intelligent detection and S5 result judgment. First, the detection area is determined, then the data of the carbon fiber in the detection area is extracted, then the environmental data is collected, and then the carbon fiber is detected for broken wires based on the data of the carbon fiber and the environmental data to obtain an output result. This embodiment includes the following steps:

[0100] S1 determines the area, and determines the detection area. The detection area is determined based on the design, usage conditions and historical failure data of the carbon fiber structure. The selection of the detection area should be able to fully reflect the health status of the carbon fiber structure. The detection area in this step refers to: an area that is susceptible to stress concentration, temperature change, humidity or vibration shock.

[0101] S2 data collection, including S21 first collection and S22 second collection.

[0102] S21 first acquisition uses non-destructive testing technology (such as X-ray, ultrasonic scanning, infrared thermal imaging, etc.) to scan the inspection area to obtain parameters such as the microstructure, density change, and thermal conductivity of the carbon fiber. These parameters can indirectly reflect the integrity of the carbon fiber. The collected carbon fiber data is recorded as the first data.

[0103] S22 Second collection, environmental factors such as temperature, humidity, stress state and vibration level will affect the performance of carbon fiber, and even directly cause broken fibers. Therefore, this step is equipped with corresponding sensors to monitor these environmental parameters in real time or regularly to capture environmental factors that affect the integrity of carbon fiber.

[0104] S3 first builds a detection model, which includes a convolutional neural network sub-model, a BI-LSTM sub-model and a decision tree sub-model.

[0105] Convolutional Neural Network sub-model (CNN sub-model): It is good at processing images or data similar to grid structures, such as carbon fiber images obtained through X-ray or ultrasound scans, and can automatically extract features and identify potential anomalies.

[0106] BI-LSTM sub-model: Suitable for processing time series data, such as historical records of temperature, humidity, stress, and vibration, and capable of capturing temporal dependencies and long-term trends in various data.

[0107] Decision tree sub-model: As the final decision-making layer, it can make comprehensive judgments based on the output results of the first two sub-models to improve the accuracy and reliability of detection.

[0108] S4 intelligent detection, including S41 first detection, S42 second detection and S43 third detection.

[0109] S41 is a first detection, in which the first data (such as the carbon fiber image) is input into the CNN sub-model to obtain a preliminary judgment on the carbon fiber structural state, i.e., a first output result.

[0110] S42 is the second test, which inputs the environmental data into the BI-LSTM sub-model, analyzes the impact of the environmental data on the carbon fiber, and obtains the second output result.

[0111] S43 is a third test, in which the first output result and the second output result are input into a decision tree sub-model, and a comprehensive analysis is performed through the decision tree sub-model to output a third output result, i.e., whether there is a broken fiber in the carbon fiber.

[0112] S5: Result judgment, judging whether the third output result meets expectations, if yes, outputting a detection result of no broken wire; if no, outputting a detection result of broken wire. The detection result of broken wire means: immediately issuing an alarm, prompting that further inspection and maintenance work is required.

[0113] This embodiment clarifies the detection area, collects the first data and environmental data (temperature, humidity, stress and vibration) of the carbon fiber, and then uses a detection model composed of a convolutional neural network sub-model (CNN sub-model), a BI-LSTM sub-model and a decision tree sub-model for intelligent analysis: first, the CNN sub-model processes the data of the carbon fiber to identify potential anomalies; then, the BI-LSTM sub-model analyzes the impact of environmental data on the state of the carbon fiber; finally, the decision tree sub-model combines the results of the two to determine whether the carbon fiber has broken fibers, and outputs the detection results accordingly to ensure the safe use of the carbon fiber structure.

[0114] like Figure 2 As shown, in other embodiments, in S2 data collection, it also includes:

[0115] S23 is the third step of collecting historical data of carbon fiber. This step aims to collect status data of carbon fiber in the past period of time, including previous inspection records of carbon fiber, performance changes during use, etc.

[0116] S24 is a fourth collection, collecting historical data of environmental data, wherein the historical data of environmental data includes: historical data of temperature data, historical data of humidity data, historical data of stress data and historical data of vibration data.

[0117] like Figure 3 As shown, after executing S2 data collection and before executing S3 first modeling, it also includes:

[0118] S25 is a quantity determination, which determines whether the quantity of various historical data is greater than a preset quantity threshold. If so, S26 is executed for the first acquisition; if not, S28 is executed for the second acquisition.

[0119] The various historical data described in this step include: the historical data of carbon fiber in the third collection in S23, and the historical data of temperature data, historical data of humidity data, historical data of stress data and historical data of vibration data in the fourth collection in S24.

[0120] Because machine learning models (i.e., the detection model in S3 first modeling) usually require a sufficient amount of data to train to ensure their generalization ability and accuracy, and the preset quantity threshold is set based on the needs of the detection model and the availability of data.

[0121] S26 first obtains historical data whose quantity is less than a preset quantity threshold, and records it as fourth data.

[0122] S27 increases the number by using a random oversampling algorithm to increase the fourth data until the number of the fourth data is not less than a preset number threshold.

[0123] S28 is a second acquisition, in which historical data whose quantity is greater than a preset quantity threshold is acquired and recorded as fifth data.

[0124] S29 reduces the number and uses a random under-sampling algorithm to perform deletion processing on the fifth data until the number of the fifth data is no greater than a preset number threshold.

[0125] For example, an automobile manufacturer needs to conduct quality inspection on newly designed carbon fiber body parts to check whether there is carbon fiber breakage. The specific steps are as follows:

[0126] 1. Determine the area and the detection area:

[0127] Based on the design drawings and simulation analysis, engineers identified several key areas in the carbon fiber body parts where stress concentration and temperature changes are large as inspection areas.

[0128] Mark the inspection points on these areas and prepare for data collection.

[0129] 2. Data collection:

[0130] 2-1 First acquisition: Use ultrasonic scanning technology to scan the carbon fiber in the detection area to obtain the first data such as the microstructure, density distribution, and physical properties of the carbon fiber.

[0131] 2-2 Second collection: Install temperature sensors, humidity sensors, stress sensors and vibration sensors near the detection area to collect environmental data in real time, including temperature, humidity, stress and vibration data.

[0132] 2-3 The third collection: retrieve the usage data of the carbon fiber component in the past period from the database or historical records, including data in the manufacturing process, early test data, etc.

[0133] 2-4 Fourth collection: Obtain historical records of environmental data related to the detection area from the database, including historical data of temperature, humidity, stress and vibration.

[0134] 2-5 Quantity determination: Perform quantity statistics on all historical data to determine whether the quantity of each type of historical data is greater than a preset quantity threshold. In this embodiment, each type of historical data has at least 1,000 data points per month, that is, the preset quantity threshold is 1,000.

[0135] 2-6 First acquisition: It is found that the data of a certain environment in some months is less than 1,000 points.

[0136] 2-7 Increase quantity: Use a random oversampling algorithm to increase the data of these months until the preset quantity threshold is reached.

[0137] 2-8 Second acquisition and 2-9 Reduction of quantity: For months with excessive amount of historical data, a random under-sampling algorithm is used to delete some of the historical data.

[0138] 3. First modeling:

[0139] A detection model is established based on the machine learning algorithm, and the data collected in the S2 data acquisition (i.e., the first data, historical data of carbon fiber, environmental data, and historical data of environmental data) is used to train the detection model.

[0140] The detection model consists of a CNN sub-model, a BI-LSTM sub-model, and a decision tree sub-model.

[0141] The CNN sub-model is used to process the ultrasonic images of carbon fiber, extract features and preliminarily judge the status of carbon fiber.

[0142] The BI-LSTM sub-model is used to process time series environmental data and analyze the impact of environmental data on carbon fiber performance.

[0143] The decision tree sub-model makes a comprehensive judgment based on the output results of the CNN sub-model and the BI-LSTM sub-model, and outputs the final detection result.

[0144] 4. Intelligent detection:

[0145] 4-1 First detection: Input the first data into the CNN sub-model to obtain the first output result.

[0146] 4-2 Second detection: Input the real-time collected environmental data into the BI-LSTM sub-model to obtain the second output result.

[0147] 4-3 Third test: Take the first output result and the second output result as input, make a comprehensive judgment through the decision tree sub-model, and output the third output result.

[0148] 5. Result judgment:

[0149] It is determined whether the third output result meets expectations (such as a preset carbon fiber health standard).

[0150] If the result is as expected, the detection result of "no broken wire" is output and monitoring continues.

[0151] If the result does not meet expectations, the detection result of "broken wire exists" is output and an alarm is triggered, prompting further inspection or maintenance.

[0152] After determining the key detection area of ​​the carbon fiber composite material component, this embodiment collects multiple rounds of data and then processes the data to prepare a sufficient and balanced data set. Then, a detection model is built based on these data to achieve intelligent detection of carbon fiber components. During the detection process, new data is input into the model to obtain the detection results, and finally the results are judged in S5 result judgment. If potential risks are found, an alarm is triggered, otherwise monitoring continues.

[0153] Example 2: Figure 4 As shown, the difference between this embodiment and embodiment 1 is that after executing S2 data collection and before executing S3 first modeling, it also includes:

[0154] S6 data fusion, such as Figure 5 As shown, the S6 data fusion includes S61 network building, S62 second modeling, S63 determining initial values, S64 determining optimal values ​​and S65 updating.

[0155] S61 builds a network to construct a wireless sensor network (WSN). A wireless sensor network is composed of a large number of stationary or mobile sensor nodes in a self-organizing and multi-hop manner, with the purpose of collaboratively detecting, processing and transmitting monitoring information of perceived objects within the network coverage area. These sensor nodes usually include sensor modules, computing modules, storage modules, communication modules and power modules, etc., which work together to achieve comprehensive monitoring of environmental data. By building a wireless sensor network, real-time and accurate collection of environmental data in the target area can be achieved.

[0156] S62: second modeling, establishing a BP neural network model.

[0157] S63 determines the initial value, and determines the initial weight and initial threshold of the BP neural network model based on the sparrow search algorithm. The traditional BP neural network model is usually trained by randomly initializing the weight and threshold, but this method is prone to fall into the local optimal solution. The sparrow search algorithm is a swarm intelligence algorithm that simulates the foraging behavior of sparrows in nature to find the global optimal solution. By optimizing the initial weight and initial threshold of the BP neural network model using the sparrow search algorithm, the prediction accuracy and training efficiency of the BP neural network model can be significantly improved.

[0158] For example, suppose there is a three-layer BP neural network model, including an input layer, a hidden layer and an output layer. The input layer has 2 neurons, the hidden layer has 3 neurons (using the Sigmoid activation function), and the output layer has 1 neuron (using the linear activation function, suitable for regression problems). The process of determining the initial weight and initial threshold using the sparrow search algorithm is as follows:

[0159] (1) Problem definition: The initial weights and initial thresholds of the BP neural network are regarded as the solution space of the optimization problem.

[0160] (2) Initialize the sparrow population: randomly generate a certain number of sparrow individuals, each of which represents a set of initial weights and initial thresholds.

[0161] Set the algorithm parameters, such as the population size of 10, the number of iterations not exceeding 50, the search space range (for each weight and threshold, set its range between [-1, 1]), etc.

[0162] Randomly generate 10 groups of initial weights and initial thresholds, each group contains 1×2+2×3+3×1+1 = 12 elements, corresponding to the weights from the input layer to the hidden layer, the weights from the hidden layer to the output layer, the threshold of the hidden layer, and the threshold of the output layer.

[0163] (3) Evaluate fitness: For each sparrow individual, use its representative initial weight and initial threshold to build a BP neural network model.

[0164] For each set of initial weights and thresholds, use them to build a BP neural network model and evaluate its performance on a small validation set. Set a regression task on the validation set and use the mean square error (MSE) as the fitness function. The smaller the MSE, the higher the fitness.

[0165] The validation set contains m samples, each of which has an input and a corresponding true output The model for each input A prediction output will be given .

[0166] ;

[0167] The smaller the MSE value is, the smaller the difference between the predicted value of the BP neural network model and the actual value is, which means that the prediction performance of the BP neural network model is better.

[0168] (4) Update the position of the sparrow: According to the rules of the sparrow search algorithm, update the position of the individual sparrow (that is, update the initial weight and initial threshold).

[0169] The update process includes the update of joiners, the update of discoverers and the update of vigilants.

[0170] The update rule of the joiner is:

[0171] ;

[0172] in, represents the dth position of the ith joiner in the tth generation in the sparrow population; represents the dth position of the ith joiner in the t+1th generation in the sparrow population; R is a random number that follows a normal distribution; is the exponential operation; is the worst position of the i-th joiner in the t-th generation sparrow population; is the population size; represents the optimal position of the i-th entrant in the t-th generation sparrow population; is a random number uniformly distributed in the interval [0,1]; is the boundary size of the search space, which is used to keep the search within a reasonable range. The L is a 1×d matrix, and each element in the matrix is ​​1.

[0173] when hour, The value of will converge to 0, indicating that the joiner needs to fly to other places to find food; when hour, The value of will be evenly distributed to each dimension. The joiner update rule can be described as: randomly find a position near the current optimal position, so that the variance of each dimension from the optimal position is as small as possible, so that The value converges to the optimal position.

[0174] The update rule of the discoverer is:

[0175] ;

[0176] in, represents the position of the i-th discoverer in the t+1 generation (i.e., the updated initial weight and initial threshold); represents the position of the i-th discoverer in the t-th generation (i.e., the initial weight and initial threshold at the beginning); is a random number uniformly distributed in the interval [0,1]; is the step size factor, which is used to control the step size of the search. The value of follows a normal distribution with a mean of 0 and a variance of 1; Refers to the position of the sparrow individual with the highest fitness in the current population (i.e. the current optimal solution); is the boundary size of the search space, used to keep the search within a reasonable range.

[0177] The update rule of the vigilant is:

[0178] ;

[0179] in, represents the dth position of the ith sentinel in the t+1th generation in the sparrow population; represents the dth position of the ith sentinel in the tth generation in the sparrow population; represents the worst position of the ith sentinel in the tth generation sparrow population; represents the optimal position of the ith sentinel in the tth generation of sparrows; C is a random number in the range of [-1,1]; is the fitness value, H is the current optimal fitness value, and F is the current worst fitness value; It is a preset value, and in this step it is taken as 0.5.

[0180] It means that the vigilant is on the edge of the group. This suggests that sentinels located among sparrow populations are aware of the danger and need to get close to other sparrows to avoid being preyed upon.

[0181] (5) Iterative optimization: Repeat the fitness evaluation and position update steps until the maximum number of iterations is reached or other stopping conditions are met.

[0182] During the iteration process, the optimal combination of initial weights and initial thresholds is gradually approached.

[0183] (6) Output the optimal solution: After the iteration, the initial weight and initial threshold represented by the sparrow individual with the highest fitness are output as the final result.

[0184] S64 determines the optimal value, adjusts the weights and thresholds of the BP neural network model through the back propagation algorithm, and during the training process, calculates the prediction error and gradient value of the model through continuous iteration, and updates the weights and bias of the model according to the gradient value until the stopping condition is met (such as reaching the maximum number of iterations, the loss of the validation set no longer decreases, etc.), obtains the optimal weights and optimal thresholds of the BP neural network model, and transmits the optimal weights and optimal thresholds of the BP neural network model to the wireless sensor network, so as to realize the real-time prediction and fusion of environmental data.

[0185] S65 update, the wireless sensor network uses the optimal weights and optimal thresholds of the received BP neural network model to fuse the environmental data. Data fusion refers to the merging and processing of data from multiple sensor nodes according to certain rules to obtain more accurate and comprehensive environmental information. In a wireless sensor network, since the monitoring range and reliability of each sensor node are limited, it is necessary to fuse the data of multiple nodes to improve the accuracy and reliability of the data. The environmental data after data fusion processing will be used as new environmental data for subsequent monitoring and analysis.

[0186] S7 initial dimensionality reduction, such as Figure 6 As shown, the S7 preliminary dimensionality reduction includes S71 correlation analysis, S72 correlation judgment, S73 retention, S74 elimination and S75 iteration.

[0187] S71 correlation analysis: based on the single factor analysis algorithm, the correlation between the i-th environmental data and the second output result is calculated and recorded as the second data. The correlation calculation method includes: Pearson correlation coefficient, Spearman rank correlation coefficient or Kendall rank correlation coefficient, etc.

[0188] S72 correlation judgment, judging whether the second data is greater than a preset correlation threshold, if so, it is considered that the environmental data is highly correlated with the second output result, and S73 is executed to retain; if not, it is considered that the correlation is insufficient, and S74 is executed to eliminate.

[0189] S73 retains, retains the i-th environmental data.

[0190] S74 removes and deletes the i-th environmental data.

[0191] S75 is an iteration, where the i+1th environmental data is used as the new ith environmental data, and steps S71 to S74 are repeated until all environmental data are processed.

[0192] S8 deep dimensionality reduction, such as Figure 7 As shown, the S8 deep dimensionality reduction includes S81 integration, S82 distance calculation, S83 centering, S84 inner product matrix construction, S85 feature decomposition, S86 feature selection and S87 result acquisition.

[0193] In S81, all environmental data retained in the preliminary dimensionality reduction in S7 are integrated into a single environmental data set, and each environmental data is a sample in the environmental data set.

[0194] S82 calculates the distance, calculates the distance between all sample pairs, such as Euclidean distance, Manhattan distance or cosine similarity, and integrates all distances into a first distance matrix, where each element of the first distance matrix represents the distance between two samples.

[0195] S83 Centering: Centering the first distance matrix (subtracting the mean or median from each distance value in the first distance matrix to ensure that the data is distributed around a central point) to obtain a second distance matrix.

[0196] S84 constructs an inner product matrix and calculates the inner product matrix G based on the second distance matrix. The calculation model is as follows:

[0197] ;

[0198] ;

[0199] in, is the i-th vector in the second distance matrix and the jth vector The inner product of is a diagonal matrix constructed according to the second distance matrix; is a diagonal matrix The element at row i and column i; is a diagonal matrix The element at row j and column j; is the distance between the i-th sample and the j-th sample in the environment dataset.

[0200] S85 eigendecomposition, perform eigendecomposition on the inner product matrix to obtain eigenvalues ​​and eigenvectors.

[0201] S86 Feature selection, selecting the first n eigenvalues ​​and the eigenvectors corresponding to the eigenvalues.

[0202] S87 obtains the result, maps the sample with the corresponding feature vector, and obtains the reduced-dimensional environment data. The calculation model is as follows:

[0203] ;

[0204] Among them, X is the environmental data after dimension reduction; is an eigenvector matrix composed of n eigenvectors; is a diagonal matrix consisting of n eigenvalues.

[0205] S9 expansion: Generate adversarial networks to learn the distribution characteristics of the reduced-dimensional environmental data and generate new environmental data, and use the reduced-dimensional environmental data and the generated environmental data as new environmental data. These new environmental data should be similar to the original environmental data in statistical characteristics, but have more quantity.

[0206] This embodiment screens out environmental data that is highly correlated with specific detection results through correlation analysis and integrates them into an environmental data set. Then, the distances between all sample pairs in the environmental data set are calculated to construct a first distance matrix, and the matrix is ​​centered or adjusted to be suitable for subsequent analysis. Then, an inner product matrix is ​​constructed based on the adjusted first distance matrix, the most important eigenvectors are selected through eigendecomposition, and the original samples are mapped to these eigenvectors to obtain the reduced-dimensional environmental data, thereby reducing the dimension of the environmental data and improving the efficiency and effect of data analysis or model training. By introducing a generative adversarial network (GAN) to learn the distribution characteristics of the reduced-dimensional environmental data, new environmental data samples are successfully generated, thereby expanding the environmental data set and increasing the amount of environmental data.

[0207] Example 3: Figure 8 As shown, the difference between this embodiment and embodiment 2 is that after executing S2 data collection and before executing S3 first modeling, it also includes:

[0208] S10 data processing, including S101 preprocessing and S102 time alignment.

[0209] S101 preprocessing, for environmental data (such as temperature, humidity, stress, vibration, etc.), preprocessing includes removing outliers, smoothing, interpolating to fill missing data, etc. The preprocessed environmental data is recorded as third data.

[0210] S102 Time alignment: Since the environmental data come from different sensors or data sources, there may be differences in the timestamps between these environmental data. In order to accurately reflect the impact of environmental conditions at the same time on carbon fiber components during modeling, this step performs time alignment on all third data. The process of time alignment is to extract all third data at the same time (or the closest timestamp) and combine them into a new environmental data set. In this way, each time point corresponds to a complete set of environmental data, which is convenient for subsequent analysis and modeling.

[0211] S11 extracts the time interval, and obtains the time interval. The time interval refers to the time interval for extracting the third data, and this time interval reflects the frequency or rate of data collection.

[0212] S12 determines the time interval to determine whether the time interval is less than the preset time interval threshold. If so, it means that the frequency of data collection is high enough to meet the modeling requirements, and execute S3 for the first modeling; if not, it means that the frequency of data collection is not high enough, which may cause data loss or fail to accurately reflect the rapid changes in environmental conditions, and execute S13 for adjustment.

[0213] S13 adjusts to reduce the time interval for collecting environmental data. This embodiment reduces the time interval for collecting environmental data to improve the frequency and real-time performance of data collection. In this process, it is necessary to adjust the sampling rate of the sensor, optimize the data collection algorithm, or increase the number of data collection devices. By adjusting the data collection strategy, enough and accurate data points can be obtained during modeling, thereby improving the accuracy and reliability of the model.

[0214] After S2 data collection is completed, in order to ensure data quality and modeling accuracy, S10 data processing needs to be performed, followed by S11 extracting the time interval to evaluate the data collection frequency. S12 determines whether the time interval meets the preset time interval threshold. If not, S13 adjusts the data collection strategy to increase the collection frequency. This series of steps provides a high-quality data foundation for the first modeling in S3.

[0215] This embodiment performs S10 data processing (preprocessing to optimize data quality and time alignment to ensure data synchronization), then extracts the time interval and evaluates whether it meets the preset frequency threshold required for modeling. If not, the data collection strategy is adjusted to increase the frequency. This series of preprocessing and evaluation adjustment steps lays a solid data foundation for the subsequent S3 first modeling, ensuring the accuracy and reliability of the model.

[0216] Embodiment 4: This embodiment discloses a carbon fiber broken wire intelligent detection system, the detection system comprising:

[0217] The area determination module is used to determine the detection area.

[0218] The data acquisition module is communicatively connected with the area determination module and the intelligent detection module, and the data acquisition module includes a first acquisition unit and a second acquisition unit.

[0219] The first acquisition unit is used to extract data of carbon fiber in the detection area, which is recorded as first data. The first data includes physical properties of the carbon fiber such as fiber arrangement, density, strength, etc. The first data also includes: microstructure, density distribution, etc.

[0220] The second acquisition unit is used to acquire environmental data, and the environmental data includes: temperature data, humidity data, stress data and vibration data.

[0221] The first modeling module is communicatively connected with the intelligent detection module and is used to establish a detection model, wherein the detection model includes: a convolutional neural network sub-model, a BI-LSTM sub-model and a decision tree sub-model.

[0222] The intelligent detection module includes a first detection unit, a second detection unit and a third detection unit.

[0223] The first detection unit is used to input the first data into the convolutional neural network sub-model, perform feature extraction and preliminary detection through the CNN sub-model, and obtain a first output result.

[0224] The second detection unit is used to input the environmental data into the BI-LSTM sub-model, analyze the potential impact of the environment on the carbon fiber component, and obtain a second output result.

[0225] The third detection unit is used to input the first output result and the second output result into the decision tree sub-model, and output a third output result.

[0226] The result judgment module is used to judge whether the third output result meets expectations.

[0227] This system integrates multiple modules such as determining the detection area, data collection (including carbon fiber characteristics and environmental data), modeling (using CNN sub-model to process carbon fiber data, BI-LSTM sub-model to process environmental data, and making comprehensive decisions through decision tree sub-model), intelligent detection and result judgment, etc., to achieve all-round and intelligent detection of carbon fiber components, aiming to improve detection efficiency and accuracy and ensure product quality and safety.

[0228] The above are all preferred embodiments of the present application, and the protection scope of the present application is not limited thereto. Therefore, any equivalent changes made according to the structure, shape, and principle of the present application should be included in the protection scope of the present application.

Claims

1. A carbon fiber broken wire intelligent detection method, characterized in that: include: Determine the area: Determine the detection area; Data collection: including the first collection and the second collection; First acquisition: extracting the data of carbon fiber in the detection area, which is recorded as the first data; Second collection: collecting environmental data; First modeling: establishing a detection model, wherein the detection model includes: a convolutional neural network sub-model, a BI-LSTM sub-model and a decision tree sub-model; Intelligent detection: including the first detection, the second detection and the third detection; First detection: inputting the first data into the convolutional neural network sub-model to obtain a first output result; Second detection: input the environmental data into the BI-LSTM sub-model to obtain the second output result; Third test: input the first output result and the second output result into the decision tree sub-model, and output the third output result; Result judgment: judge whether the third output result meets expectations. If so, output a detection result indicating no broken wires; if not, output a detection result indicating broken wires.

2. The carbon fiber broken wire intelligent detection method according to claim 1, characterized in that: After executing the step of data collection and before executing the step of first modeling, the method further includes: Data fusion: including network construction, second modeling, initial value determination, optimal value determination and update; Network construction: build wireless sensor network; Second modeling: establish BP neural network model; Determine the initial value: Based on the sparrow search algorithm, determine the initial weight and initial threshold of the BP neural network model; Determine the optimal value: adjust the weights and thresholds of the BP neural network model through the back propagation algorithm, obtain the optimal weights and thresholds of the BP neural network model, and transmit the optimal weights and thresholds of the BP neural network model to the wireless sensor network; Update: The environmental data is fused using a wireless sensor network, and the fused environmental data is used as new environmental data.

3. The carbon fiber broken wire intelligent detection method according to claim 2 is characterized in that: After performing the step of data fusion and before performing the step of first modeling, the method further includes: Initial dimensionality reduction: including relevance analysis, relevance judgment, retention, elimination and iteration; Correlation analysis: Based on the single factor analysis algorithm, the correlation between the i-th environmental data and the second output result is calculated and recorded as the second data; Relevance judgment: judging whether the second data is greater than a preset relevance threshold, if so, executing the retaining step; if not, executing the eliminating step; Keep: keep the i-th environment data; Eliminate: delete the i-th environment data; Iteration: Take the i+1th environmental data as the new ith environmental data and perform the correlation analysis step.

4. The carbon fiber broken wire intelligent detection method according to claim 3 is characterized in that: After performing the preliminary dimensionality reduction step and before performing the first modeling step, the method further includes: Deep dimensionality reduction: including integration, distance calculation, centering, inner product matrix construction, eigendecomposition, feature selection and obtaining results; Integration: All environmental data in the retained steps are integrated into an environmental data set, and one environmental data is recorded as a sample in the environmental data set; Calculate distance: Calculate the distance between all sample pairs and integrate all distances into the first distance matrix; Centralization: Centralize the first distance matrix to obtain the second distance matrix; Construct the inner product matrix: Based on the second distance matrix, calculate the inner product matrix G. The calculation model is as follows: ； ； in, is the i-th vector in the second distance matrix and the jth vector The inner product of is a diagonal matrix constructed according to the second distance matrix; is a diagonal matrix The element at row i and column i; is a diagonal matrix The element at row j and column j; is the distance between the i-th sample and the j-th sample in the environmental data set; Eigendecomposition: Perform eigendecomposition on the inner product matrix to obtain eigenvalues ​​and eigenvectors; Feature selection: select the first n eigenvalues ​​and the eigenvectors corresponding to the eigenvalues; Obtain results: Map the samples to the corresponding feature vectors to obtain the reduced-dimensional environmental data. The calculation model is as follows: ； Among them, X is the environmental data after dimension reduction; is an eigenvector matrix composed of n eigenvectors; is a diagonal matrix consisting of n eigenvalues.

5. The carbon fiber broken wire intelligent detection method according to claim 4, characterized in that: After performing the step of deep dimensionality reduction and before performing the step of first modeling, the method further includes: Expansion: Generate adversarial networks to learn the distribution characteristics of the reduced-dimensional environmental data and generate new environmental data. The reduced-dimensional environmental data and the generated environmental data are used as new environmental data.

6. The carbon fiber broken wire intelligent detection method according to claim 1, characterized in that: After executing the step of data collection and before executing the step of first modeling, the method further includes: Data processing: including preprocessing and time alignment; Preprocessing: preprocessing the environmental data, and recording the preprocessed environmental data as the third data; Time alignment: extract the third data at the same time and use the extracted third data as new environment data.

7. The carbon fiber broken wire intelligent detection method according to claim 6, characterized in that: After performing the time alignment step and before performing the first modeling step, the method further includes: Extracting time interval: extracting the time interval of the third data in the step of obtaining time alignment; Determine the time interval: determine whether the time interval is less than a preset time interval threshold, if so, execute the first modeling step; if not, execute the adjustment step; Adjustment: Reduce the time interval for collecting environmental data.

8. The carbon fiber broken wire intelligent detection method according to claim 1, characterized in that: The data collection steps also include: The third collection: collect historical data of carbon fiber; Fourth collection: collect historical data of environmental data.

9. The carbon fiber broken wire intelligent detection method according to claim 8, characterized in that: After executing the step of data collection and before executing the step of first modeling, the method further includes: Quantity judgment: judging whether the quantity of various historical data is greater than a preset quantity threshold, if so, executing the first acquisition step; if not, executing the second acquisition step; First acquisition: acquiring historical data whose quantity is less than a preset quantity threshold, recorded as fourth data; Increasing the number: using a random oversampling algorithm to increase the fourth data until the number of the fourth data is not less than a preset number threshold; Second acquisition: acquiring historical data whose quantity is greater than a preset quantity threshold, and recording it as fifth data; Reduce the number: Use a random undersampling algorithm to delete the fifth data until the number of the fifth data is no more than a preset number threshold.

10. An intelligent carbon fiber broken wire detection system, the detection system is applicable to the detection method according to any one of claims 1 to 9, characterized in that: include: A region determination module is used to determine a detection region; A data acquisition module, which is in communication with the area determination module and the intelligent detection module, and the data acquisition module includes a first acquisition unit and a second acquisition unit; A first acquisition unit is used to extract data of carbon fibers in a detection area, which is recorded as first data; A second collection unit, used for collecting environmental data; A first modeling module is communicatively connected with the intelligent detection module and is used to establish a detection model, wherein the detection model includes: a convolutional neural network sub-model, a BI-LSTM sub-model and a decision tree sub-model; An intelligent detection module, comprising a first detection unit, a second detection unit and a third detection unit; A first detection unit, used to input first data into a convolutional neural network sub-model to obtain a first output result; A second detection unit is used to input the environmental data into the BI-LSTM sub-model to obtain a second output result; A third detection unit, used to input the first output result and the second output result into the decision tree sub-model, and output a third output result; The result judgment module is used to judge whether the third output result meets expectations.

Images