A high-strength screw strength prediction analysis method based on deep learning

Abstract

The application discloses a high-strength screw strength prediction analysis method based on deep learning, and particularly relates to the field of deep learning, and comprises the following steps: S1, multi-dimensional data acquisition; S2, coupling feature matrix construction; S3, multi-output deep learning model training; S4, strength prediction; S5, comprehensive strength scoring; S6, strength grade division; and S7, decision output. According to the application, multi-dimensional data acquisition, coupling feature matrix construction and output model prediction are carried out to identify strength abnormalities, and when there is a strength deviation, targeted grading determination is carried out based on a strength threshold formed by a static threshold and a dynamic baseline, so that the quality continuity of the construction screw is ensured to a certain extent, the connection failure risk caused by the substandard strength is reduced, the misjudgment waste caused by traditional fragmented data acquisition is avoided, the equipment wear caused by frequent manual rechecking is reduced, and the quality control cost is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning, more particularly, the present application relates to a high-strength screw strength prediction analysis method based on deep learning. BACKGROUND

[0002] High-strength screws are widely used in aerospace, automobiles, bridges, wind power and other fields, and their failure may cause major safety accidents. As the core components of mechanical connection and steel structure node fastening, the strength performance directly determines the equipment operation safety and service life.

[0003] The traditional high-strength screw strength analysis method generally uses a multi-physical field coupling model to predict the life of high-strength screws under extreme conditions, but it still has some shortcomings in actual use. First, the existing method predicts the strength of the screw with single data collection dimension and isolated analysis method. In this case, it is difficult to accurately reflect the true comprehensive performance of the screw. In order to ensure the reliability of the prediction results, the traditional method often relies on independent testing and acceptance of single strength indicators. This decentralized testing mode leads to long testing period and high cost. More importantly, the internal coupling relationship between the strength indicators is ignored during the testing process, which cannot provide global guidance for the optimization of the production process of the screw and may cause potential quality and safety hazards.

[0004] Second, after completing the strength prediction, it is difficult to conduct scientific and quantitative comprehensive evaluation. The traditional empirical formula method and pass / fail judgment method are strongly influenced by subjectivity and experience, and the evaluation results do not match the actual service performance of the screw, which may lead to two extreme cases: one is over-conservative, which judges the screw that can meet the specific scene use as unqualified, causing material waste and increasing production cost; the other is insufficient evaluation, which allows the screw with hidden dangers to flow into the market, increasing the risk of equipment operation and the cost of use cycle. SUMMARY

[0005] Therefore, the embodiments of the present application provide a high-strength screw strength prediction analysis method based on deep learning, which accurately predicts the tensile strength, yield strength, fatigue strength and pre-tightening force retention rate of high-strength screws and comprehensively scores the strength, effectively solving the problems raised in the background art.

[0006] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0007] S1: Multi-dimensional data collection: collect the material, process, structure and environment original data of the high-strength screw according to the predefined data specification, and construct an original data matrix therefrom;

[0008] S2: Coupling feature matrix construction: preprocessing the original data matrix, sequentially coupling feature extraction, and thus constructing a coupling feature matrix;

[0009] S3: Multi-output deep learning model training: constructing a feature fusion model, taking the coupling feature matrix as input, simultaneously outputting the predicted values of the tensile strength index, yield strength index, fatigue strength index and pre-tightening force retention index, and then obtaining the loss function for feature fusion model optimization;

[0010] S4: Strength prediction: processing the multi-dimensional data corresponding to the high-strength screw to be predicted to obtain an input feature vector, and sequentially outputting the predicted values of the four strength indexes according to the feature fusion model;

[0011] S5: Comprehensive strength score: based on the strength prediction result, calculating the strength score of each index, using dynamic parameter adjustment to weight the strength score corresponding to each index, and then outputting the comprehensive strength index;

[0012] S6: Strength grade division: the strength threshold is composed of the static threshold established according to historical data and the dynamic baseline generated by the event sequence prediction, and then the screw strength is divided into three grades based on the comprehensive strength index, including excellent, qualified and to be optimized;

[0013] S7: Decision output: outputting the corresponding decision support information according to the comprehensive strength index and the strength grade.

[0014] Technical effects and advantages of the present application:

[0015] 1. The present application identifies strength anomalies through multi-dimensional data acquisition, coupling feature matrix construction and output model prediction, and makes targeted grading determination based on the strength threshold composed of static threshold and dynamic baseline when identifying strength deviation, which is not limited to traditional single-index detection or qualitative judgment, and can realize real-time strength prediction and grading in the screw production process in some cases, thereby ensuring the quality continuity of construction screws to a certain extent, better controlling the matching degree of screw strength and application scenarios, reducing the risk of connection failure caused by insufficient strength, and avoiding the waste caused by misjudgment of traditional fragmented data acquisition and the equipment wear caused by frequent manual reinspection, thereby reducing the quality control cost;

[0016] 2. The present application deeply relates the comprehensive strength index and the strength grade in the whole life cycle of screw production, thereby obtaining an adaptive scheme by using quantitative decision suggestions to accurately match the production process and selection scene in the whole life cycle management of screw, realizing high-precision grasping of the optimization direction of each link of screw, avoiding unnecessary waste of raw materials and reducing the equipment failure handling cost caused by strength failure in the later period, achieving the goals of cost reduction and efficiency improvement, ensuring equipment safety and improving engineering reliability. Attached Figure Description

[0017] Fig. 1 This is a schematic diagram of the overall structure of the present invention.

[0018] Fig. 2 This is a flowchart illustrating the S5 implementation of the present invention. Detailed Implementation

[0019] 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.

[0020] like Figs. 1-2 The invention illustrates a high-strength screw strength prediction and analysis method based on deep learning. The specific implementation of this invention includes the following steps:

[0021] S1: Multi-dimensional data acquisition: Collect raw data on the material, process, structure and environment of high-strength screws according to predefined data specifications, thereby constructing a raw data matrix.

[0022] In this embodiment, it should be specifically noted that the predefined data specification is a detailed and clear set of data collection standards, the specific contents of which include:

[0023] The raw data of high-strength screw materials are obtained by deploying professional acquisition equipment in the testing process, including a spectrometer, metallographic microscope and hardness tester. The chemical composition data of the material is collected through elemental spectral analysis protocol. The chemical composition includes the content of carbon C, manganese Mn, silicon Si, chromium Cr, molybdenum Mo, phosphorus P and sulfur S. Metallographic imaging technology is used to analyze the microstructure to obtain information on grain size and martensite content. Brinell hardness is used to test the surface hardness value of the material.

[0024] The raw data for high-strength screw manufacturing processes are obtained by deploying dedicated monitoring equipment at key nodes in the production process, including furnace temperature trackers, pressure sensors, and eddy current thickness gauges. These key nodes include heat treatment, cold working, and surface treatment. During heat treatment, the furnace temperature tracker collects data on quenching temperature, holding time, cooling rate, tempering temperature, and tempering time. During cold treatment, the pressure sensor collects data on thread rolling reduction, thread rolling speed, and upsetting pressure. During surface treatment, the eddy current thickness gauge collects data on the thickness of the galvanized layer and the hardness of the nitrided layer.

[0025] The high-strength screw structure original data is collected by deploying high-precision measuring equipment in the size detection link, specifically including a three-coordinate measuring instrument and a thread micrometer, to collect thread parameters and head parameters, wherein the thread parameters include nominal diameter, pitch, thread stress cross-sectional area, and root fillet radius, and the head parameters include head thickness, hexagon side distance, and transition fillet;

[0026] The high-strength screw environment original data is collected by deploying an environmental sensing device in the service link, specifically a strain gauge sensor, to obtain a strain gauge signal, extract micro-strain therefrom, convert the micro-strain into surface stress according to Hooke's law, and then multiply the surface stress by the effective cross-sectional area of the screw to calculate the pre-tightening force and obtain the number of working cycles.

[0027] Further, the material original data is the root of screw strength, the chemical composition determines the hardenability and matrix strength of the material, for example, the C content directly affects the hardness of the martensite after quenching, thereby improving the tensile and yield strength; Cr and Mo elements can enhance the tempering stability of the material and reduce the strength decay at high temperatures; P and S are harmful elements, and excessive amounts will cause grain boundary embrittlement, reducing fatigue strength and impact resistance; grain size and martensite content are micro parameters of the screw, the smaller the grain size, the more the grain boundaries, and the higher the martensite content, the higher the tensile strength of the material; the surface hardness can timely find problems such as uneven composition and abnormal structure of the raw material, and is helpful for screening qualified raw materials, providing a high-quality matrix guarantee for subsequent process execution, and avoiding strength failure due to raw material problems;

[0028] The process original data is a key basis for converting the potential of the material into the actual strength of the screw, and data anomalies usually mean that the screw strength performance cannot meet the standards, for example, a quenching temperature that is too low will result in insufficient martensite content and low tensile / yield strength; by collecting heat treatment, cold working, and surface treatment processes, pre-tightening force anomalies in process execution can be found in time, providing data support for the consistency of strength in stable batch production and reducing the risk of strength failure due to process deviations;

[0029] The structure original data can reflect whether the local strength of the high-strength screw fails, and data anomalies usually mean that the screw has stress concentration and assembly failure risks; by collecting thread parameters and head parameter data, size forming defects can be found in time, and it is helpful to verify the rationality of the structure design, providing a guarantee for the reliability of screw assembly and preventing assembly breakage and insufficient pre-tightening force caused by structural problems;

[0030] The environment original data reflects the use of the screw, and by collecting pre-tightening force and working cycle number, strength decay problems in service can be found in time, improving the accuracy of strength prediction, providing a basis for developing operation and maintenance schemes, and avoiding equipment safety accidents caused by service strength decay.

[0031] It should be explained that the original data on the materials, processes, structures, and environment of high-strength screws are all obtained from historical data stored in each production cycle. Historical data is used because it can accurately reflect the production and testing process of high-strength screws within the production cycle. Therefore, the feature matrix obtained from historical data can ensure that the predicted indicators can reflect the condition of the batch of high-strength screws during model training, thereby improving the service life of the screws.

[0032] It should be further explained that the process of constructing the original data matrix is ​​as follows:

[0033] A1: Collect N high-strength screw samples. For each screw sample, according to the predefined data specifications, aggregate and collect all M original data items in the dimensions of material, process, structure and environment through each batch production cycle. Each sample corresponds to the data of each batch production cycle.

[0034] It should be added that the high-strength screw samples should cover different production batches, process parameters, and suppliers to ensure the representativeness of the data.

[0035] A2: Create an N-row, M-column database with high-strength screw samples as rows and feature data as columns. This database is the original data matrix X, where each row is a row vector consisting of all the original data of a sample.

[0036] S2: Construction of Coupled Feature Matrix: The original data matrix is ​​preprocessed, and coupled features are extracted sequentially to construct the coupled feature matrix.

[0037] In this embodiment, it is necessary to specifically explain the preprocessing of the original data matrix, which includes:

[0038] Outlier removal: For the original data x n，m The mean Ax and standard deviation Sx of each sample are obtained using the formulas for calculating the mean and standard deviation, respectively. If |x| = ... n，m -Ax|>G 0.95，N ×Sx, where x n，m G represents the original data x in the nth row and mth column. 0.95，N The value representing the Grubbs threshold indicates that the sample should be removed.

[0039] Normalization: For the original data x n，m Normalization is performed to obtain x′ n，mSpecifically, normalization can be achieved by subtracting the original data from the mean and then comparing it to the standard deviation. The purpose of normalization is to eliminate feature differences between different units and ensure that each feature is compared on the same scale. After processing, the mean of each original column is 0 and the standard deviation is 1, resulting in a normalized data matrix X′∈R. N′×M .

[0040] It should be further explained that the construction of the coupling feature matrix specifically includes:

[0041] B1: Based on domain knowledge, feature design traverses all parameter pairs that may interact, generating a large number of candidate coupled feature forms.

[0042] In an example of the above scheme, the domain knowledge obtained is that the hardenability of steel depends on the carbon content and the cooling rate. The carbon content determines the theoretically achievable maximum hardness, while the quenching temperature is the process condition for achieving this theory. Therefore, their product (quenching temperature × carbon content) can be used as a coupling feature to quantify the effective quenching energy.

[0043] Another example of the above scheme is to obtain domain knowledge that the stress concentration factor Kt is closely related to the geometry, and fatigue failure often originates from stress concentration points. For the same fillet radius, the degree of danger is different under different load stress levels. By using the ratio (thread root fillet radius / maximum working stress) as a coupling feature, the competitive relationship between geometric stress release effect and external load stress can be simulated.

[0044] B2: Calculate the correlation coefficient between each candidate coupling feature and the four target strength indicators through correlation analysis, and thereby select the K coupling features with the most significant correlation coefficients. For example, K=12.

[0045] It should be added that the four target strength indicators, including the target tensile strength, target yield strength, target fatigue strength, and target preload retention, were set from the beginning. The actual values ​​of the four strength indicators were obtained through physical experiments. For example, the tensile testing machine measured the target tensile strength and target yield strength, the fatigue testing machine measured the target fatigue strength, and the long-term durability test measured the target preload retention.

[0046] It should be explained that the correlation coefficient can specifically be the Pearson correlation coefficient. When calculating the correlation coefficient between each candidate coupling feature and the four target strength indicators, each candidate coupling feature is precisely matched with the four target strength indicators. This results in each candidate coupling feature and the target matrix being divided into two arrays. The mean and standard deviation of each array are then calculated separately. Finally, the mean and standard deviation are used to calculate the Pearson correlation coefficient. The calculated correlation coefficient is denoted as r. If r > 0, it indicates a positive correlation between the two; if r < 0, it indicates a negative correlation between the two; the closer |r| is to 1, the higher the correlation.

[0047] B3: The selected K coupled features are taken as new columns and horizontally concatenated to the right side of the preprocessed original feature matrix X′. Principal component analysis is then used to reduce its dimensionality, thereby constructing the final coupled feature matrix F∈R. N′×D .

[0048] It should be added that principal component analysis retains principal components with a cumulative variance contribution rate ≥ 95% of the matrix. The variance contribution rate is calculated by transforming the original matrix into a covariance matrix, extracting its eigenvalues ​​and arranging them in descending order. The variance contribution rate of a single principal component is equal to the proportion of its corresponding eigenvalue to the sum of all eigenvalues. The cumulative variance contribution rate is then calculated by summing the variance contribution rates of individual principal components. When the cumulative variance contribution rate first reaches 95%, the principal component is retained, and the number of principal components is counted as the new column dimension of the final coupled feature matrix.

[0049] S3: Multi-output deep learning model training: Construct a feature fusion model, taking the coupled feature matrix as input, and simultaneously outputting the predicted values ​​of tensile strength index, yield strength index, fatigue strength index and preload retention index, and then obtain the loss function to optimize the feature fusion model.

[0050] In this embodiment, it should be specifically explained that the feature fusion model includes:

[0051] CNN Feature Extraction: Consists of 2 convolutional layers and 1 max pooling layer, used to extract the spatial correlation of features and output a sequence containing feature maps;

[0052] LSTM temporal modeling: It consists of two LSTM layers, which are used to capture the temporal dependencies of features and output an output sequence containing the hidden states of all time steps. The output vector of each time step contains the information context of all previous features.

[0053] Attention: Used to assign weights to the output features of LSTM, highlighting key features;

[0054] Fully connected output layer: Composed of two fully connected layers, which map the weighted high-level features extracted by the Attention mechanism to the predicted values ​​of the final tensile strength index, yield strength index, fatigue strength index and preload retention index.

[0055] The loss function is implemented using weighted mean squared error, specifically expressed as:

[0056] ,

[0057] Where L represents the loss function. This represents the true value of the k-th strength index for the n-th sample, where k=1 corresponds to the tensile strength index, k=2 to the yield strength index, k=3 to the fatigue strength index, and k=4 to the preload retention index. Let N′ represent the predicted value of the k-th intensity index for the n-th sample, and let N′ be the number of rows in the final coupled feature matrix. k The weighting coefficients for the four strength indicators can be adjusted according to the application scenario. For example, if the screw type is a steel structure node screw, then w1, w2, w3 and w4 are 0.2, 0.2, 0.3 and 0.3 respectively. If the screw type is a static fastening screw, then w1, w2, w3 and w4 are 0.3, 0.3, 0.2 and 0.2 respectively.

[0058] It should be added that model optimization is performed after the loss function is calculated. Specifically, the model is optimized based on the weight coefficients in the loss function to minimize the loss function of the predicted values.

[0059] It should be explained that the tensile strength index represents the maximum stress that a material can withstand under tensile load until it breaks; the yield strength index represents the stress at which a material begins to undergo significant plastic deformation; the fatigue strength index represents the maximum stress that a material can withstand an infinite number of cycles under alternating load (load with periodically changing magnitude and direction) without breaking; and the preload retention index represents the percentage of the initial preload that a screw can retain during long-term service after being tightened.

[0060] S4: Strength Prediction: The multi-dimensional data corresponding to the high-strength screw to be predicted is processed to obtain the input feature vector, and the predicted values ​​of the four major strength indicators are output in sequence according to the feature fusion model.

[0061] In this embodiment, the specific process of S4 needs to be explained in detail as follows:

[0062] The multi-dimensional data of the high-strength screw to be predicted is processed according to S1 and S2 to obtain the input feature vector f. test The data is then input into the feature fusion model, which outputs predicted values ​​for the four intensity indicators, specifically as follows: , These are the predicted values ​​for tensile strength, yield strength, fatigue strength, and preload retention, respectively.

[0063] S5: Comprehensive Intensity Score: Based on the intensity prediction results, the intensity score of each indicator is calculated, and the corresponding intensity scores of each indicator are weighted by dynamic parameter adjustment, thereby outputting the comprehensive intensity index.

[0064] In this embodiment, the specific calculation process for the intensity score needs to be explained in detail as follows:

[0065] For each intensity index, a score is awarded based on the deviation rate between the predicted value and the standard value, specifically expressed as follows:

[0066] ,

[0067] Where Sk represents the intensity score corresponding to the k-th intensity index, δ k This represents the deviation rate between the predicted value and the standard value corresponding to the k-th intensity index, and Y st，k This represents the standard value of each strength index, which can be specifically set through industry standards, δ k，max This indicates the maximum allowable deviation rate. For example, the maximum allowable deviation rates for tensile strength, yield strength, fatigue strength, and preload retention are 5%, 5%, 10%, and 20%, respectively.

[0068] Furthermore, the comprehensive intensity index is calculated based on each intensity score and weighting coefficient, and is specifically expressed as follows:

[0069] ,

[0070] Where HS represents the comprehensive strength index, and S k w represents the k-th intensity rating. k This represents the weight coefficient corresponding to the k-th intensity index, which is consistent with the weight coefficient in the loss function.

[0071] S6: Strength level classification: The strength threshold is composed of a static threshold established based on historical data and a dynamic baseline generated by event sequence prediction. Then, based on the comprehensive strength index, the screw strength is classified into three levels, including excellent, qualified and needing optimization.

[0072] In this embodiment, the specific implementation process of S6 is as follows:

[0073] C1: By collecting and storing at least 30 days of historical data on high-strength screws, including the predicted values ​​of their final calculated tensile strength, yield strength, fatigue strength, and preload retention rate;

[0074] The comprehensive strength index was calculated and statistically analyzed based on 30 days of historical data. The 95th percentile was taken as the theoretical performance extreme value of the comprehensive strength index and used as the static threshold. This baseline represents the threshold of the excellent level that can be achieved under normal working conditions.

[0075] The Holt-Winters triple exponential smoothing model is used to learn the comprehensive intensity index series corresponding to 30-day historical data, predict the expected value of the comprehensive intensity index in the next period, and use it as a dynamic baseline. This baseline reflects the normal fluctuation trend and level of the intensity index in the near future.

[0076] C2: Based on the comprehensive intensity index of S6, calculate its relative percentage deviation from the dynamic baseline. Specifically, calculate the absolute difference between the current comprehensive intensity index and the dynamic baseline and compare it with the dynamic baseline.

[0077] If the current comprehensive strength index is greater than or equal to the static threshold, the excellent level is triggered, indicating that the screw strength has exceeded the excellent range.

[0078] C3: When the relative deviation percentage Δ of the comprehensive strength index is greater than 30%, the qualified level is triggered;

[0079] When the relative deviation percentage Δ of the comprehensive strength index is greater than 50%, optimization is triggered.

[0080] S7: Decision Output: Based on the comprehensive intensity index and intensity level, output corresponding decision support information.

[0081] In this embodiment, it should be specifically explained that the decision support information includes providing production process adjustment suggestions for screws of the desired optimization grade and recommending applicable screw grades based on application scenarios. The production process adjustment suggestions include, but are not limited to, key influencing factors identified by the system through feature fusion models that lead to a low overall strength index, especially causing a low score for a particular individual strength item. The applicable screw grade recommendation matches the screw strength grade with specific application scenarios based on the pre-set performance requirements of different application scenarios in the knowledge base. For example, in the steel structure nodes of an offshore platform, the overall strength index of this batch of screws is 84.91, which is a good grade, but its preload retention rate score does not meet the requirements of the marine environment. Therefore, it is not recommended for use on offshore platforms, but is recommended for use in general industrial plant steel structures.

[0082] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other.

[0083] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting and analyzing the strength of high-strength screws based on deep learning, characterized in that, include: S1: Multi-dimensional data acquisition: Collect raw data on the material, process, structure and environment of high-strength screws according to predefined data specifications, thereby constructing a raw data matrix; S2: Construction of Coupled Feature Matrix: The original data matrix is ​​preprocessed, and coupled features are extracted sequentially to construct the coupled feature matrix. S3: Multi-output deep learning model training: Construct a feature fusion model, taking the coupled feature matrix as input, and simultaneously outputting the predicted values ​​of tensile strength index, yield strength index, fatigue strength index and preload retention index, and then obtain the loss function to optimize the feature fusion model; S4: Strength Prediction: The multi-dimensional data corresponding to the high-strength screw to be predicted is processed to obtain the input feature vector, and the predicted values ​​of the four major strength indicators are output in sequence according to the feature fusion model. S5: Comprehensive Intensity Score: Based on the intensity prediction results, the intensity score of each indicator is calculated, and the corresponding intensity scores of each indicator are weighted by dynamic parameter adjustment, and then the comprehensive intensity index is output. S6: Strength level classification: The strength threshold is composed of a static threshold established based on historical data and a dynamic baseline generated by event sequence prediction. The static threshold is determined based on the 95th percentile of historical data, and the dynamic baseline is predicted based on the Holt-Winters triple exponential smoothing model. Then, based on the comprehensive strength index, the screw strength is classified into three levels, including excellent, qualified and needing optimization. S7: Decision Output: Based on the comprehensive intensity index and intensity level, output corresponding decision support information.

2. The high-strength screw strength prediction and analysis method based on deep learning according to claim 1, characterized in that: The specific process for constructing the original data matrix is ​​as follows: A1: Collect N high-strength screw samples. For each screw sample, according to the predefined data specifications, aggregate and collect all M original data in the dimensions of material, process, structure and environment through each batch production cycle. Each sample corresponds to the data of each batch production cycle. A2: Create an N-row, M-column database with high-strength screw samples as rows and feature data as columns. This database is the original data matrix X, where each row is a row vector consisting of all the original data of a sample.

3. The high-strength screw strength prediction and analysis method based on deep learning according to claim 1, characterized in that: The construction of the coupling feature matrix specifically includes: B1: Based on domain knowledge, feature design traverses all parameter pairs that may interact, and designs a large number of candidate coupled feature forms. B2: Calculate the correlation coefficient between each candidate coupling feature and the four target strength indicators through correlation analysis, and select the K coupling features with the most significant correlation coefficients. B3: The selected K coupled features are taken as new columns and horizontally concatenated to the right side of the preprocessed original feature matrix X′. Principal component analysis is then used to reduce its dimensionality, thereby constructing the final coupled feature matrix F∈R. N′×D .

4. The high-strength screw strength prediction and analysis method based on deep learning according to claim 1, characterized in that: The loss function is implemented using weighted mean square error, specifically expressed as follows: ， Where L represents the loss function. This represents the true value of the k-th strength index for the n-th sample, where k=1 corresponds to the tensile strength index, k=2 to the yield strength index, k=3 to the fatigue strength index, and k=4 to the preload retention index. Let N′ represent the predicted value of the k-th intensity index for the n-th sample, and let N′ be the number of rows in the final coupled feature matrix. k This represents the weighting coefficients corresponding to the four strength indicators.

5. The high-strength screw strength prediction and analysis method based on deep learning according to claim 1, characterized in that: The specific process of S4 is as follows: The multi-dimensional data of the high-strength screw to be predicted is processed according to S1 and S2 to obtain the input feature vector f. test The data is then input into the feature fusion model, which outputs predicted values ​​for the four intensity indicators, specifically as follows: , These are the predicted values ​​for tensile strength, yield strength, fatigue strength, and preload retention, respectively.

6. The high-strength screw strength prediction and analysis method based on deep learning according to claim 4, characterized in that: The specific calculation process for the strength score corresponding to the strength index is as follows: For each intensity index, a score is awarded based on the deviation rate between the predicted value and the standard value, specifically expressed as follows: ， Where S k δ represents the intensity score corresponding to the k-th intensity index. k This represents the deviation rate between the predicted value and the standard value corresponding to the k-th intensity index, and Y st，k This represents the standard value of each strength index, which can be specifically set through industry standards, δ k，max This indicates the maximum allowable deviation rate, and the subscript n represents the number of the screw sample to be predicted, and its predicted value. The calculation method is consistent with claim 4 and the aforementioned model training process.

7. The high-strength screw strength prediction and analysis method based on deep learning according to claim 1, characterized in that: The comprehensive strength index is calculated based on each strength score and weighting coefficient, and is specifically expressed as follows: ， Where HS represents the comprehensive strength index, and S k w represents the k-th intensity rating. k This represents the weight coefficient corresponding to the k-th intensity index, which is consistent with the weight coefficient in the loss function.

8. The high-strength screw strength prediction and analysis method based on deep learning according to claim 1, characterized in that: The specific implementation process of S6 is as follows: C1: By collecting and storing at least 30 days of historical data on high-strength screws, including the predicted values ​​of their final calculated tensile strength, yield strength, fatigue strength, and preload retention rate; The comprehensive strength index was calculated and statistically analyzed based on 30 days of historical data. The 95th percentile was taken as the theoretical performance extreme value of the comprehensive strength index and used as the static threshold. This baseline represents the threshold of the excellent level that can be achieved under normal working conditions. The Holt-Winters triple exponential smoothing model is used to learn the comprehensive intensity index series corresponding to 30-day historical data, predict the expected value of the comprehensive intensity index in the next period, and use it as a dynamic baseline. This baseline reflects the normal fluctuation trend and level of the intensity index in the near future. C2: Based on the comprehensive strength index of S6, calculate its relative percentage deviation from the dynamic baseline. Specifically, calculate the absolute difference between the current comprehensive strength index and the dynamic baseline and compare it with the dynamic baseline. If the current comprehensive strength index is greater than or equal to the static threshold, the excellent level is triggered, indicating that the screw strength has exceeded the excellent range. C3: When the relative deviation percentage Δ of the comprehensive strength index is greater than 30%, the qualified level is triggered; When the relative deviation percentage Δ of the comprehensive strength index is greater than 50%, optimization is triggered.

Images