Compressor diaphragm fatigue life evaluation method based on multi-source data

Abstract

The present application belongs to the technical field of evaluation, and particularly relates to a compressor diaphragm fatigue life evaluation method based on multi-source data. The method comprises the following steps: obtaining diaphragm parameters and historical life data sets; calculating a feature difference vector p norm reciprocal as a similarity weight; using KD tree to query the average Euclidean distance reciprocal of the near neighbor data as a local density index; adjusting a basic attenuation factor to obtain an information attenuation factor by using the local density index; multiplying the life variance reciprocal and the information attenuation factor to obtain a confidence coefficient; fusing the similarity weight and the confidence coefficient to obtain a weight; and constructing a weighted least squares support vector machine prediction model according to the weight and outputting a result. That is, the scheme of the present application can filter the interference of poor data, focus the model on high-value samples, and realize accurate evaluation of fatigue life.

Description

Technical Field

[0001] This invention belongs to the field of evaluation technology, and in particular relates to a method for evaluating the fatigue life of compressor diaphragms based on multi-source data. Background Technology

[0002] Compressors, as core equipment in many industrial sectors, play an irreplaceable role in industries such as petrochemicals, refrigeration, and aerospace. The diaphragm is a key moving component in the compressor, changing the cylinder volume through high-frequency reciprocating deformation to achieve gas compression and discharge. In actual operation, compressor diaphragms operate in extremely harsh environments, enduring not only alternating high-pressure gas fluctuations but also the continuous effects of varying operating temperatures. This long-term reciprocating bending and stretching can easily induce diaphragm fatigue failure or even rupture. Diaphragm fatigue fracture not only causes equipment downtime and significant economic losses but may also lead to hazardous gas leaks and safety accidents. Therefore, conducting scientific fatigue life assessments of compressor diaphragms is of paramount engineering significance and application value for guiding preventative equipment maintenance and ensuring the safe and stable operation of industrial production.

[0003] Current data-driven fatigue life prediction methods typically rely on a large amount of accumulated historical fatigue test data to build models. However, diaphragms in different scenarios vary in properties such as elastic modulus, Poisson's ratio, coefficient of thermal expansion, and surface roughness, and their operating conditions, including temperature and pressure fluctuations, differ, leading to inconsistencies in the reference value of massive amounts of historical data for the diaphragm currently being evaluated. Furthermore, historical data itself possesses multiple dimensions: the variance of lifespan in repeated tests indicates data reliability, the time of data generation relates to information timeliness, and the data distribution density in the parameter space affects the representativeness of the features. Traditional evaluation methods often fail to consider these factors, typically treating historical data indiscriminately or using simple, uniform weighting. This makes prediction models susceptible to interference from low-correlation, high-dispersion, or outdated data, severely limiting the accuracy and reliability of the evaluation results. Therefore, establishing a scientifically sound data weighting mechanism that integrates multidimensional differences to construct a high-precision fatigue life prediction model is a critical issue that urgently needs to be addressed. Summary of the Invention

[0004] To improve the accuracy and reliability of diaphragm fatigue life assessment results, one or more embodiments of the present invention provide a method for assessing compressor diaphragm fatigue life based on multi-source data, including the following steps:

[0005] The elastic modulus, Poisson's ratio, coefficient of thermal expansion, surface roughness, operating temperature, and pressure fluctuation parameters of the compressor diaphragm to be evaluated are obtained. A historical fatigue life dataset containing multiple historical data points is also obtained, with each historical data point containing fatigue life data from multiple repeated tests. A normalized feature difference vector is constructed between each historical data point and the compressor diaphragm to be evaluated in the dimensions of elastic modulus, Poisson's ratio, coefficient of thermal expansion, and surface roughness. The p-norm of the feature difference vector is calculated, and the reciprocal of the p-norm is determined as the basic similarity weight. In the parameter space composed of the normalized operating temperature and pressure fluctuation parameters, the average Euclidean distance between each historical data point and its k nearest neighbor data points is queried using a KD-tree, and the reciprocal of the average Euclidean distance is set as the local density index. The statistical variance of fatigue life data for each historical data point is extracted, a basic decay factor related to the data timestamp is set, and the information decay factor is obtained by power function adjustment of the basic decay factor using the normalized local density index. The data confidence coefficient is determined by multiplying the inverse of the statistical variance with the information decay factor. The unnormalized weight is obtained by multiplying the basic similarity weight with the data confidence coefficient and then normalized. A weighted least squares support vector machine prediction model is constructed based on weights; the operating temperature and pressure fluctuation parameters of the compressor diaphragm to be evaluated are input into the prediction model, and the fatigue life assessment results are output.

[0006] Preferably, the construction of a normalized feature difference vector between each historical data point and the compressor diaphragm to be evaluated in the dimensions of elastic modulus, Poisson's ratio, coefficient of thermal expansion, and surface roughness, and the calculation of the p-norm of the feature difference vector, with the reciprocal of the p-norm determined as the basic similarity weight, includes: calculating the absolute difference between the historical data point and the compressor diaphragm to be evaluated in the four feature dimensions of elastic modulus, Poisson's ratio, coefficient of thermal expansion, and surface roughness; using the difference between the maximum and minimum values ​​of the corresponding feature dimensions in the historical dataset, performing division normalization on the absolute difference respectively, and combining them to form a four-dimensional normalized feature difference vector; calculating the p-norm of the normalized feature difference vector, and determining the basic similarity weight by the reciprocal of the sum of the p-norm and a preset positive smoothing term, where p is a constant greater than or equal to 1.

[0007] Preferably, the step of querying the average Euclidean distance between each historical data point and its k nearest neighbor data points in the parameter space composed of normalized operating temperature and pressure fluctuation parameters, and setting the reciprocal of the average Euclidean distance as the local density index, includes: constructing a KD-tree spatial index structure using the normalized operating temperature and pressure fluctuation parameters of all historical data points as spatial coordinate features; traversing the historical dataset, and for each target historical data point, retrieving the k nearest neighbor historical data points through the KD-tree spatial index structure; calculating the sum of the Euclidean distances between the target historical data point and the k neighboring historical data points, and dividing by k to obtain the average Euclidean distance; and calculating the reciprocal of the sum of the average Euclidean distance and a preset positive smoothing term as the local density index of the target historical data point.

[0008] Preferably, the step of extracting the statistical variance of fatigue life data for each historical data point, setting a basic attenuation factor related to the data timestamp, and using a normalized local density index to adjust the basic attenuation factor using a power function to obtain the information attenuation factor includes: extracting the time difference between the timestamp of each historical data point and the current evaluation time; uniformly converting the time difference to a preset time unit, wherein the preset time unit is a month; the attenuation base represents the information retention ratio after each preset time unit, and the basic attenuation factor is calculated based on the uniformly converted time difference; dividing the preset adjustment coefficient by the normalized local density index after lower limit constraint processing to obtain the adjustment index; and performing a power operation with the basic attenuation factor as the base and the adjustment index as the exponent to calculate the information attenuation factor.

[0009] Preferably, the step of multiplying the reciprocal of the statistical variance by the information decay factor to determine the data confidence coefficient includes: calculating the statistical variance of fatigue life data for each historical data point; dividing 1 by the sum of the statistical variance and a preset positive smoothing term to obtain the reciprocal of the variance; multiplying the reciprocal of the variance by the information decay factor of the corresponding historical data point, and determining the value of the result as the data confidence coefficient of the historical data point.

[0010] Preferably, the step of multiplying the basic similarity weights by the data confidence coefficients to obtain non-normalized weights, and then normalizing them to obtain the weights, includes: for each historical data point, performing a scalar product of the basic similarity weights corresponding to the data point and the data confidence coefficients to obtain non-normalized weights; summing the non-normalized weights of all historical data points in the historical dataset to obtain a total weight; and dividing the non-normalized weight of each historical data point by the total weight to obtain the normalized weight of each historical data point.

[0011] Preferably, the step of constructing a weighted least squares support vector machine prediction model based on weights includes: using the operating temperature and pressure fluctuation parameters of each historical data point in the historical dataset as input features, and using the mean of fatigue life data from multiple repeated tests as output labels to establish an optimization objective function for the least squares support vector machine; using the weight of each historical data point as a weight factor for the penalty coefficient and substituting it into the optimization objective function to form a weighted optimization objective function; solving the weighted optimization objective function to obtain model parameters, and constructing the weighted least squares support vector machine prediction model.

[0012] Preferably, the step of inputting the operating temperature and pressure fluctuation parameters of the compressor diaphragm to be evaluated into the prediction model and outputting the fatigue life assessment result includes: extracting the current operating temperature and pressure fluctuation parameters of the compressor diaphragm to be evaluated; converting the extracted operating temperature and pressure fluctuation parameters into a test feature vector consistent with the input dimension of the prediction model; inputting the test feature vector into the prediction model for forward inference calculation, and outputting the fatigue life prediction value of the compressor diaphragm to be evaluated as the assessment result.

[0013] Preferably, the step of obtaining a historical fatigue life dataset containing multiple historical data points includes: extracting an initial dataset from a historical database using a data reading program; grouping and aggregating the initial dataset using material batches and test conditions as keys to obtain multiple historical data points; and extracting the number of repeated fatigue fracture cycles under each historical data point after validity screening as fatigue life data of the multiple repeated tests for the corresponding historical data point.

[0014] Preferably, the step of solving the weighted optimization objective function to obtain model parameters and constructing a weighted least squares support vector machine prediction model includes: setting a regularization penalty coefficient, and determining the product of the weight of each historical data point and the regularization penalty coefficient as the error squared term penalty coefficient for the corresponding historical data point; using the reciprocal of the error squared term penalty coefficient for each historical data point as a diagonal correction term, and forming a matrix to be solved under the Caro-Kun-Tucker optimality condition together with the kernel matrix; solving a system of linear algebraic equations containing the matrix to be solved to obtain a Lagrange multiplier coefficient vector and a scalar bias term; and substituting the Lagrange multiplier coefficient vector and the scalar bias term into the kernel function expansion to construct the weighted least squares support vector machine prediction model.

[0015] The beneficial effects of this invention are as follows: This invention improves the accuracy and reliability of prediction results. By constructing a multi-dimensional feature difference vector and calculating the reciprocal of the sum of the p-norm and a preset positive smoothing term, the basic similarity between historical samples and the target to be evaluated at the material level is represented. In the working condition parameter space, the KD-tree algorithm is used to calculate the local density index, and combined with the statistical variance of historical samples and a timestamp-based decay factor, the confidence coefficient of the data is defined. This filters and weakens the interference of poor-quality data with high dispersion, long history, and isolated distribution on the model. The similarity and confidence are deeply integrated to generate weights, which guide the construction of a weighted least squares support vector machine, enabling the prediction model to highly focus on high-value and highly correlated samples, thus achieving accurate assessment of the fatigue life of compressor diaphragms. Attached Figure Description

[0016] Figure 1 This is a flowchart of the compressor diaphragm fatigue life assessment method based on multi-source data in an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the differences in normalized features in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the change in information attenuation factor in an embodiment of the present invention; Figure 4 This is a schematic diagram comparing the ablation experiment results in an embodiment of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0018] One or more embodiments of the present invention provide a method for assessing the fatigue life of compressor diaphragms based on multi-source data, such as... Figure 1 As shown, it includes the following steps: S1, obtain the elastic modulus, Poisson's ratio, coefficient of thermal expansion, surface roughness, operating temperature and pressure fluctuation parameters of the compressor diaphragm to be evaluated, and obtain a historical fatigue life dataset containing multiple historical data points. Each historical data point contains fatigue life data from multiple repeated tests.

[0019] The elastic modulus and Poisson's ratio of the compressor diaphragm under evaluation were measured using a tensile testing machine. The coefficient of thermal expansion was measured using a thermal expansion meter. Surface roughness was obtained by scanning with a white light interferometer. Operating temperature and pressure fluctuation parameters were collected in real time using thermocouple sensors and pressure transmitters. The collected parameters were converted into floating-point data and stored on a local server. A historical fatigue life dataset was extracted from a historical database using a data reading program. This dataset contains multiple historical data points, which were grouped and aggregated using material batches and test conditions as keys. The number of repeated fatigue fracture cycles for each historical data point, after validity screening, and suitable for Weibull distribution fitting, was extracted as fatigue life data. Weibull distribution fitting refers to using a Weibull probability model to estimate parameters such as life, failure time, or number of fatigue cycles to characterize the failure pattern of the samples and predict reliable life.

[0020] S2, construct a normalized feature difference vector between each historical data point and the compressor diaphragm to be evaluated in terms of elastic modulus, Poisson's ratio, coefficient of thermal expansion, and surface roughness. Calculate the p-norm of the feature difference vector and determine the reciprocal of the p-norm as the basic similarity weight.

[0021] The normalization process, or an equivalent normalization algorithm, is invoked to first determine the maximum, minimum, and characteristic span of the four dimensions—elastic modulus, Poisson's ratio, coefficient of thermal expansion, and surface roughness—from the historical dataset. Then, the absolute difference between each historical data point and the compressor diaphragm to be evaluated across these four dimensions is calculated, and each absolute difference is divided by the characteristic span of the corresponding dimension to generate a four-dimensional normalized feature difference vector. This approach uses the uniform characteristic span of the historical dataset as the scaling benchmark, rather than independently normalizing the samples to be evaluated and the historical samples before subtracting them.

[0022] The p-norm of the normalized feature difference vector is calculated using a norm calculation program; when p is 2, the p-norm is the Euclidean norm. In the algorithm logic, 1 × 10⁻⁶ is added to the calculated p-norm. -6 To prevent the denominator from being zero, a minimum constant is used. The reciprocal mathematical operation is performed, and the resulting floating-point value is used as the basic similarity weight of the compressor diaphragm to be evaluated corresponding to the historical data point. The values ​​are then stored in a floating-point one-dimensional array.

[0023] In one implementation, the construction of a normalized feature difference vector between each historical data point and the compressor diaphragm to be evaluated in terms of elastic modulus, Poisson's ratio, coefficient of thermal expansion, and surface roughness, and the calculation of the p-norm of the feature difference vector, with the reciprocal of the p-norm determining the basic similarity weight, includes: Calculate the absolute differences between historical data points and the compressor diaphragm to be evaluated in four characteristic dimensions: elastic modulus, Poisson's ratio, coefficient of thermal expansion, and surface roughness; By using the difference between the maximum and minimum values ​​of the corresponding feature dimensions in the historical dataset, the absolute difference is normalized by division, and then combined to form a four-dimensional normalized feature difference vector. Calculate the p-norm of the normalized feature difference vector, and determine the basic similarity weight by the reciprocal of the sum of the p-norm and the preset positive smoothing term, where p is a constant greater than or equal to 1.

[0024] Obtain the elastic modulus (e.g., 200 GPa), Poisson's ratio (e.g., 0.3), and coefficient of thermal expansion (e.g., 1.2 × 10⁻⁶) of the compressor diaphragm to be evaluated. -5 ) / K and surface roughness, for example For a length of 0.8 micrometers, the four feature values ​​corresponding to the i-th historical data point in the historical dataset are read, and the absolute differences in the four dimensions are calculated using the absolute difference calculation formula. To eliminate the scale effect caused by different units, the maximum and minimum values ​​of all data points in the historical dataset for each dimension are extracted, and the feature span interval is calculated. This interval value is used to perform division normalization on the absolute difference. For example, when the maximum elastic modulus of the historical dataset is 220 GPa and the minimum is 180 GPa, the span is 40 GPa. If the absolute difference between the membrane to be evaluated and the historical data point is 10 GPa, then the normalization value for this dimension is 0.25. The normalized values ​​of the above four dimensions are combined in a fixed order to form a four-dimensional normalized feature difference vector, and the p-norm of this feature difference vector is calculated. The constant p can range from 1 to 3, for example, p=2, i.e., the Euclidean norm, to balance the weights of the differences in each dimension. A very small positive number is added to this p-norm, such as... 10 -6 To prevent anomalies caused by a zero denominator, the reciprocal of the p-norm is calculated as the basic similarity weight for that historical data point. A larger weight value indicates that the historical data is closer to the characteristics of the membrane being evaluated. The normalized feature differences are illustrated below. Figure 2 As shown.

[0025] S3. In the parameter space composed of the normalized working temperature and pressure fluctuation parameters, the average Euclidean distance between each historical data point and the k nearest neighbor data points is queried using a KD tree, and the reciprocal of the average Euclidean distance is set as the local density index.

[0026] The maximum-minimum normalization algorithm is called again to normalize the operating temperature and pressure fluctuation parameters of all historical data points, and the data is then concatenated column-wise into a two-dimensional parameter space feature matrix. The KD-tree spatial indexing algorithm is then called, using the constructed two-dimensional parameter space feature matrix as input parameters to instantiate the KD-tree model. Each historical data point in the two-dimensional parameter space feature matrix is ​​iterated through, and the built-in query function of the KD-tree model is called, setting the query quantity to k+1. After removing the target historical data point itself from the query results, the k nearest neighbor historical data points are retained, and nearest neighbor retrieval is performed using the Euclidean distance metric. The arithmetic mean of the Euclidean distances from the current historical data point to the k nearest neighbor historical data points is calculated. To prevent division by zero from causing computational crashes, 1×10⁻⁶ is added to the calculated average Euclidean distance value. -8 After the numerical smoothing term, the reciprocal operation is performed, and the resulting numerical result is set as the local density index of the corresponding data point.

[0027] In one implementation, the step of querying the average Euclidean distance between each historical data point and its k nearest neighbor data points using a KD tree in the parameter space formed by the normalized operating temperature and pressure fluctuation parameters, and setting the reciprocal of the average Euclidean distance as the local density exponent, includes: Using the normalized operating temperature and pressure fluctuation parameters of all historical data points as spatial coordinate features, a KD tree spatial index structure is constructed. Traverse the historical dataset, and for each target historical data point, retrieve the k nearest neighboring historical data points using the KD-tree spatial index structure; Calculate the sum of the Euclidean distances between the target historical data point and the k adjacent historical data points, and divide by k to obtain the average Euclidean distance; The reciprocal of the sum of the average Euclidean distance and a preset positive smoothing term is calculated and used as the local density index of the target historical data point.

[0028] For historical data points, the operating temperature (e.g., 20℃ to 120℃ range) and pressure fluctuation parameters (e.g., pressure pulsation peak-to-peak value, ranging from 0.5MPa to 2.0MPa) are mapped to the [0,1] interval using a maximum-minimum normalization method. A two-dimensional feature space is constructed using these two-dimensional normalized parameters as coordinates. Within this space, a KD-tree spatial index structure is constructed based on a spatial partitioning recursive algorithm. Specifically, at each node, the dimension with the largest variance is selected as the partition axis, and the median of the data in that dimension is used as the split surface, until all historical data points are assigned to leaf nodes, thus enabling spatial feature lookup. For the target historical data point in the historical dataset, a backtracking search algorithm is used with the constructed KD-tree to retrieve the k nearest neighboring historical data points. The selection of parameter k affects the smoothness of the density assessment; its value can range from 5 to 20, for example, k=10. The Euclidean distance between the target point and the k nearest neighbors is calculated, and the sum of these distances is divided by k to obtain the average Euclidean distance. For example, if the sum of the Euclidean distances of the 10 nearest neighbors is 0.5, then the average Euclidean distance is 0.05. A small amount is added to this average Euclidean distance to prevent overflow during division by zero, such as... =10 -5 Taking the reciprocal yields the local density index. This local density index is approximately 20, a value that represents the density of the data distribution in the parameter space.

[0029] S4. Extract the statistical variance of fatigue life data for each historical data point, set a basic attenuation factor related to the data timestamp, and use the normalized local density index to adjust the basic attenuation factor by a power function to obtain the information attenuation factor.

[0030] For each historical data point containing data on the number of fatigue life fractures from multiple repeated tests, the statistical variance of the dispersion is calculated. The timestamp when each historical data point was written to the database and the system timestamp of the current model evaluation task are extracted, the time difference between the two is calculated, and then uniformly converted to a time difference value in months. A preset attenuation base b is set, where 0 < b < 1, and the basic attenuation factor is calculated based on the time difference Δt, for example, calculating... Alternatively, the basic decay factor can be calculated using an equivalent exponential decay form.

[0031] The maximum-minimum normalization algorithm is then applied to all the local density indices obtained above. To avoid division by zero anomalies, the normalized local density indices are limited to a preset lower limit δ, for example, δ is set to 0.1, resulting in the normalized local density indices. Set the preset adjustment coefficient. Divided by the normalized local density index The adjustment index is obtained. Perform a power function operation, using the calculated base decay factor as the base, to adjust the exponent. A nonlinear power operation is performed to calculate the exponent, and the output value is the information decay factor for that data point. Since the basic decay factor is less than 1, the larger the normalized local density exponent, the smaller the adjustment exponent, and the larger the information decay factor, thus preserving high reference value for dense data areas; conversely, the smaller the normalized local density exponent, the larger the adjustment exponent, and the smaller the information decay factor, thus accelerating the decay of sparse data areas.

[0032] In one implementation, the step of extracting the statistical variance of fatigue life data for each historical data point, setting a basic attenuation factor related to the data timestamp, and using a power function to adjust the basic attenuation factor using a normalized local density index to obtain an information attenuation factor includes: Extract the time difference between the timestamp of each historical data point and the current evaluation time, and convert the time difference into a preset time unit, which is a month; the decay base represents the information retention ratio after each preset time unit, and calculate the basic decay factor based on the converted time difference; The preset adjustment coefficient is divided by the normalized local density index after the lower limit constraint processing to obtain the adjustment index; The information attenuation factor is calculated by exponentiation with the base attenuation factor as the base and the adjustment index as the exponent.

[0033] Extract the timestamps of each historical data point, such as May 1, 2022, and compare them with the current evaluation time, such as October 1, 2023, and convert the time difference to 17 months. Set a preset attenuation base, which can range from 0.80 to 0.99, for example, a value of 0.95. Calculate the basic attenuation factor using a formula with the attenuation base as the base and the time difference as the exponent. In this example, it is 0.95 to the power of 17, approximately equal to 0.418. This basic attenuation factor represents the attenuation law of data reliability over time. Normalize the previously obtained local density index using the maximum-minimum method to obtain a normalized local density index in the range [0.1, 1.0], with the lower limit set to 0.1 to avoid triggering division by zero anomalies. Set a preset adjustment coefficient to control the intensity of density intervention on attenuation, with a value range from 0.5 to 2.0, for example, 1.0. Divide the coefficient by the normalized local density index to obtain the adjustment index. If the normalized local density index is 0.5, then the adjustment index is 2.0. Performing a power operation, i.e., using 0.418 as the base and 2.0 as the exponent, yields an information decay factor of approximately 0.175. If the normalized local density index of another historical data point is 1.0, then the adjustment index is 1.0, and the information decay factor is approximately 0.418, higher than the information decay factor when the normalized local density index is 0.5, indicating that data in dense regions retains higher reference value. This power function adjustment mechanism achieves the adjustment of data weights: accelerating weight decay for data in sparse regions and retaining higher reference value for data in dense regions. The change in the information decay factor is illustrated in the diagram below. Figure 3 As shown.

[0034] S5. Multiply the inverse of the statistical variance by the information decay factor to determine the data confidence coefficient; multiply the basic similarity weight by the data confidence coefficient to obtain the non-normalized weight, and then normalize it to obtain the weight.

[0035] The statistical variance calculated in the previous step is added to a preset positive smoothing term, and then its reciprocal is taken. The reciprocal of the statistical variance is then multiplied element-wise by the corresponding information decay factor to obtain the data confidence coefficient representing the reliability of each historical data point. The one-dimensional array of basic similarity weights calculated in the previous step is multiplied at corresponding positions by the one-dimensional array of data confidence coefficients to obtain the non-normalized weight set for each historical data point. A linear summation normalization algorithm is used to calculate the sum of the non-normalized weight elements of all historical data points. The non-normalized weights of each historical data point are then divided by this sum to obtain a one-dimensional floating-point array of weights that sums to 1.

[0036] In one implementation, determining the data confidence coefficient by multiplying the inverse of the statistical variance by the information decay factor includes: Calculate the statistical variance of fatigue life data for each historical data point; Divide 1 by the sum of the statistical variance and the preset positive smoothing term to obtain the reciprocal of the variance; The inverse of the variance is multiplied by the information decay factor of the corresponding historical data point, and the result is determined as the data confidence coefficient of the historical data point.

[0037] For a given historical data point, N repeated tests are conducted under the same operating conditions, where N can range from 3 to 10 (e.g., N=5). This yields a set of fatigue life measurements. The arithmetic mean of all life values ​​within this set is calculated, and the statistical variance is then solved using the statistical variance formula. For example, the life values ​​from 5 repeated tests might be 1.02, 0.98, 1.05, 1.00, and 0.95, in units of 10⁻⁶ cycles. 6 In the next iteration, the mean is calculated to be 1.00, and the statistical variance is 0.00145. To represent the central tendency of this data set, a smoothing parameter is input, for example, a value of 0.01 to prevent the variance from being zero due to perfect data uniformity. 1 is divided by the statistical variance after adding the smoothing parameter to calculate the reciprocal of the variance. Using the example value, the reciprocal of the variance is approximately 87.34. Multiplying this reciprocal of the variance by the information decay factor for the corresponding historical data point, for example, a calculated value of 0.175, yields a confidence coefficient of approximately 15.28. This step combines the stability of the experimental data with the spatiotemporal decay characteristics, representing the confidence level of a single historical data point when used for current prediction and evaluation.

[0038] In one implementation, the step of multiplying the basic similarity weight by the data confidence coefficient to obtain a non-normalized weight, and then performing normalization processing to obtain the weight, includes: For each historical data point, the basic similarity weight corresponding to the data point is multiplied by the data confidence coefficient by a scalar to obtain the non-normalized weight. The non-normalized weights of all historical data points in the historical dataset are summed to obtain the total weights; Divide the unnormalized weight of each historical data point by the sum of the weights to obtain the normalized weight of each historical data point.

[0039] Extract the basic similarity weight (e.g., 2.5 as calculated above) and the data confidence coefficient (e.g., 15.28 as calculated above) from each historical data point in the historical dataset. Use scalar multiplication to calculate the unnormalized weight of that data point; in the example, the product is 38.2. This value represents the reference value of historical data for the current evaluation object in terms of attribute similarity, time decay, and experimental reliability. Iterate through all data points in the historical dataset globally (e.g., a total of 1000 samples), summing the unnormalized weights of each data point to calculate the global weight sum (assuming the total sum in the example is 12500). Divide the unnormalized weight of each historical data point by this global weight sum; in the example, divide 38.2 by the total sum of 12500 to obtain a normalized weight of approximately 0.003056. This normalization process ensures that the weights of all historical data are distributed in the [0,1] interval and the total sum is 1, satisfying the numerical input requirements of the sample weight factors for subsequent weighted machine learning models.

[0040] S6. Construct a weighted least squares support vector machine prediction model based on weights; input the working temperature and pressure fluctuation parameters of the compressor diaphragm to be evaluated into the prediction model, and output the fatigue life assessment results.

[0041] A weighted least squares support vector machine regression prediction model is constructed based on the aforementioned one-dimensional array of weights summing to 1. A Gaussian radial basis function kernel is selected as the nonlinear tool for feature mapping, and a linear equation solver is used to solve the system of linear equations based on the KKT optimality conditions. In the objective function of the weighted least squares support vector machine, the penalty coefficient of the squared error term for the i-th historical data point is set as the product of the regularization penalty coefficient and the weight of that historical data point. In the equivalent KKT linear equation system, the diagonal correction term is set as the reciprocal of the penalty coefficient of the squared error term, and together with the kernel matrix, forms the matrix to be solved. The Lagrange multiplier coefficient vector and scalar bias term are obtained through solving the linear equation system, completing the closed-loop training of the prediction model. The KKT linear equation system refers to transforming the optimality conditions of a constrained optimization problem into a set of linear algebraic equations. Solving this system yields the model parameters and Lagrange multipliers. The regularization penalty coefficient is preset based on the number of historical samples and the normalized weight scale. The current operating temperature and pressure fluctuation parameters of the compressor diaphragm to be evaluated are concatenated into a one-dimensional test input vector. The inner product matrix of the radial basis kernel function of this input vector with all historical training data is calculated. After matrix multiplication with the calculated Lagrange multiplier coefficient vector, a scalar bias term is added. The output model prediction scalar value is used as the fatigue life evaluation result of the compressor diaphragm to be evaluated. When it is necessary to output the number of fracture cycles in integer form, rounding or unit conversion is performed according to the preset accuracy.

[0042] In one implementation, the construction of the weighted least squares support vector machine prediction model based on weights includes: Using the operating temperature and pressure fluctuation parameters of each historical data point in the historical dataset as input features, and the mean value of fatigue life data from multiple repeated tests as output labels, an optimization objective function of least squares support vector machine is established. The weight of each historical data point is used as a weight factor of the penalty coefficient and substituted into the optimization objective function to form a weighted optimization objective function; Solve the weighted optimization objective function to obtain the model parameters, and construct the weighted least squares support vector machine prediction model.

[0043] A two-dimensional input feature vector is constructed using the operating temperature and pressure fluctuation parameters of historical data points, and a training sample set is built using the average fatigue life of the corresponding repeated tests as the output label. For the least squares support vector machine, the hyperplane parameters are set as the weight vector and bias, and the input error variable is used to construct an optimization objective function with an error squared term. To achieve differentiated learning for different samples, a regularization penalty coefficient is used, with a value ranging from 10 to 1000, for example, a value of 100. The weights of each historical data point calculated earlier are multiplied and added to the error squared penalty term of the corresponding sample to construct a weighted optimization objective function. In this function, the feature mapping kernel function can be a radial basis function, with kernel parameters ranging from 1.0 to 5.0, for example, set to 2.0. A Lagrange function is constructed using the Lagrange multiplier method, and the equality constraints are integrated into the objective function. The partial derivatives with respect to the weight vector, bias, error variable, and Lagrange multipliers are set to zero, transforming the optimization problem into solving a system of linear matrix equations. Solving the matrix equations yields the Lagrange multiplier vectors and bias constants. Substituting the obtained model parameters back into the kernel function expansion, a nonlinear prediction model expression for the weighted least squares support vector machine is constructed. This model structure takes into account the imbalance in information quality and reference value of historical data during training.

[0044] In one implementation, the step of inputting the operating temperature and pressure fluctuation parameters of the compressor diaphragm to be evaluated into the prediction model and outputting fatigue life assessment results includes: Extract the current operating temperature and pressure fluctuation parameters of the compressor diaphragm to be evaluated; The extracted operating temperature and pressure fluctuation parameters are converted into test feature vectors that are consistent with the input dimensions of the prediction model. The test feature vector is input into the prediction model for forward inference calculation, and the fatigue life prediction value of the compressor diaphragm to be evaluated is output as the evaluation result.

[0045] By deploying a sensor network or monitoring bus at the compressor diaphragm's service site, the current operating temperature of the compressor diaphragm under evaluation is collected, for example, a sensor reading of 85.5℃. Pressure fluctuation data is simultaneously acquired, and the peak-to-peak value of the pressure pulsation is extracted as the pressure fluctuation parameter, for example, 1.25 MPa. To ensure consistency between the test data distribution and the training data distribution, the two parameters are normalized according to the standardized formula and scaling ratio used during historical model training, for example, mapped to 0.65 and 0.50 respectively. They are then concatenated in a predetermined order: operating temperature first, pressure fluctuation parameter second, to generate a 1×2 two-dimensional test feature vector. This normalized test feature vector is then fed into a weighted least squares support vector machine prediction model with completed parameter solving for forward inference calculation. Inside the model, the RBF kernel function response value between the test feature vector and all historical support vector data points is calculated. This response value is multiplied and accumulated with the corresponding Lagrange multipliers, and a bias constant is added to the end of the result. This calculation process outputs a specific floating-point value, such as 1.08×10⁻⁶. 6 In the next cycle, this value is the predicted fatigue life of the compressor diaphragm under the current specified operating conditions. The evaluation system outputs the predicted value as the fatigue life evaluation result to guide preventive maintenance and replacement decisions.

[0046] The ablation experiment was conducted based on a dataset of 1500 real compressor diaphragm operation data sets, with 1000 sets used as the training set and 500 sets used as an independent test set for fatigue life prediction verification. The experimental hardware platform used a workstation with a 64-core CPU and 128GB of memory. The feature dimensions were uniformly set using operating temperature and pressure fluctuation parameters. The evaluation metrics were the mean absolute error and prediction accuracy. Prediction accuracy represents the percentage of test samples whose predicted fatigue life value does not exceed 10% of the true value. The experiment included four comparison groups: a traditional least squares support vector machine basic model without any sample weights, Model 1 with only basic similarity weights, Model 2 combining basic similarity weights and information decay factors, and a complete implementation model with weights obtained by superimposing data confidence coefficients.

[0047] By performing forward inference computation on the same test set, the mean absolute error of the base model was 85,000 iterations, with a prediction accuracy of 82.3%. After adding basic similarity weights, the mean absolute error of Model 1 decreased to 62,000 iterations, and the prediction accuracy improved to 88.5%. Model 2, constructed by combining temporal and spatial density decay features, reduced the mean absolute error to 45,000 iterations, achieving a prediction accuracy of 92.1%. The weighted least squares support vector machine prediction model, which uses the complete implementation to calculate the comprehensive weights, performed best, reducing the mean absolute error to 21,000 iterations and increasing the prediction accuracy to 97.4%. The ablation experiment results are illustrated below. Figure 4 As shown.

[0048] The calculation of basic similarity weights filters reference samples with similar characteristics, while the information decay mechanism filters out low-value data that is old in time or spatially isolated. The addition of the data confidence coefficient suppresses random interference caused by the variance of repeated trials. The complete scheme assigns reference value to historical samples through a multi-dimensional weighting system, eliminating the negative impact of heterogeneous noise data on the training process and improving the reliability and inference accuracy of compressor diaphragm fatigue life assessment.

[0049] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A method for assessing the fatigue life of compressor diaphragms based on multi-source data, characterized in that, Includes the following steps: The elastic modulus, Poisson's ratio, coefficient of thermal expansion, surface roughness, operating temperature, and pressure fluctuation parameters of the compressor diaphragm to be evaluated are obtained. A historical fatigue life dataset containing multiple historical data points is also obtained, with each historical data point containing fatigue life data from multiple repeated tests. A normalized feature difference vector is constructed between each historical data point and the compressor diaphragm to be evaluated in the dimensions of elastic modulus, Poisson's ratio, coefficient of thermal expansion, and surface roughness. The p-norm of the feature difference vector is calculated, and the reciprocal of the p-norm is determined as the basic similarity weight. In the parameter space composed of the normalized operating temperature and pressure fluctuation parameters, the average Euclidean distance between each historical data point and its k nearest neighbor data points is queried using a KD-tree, and the reciprocal of the average Euclidean distance is set as the local density index. The statistical variance of fatigue life data for each historical data point is extracted, a basic decay factor related to the data timestamp is set, and the information decay factor is obtained by power function adjustment of the basic decay factor using the normalized local density index. The data confidence coefficient is determined by multiplying the inverse of the statistical variance with the information decay factor. The unnormalized weight is obtained by multiplying the basic similarity weight with the data confidence coefficient and then normalized. A weighted least squares support vector machine prediction model is constructed based on weights; the operating temperature and pressure fluctuation parameters of the compressor diaphragm to be evaluated are input into the prediction model, and the fatigue life assessment results are output.

2. The method for assessing the fatigue life of a compressor diaphragm based on multi-source data according to claim 1, characterized in that, The process involves constructing a normalized feature difference vector between each historical data point and the compressor diaphragm to be evaluated in terms of elastic modulus, Poisson's ratio, coefficient of thermal expansion, and surface roughness. The p-norm of this feature difference vector is calculated, and the reciprocal of the p-norm is used to determine the basic similarity weight, including: Calculate the absolute differences between historical data points and the compressor diaphragm to be evaluated in four characteristic dimensions: elastic modulus, Poisson's ratio, coefficient of thermal expansion, and surface roughness; By using the difference between the maximum and minimum values ​​of the corresponding feature dimensions in the historical dataset, the absolute difference is normalized by division, and then combined to form a four-dimensional normalized feature difference vector. Calculate the p-norm of the normalized feature difference vector, and determine the basic similarity weight by the reciprocal of the sum of the p-norm and the preset positive smoothing term, where p is a constant greater than or equal to 1.

3. The method for assessing the fatigue life of a compressor diaphragm based on multi-source data according to claim 2, characterized in that, In the parameter space formed by the normalized operating temperature and pressure fluctuation parameters, the average Euclidean distance between each historical data point and its k nearest neighbor data points is queried using a KD tree. The reciprocal of the average Euclidean distance is set as the local density exponent, including: Using the normalized operating temperature and pressure fluctuation parameters of all historical data points as spatial coordinate features, a KD tree spatial index structure is constructed. Traverse the historical dataset, and for each target historical data point, retrieve the k nearest neighboring historical data points using the KD-tree spatial index structure; Calculate the sum of the Euclidean distances between the target historical data point and the k adjacent historical data points, and divide by k to obtain the average Euclidean distance; The reciprocal of the sum of the average Euclidean distance and a preset positive smoothing term is calculated and used as the local density index of the target historical data point.

4. The method for assessing the fatigue life of a compressor diaphragm based on multi-source data according to claim 1, characterized in that, The process involves extracting the statistical variance of fatigue life data for each historical data point, setting a basic attenuation factor related to the data timestamp, and using a power function adjustment of the basic attenuation factor with a normalized local density index to obtain the information attenuation factor, including: Extract the time difference between the timestamp of each historical data point and the current evaluation time, and convert the time difference into a preset time unit, which is a month; preset the attenuation base, which represents the information retention ratio after each preset time unit, and calculate the basic attenuation factor based on the converted time difference; The preset adjustment coefficient is divided by the normalized local density index after the lower limit constraint processing to obtain the adjustment index; The information attenuation factor is calculated by exponentiation with the base attenuation factor as the base and the adjustment index as the exponent.

5. The method for assessing the fatigue life of a compressor diaphragm based on multi-source data according to claim 1 or 4, characterized in that, The step of determining the data confidence coefficient by multiplying the inverse of the statistical variance by the information decay factor includes: Calculate the statistical variance of fatigue life data for each historical data point; Divide 1 by the sum of the statistical variance and the preset positive smoothing term to obtain the reciprocal of the variance; The inverse of the variance is multiplied by the information decay factor of the corresponding historical data point, and the result is determined as the data confidence coefficient of the historical data point.

6. The method for assessing the fatigue life of a compressor diaphragm based on multi-source data according to claim 1, characterized in that, The step of multiplying the basic similarity weights by the data confidence coefficients to obtain non-normalized weights, and then normalizing them to obtain the final weights, includes: For each historical data point, the basic similarity weight corresponding to the data point is multiplied by the data confidence coefficient by a scalar to obtain the non-normalized weight. The non-normalized weights of all historical data points in the historical dataset are summed to obtain the total weights; Divide the unnormalized weight of each historical data point by the sum of the weights to obtain the normalized weight of each historical data point.

7. The method for assessing the fatigue life of a compressor diaphragm based on multi-source data according to claim 1, characterized in that, The weighted least squares support vector machine prediction model based on weights includes: Using the operating temperature and pressure fluctuation parameters of each historical data point in the historical dataset as input features, and the mean value of fatigue life data from multiple repeated tests as output labels, an optimization objective function of least squares support vector machine is established. The weight of each historical data point is used as a weight factor of the penalty coefficient and substituted into the optimization objective function to form a weighted optimization objective function; Solve the weighted optimization objective function to obtain the model parameters, and construct the weighted least squares support vector machine prediction model.

8. The method for assessing the fatigue life of a compressor diaphragm based on multi-source data according to claim 1, characterized in that, The process involves inputting the operating temperature and pressure fluctuation parameters of the compressor diaphragm to be evaluated into the prediction model, and outputting fatigue life assessment results, including: Extract the current operating temperature and pressure fluctuation parameters of the compressor diaphragm to be evaluated; The extracted operating temperature and pressure fluctuation parameters are converted into test feature vectors that are consistent with the input dimensions of the prediction model. The test feature vector is input into the prediction model for forward inference calculation, and the fatigue life prediction value of the compressor diaphragm to be evaluated is output as the evaluation result.

9. The method for assessing the fatigue life of a compressor diaphragm based on multi-source data according to claim 1, characterized in that, The process of obtaining a historical fatigue life dataset containing multiple historical data points includes: The initial dataset is extracted from the historical database using a data reading program; The initial dataset is grouped and aggregated using material batch and test condition as keys to obtain multiple historical data points; The number of repeated fatigue fracture cycles at each historical data point, after validity screening, is extracted and used as the fatigue life data of the repeated tests at the corresponding historical data point.

10. The method for assessing the fatigue life of a compressor diaphragm based on multi-source data according to claim 7, characterized in that, The process of solving the weighted optimization objective function to obtain model parameters and constructing a weighted least squares support vector machine prediction model includes: Set a regularization penalty coefficient, and multiply the weight of each historical data point by the regularization penalty coefficient to determine the error squared term penalty coefficient for the corresponding historical data point; The reciprocal of the penalty coefficient of the squared error term for each historical data point is used as the diagonal correction term, and together with the kernel matrix, they form the matrix to be solved under the Caro-Kun-Tucker optimality condition. Solve the system of linear algebraic equations containing the matrix to be solved to obtain the Lagrange multiplier coefficient vector and scalar bias term; Substituting the Lagrange multiplier coefficient vector and scalar bias term into the kernel function expansion, the weighted least squares support vector machine prediction model is constructed.

Images