Cross shaft universal joint fracture early warning method and system based on fatigue detection

Abstract

The invention relates to the technical field of universal joint fatigue detection, in particular to a cross shaft universal joint fracture early warning method based on fatigue detection, which comprises the following steps of: constructing an original data group, preprocessing detection data in the original data group to obtain a preprocessing array, and mapping an original fatigue damage value of the original data group to the preprocessing array; defining a training sample, dividing a training set and a test set, training the fatigue damage analysis model through the training set, and evaluating and updating the fatigue damage analysis model through the test set; when the fatigue damage analysis model is trained, feature analysis logic of hierarchical association is constructed, and the layering rule of the child nodes is analyzed; the current data set is preprocessed and input into a fatigue damage analysis model, and the fatigue damage analysis model outputs a current fatigue damage value; setting different fatigue intervals, comparing the different fatigue intervals with current fatigue damage values corresponding to the current data set, and outputting corresponding signals matched with the fatigue intervals.

Description

Technical Field

[0001] This invention relates to the field of universal joint fatigue detection technology, and more specifically, to a method and system for early warning of cross shaft universal joint fracture based on fatigue detection. Background Technology

[0002] As a core component of transmission systems in automobiles, construction machinery, and other industries, the fracture of the universal joint can directly lead to transmission failure, resulting in serious consequences such as equipment downtime and safety accidents. Fatigue damage is the main cause of universal joint failure (accounting for over 70%). Therefore, building a fracture early warning system based on fatigue detection is of great significance for ensuring equipment operation safety and reducing maintenance costs. Specifically, fatigue testing of universal joints includes two core methods: static testing and dynamic testing. While dynamic testing can capture fatigue damage trends under load in real time, the alternating torque, impact load, and vibration transmitted can generate interference signals. These signals can overlap with damage characteristics (such as microcracks and abnormal vibration), making it difficult to distinguish between fluctuations in operating conditions and actual damage. It may also mask early weak points such as microcracks in the journal transition fillet and internal hidden damage. If such damage is not identified in time, it can easily propagate rapidly and lead to fracture failure. At the same time, fatigue damage of universal joints is progressive and insidious. Damage such as uneven hardness of the quenched layer leading to a decrease in fatigue resistance and micro-scratches on the surface becoming crack initiation sources can only be clearly presented in a static environment without load or vibration. To avoid inaccurate or incomplete fatigue testing of the universal joint due to the superposition of dynamic detection interference and the omission of early hidden damage, which could even lead to safety hazards, offline testing methods are typically used when the universal joint is not in operation (equipment is stopped and there is no load transmission). These methods include magnetic particle testing to check for microcracks on the journal transition fillet surface, 2-5MHz ultrasonic testing to check for internal hidden damage, a roughness tester with a 0.8mm sampling length to measure the journal surface finish, and a portable hardness tester to check whether the hardness uniformity of the quenched layer meets the HRC58-62 standard. This eliminates dynamic vibration and load interference, allowing for accurate and comprehensive fatigue testing of the universal joint and providing a basis for subsequent evaluation and maintenance.

[0003] When performing fatigue testing on a cross shaft universal joint using offline testing methods, the core objective is to obtain a series of raw test data that reflects the condition of the cross shaft universal joint. Then, these data must be compared one by one with preset standard data or normal condition data. However, in scenarios with a large workload of testing (such as automobile manufacturers mass-producing universal joints, requiring testing of each batch of products) and rich data dimensions (simultaneously testing multiple indicators such as microcracks on the journal transition fillet surface, internal hidden damage, journal surface finish, and uniformity of hardened layer hardness, with each indicator having multiple data points), relying solely on manual comparison and analysis of the raw test data would significantly increase the workload. Specifically, manual work requires checking, calculating differences, and judging whether a large number of test data items meet the standards one by one, which not only consumes a lot of time and manpower but is also prone to errors in data comparison due to human negligence. This can affect the accuracy of subsequent fatigue damage assessment of the universal joint, potentially leading to the failure to detect potential fatigue damage in a timely manner and posing a risk to equipment operation safety. Therefore, we propose a fatigue detection-based method and system for early warning of universal joint fracture. Summary of the Invention

[0004] The purpose of this invention is to solve the problems in fatigue testing of universal joints, such as the susceptibility of dynamic testing to interference leading to missed early damage detection, and the large workload and insufficient accuracy of manual analysis of offline testing data, which in turn affect the accuracy of fatigue damage assessment and the safety of equipment operation.

[0005] To achieve the above objectives, this invention provides a method for early warning of fracture of a universal joint based on fatigue detection, comprising the following steps: S1. Construct the original data set, preprocess the detection data within the original data set to obtain the preprocessed array, and map the original fatigue damage values ​​of the original data set to the preprocessed array, which is the training sample. S2. Define training samples and divide them into training set and test set. Train the fatigue damage analysis model using the training set and evaluate and update the fatigue damage analysis model using the test set. The fatigue damage analysis model includes child nodes at different levels and a leaf node. Each level includes two child nodes, left and right. Each child node is assigned a corresponding hierarchical rule. The leaf node includes multiple training samples. The predicted value of the leaf node is the average of the original fatigue damage values ​​of all training samples in that node. During training, the detection data of each detection feature is selected as the partition value, and then the left and right child nodes are determined by different partition values. An initial proportional coefficient is set to extract the common loss, and then the hierarchical rules of the child nodes are analyzed again by the training samples after extracting the common loss. When dividing child nodes at the next level, the common loss from the previous level is introduced for the division, and the initial scaling factor is dynamically adjusted to select the common loss until the leaf node is reached; S3. Preprocess the current data set input to the fatigue damage analysis model, and the fatigue damage analysis model outputs the current fatigue damage value; then set different fatigue intervals, compare the current fatigue damage value corresponding to the current data set with the different fatigue intervals, and output the signal corresponding to the matching fatigue interval.

[0006] As a further improvement to this technical solution, in step S2, a preprocessing array is defined, and the corresponding original fatigue damage values ​​are used as training samples: the detection data corresponding to different detection features under the static detection state are perceived, and the detection data are sorted according to a fixed order to construct an original data group. The detection data in the original data group is preprocessed and the output is a preprocessing array. The original fatigue damage values ​​are mapped to the preprocessing array to form training samples. Multiple training samples are divided into a training set and a test set, and the number of training samples corresponding to the training set is greater than the number of training samples corresponding to the test set. The sum of the number of samples corresponding to the training set and the test set is the total number of training samples.

[0007] As a further improvement to this technical solution, each sub-node in the fatigue damage analysis model is assigned a corresponding hierarchical rule, which includes detection features and partitioning values. When training the fatigue damage analysis model, the original fatigue damage values ​​of all training samples are first added together, and then divided by the number of training samples to obtain the mean of the original fatigue damage values. For each training sample in the training set, the squared difference between the original fatigue damage value and the mean is calculated. Then, all squared differences are summed and divided by the number of training samples to obtain the basic variance. As a further improvement to this technical solution, the detection features of each preprocessed array in the corresponding training samples of the training set are traversed sequentially, and the detection data of each with the same detection feature are selected as the splitting value in turn. The training samples are then recursively divided into left child nodes and right child nodes based on the splitting value. The left child node satisfies the condition that the detection data is less than or equal to the partition value; the right child node satisfies the condition that the detection data is greater than or equal to the partition value; the variances of the left and right child nodes are calculated again using the same method as the calculation of the basic variance, to obtain the variances of the left and right child nodes; The left child node weight is obtained by dividing the number of training samples in the left child node by the total number of training samples. The first variance is obtained by multiplying the left child node weight by the variance of the left child node. The right child node weight is obtained by dividing the number of training samples in the right child node by the total number of training samples. The second variance is obtained by multiplying the right child node weight by the variance of the right child node. The sum of the first and second variances is the total variance after dividing all training samples. The variance reduction is obtained by subtracting the total variance from the base variance. By using the test data as partition values ​​in sequence, multiple variance reductions are obtained. The test data that results in the largest variance reduction is selected as the stratification rule for the child nodes. The above test data selection and node stratification operations are repeated to obtain a fatigue damage analysis model that includes child nodes at different levels. As a further improvement to this technical solution, in the process of constructing the fatigue damage analysis model, a reduction threshold is set. If the maximum variance reduction of the left child node and the right child node is less than or equal to the reduction threshold, the preprocessing arrays corresponding to the left child node and the right child node are merged. The merged left child node and the right child node are defined as leaf nodes. The leaf nodes contain multiple training samples with similar preprocessing arrays. The predicted value of the leaf node is the average value of the original fatigue damage values ​​of all training samples in the node, thereby completing the training of the fatigue damage analysis model.

[0008] As a further improvement to this technical solution, in step S2, the variance of the left child node and the variance of the right child node corresponding to the maximum variance reduction are received, an initial scaling factor is set, and the left child node variance is multiplied by the initial scaling factor to obtain the left reduced variance, and the right child node variance is multiplied by the initial scaling factor to obtain the right reduced variance. After dividing the left child node, different numbers of training samples are retrieved from the left child node in turn, and the variance of the left child node corresponding to the left child node is calculated again. The left child node is selected to retrieve training samples with the largest number of training samples retrieved and the variance of the left child node being closest to the variance of the left child node. The same method is used to retrieve some training samples from the right child node. The common loss is defined by the partial training samples retrieved from the left and right child nodes respectively, and then the hierarchical rules of the left and right child nodes are set by the detection results that have not yet been retrieved. Before setting the hierarchical rules for the next level, add the common loss of the previous level, and then select the corresponding common loss until the division reaches the leaf node.

[0009] As a further improvement to this technical solution, in step S2, the initial proportional coefficient is dynamically adjusted by subtracting the reduction threshold from the maximum variance reduction to obtain the variance difference, an amplification coefficient is set, and the amplification coefficient is multiplied by the variance difference and added to the initial proportional coefficient to obtain the increased proportional coefficient.

[0010] As a further improvement to this technical solution, the fatigue damage analysis model is evaluated and updated: Retrieve all preprocessed arrays in the test set, along with their corresponding original fatigue damage values; sequentially input the preprocessed arrays into the fatigue damage analysis model trained on the training set, retrieve the hierarchical rules of the child nodes, including detection features and corresponding partition values, retrieve the detection data in the preprocessed array that have the same detection features as the child nodes, compare the detection data with the partition values ​​of the corresponding child nodes, recursively partition the preprocessed array until it is partitioned to the leaf nodes, and output the predicted value corresponding to the leaf nodes, which is the current fatigue damage value corresponding to the preprocessed array; Retrieve the original fatigue damage values ​​from the preprocessing array and calculate the difference between the current fatigue damage value and the original fatigue damage value. The gradient descent strategy is used to iteratively optimize the reduction threshold, initial proportional coefficient, and amplification coefficient until the difference in damage value no longer increases.

[0011] As a further improvement to this technical solution, in step S3, a trained fatigue damage analysis model is used to perform static detection of the universal joint. The current data group corresponding to the current offline detection method is received, and the current data group is preprocessed to obtain a preprocessed array, which is then input into the fatigue damage analysis model. Each detection feature in the preprocessed array is compared with its corresponding partition value, and the preprocessed array is recursively partitioned until it is partitioned to a leaf node. The predicted value corresponding to the leaf node is the current fatigue damage value corresponding to the current data group.

[0012] A fatigue-detection-based early warning system for fracture of a universal joint includes: The data preprocessing module constructs the original data set, preprocesses the detection data within the original data set to obtain the preprocessed array, and maps the original fatigue damage values ​​of the original data set to the preprocessed array. The model training and optimization module defines training samples and divides them into training and test sets. The fatigue damage analysis model is trained using the training set and evaluated and updated using the test set. When training the fatigue damage analysis model, a hierarchical feature analysis logic is constructed to analyze the hierarchical rules of the child nodes. The fatigue detection and early warning module preprocesses the current data set input to the fatigue damage analysis model, which then outputs the current fatigue damage value and the corresponding signal.

[0013] In addition to the objectives, features, and advantages described above, the present invention has other objectives, features, and advantages. The invention will now be described in further detail with reference to the figures. Attached Figure Description

[0014] Figure 1 This is a schematic diagram illustrating the origin of the training samples for this invention; Figure 2 This is a schematic diagram illustrating the division of the training and test sets in this invention; Figure 3 This is a schematic diagram of the fatigue damage analysis model of the present invention; Figure 4 This is a schematic diagram of the hierarchical rules of each sub-node in the fatigue damage analysis model of this invention; Figure 5 This is a schematic diagram illustrating the origin of the common loss introduced when training the fatigue damage analysis model of this invention. Figure 6 This is a flowchart illustrating the working principle of the fatigue damage analysis model of this invention.

[0015] The meanings of the labels in the diagram are as follows: 100. Data preprocessing module; 200. Model training and optimization module; 300. Fatigue detection and early warning module. Detailed Implementation

[0016] The technical solutions in 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.

[0017] refer to Figures 1-6 As shown, a method for early warning of fracture of a universal joint based on fatigue detection includes the following steps: S1. Construct the original data set, preprocess the detection data within the original data set to obtain a preprocessed array, and map the original fatigue damage values ​​of the original data set to the preprocessed array. Specifically: In the static inspection of the cross shaft universal joint, each time an offline inspection is carried out, the same offline inspection method (such as magnetic particle inspection, ultrasonic inspection, surface roughness tester inspection, portable hardness tester inspection, etc.) is used to collect inspection data with different inspection features (each different offline inspection method is a detection feature) (such as the frequency domain characteristics of vibration signal, the amplitude and frequency of torque fluctuation, journal surface roughness, hardness value, etc.). S1.1, Perceive the detection data corresponding to different detection features under static detection conditions, and construct the original data group by sorting the detection data in a fixed order: ; in To measure the amount of data detected, For the first Detection data corresponding to each detection feature; S1.2. Preprocess each detection data within each raw data group sequentially to obtain all raw data groups. Corresponding preprocessing array , of which preprocessed arrays , For preprocessed arrays The Middle Detection data after feature preprocessing for class detection. This is a preprocessing function; Preprocessing functions Used for data cleaning, data normalization, feature extraction, data transformation, and other processing of detection data: Data cleaning includes outlier handling and missing value handling: Outlier handling: A statistical method is used to calculate the mean and standard deviation of the test data. Test data that deviate from the mean by more than a certain multiple (such as 3 times) of the standard deviation are identified as outliers, and then filled using the mean, median, etc. Missing value handling: For continuous detection data, if the missing proportion is small, the mean can be used for imputation; for categorical data, the mode can be used for imputation. Data normalization: Mapping detection data to... The interval, the formula is ,in and The first The minimum value of the maximum value of the detection data corresponding to the class detection feature; Feature extraction: For time-series data such as vibration signals, statistical features such as mean, variance, peak value, and root mean square can be extracted to reflect the overall characteristics of the detected data in the time domain; the time-domain signal is converted to the frequency domain by Fourier transform to extract features such as frequency components, spectral peak value, and bandwidth for analyzing the frequency characteristics of the signal; wavelet transform is used to perform multi-scale analysis of the signal to extract features such as modulus maxima and energy of wavelet coefficients, which is suitable for analyzing non-stationary signals.

[0018] Data transformation: Continuous data can be transformed into discrete categorical data. This can be achieved through equal-width binning (dividing the data range into equally wide intervals), equal-frequency binning (ensuring that each interval contains the same number of samples), or cluster-based binning methods. For categorical data, one-hot encoding can be used to convert categorical variables into binary vectors, or label encoding can be used to map categories to integer labels.

[0019] S1.3, Establish the original data group With preprocessed array The mapping relationship between them maps the original data groups. Original fatigue damage value To preprocessed array (Original data set) and preprocessed array Original fatigue damage value (Same), define the preprocessing array and corresponding original fatigue damage value These are training samples.

[0020] S2. Define training samples and divide them into training and test sets. Train the fatigue damage analysis model using the training set. The fatigue damage analysis model includes child nodes at different levels and a leaf node. Each level includes two child nodes, left and right. Each child node is assigned a corresponding hierarchical rule. The leaf node includes multiple training samples. The predicted value of the leaf node is the average of the original fatigue damage values ​​of all training samples in that node. During training, the detection data of each detection feature is selected as the partition value, and then the left and right child nodes are determined by different partition values. An initial proportional coefficient is set to extract the common loss, and then the hierarchical rules of the child nodes are analyzed again by the training samples after extracting the common loss. When dividing child nodes at the next level, the common loss from the previous level is introduced for the division, and the initial scaling factor is dynamically adjusted to select the common loss until the leaf node is reached; S2.1, Define the preprocessing array and the corresponding original fatigue damage value The training samples specifically include: training samples; partitioning Each training sample consists of a training set (used to train the fatigue damage analysis model) and a test set (used to evaluate the fatigue damage analysis model) (e.g., a 7:3 training / test set). The number of training samples corresponds to the number of training sets. > Number of training samples corresponding to the test set And the number of training samples and number of training samples The sum is the total number of training samples. .

[0021] S2.2 Training the fatigue damage analysis model using the training set: The fatigue damage analysis model includes child nodes at different levels and a leaf node. Each level includes two child nodes, left and right. Each child node is assigned a corresponding hierarchical rule (detection features and partitioning values). The leaf node contains multiple training samples, and the predicted value of the leaf node is the average of the original fatigue damage values ​​of all training samples in that node. During the training of the fatigue damage analysis model, S2.3, the detection data for each detection feature is defined as the partition value. Then, through different division values Determine the left and right child nodes and set the initial scaling factor. Extract the common loss, and then analyze the hierarchical rules of the sub-nodes again using the training samples after extracting the common loss; When partitioning child nodes at the next level, the common loss from the previous level is introduced for partitioning, and the initial scaling factor is dynamically adjusted. Select common loss values ​​until leaf nodes are reached; S2.4. Use the test set to evaluate the fatigue damage analysis model, and update the fatigue damage analysis model based on the evaluation results.

[0022] The steps for training the fatigue damage analysis model using the training set in S2.2 are as follows: Calculate the original fatigue damage values ​​corresponding to all training samples in the training set. basic variance ,in The mean of the original fatigue damage values ​​in the training samples of the training set; Iterate through each preprocessed array in the corresponding training sample of the training set. The detection features are selected, and the detection data for each identical detection feature are chosen. In turn, they are used as the dividing values. By dividing values The training samples are recursively divided into left and right child nodes: The left child node satisfies the detection data. ≤ partition value including Each training sample corresponds to an original fatigue damage value. The right child node satisfies the detection data. > Division value including Each training sample corresponds to an original fatigue damage value. Number of training samples for the left child node and the number of training samples for the right child node The sum is the number of training samples in the training set. Calculate the variances of the left and right child nodes separately: The original fatigue damage value corresponding to the left child node is The variance of the left child node is ,in This represents the average value of the original fatigue damage values ​​of the left child node; The original fatigue damage value corresponding to the right child node is The variance of the right node is ,in This represents the average value of the original fatigue damage values ​​of the right child node. Number of samples trained using the left child node Divide by the total number of training samples Obtain the weight of the left child node. Using the weight of the left child node Multiply by the variance of the left child node , as the first variance; the number of samples trained through the right child node. Divide by the total number of training samples Obtain the weight of the right child node. Using the weight of the right child node Multiply by the variance of the right node , as the second variance; summing the first and second variances yields the total variance after dividing all training samples. ; To quantify the improvement effect of feature segmentation on the concentration of original fatigue damage values, and to screen out the segmentation rule that maximizes the concentration of damage values ​​within child nodes, the basic variance was used. Subtract total variance The variance reduction was obtained. ; If the variance decreases The larger the value, the higher the original fatigue damage value in the training samples within the divided child nodes. The more concentrated the data, the better the segmentation effect; therefore, different detection features correspond to different detection data. As the dividing value At that time, the detection data of all detection features in the training samples As the dividing value After performing the above calculations, several variance reductions were obtained. Screening reduces variance The largest detection feature and the corresponding detection data , will detect data The hierarchical rules are based on the detection features as child nodes, and then the above detection data is repeated. By selecting operations that layer nodes, a fatigue damage analysis model including sub-nodes at different levels is obtained, so that the similarity of training samples within each sub-node gradually increases with recursive partitioning. In the process of constructing the fatigue damage analysis model, a reduction threshold is set. If the maximum variance of the left child node decreases by... ≤Reduction threshold If the left and right child nodes are in the same preprocessed array, then the preprocessed arrays corresponding to the left and right child nodes are merged. The merged left and right child nodes are defined as leaf nodes, and each leaf node contains multiple preprocessed arrays. For similar training samples, the predicted value of a leaf node is the average of the original fatigue damage values ​​of all training samples within that node. ,in This represents the number of training samples within a leaf node. The first leaf node The original fatigue damage values ​​of each training sample; Specifically: Taking the surface roughness, hardened layer hardness, and internal hidden damage depth detection features (each corresponding to a unique detection data) in the static inspection of a cross-shaped universal joint as an example: when recursively dividing the sub-nodes into different levels, the training samples cover various cases of surface roughness, hardened layer hardness, and internal hidden damage depth. At this time, the detection data of the above three detection features are selected sequentially. Calculate the variance reduction when using them as the basis for partitioning. From this, we can select the amount that reduces variance. The largest detection feature and its corresponding detection data Based on this detection feature and detection data The process involves partitioning the nodes, and the resulting child nodes are designated as the first-level child nodes. Then, for each newly generated first-level child node, the process of selecting detection features and calculating the variance reduction is repeated. The steps of selecting the optimal partitioning criteria and partitioning are repeated until the variance of the left and right child nodes decreases. ≤Reduction threshold This completes the training of the fatigue damage analysis model.

[0023] In the training process of traditional fatigue damage analysis models, relying solely on the reduction of variance through maximum variance is insufficient. Screening detection features and division values This method can easily lead to confusion in the detection features. For example, when performing fatigue damage analysis on a universal joint, multiple detection data are obtained, such as journal surface roughness, hardness of the quenched layer, and depth of internal hidden damage. Some detection features with low correlation to fatigue damage, such as the temperature and humidity of the detection environment (in static testing, normal fluctuations in temperature and humidity have no direct impact on fatigue damage), may be affected by the randomness of the detection data distribution, resulting in a reduction in the variance calculation. It exhibits a significant "advantage" and is mistakenly selected as a detection feature, thereby interfering with the identification of truly critical fatigue damage features (such as fatigue damage caused by stress concentration due to uneven hardness of the quenched layer). S2.3, Received maximum variance reduction The corresponding left child node variance right child node variance The initial proportional coefficient can be set based on historical data and industry experience from offline testing of the cross-axis universal joint. This is used to extract a portion of the training samples, specifically: Using initial scaling factors respectively Reduce the variance of the left child node To reduce variance on the left Reduce the variance of the right child node To reduce variance on the right ; After dividing the left child nodes, a portion of the training samples within the left child nodes is retrieved, and then the variance of the corresponding left child nodes after retrieval is calculated. The one that retrieves the most training samples and has the highest variance in the left child node. The closest to left-lowering variance As a constraint, select and retrieve training samples from the left child node; and use the same method to retrieve some training samples from the right child node. The training samples retrieved from the left and right child nodes are defined as the common loss, and then the hierarchical rules for the left and right child nodes are set using the detection results that have not yet been retrieved. Before setting the hierarchical rules for the next level, add the common loss of the previous level, and then select the corresponding common loss until the division reaches the leaf node; This allows for the precise selection and elimination of non-critical detection results that have minimal impact on the original fatigue damage value during the left and right child node division process, enabling the fatigue damage analysis model to focus on key detection results. Simultaneously, it establishes the next level of stratification rules based on the exclusion of non-critical detection results from the upper level, forming a hierarchical analysis path from coarse to fine screening. This avoids logical breaks in detection results between levels, making the fatigue damage analysis model's division of training samples more closely aligned with the actual evolution of fatigue damage. Furthermore, it constructs a hierarchical feature analysis logic, improving the accuracy of the fatigue damage analysis model in fatigue damage assessment and providing more reliable support for fatigue damage value prediction in static detection.

[0024] As the recursive division of left and right child nodes gradually reduces the number of corresponding training samples, thus completing the subdivision, although an initial scaling factor is used... This can increase the number of training samples during the recursive partitioning of child nodes at each level, thus optimizing the training process. However, due to the initial scaling factor... It is a fixed value and cannot adapt to the differences in feature distribution of training samples in different levels of child nodes—for example, upper-level child node samples contain a large amount of non-critical feature data, requiring a lower scaling factor. To filter out more common losses and eliminate interference, while deep sub-node samples have already focused on core damage features, if the same scaling factor is still used... Over-screening can lead to the loss of key features, resulting in a mismatch between the selected common loss characteristics and the actual damage feature requirements. This can either result in insufficient screening, retaining too much irrelevant data that interferes with the model, or excessive screening, losing key features and reducing evaluation accuracy. Therefore: S2.3, Introducing the initial scaling factor Dynamic adjustment mechanism, dynamically adjusting the initial proportional coefficient. Specifically, based on the reduction in maximum variance With reduction threshold Variance differences between Gradually increase the initial proportional coefficient The proportional coefficient after gradual increase Reduce the number of common loss values ​​selected, among which To the variance difference The relevant amplification coefficient is used to dynamically adjust the selection scale of common loss, thereby avoiding excessive interference of irrelevant common loss on the training of fatigue damage analysis model.

[0025] S2.4. Use a test set to evaluate the fatigue damage analysis model: Retrieve all preprocessed arrays in the test set and the corresponding original fatigue damage value ; The preprocessed array is sequentially input into the fatigue damage analysis model trained on the training set to evaluate the accuracy of the fatigue damage analysis model, and to test the preprocessed array within the test set. For example: retrieve the hierarchical rules of the child nodes, which include the detection features and their corresponding partition values. Retrieve the preprocessing array Detection data with the same detection features as child nodes , will detect data The partition value of the corresponding child node In comparison, recursively partitioning the preprocessing array until the preprocessing array is divided. Reach the leaf node and output the predicted value corresponding to the leaf node. That is, the preprocessing array Corresponding current fatigue damage value ; Retrieve the preprocessing array Original fatigue damage value Calculate the current fatigue damage value Compared with the original fatigue damage value Difference in damage values ; Iterative optimization of the reduction threshold using a gradient descent strategy Initial scaling factor Growth rate coefficient Until the difference in damage values No longer increase, reduce the threshold amount Initial scaling factor Growth rate coefficient Capable of adapting to the training requirements of fatigue damage analysis models; in order to reduce the threshold. For example, its update formula is: ,in For learning rate, For loss difference For the reduction threshold The partial derivatives of .

[0026] S3. Perform static testing of the universal joint using the trained fatigue damage analysis model. After the fatigue damage analysis model completes the testing, output the current fatigue damage value. The specific working principle is as follows: Receive the current data group corresponding to the current offline detection method. And through preprocessing functions Preprocess the current data set , obtain the preprocessed array And input into the fatigue damage analysis model, preprocessing array Each detection feature Each with the corresponding division value In comparison, recursively partitioning the preprocessing array until the preprocessing array is divided. Reach the leaf node and output the predicted value corresponding to the leaf node. This is the current data group. Corresponding current fatigue damage value Thus, the fatigue damage analysis model can accurately and efficiently detect the current fatigue damage value of the universal joint. It can detect potential fatigue damage in advance, making it easier to take timely maintenance or replacement measures, effectively avoiding transmission failure caused by fatigue fracture of the universal joint, which could lead to serious consequences such as equipment downtime and safety accidents, ensuring equipment operation safety and reducing operation and maintenance costs.

[0027] Furthermore, S3 sets multiple different fatigue ranges, each representing a normal and safe range. Mild fatigue range Severe fatigue range ; Furthermore, each fatigue range corresponds to a set of early warning signals. Specifically: normal safe range There is a corresponding normal signal. Mild fatigue range There is a corresponding mild signal Severe fatigue range There is an emergency signal. ; Receive the current data set output by the fatigue damage analysis model. Corresponding current fatigue damage value Compare the current fatigue damage values ​​sequentially. Matching corresponding signal outputs to multiple different fatigue ranges: If the lower limit of the normal safety range ≤Current fatigue damage value ≤ Upper limit of normal safety range Then the corresponding normal signal will be output. Normal signal This is used to confirm that the cross-shaft universal joint currently has no fatigue risk and is in a safe operating condition; If the lower limit of the mild fatigue range <Current fatigue damage value ≤ Upper limit of mild fatigue range Then the corresponding mild signal will be output. The indicator suggests early fatigue damage in the universal joint (such as minor surface scratches and hardness slightly below standard). While this does not affect short-term operation, it warrants monitoring. If the lower limit of the severe fatigue range ≤Current fatigue damage value ≤ Upper limit of severe fatigue range Then the corresponding emergency signal will be output. This is used to warn that the universal joint is nearing the critical fracture state (such as excessive internal hidden damage depth or cracks extending to critical stress areas), and continued use will directly lead to fracture.

[0028] A fatigue-detection-based early warning system for fracture of a universal joint includes: The data preprocessing module 100 constructs the original data set, preprocesses the detection data within the original data set to obtain a preprocessed array, and maps the original fatigue damage values ​​of the original data set to the preprocessed array. The model training and optimization module 200 defines training samples and divides them into training and test sets. The fatigue damage analysis model is trained using the training set and evaluated and updated using the test set. When training the fatigue damage analysis model, a hierarchical feature analysis logic is constructed to analyze the hierarchical rules of the child nodes. The fatigue detection and early warning module 300 preprocesses the current data set input to the fatigue damage analysis model, and the fatigue damage analysis model outputs the current fatigue damage value and the corresponding signal.

[0029] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for early warning of fracture of a universal joint based on fatigue detection, characterized in that, Includes the following steps: S1. Construct the original data set, preprocess the detection data within the original data set, output the preprocessed array, map the original fatigue damage values ​​to the preprocessed array, and form training samples. S2. Divide the training samples into a training set and a test set to train the fatigue damage analysis model; The fatigue damage analysis model includes child nodes at different levels and a leaf node. Each child node is assigned a corresponding hierarchical rule. The leaf node includes multiple training samples. The average value of the original fatigue damage value corresponding to the training samples is the output of the leaf node. During training, the detection data of each detection feature is selected as the partition value. Then, the child nodes are divided by different partition values, and the common loss is selected by setting an initial scaling factor. When dividing the child nodes in the next level, the common loss of the previous level is introduced for the division, and the initial scaling factor is dynamically adjusted to select the common loss until the leaf node. S3. Receive the current data group, preprocess the current data group and input it into the fatigue damage analysis model; the fatigue damage analysis model divides the current data group into leaf nodes and outputs the current fatigue damage value corresponding to the current data group; Different fatigue ranges are set, and the different fatigue ranges are compared with the current fatigue damage value. The fatigue range corresponding to the current fatigue damage value is matched, and the warning signal corresponding to the matched fatigue range is output.

2. The method for early warning of cross-shaft universal joint fracture based on fatigue detection according to claim 1, characterized in that: S2 defines a preprocessing array and corresponding original fatigue damage values ​​as training samples: sensing the detection data corresponding to different detection features under static detection state, and constructing an original data group by sorting the detection data according to a fixed order, preprocessing the detection data in the original data group to output the preprocessing array, mapping the original fatigue damage values ​​to the preprocessing array, and forming training samples. The training samples are divided into a training set and a test set, and the number of training samples in the training set is greater than the number of training samples in the test set. The sum of the number of training samples in the training set and the test set is the total number of training samples.

3. The method for early warning of cross-shaft universal joint fracture based on fatigue detection according to claim 2, characterized in that: Each sub-node in the fatigue damage analysis model is assigned a corresponding hierarchical rule, which includes detection features and division values. When training the fatigue damage analysis model, the original fatigue damage values ​​of all training samples are first summed, and then divided by the number of training samples to obtain the mean of the original fatigue damage values. For each training sample in the training set, the squared difference between the original fatigue damage value and the mean is calculated. Then, all squared differences are summed and divided by the number of training samples to obtain the basic variance.

4. The method for early warning of cross-shaft universal joint fracture based on fatigue detection according to claim 3, characterized in that: The detection features of each preprocessed array in the training samples are traversed sequentially. Detection data with the same detection feature are selected as the splitting value, and the training samples are recursively divided into left and right child nodes based on the splitting value. The left child node satisfies the condition that the detection data is less than or equal to the partition value; the right child node satisfies the condition that the detection data is greater than or equal to the partition value; the variances of the left and right child nodes are calculated again using the same method as the calculation of the basic variance, to obtain the variances of the left and right child nodes; The weight of the left child node is obtained by dividing the number of training samples in the left child node by the total number of training samples. The first variance is obtained by multiplying the left child node weight by the variance of the left child node. The weight of the right child node is obtained by dividing the number of training samples in the right child node by the total number of training samples. The second variance is obtained by multiplying the right child node weight by the variance of the right child node. The sum of the first and second variances is the total variance after dividing all training samples. The reduction in variance is obtained by subtracting the total variance from the basic variance. By using the test data as partition values ​​in sequence, multiple variance reductions are obtained. The test data that results in the largest variance reduction is selected as the stratification rule for the child nodes. The above test data selection and node stratification operations are repeated to obtain a fatigue damage analysis model that includes child nodes at different levels.

5. The method for early warning of cross-shaft universal joint fracture based on fatigue detection according to claim 4, characterized in that: In the process of constructing the fatigue damage analysis model, a reduction threshold is set. If the maximum variance reduction of the left child node and the right child node is less than or equal to the reduction threshold, the preprocessing arrays corresponding to the left child node and the right child node are merged. The merged left child node and the right child node are defined as leaf nodes. Each leaf node contains multiple training samples with similar preprocessing arrays. The predicted value of the leaf node is the average value of the original fatigue damage values ​​of all training samples in that node, thereby completing the training of the fatigue damage analysis model.

6. The method for early warning of cross-shaft universal joint fracture based on fatigue detection according to claim 4, characterized in that: S2 receives the left child node variance and right child node variance corresponding to the maximum variance reduction, sets an initial scaling factor, and sequentially multiplies the left child node variance by the initial scaling factor to obtain the left reduced variance and the right child node variance by the initial scaling factor to obtain the right reduced variance. After dividing the left child node, different numbers of training samples are retrieved from the left child node in turn, and the variance of the left child node corresponding to the left child node is calculated again. The left child node is selected to retrieve training samples with the largest number of training samples retrieved and the variance of the left child node being closest to the variance of the left child node. The same method is used to retrieve some training samples from the right child node. The common loss is defined by the partial training samples retrieved from the left and right child nodes respectively, and then the hierarchical rules of the left and right child nodes are set by the detection results that have not yet been retrieved. Before setting the hierarchical rules for the next level, add the common loss of the previous level, and then select the corresponding common loss until the division reaches the leaf node.

7. The method for early warning of cross-shaft universal joint fracture based on fatigue detection according to claim 6, characterized in that: In step S2, the initial proportional coefficient is dynamically adjusted by subtracting the reduction threshold from the maximum variance reduction to obtain the variance difference. An amplification coefficient is set, and the amplification coefficient is multiplied by the variance difference and added to the initial proportional coefficient to obtain the increased proportional coefficient.

8. The method for early warning of cross-shaft universal joint fracture based on fatigue detection according to claim 1, characterized in that: Evaluate and update the fatigue damage analysis model: Retrieve all preprocessed arrays in the test set, along with their corresponding original fatigue damage values; sequentially input the preprocessed arrays into the fatigue damage analysis model trained on the training set, retrieve the hierarchical rules of the child nodes, including detection features and corresponding partition values, retrieve the detection data in the preprocessed array that have the same detection features as the child nodes, compare the detection data with the partition values ​​of the corresponding child nodes, recursively partition the preprocessed array until it is partitioned to the leaf nodes, and output the predicted value corresponding to the leaf nodes, which is the current fatigue damage value corresponding to the preprocessed array; Retrieve the original fatigue damage values ​​from the preprocessing array and calculate the difference between the current fatigue damage value and the original fatigue damage value. The gradient descent strategy is used to iteratively optimize the reduction threshold, initial proportional coefficient, and amplification coefficient until the difference in damage value no longer increases.

9. The method for early warning of cross-shaft universal joint fracture based on fatigue detection according to claim 1, characterized in that: In step S3, a trained fatigue damage analysis model is used to perform static detection of the universal joint. The current data group corresponding to the current offline detection method is received and preprocessed to obtain a preprocessed array, which is then input into the fatigue damage analysis model. Each detection feature in the preprocessed array is compared with its corresponding partition value. The preprocessed array is recursively partitioned until it is partitioned to a leaf node. The predicted value corresponding to the leaf node is the current fatigue damage value corresponding to the current data group.

10. A fatigue-detection-based early warning system for universal joint fracture, applied to the fatigue-detection-based early warning method for universal joint fracture according to any one of claims 1-9, characterized in that, include: The data preprocessing module (100) constructs the original data set, preprocesses the detection data within the original data set, obtains the preprocessed array, and maps the original fatigue damage values ​​of the original data set to the preprocessed array. The model training and optimization module (200) defines training samples and divides them into training and test sets. The fatigue damage analysis model is trained using the training set and evaluated and updated using the test set. When training the fatigue damage analysis model, the module constructs hierarchical feature analysis logic and analyzes the hierarchical rules of the child nodes. The fatigue detection and early warning module (300) preprocesses the current data set input to the fatigue damage analysis model, and the fatigue damage analysis model outputs the current fatigue damage value and the corresponding signal.

Images