Titanium alloy wire performance detection method, equipment and system

Abstract

The invention relates to the technical field of titanium alloy detection, and discloses a titanium alloy wire performance detection method, equipment and system. The method comprises the following steps: acquiring real-time force value displacement data of a tensile test and echo signal data of ultrasonic detection, and processing to obtain smooth force value displacement data and spectrum characteristic parameters; performing feature layer fusion on the yield strength, the tensile strength, the peak frequency and the amplitude attenuation rate to construct a first multi-dimensional feature matrix; inputting a pre-trained deep belief network model, and predicting mechanical properties such as elongation and hardness value; and comparing with a preset threshold value, marking unqualified items and generating a detection report. According to the method, through multi-source data fusion and advanced algorithm processing, the comprehensiveness and accuracy of detection are improved, efficient evaluation of the performance of the titanium alloy wire is achieved, and the method is suitable for the strict detection requirements for the quality of the titanium alloy wire in the fields of aerospace, medical instruments and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of titanium alloy detection, and in particular to a method, equipment and system for detecting the performance of a titanium alloy wire. Background Art

[0002] Titanium alloy wire, thanks to its high strength, corrosion resistance, and lightweight properties, has found widespread application in aerospace, medical devices, high-end manufacturing, and other fields. These applications place extremely high demands on the mechanical properties of titanium alloy wire, and its stability and reliability are directly related to the quality and safety of the end product. Therefore, accurate and efficient testing of titanium alloy wire properties has become a crucial step in its production and application. At present, the performance testing of titanium alloy wire mainly relies on traditional physical testing methods, such as tensile testing and hardness testing. Tensile testing is a common means of evaluating the mechanical properties of materials by applying axial tension to the wire and measuring its parameters such as yield strength, tensile strength and elongation before fracture. However, this method has obvious limitations: the tensile test is a destructive test, and the wire cannot be put into use after the test, resulting in a waste of raw materials. This will significantly increase production costs, especially for high-value titanium alloy materials; the test process requires manual operation and data recording, which is not only time-consuming, but also easily affected by the operator's skill level and environmental factors. The consistency and accuracy of the test results are difficult to guarantee. In addition to tensile testing, ultrasonic testing technology is also used to assess the quality of titanium alloy wire, primarily for detecting internal defects such as cracks and pores. Traditional ultrasonic testing focuses primarily on the time domain characteristics of the signal, analyzing the amplitude and propagation time of the echo signal to determine the presence of defects. However, this method lacks sufficient correlation analysis with the material's mechanical properties and cannot directly reflect key mechanical indicators such as the wire's yield strength and hardness, limiting the application of the test results. With the development of industrial intelligence, some detection methods that combine sensor technology and data processing algorithms have gradually emerged. However, in existing technologies, most methods rely on only a single type of detection data for performance evaluation. For example, only tensile test data or ultrasonic detection signals are used, which makes it difficult to fully reflect the comprehensive performance of titanium alloy wire. At the same time, the noise suppression effect during data processing is poor, and the feature parameter extraction is not comprehensive, resulting in low performance prediction accuracy. In addition, traditional machine learning models often have insufficient generalization capabilities when processing multi-dimensional and nonlinear feature data, and cannot meet the needs of high-precision detection. In actual production, due to the complex production process of titanium alloy wire, which involves multiple links such as melting, rolling, and heat treatment, parameter fluctuations in any link may lead to differences in material properties. Existing detection methods are difficult to achieve rapid and comprehensive evaluation of wire properties, and often miss detection or misjudgment, which not only affects production efficiency but may also cause unqualified products to enter the market, posing a safety hazard. Therefore, developing a titanium alloy wire performance detection method that can integrate multiple detection data, improve detection accuracy and efficiency, and achieve non-destructive or semi-destructive detection has become a problem that current technicians in this field need to solve. Summary of the Invention

[0003] The object of the present invention is to provide a method, device and system for detecting the properties of titanium alloy wire to solve the problems raised in the above background technology.

[0004] To achieve the above object, the present invention provides a method for detecting the performance of titanium alloy wire, the method comprising: Collecting real-time force-displacement data of the titanium alloy wire to be tested during the tensile test and echo signal data returned by the ultrasonic testing device, and transmitting the real-time force-displacement data and the echo signal data to a data processing unit; The data processing unit performs denoising on the real-time force displacement data using an adaptive median filter algorithm to obtain smoothed force displacement data, and simultaneously performs fast Fourier transform on the echo signal data to extract the amplitude spectrum and phase spectrum of the signal as spectral characteristic parameters; Performing feature layer fusion on the yield strength and tensile strength parameters in the smoothed force displacement data and the peak frequency and amplitude attenuation rate parameters in the spectrum feature parameters to construct a first multi-dimensional feature matrix; Inputting the first multi-dimensional feature matrix into a pre-trained deep belief network model, and predicting the elongation and hardness values ​​of the titanium alloy wire using the deep belief network model to obtain mechanical property prediction results; The mechanical property prediction result is compared with a preset performance threshold range. When the mechanical property prediction result exceeds the preset performance threshold range, it is marked as unqualified, and a titanium alloy wire performance test report containing unqualified items and corresponding parameters is generated.

[0005] Preferably, the first multi-dimensional feature matrix is ​​input into a pre-trained deep belief network model, and the elongation and hardness values ​​of the titanium alloy wire are predicted by the deep belief network model to obtain the mechanical property prediction results, including: Obtaining a multi-dimensional feature matrix and measured mechanical property data corresponding to historical titanium alloy wire samples from a database, and dividing the historical titanium alloy wire samples into a training set and a validation set; Using the training set to perform unsupervised pre-training on the initial deep belief network model, extracting features layer by layer through stacked restricted Boltzmann machines, and fine-tuning the parameters of the initial deep belief network model using a backpropagation algorithm; Using the validation set to perform performance evaluation on the fine-tuned deep belief network model, calculating the prediction error, and stopping training when the prediction error is less than a preset error threshold to obtain the pre-trained deep belief network model; The first multi-dimensional feature matrix is ​​input into the pre-trained deep belief network model, and the elongation prediction value and hardness prediction value of the titanium alloy wire are output as the mechanical property prediction result.

[0006] Preferably, the unsupervised pre-training of the initial deep belief network model using the training set, extracting features layer by layer by stacking restricted Boltzmann machines, and fine-tuning the parameters of the initial deep belief network model using a backpropagation algorithm includes: Initializing the number of neurons in the input layer, hidden layer, and output layer of the initial deep belief network model, and setting the learning rate and number of iterations of the restricted Boltzmann machine; Inputting the multi-dimensional feature matrix in the training set into the first layer of restricted Boltzmann machine, training it through the contrastive divergence algorithm to obtain the first layer feature map; The first layer of feature maps is used as the input of the second layer of restricted Boltzmann machine for training, and stacked to the preset number of layers in sequence to complete the unsupervised pre-training process; The measured mechanical property data of the training set is used as a label, and the mean square error between the predicted value and the label is calculated by the back propagation algorithm. The connection weights and bias values ​​between the neurons in each layer are adjusted according to the mean square error to complete parameter fine-tuning.

[0007] Preferably, the measured mechanical property data of the training set is used as a label, the mean square error between the predicted value and the label is calculated by a back propagation algorithm, and the connection weights and bias values ​​between neurons in each layer are adjusted according to the mean square error to complete parameter fine-tuning, including: Inputting the multi-dimensional feature matrix of the training set into the deep belief network model that has completed unsupervised pre-training to obtain the predicted mechanical property data of each sample; Calculating the mean square error between the predicted mechanical property data and the measured mechanical property data, and using the mean square error as a loss function; The loss function is derived using the gradient descent method to obtain the gradient direction of the connection weights and bias values ​​of each layer; The connection weights and bias values ​​are updated at a preset learning rate according to the gradient direction, and iteration is repeated until the loss function converges to a preset minimum value range.

[0008] Preferably, the yield strength and tensile strength parameters in the smoothed force displacement data are fused with the peak frequency and amplitude attenuation rate parameters in the spectrum feature parameters to construct a first multi-dimensional feature matrix, including: Normalizing the yield strength parameter and the tensile strength parameter in the smoothed force displacement data to obtain standardized mechanical parameters; Normalizing the peak frequency parameter and the amplitude attenuation rate parameter in the spectrum characteristic parameters to obtain standardized spectrum parameters; splicing the standardized mechanical parameters and the standardized spectral parameters in columns to form an initial characteristic matrix; An independent component analysis algorithm is used to perform dimensionality reduction processing on the initial feature matrix to remove redundant features, thereby obtaining a first multi-dimensional feature matrix.

[0009] Preferably, the independent component analysis algorithm is used to perform dimensionality reduction processing on the initial feature matrix to remove redundant features to obtain a first multi-dimensional feature matrix, including: Centralizing the initial feature matrix so that the mean of each feature is zero; Calculating the covariance matrix of the initial characteristic matrix after the centralization process, and performing eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​and eigenvectors; Sorting by the size of the eigenvalues, selecting the first k eigenvectors whose cumulative contribution rate reaches a preset threshold, and forming a projection matrix; The initial feature matrix after the centralization process is multiplied by the projection matrix to obtain a multi-dimensional feature matrix after dimensionality reduction.

[0010] Preferably, the spectrum characteristic parameters include: performing truncation processing on the echo signal data to remove the noise segments at the beginning and end of the signal to obtain a valid signal segment; Adding a Hanning window to the valid signal segment to reduce spectrum leakage, thereby obtaining a windowed signal; Performing a fast Fourier transform on the windowed signal to convert the time domain signal into a frequency domain signal; The frequency corresponding to the maximum peak of the amplitude spectrum is extracted from the frequency domain signal as a peak frequency parameter, and the frequency range when the amplitude decays from the peak to half is calculated as an amplitude decay rate parameter.

[0011] Preferably, the collecting of real-time force displacement data of the titanium alloy wire to be tested during the tensile test and echo signal data returned by the ultrasonic testing device includes: The real-time force data and displacement data of the titanium alloy wire during the tensile test are synchronously collected by sensors, wherein the real-time force data is obtained by a tension and pressure sensor, and the displacement data is collected by a laser rangefinder; Controlling the ultrasonic detection device to emit ultrasonic pulses at a preset frequency to different radial cross-sections of the titanium alloy wire, and receiving echo signal data returned from each cross-section, wherein the preset frequency is dynamically adjusted according to the diameter specification of the titanium alloy wire; Timestamp tags are added to the real-time force data, displacement data, and echo signal data to establish a time-correlated index for multi-source data.

[0012] Preferably, the present invention further includes a titanium alloy wire performance detection system for implementing the above-mentioned titanium alloy wire performance detection method, the system comprising: An acquisition module is used to collect real-time force and displacement data of the titanium alloy wire to be tested during the tensile test and echo signal data returned by the ultrasonic detection device, and transmit the real-time force and displacement data and the echo signal data to a data processing unit; a preprocessing module, configured to perform denoising processing on the real-time force displacement data by using an adaptive median filtering algorithm through the data processing unit to obtain smoothed force displacement data, and simultaneously perform fast Fourier transform on the echo signal data to extract spectral characteristic parameters; a fusion module, configured to perform feature-layer fusion of the yield strength and tensile strength parameters in the smoothed force-displacement data with the peak frequency and amplitude attenuation rate parameters in the spectrum feature parameters to construct a first multi-dimensional feature matrix; A prediction module is used to input the first multi-dimensional feature matrix into a pre-trained deep belief network model, and predict the elongation and hardness values ​​of the titanium alloy wire through the deep belief network model to obtain a mechanical property prediction result; A report generation module is used to compare the mechanical property prediction results with a preset performance threshold range, mark the mechanical property prediction results as unqualified when they exceed the preset performance threshold range, and generate a titanium alloy wire performance test report containing unqualified items and corresponding parameters.

[0013] Preferably, the present invention also includes a computing device, which includes a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement a titanium alloy wire performance detection method as described above.

[0014] Compared with the prior art, the present invention has the following beneficial effects: By integrating real-time force-displacement data from tensile testing and echo signal data from ultrasonic testing, this method achieves effective fusion of multi-source information, overcoming the limitations of traditional single-source testing methods. Traditional testing methods often rely on only one type of data, either from tensile testing or ultrasonic testing, resulting in an incomplete assessment of material properties. This method, however, fuses both data types at the feature level to construct a first multi-dimensional feature matrix, which more comprehensively reflects the physical properties and internal structure of titanium alloy wire, providing a richer basis for subsequent performance predictions. During data processing, an adaptive median filter algorithm is used to denoise real-time force-displacement data. Compared to traditional filtering methods, this algorithm better preserves key features in the data while effectively eliminating noise interference, resulting in smoother and more accurate force-displacement data. This provides a reliable basis for extracting important parameters such as yield strength and tensile strength. Fast Fourier transform is performed on the echo signal data to extract the amplitude and phase spectra as spectral characteristic parameters. This allows for in-depth exploration of the internal structural information of the material contained in the ultrasonic signal, such as grain size and defect distribution. This information is closely related to the mechanical properties of the material, further enhancing the representativeness of the characteristic parameters. By integrating key mechanical parameters from smoothed force-displacement data with spectral characteristic parameters, a first multidimensional characteristic matrix is ​​constructed, leveraging the complementary strengths of different characteristic parameters. Parameters such as yield strength and tensile strength directly reflect the material's mechanical bearing capacity, while spectral parameters such as peak frequency and amplitude decay rate reflect the material's internal microstructural characteristics. The combination of these two provides a more comprehensive characteristic matrix, providing a more robust data foundation for subsequent performance predictions. The introduction of a pre-trained deep belief network model for mechanical property prediction fully leverages its strengths in handling nonlinear, high-dimensional data. Through multiple layers of nonlinear transformations, the deep belief network automatically learns the complex relationships within the feature matrix, offering higher prediction accuracy and stronger generalization capabilities than traditional machine learning models. This model, used to predict elongation and hardness values, enables rapid determination of key mechanical property parameters without destructive testing, reducing raw material waste and improving testing efficiency. During the testing process, by comparing the predicted mechanical property results with the preset threshold range, the quality of the titanium alloy wire can be quickly determined. A test report containing specific unqualified items and corresponding parameters is generated, providing a clear and intuitive reference for quality control during the production process. This method not only improves the traceability of test results, but also helps production personnel promptly identify potential problems in the production process, facilitate the adjustment and optimization of process parameters, and reduce the number of unqualified products. The entire testing process automates data collection, processing, analysis, and report generation, reducing manual intervention and the impact of human factors on test results, thereby improving the consistency and reliability of testing. Furthermore, this method is applicable to the testing of titanium alloy wires of different specifications and production batches. It is highly versatile and practical, meeting the quality testing needs of large-scale industrial production and helping to improve the overall quality level of titanium alloy wire production. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a working principle diagram of the titanium alloy wire performance testing method of the present invention; Figure 2 Flowchart for deep belief network model training and prediction; Figure 3 Flowchart for unsupervised pre-training of deep belief networks; Figure 4 Flowchart constructed for the first multidimensional feature matrix; Figure 5 Flowchart for echo signal spectrum feature extraction. DETAILED DESCRIPTION

[0016] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0017] See also Figure 1 The present invention provides a method, device and system for detecting the performance of titanium alloy wire, the method comprising: Step 1: Collect real-time force displacement data of the titanium alloy wire to be tested during the tensile test and echo signal data returned by the ultrasonic detection device, and transmit the real-time force displacement data and the echo signal data to a data processing unit.

[0018] Step 2: The data processing unit uses an adaptive median filtering algorithm to denoise the real-time force displacement data to obtain smooth force displacement data. At the same time, a fast Fourier transform is performed on the echo signal data to extract the amplitude spectrum and phase spectrum of the signal as spectral characteristic parameters.

[0019] Step 3: Perform feature layer fusion on the yield strength and tensile strength parameters in the smoothed force displacement data and the peak frequency and amplitude attenuation rate parameters in the spectrum feature parameters to construct a first multi-dimensional feature matrix.

[0020] Step 4: Input the first multi-dimensional feature matrix into a pre-trained deep belief network model, and use the deep belief network model to predict the elongation and hardness values ​​of the titanium alloy wire to obtain mechanical property prediction results.

[0021] Step 5: Compare the mechanical property prediction result with the preset performance threshold range. When the mechanical property prediction result exceeds the preset performance threshold range, mark it as unqualified, and generate a titanium alloy wire performance test report containing unqualified items and corresponding parameters.

[0022] Example 1: See Figure 2 , the first multi-dimensional feature matrix is ​​input into the pre-trained deep belief network model, and the elongation and hardness values ​​of the titanium alloy wire to be tested are predicted by the deep belief network model to obtain the mechanical property prediction results. The specific process is as follows: The multi-dimensional feature matrix and measured mechanical property data corresponding to historical titanium alloy wire samples were obtained from the database and then divided into a training set and a validation set. The partitioning process required careful consideration of the sample size and feature distribution to determine the appropriate partitioning ratio. After the partitioning was completed, the distribution of each feature parameter between the training and validation sets was checked to ensure consistency, avoiding bias in model training due to differences in data distribution.

[0023] The initial deep belief network model is pre-trained unsupervised using the training set, and features are extracted layer by layer by stacking restricted Boltzmann machines. The restricted Boltzmann machine, the fundamental unit of the deep belief network, consists of a visible layer and a hidden layer. Neurons within a layer are unconnected, while neurons between layers are fully connected. When training the first layer of the restricted Boltzmann machine, the multi-dimensional feature matrix from the training set is used as the input for the visible layer. The connection weights between the visible and hidden layers are adjusted through learning, enabling the hidden layer to capture the basic features of the input data. After training the first layer, the output of this hidden layer is used as the input for the visible layer of the second layer of the restricted Boltzmann machine, and the same training process is repeated to extract more abstract features. This method is repeated layer by layer until the preset number of layers is reached, completing the unsupervised pre-training. After the unsupervised pre-training is completed, the parameters of the initial deep belief network model are fine-tuned using the backpropagation algorithm. The backpropagation algorithm calculates the error between the model output and the actual value, transfers the error from the output layer to the input layer layer by layer, and adjusts the connection weights and bias values ​​between neurons in each layer according to the size of the error, so that the model's prediction results are closer to the actual value.

[0024] The performance of the fine-tuned deep belief network model is evaluated using the validation set, and the prediction error is calculated. The prediction error calculation requires the selection of an appropriate evaluation metric. By inputting the multi-dimensional feature matrix from the validation set into the model, the predicted mechanical properties data are obtained, which are then compared with the measured mechanical properties data from the validation set to calculate the error value. During the evaluation process, the error changes must be continuously monitored. When the prediction error is less than the preset error threshold, training is stopped, and the pre-trained deep belief network model is obtained. If the prediction error does not reach the preset error threshold, the model parameters, such as the learning rate and number of iterations, need to be readjusted, and training and evaluation are repeated until the error meets the requirements.

[0025] The first multi-dimensional feature matrix is ​​input into the pre-trained deep belief network model, and the model performs multi-layer processing on the input feature matrix. It can be understood that the first multi-dimensional feature matrix is ​​the multi-dimensional feature matrix corresponding to the titanium alloy wire to be tested. After the input layer receives the first multi-dimensional feature matrix, it passes the data to the first hidden layer. The hidden layer calculates the input data through a preset activation function, extracts features and passes them to the next hidden layer. Each hidden layer processes the data in turn, gradually extracting key features related to the elongation and hardness value of the titanium alloy wire. Finally, the output layer processes the data passed by the last hidden layer and outputs the predicted elongation and hardness value of the titanium alloy wire to be tested as the mechanical property prediction results. During the entire prediction process, the model will keep the parameters determined during training unchanged to ensure the stability and consistency of the prediction results.

[0026] During model training and prediction, attention must be paid to data integrity and accuracy. Historical sample data in the database must be rigorously screened and verified to remove missing or erroneous data to avoid negative impacts on model training. Furthermore, when dividing the training and validation sets, a random split should be used to avoid data bias caused by human factors. During unsupervised pre-training and parameter tuning, the number of training iterations should be appropriately set to prevent overfitting or underfitting of the model. Overfitting can cause the model to perform well on the training set but poorly on the validation set and new data. Underfitting prevents the model from fully learning the features in the data, resulting in inaccurate predictions. Furthermore, during performance evaluation, sufficient validation set data must be used to accurately reflect the model's generalization ability. Excessive validation set data may result in significant fluctuations in evaluation results, making it difficult to reliably assess model performance. Through the above process, a stable and reliable deep belief network model can be constructed for predicting the elongation and hardness of titanium alloy wire, providing effective support for performance testing of titanium alloy wire.

[0027] Example 2: See Figure 3, the initial deep belief network model is unsupervised pre-trained using the training set, features are extracted layer by layer through stacked restricted Boltzmann machines, and the parameters of the initial deep belief network model are fine-tuned using the backpropagation algorithm, as follows: Initialize the number of neurons in the input layer, hidden layer, and output layer of the initial deep belief network model. The number of neurons in the input layer must match the dimension of the first multi-dimensional feature matrix, that is, the number of neurons in the input layer must match the number of feature parameters contained in the feature matrix. The setting of the hidden layer must take into account the complexity of the features. Multiple hidden layers can be set, and the number of neurons in each layer can be adjusted according to the number of neurons in the previous layer. Usually, it gradually decreases or remains stable as the number of layers increases. The number of neurons in the output layer must be consistent with the number of mechanical performance parameters to be predicted. Since two parameters, elongation and hardness, need to be predicted, two neurons are set in the output layer. At the same time, determine the learning rate and number of iterations of the restricted Boltzmann machine. The value of the learning rate must take into account the training speed and stability, and the number of iterations is set according to the convergence of the model during training.

[0028] The multi-dimensional feature matrix from the training set is input to the first-layer restricted Boltzmann machine and trained using the contrastive divergence algorithm. The contrastive divergence algorithm operates as follows: First, the data in the multi-dimensional feature matrix is ​​used as the input to the visible layer. Based on the states of the visible layer neurons, the activation probabilities of the hidden layer neurons are calculated, thereby determining the states of the hidden layer. Then, the visible layer state is reconstructed based on the hidden layer state to obtain reconstructed data. By comparing the difference between the original input data and the reconstructed data, the connection weights and bias values ​​between the visible and hidden layers are adjusted. This process is repeated until the difference between the reconstructed data and the original input data reaches a stable state. At this point, the training of the first-layer restricted Boltzmann machine is completed, and the feature map of the first layer is obtained. The feature map of the first layer is used as the input to the second-layer restricted Boltzmann machine, which is trained using the same contrastive divergence algorithm. The connection weights and bias values ​​of this layer are adjusted to obtain the feature map of the second layer. The restricted Boltzmann machine is stacked layer by layer in the same manner until the preset number of layers are trained, and the unsupervised pre-training process ends.

[0029] The measured mechanical properties data of the training set are used as labels, and the mean square error between the predicted value and the label is calculated through the back-propagation algorithm. The connection weights and bias values ​​between the neurons in each layer are adjusted according to the mean square error to complete the parameter fine-tuning. The specific operation is as follows: the multi-dimensional feature matrix of the training set is input into the deep belief network model that has completed unsupervised pre-training. The model outputs the predicted mechanical properties data of each sample through calculation and processing of neurons in each layer. The predicted mechanical properties data are compared with the corresponding measured mechanical properties data, and the mean square error between the two is calculated, which is used as the loss function to measure the prediction effect of the model. The loss function is derived using the gradient descent method to obtain the gradient change direction of the connection weights and bias values ​​of each layer. According to the gradient direction, the connection weights and bias values ​​are updated according to the preset learning rate. After the update is completed, the multi-dimensional feature matrix of the training set is input into the model again to calculate the new loss function value. Repeat the above process of calculating the loss function, deriving, and updating the parameters until the value of the loss function converges to the preset minimum value range. During the parameter fine-tuning process, it is necessary to continuously pay attention to the changing trend of the loss function. If the loss function value no longer decreases significantly or fluctuates, the learning rate can be appropriately adjusted or the number of iterations can be increased to ensure that the parameter adjustment can make the model's prediction results closer to the measured values.

[0030] During the parameter fine-tuning process, attention should be paid to the coordination of parameter adjustments in each layer to avoid excessive adjustment of parameters in a certain layer affecting the feature extraction effects of other layers. At the same time, in order to prevent the model from overfitting the training set data during training, the validation set data can be regularly introduced for testing during training to observe the performance of the model on the validation set. If the loss function of the model on the training set continues to decrease, but the loss function on the validation set begins to rise, it means that the model is overfitting, and the current parameter adjustment needs to be stopped, and the complexity of the model should be appropriately reduced or regularization processing should be added. Through the above unsupervised pre-training and parameter fine-tuning process, the deep belief network model can fully learn the relationship between the features in the training set and the measured mechanical properties data, forming a stable prediction capability.

[0031] Example 3: See Figure 4 , the yield strength and tensile strength parameters in the smoothed force displacement data are fused with the peak frequency and amplitude attenuation rate parameters in the spectrum feature parameters at the feature layer to construct the first multi-dimensional feature matrix. The specific process is as follows: The yield strength and tensile strength parameters in the smoothed force-displacement data are normalized to obtain standardized mechanical parameters. The normalization process involves first calculating the maximum and minimum yield strength and tensile strength values ​​for all samples in the training set. Then, the yield strength and tensile strength of each sample are converted according to a unified rule, ensuring that the converted parameter values ​​fall within a specific range. This process allows yield strength and tensile strength parameters of different magnitudes to be compared on the same numerical scale, preventing the impact of differences in the magnitude of the original parameters on subsequent feature fusion and model training.

[0032] The peak frequency and amplitude decay parameters in the spectral feature parameters are normalized to obtain standardized spectral parameters. This standardization is similar to the normalization of yield strength and tensile strength. First, the maximum and minimum values ​​of the peak frequency and amplitude decay rate in the training set are determined. Then, the peak frequency and amplitude decay rate of each sample are converted to a numerical range consistent with the standardized mechanical parameters according to the same rules. This ensures that the standardized mechanical parameters and the standardized spectral parameters maintain a consistent numerical range, providing a unified foundation for subsequent feature splicing.

[0033] The standardized mechanical parameters and standardized spectral parameters are spliced ​​by column to form an initial characteristic matrix. During splicing, for each titanium alloy wire sample, its corresponding standardized yield strength, standardized tensile strength, standardized peak frequency, and standardized amplitude attenuation rate are arranged in order into a row of data, with each data point serving as an element in the matrix. A row of data from all samples is combined to form a two-dimensional matrix, the initial characteristic matrix. The number of rows in the initial characteristic matrix is ​​equal to the number of samples, and the number of columns is equal to the total number of standardized parameters. Each element represents the value of a sample on a certain standardized parameter.

[0034] The independent component analysis algorithm is used to reduce the dimensionality of the initial feature matrix and remove redundant features, resulting in an optimized multidimensional feature matrix. This optimized multidimensional feature matrix is ​​referred to as the first multidimensional feature matrix. The specific steps are: centering the initial feature matrix, calculating the average value of each feature across all samples, and then subtracting the corresponding average value from the eigenvalue of each sample to achieve a mean of zero for each feature. Centering eliminates numerical deviations in features, making comparisons between different features more targeted.

[0035] Calculate the covariance matrix of the initial feature matrix after centralization. Each element in the covariance matrix reflects the degree of linear correlation between two different features, with larger values ​​indicating stronger correlation. Perform eigenvalue decomposition on the covariance matrix to obtain multiple eigenvalues ​​and corresponding eigenvectors. The size of the eigenvalue indicates the amount of information contained in the corresponding eigenvector; larger eigenvalues ​​indicate richer information carried by the corresponding eigenvector.

[0036] The eigenvalues ​​are sorted from largest to smallest. The cumulative contribution ratio (the ratio of the cumulative sum of the first k eigenvalues ​​to the total sum of all eigenvalues) is then calculated. The first k eigenvectors whose cumulative contribution ratios reach a preset threshold are selected and combined to form a projection matrix. The threshold setting requires a comprehensive consideration of both feature integrity and dimensionality reduction efficiency. By selecting an appropriate k value, the key information from the initial features can be retained while reducing the dimensionality of the features.

[0037] The initial feature matrix after centralization is multiplied by the projection matrix to obtain a multi-dimensional feature matrix after dimensionality reduction, that is, an optimized multi-dimensional feature matrix, that is, a first multi-dimensional feature matrix.

[0038]

[0039] in, Represents the multi-dimensional feature matrix after dimensionality reduction, represents the initial feature matrix after centralization, Represents the projection matrix.

[0040] The reduced multidimensional feature matrix has the same number of rows as the initial feature matrix, and has k columns, which is the number of selected eigenvectors. This process removes redundant information from the initial feature matrix, reducing the feature dimension while retaining key information crucial for predicting titanium alloy wire properties. This optimized multidimensional feature matrix, when subsequently fed into a deep belief network model, reduces the computational complexity and improves the model's prediction speed.

[0041] It can be understood that when the multi-dimensional feature matrix corresponding to the historical titanium alloy wire samples is used to train the initial deep belief network model, the reduced-dimensional multi-dimensional feature matrix corresponding to the historical titanium alloy wire samples can also be obtained through the aforementioned steps, and the reduced-dimensional multi-dimensional feature matrix corresponding to the historical titanium alloy wire samples can be used to train the initial deep belief network model, thereby reducing the computational complexity of the model, improving the training efficiency of the model, and avoiding the risk of model overfitting due to redundant features.

[0042] Throughout the feature fusion and dimensionality reduction process, it is crucial to ensure that all processing steps are based on the parameters in the training set to ensure consistent processing rules. The feature parameters of newly collected titanium alloy wire samples (i.e., the titanium alloy wire to be tested) must be processed identically to the maximum and minimum values ​​and projection matrix determined in the training set. This ensures that the first multidimensional feature matrix of the new sample maintains the same format and scale as the feature matrix used during model training, ensuring that the model can accurately perform predictions.

[0043] Example 4: See Figure 5 , perform fast Fourier transform on the echo signal data, and extract the amplitude spectrum and phase spectrum of the signal as spectral characteristic parameters, as follows: The echo signal data is truncated to remove the noise segments at the beginning and end of the signal to obtain the valid signal segment. During the echo signal data acquisition process, the beginning and end of the signal often contain a lot of noise due to electromagnetic interference when the detection device is started and shut down. This noise will affect the accuracy of subsequent signal analysis. During processing, the time domain waveform of the echo signal is first observed. The beginning and end of the signal with amplitude significantly lower than the middle stable segment are identified. These segments are determined to be noise segments and removed. For example, when the time domain waveform of the echo signal shows rapid amplitude fluctuations at both ends and an overall low amplitude, while the middle portion has a stable amplitude with less fluctuation, the middle stable portion is retained as the valid signal segment.

[0044] A Hanning window is added to the valid signal segment to reduce spectral leakage. The valid signal segment is finite in the time domain. Direct Fourier transforms can cause spectral leakage due to abrupt signal truncation, leading to deviations in the frequency domain analysis results. A Hanning window is a window function that allows the signal to smoothly transition to zero at both ends. Multiplying it by the valid signal segment can smooth out the signal truncation in the time domain. The Hanning window length is determined based on the valid signal segment length, ensuring consistency. Each data point in the valid signal segment is then multiplied by the corresponding Hanning window function value to produce a windowed signal. After windowing, the signal amplitude at both ends gradually approaches zero, reducing the sidelobe interference in the frequency domain caused by abrupt signal truncation.

[0045] A fast Fourier transform is performed on the windowed signal to convert the time domain signal into a frequency domain signal. The time domain signal reflects how the signal amplitude changes over time, while the frequency domain signal reflects the distribution characteristics of the different frequency components in the signal. The fast Fourier transform uses mathematical operations to convert the windowed time domain discrete signal into a frequency domain signal, obtaining the amplitude and phase information corresponding to each frequency point. During the conversion process, the number of Fourier transform points is determined based on the length of the windowed signal to ensure that the transformed frequency domain signal can fully reflect the frequency characteristics of the original signal. After the transformation is completed, a three-dimensional frequency domain signal data containing frequency, amplitude, and phase is obtained, where frequency represents the different frequency components contained in the signal, amplitude represents the intensity of each frequency component, and phase represents the phase offset of each frequency component.

[0046] The frequency corresponding to the maximum peak of the amplitude spectrum is extracted from the frequency domain signal as the peak frequency parameter, and the frequency range where the amplitude decays from the peak to half is calculated as the amplitude decay rate parameter. The amplitude spectrum is a curve showing how the amplitude of a frequency domain signal changes with frequency. By observing the amplitude spectrum, the maximum peak point can be found, and the frequency corresponding to this peak point is the peak frequency parameter. For example, if the amplitude of a frequency point in the amplitude spectrum is significantly higher than that of other frequency points and is the maximum value in the entire spectrum, the value at this frequency point is the peak frequency parameter. To extract the amplitude decay rate parameter, after determining the maximum peak, find the two frequency points where the amplitude drops to half of the peak value: a low frequency point below the peak frequency and a high frequency point above the peak frequency. The difference between these two frequency points is the amplitude decay rate parameter. For example, if the peak frequency is 5 MHz, the low frequency point where the amplitude drops to half of the peak value is 4.8 MHz and the high frequency point is 5.3 MHz, the amplitude decay rate parameter is 0.5 MHz.

[0047] Throughout the entire processing process, attention must be paid to the relationship between the sampling frequency of the echo signal data and the frequency resolution of the frequency domain signal after Fourier transform. The higher the sampling frequency, the higher the frequency resolution of the frequency domain signal, and the finer the frequency components that can be distinguished, but this will also increase the workload of data processing; if the sampling frequency is too low, it may not be possible to accurately capture the high-frequency components in the signal, resulting in incomplete frequency feature extraction. In addition, when extracting the peak frequency and amplitude attenuation rate, it is necessary to exclude abnormal peak interference in the frequency domain signal. These abnormal peaks may be caused by sudden noise during the signal acquisition process and are not a reflection of the characteristics of the titanium alloy wire itself. By comparing the frequency domain signals of multiple titanium alloy wires of the same specifications, common peak features can be identified to distinguish between valid peaks and abnormal peaks, ensuring that the extracted spectral feature parameters can truly reflect the internal structural characteristics of the titanium alloy wire.

[0048] Example 5: Collecting real-time force displacement data of the titanium alloy wire to be tested during the tensile test and echo signal data returned by the ultrasonic testing device, as follows: Sensors simultaneously collect real-time force and displacement data from titanium alloy wire during tensile testing. The tensile test is conducted on a dedicated tensile testing machine. The ends of the titanium alloy wire are fixed to the machine's chucks. As the chuck moves, the wire is subjected to axial tension and deformation. Real-time force data is acquired using a tension and pressure sensor mounted on the chuck. This sensor, aligned with the force acting on the wire, senses the tensile force in real time and converts the mechanical signal into an electrical output. Displacement data is collected using a laser rangefinder, mounted on one side of the testing machine. Its laser emission direction is parallel to the wire's tensile direction. The laser beam is directed directly onto the wire surface near the chuck or onto a moving component connected to the chuck. The displacement change is calculated by measuring the time difference between laser emission and reflection. The sensor's data acquisition frequency is set according to the test requirements. A lower acquisition frequency is used during the elastic deformation phase of the wire, while a higher acquisition frequency is used near the yield point and fracture phase to capture critical mechanical details.

[0049] The ultrasonic testing device is controlled to emit ultrasonic pulses at a preset frequency at different radial cross-sections of the titanium alloy wire and receive echo signals from each cross-section. The ultrasonic testing device consists of an ultrasonic probe, a pulse generator, and a signal receiver. The ultrasonic probe is mounted on a robotic arm that can move along the length of the wire, enabling scanning and testing of multiple radial cross-sections of the wire during the tensile test. The preset frequency is determined based on the diameter of the titanium alloy wire. For smaller diameter wires, a higher ultrasonic frequency is used, resulting in a shorter wavelength of the ultrasonic pulse, which can detect smaller defects or structural changes within the wire. For larger diameter wires, the ultrasonic frequency is appropriately lowered to increase the propagation depth of the ultrasonic pulse within the wire, ensuring that features in the wire's center can be detected. During the testing process, the robotic arm moves the ultrasonic probe according to a preset step size. At each preset position, an ultrasonic pulse is emitted at the radial cross-section at that position. The pulse passes through the wire and is reflected by the opposite reflective surface, generating an echo signal. The echo signal is received by the ultrasonic probe, converted into an electrical signal, and transmitted to the signal receiver for preliminary processing.

[0050] Timestamp tags are added to real-time force data, displacement data, and echo signal data to establish a time-correlated index for multi-source data. Timestamp tags are generated by the system clock and are accurate to the millisecond level. Each data point is automatically timestamped by the system as it is collected. For real-time force and displacement data, since both are collected synchronously by the sensors, each time point corresponds to a set of force and displacement data. For echo signal data, since the ultrasonic detection device detects different radial sections at different times, each echo signal data point is timestamped at the time of collection. Timestamp tags allow force data, displacement data, and echo signal data at the same moment to be correlated. For example, when a wire is stretched to a certain displacement, the corresponding force magnitude and the echo signal characteristics of a specific radial section at that moment can form a complete set of correlated data. This time-correlated index enables subsequent data processing to accurately analyze the characteristic changes in the wire's internal structure under specific mechanical conditions, providing a temporal correspondence for the fusion of force and displacement data with echo signal data.

[0051] During the data collection process, it is necessary to ensure the synchronization between the various devices to avoid deviations in the timestamp labels due to device response delays. A hardware synchronization trigger mechanism can be used to enable the tension and pressure sensors, laser rangefinders, and ultrasonic detection devices to start data collection under the same trigger signal, reducing synchronization errors between devices. At the same time, the collected data is cached in real time to prevent data loss due to delayed data transmission speeds. For ultrasonic detection devices, their transmission frequency and probe position need to be calibrated regularly to ensure that the detection conditions of different radial sections are consistent and to reduce the impact of the device's own errors on the echo signal data. In addition, before the tensile test begins, the sensors and ultrasonic detection devices need to be preheated and calibrated to check whether the equipment is working properly and ensure that the collected data is accurate and reliable. Through the above data acquisition process, complete mechanical change data and internal structural characteristic data of the titanium alloy wire during the tensile test can be obtained, providing original data support for subsequent performance testing.

[0052] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0053] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for detecting the performance of titanium alloy wire, characterized in that: include: Collecting real-time force-displacement data of the titanium alloy wire to be tested during the tensile test and echo signal data returned by the ultrasonic testing device, and transmitting the real-time force-displacement data and the echo signal data to a data processing unit; The data processing unit performs denoising on the real-time force displacement data using an adaptive median filter algorithm to obtain smoothed force displacement data, and simultaneously performs fast Fourier transform on the echo signal data to extract the amplitude spectrum and phase spectrum of the signal as spectral characteristic parameters; Performing feature layer fusion on the yield strength and tensile strength parameters in the smoothed force displacement data and the peak frequency and amplitude attenuation rate parameters in the spectrum feature parameters to construct a first multi-dimensional feature matrix; Inputting the first multi-dimensional feature matrix into a pre-trained deep belief network model, and using the deep belief network model to predict the elongation and hardness values ​​of the titanium alloy wire to be tested to obtain a mechanical property prediction result; The mechanical property prediction result is compared with a preset performance threshold range. When the mechanical property prediction result exceeds the preset performance threshold range, it is marked as unqualified, and a titanium alloy wire performance test report containing unqualified items and corresponding parameters is generated.

2. The method for detecting the performance of a titanium alloy wire according to claim 1, wherein: The first multi-dimensional feature matrix is ​​input into a pre-trained deep belief network model, and the elongation and hardness values ​​of the titanium alloy wire are predicted by the deep belief network model to obtain a mechanical property prediction result, including: Obtaining a multi-dimensional feature matrix and measured mechanical property data corresponding to historical titanium alloy wire samples from a database, and dividing the historical titanium alloy wire samples into a training set and a validation set; Using the training set to perform unsupervised pre-training on the initial deep belief network model, extracting features layer by layer through stacked restricted Boltzmann machines, and fine-tuning the parameters of the initial deep belief network model using a backpropagation algorithm; Using the validation set to perform performance evaluation on the fine-tuned deep belief network model, calculating the prediction error, and stopping training when the prediction error is less than a preset error threshold to obtain the pre-trained deep belief network model; The first multi-dimensional feature matrix is ​​input into the pre-trained deep belief network model, and the elongation prediction value and hardness prediction value of the titanium alloy wire are output as the mechanical property prediction result.

3. The method for detecting the performance of a titanium alloy wire according to claim 2, wherein: The unsupervised pre-training of the initial deep belief network model using the training set, extracting features layer by layer by stacking restricted Boltzmann machines, and fine-tuning the parameters of the initial deep belief network model using a back propagation algorithm include: Initializing the number of neurons in the input layer, hidden layer, and output layer of the initial deep belief network model, and setting the learning rate and number of iterations of the restricted Boltzmann machine; Inputting the multi-dimensional feature matrix in the training set into the first layer of restricted Boltzmann machine, training it through the contrastive divergence algorithm to obtain the first layer feature map; The first layer of feature maps is used as the input of the second layer of restricted Boltzmann machine for training, and stacked to the preset number of layers in sequence to complete the unsupervised pre-training process; The measured mechanical property data of the training set is used as a label, and the mean square error between the predicted value and the label is calculated by the back propagation algorithm. The connection weights and bias values ​​between the neurons in each layer are adjusted according to the mean square error to complete parameter fine-tuning.

4. The method for detecting properties of titanium alloy wire according to claim 3, wherein: The measured mechanical property data of the training set is used as a label, the mean square error between the predicted value and the label is calculated by the back propagation algorithm, and the connection weights and bias values ​​between neurons in each layer are adjusted according to the mean square error to complete parameter fine-tuning, including: Inputting the multi-dimensional feature matrix of the training set into the deep belief network model that has completed unsupervised pre-training to obtain the predicted mechanical property data of each sample; Calculating the mean square error between the predicted mechanical property data and the measured mechanical property data, and using the mean square error as a loss function; The loss function is derived using the gradient descent method to obtain the gradient direction of the connection weights and bias values ​​of each layer; The connection weights and bias values ​​are updated at a preset learning rate according to the gradient direction, and iteration is repeated until the loss function converges to a preset minimum value range.

5. The method for detecting properties of titanium alloy wire according to claim 1, wherein: The yield strength and tensile strength parameters in the smoothed force displacement data are fused with the peak frequency and amplitude attenuation rate parameters in the spectrum feature parameters to construct a first multi-dimensional feature matrix, including: Normalizing the yield strength parameter and the tensile strength parameter in the smoothed force displacement data to obtain standardized mechanical parameters; Normalizing the peak frequency parameter and the amplitude attenuation rate parameter in the spectrum characteristic parameters to obtain standardized spectrum parameters; splicing the standardized mechanical parameters and the standardized spectral parameters in columns to form an initial characteristic matrix; An independent component analysis algorithm is used to perform dimensionality reduction processing on the initial feature matrix to remove redundant features, thereby obtaining the first multi-dimensional feature matrix.

6. The method for detecting properties of titanium alloy wire according to claim 5, characterized in that: The independent component analysis algorithm is used to perform dimensionality reduction processing on the initial feature matrix to remove redundant features to obtain the first multi-dimensional feature matrix, including: Centralizing the initial feature matrix so that the mean of each feature is zero; Calculating the covariance matrix of the initial feature matrix after centralization, and performing eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​and eigenvectors; Sorting by the size of the eigenvalues, selecting the first k eigenvectors whose cumulative contribution rate reaches a preset threshold, and forming a projection matrix; The initial feature matrix after centralization is multiplied by the projection matrix to obtain the first multi-dimensional feature matrix.

7. The method for detecting properties of titanium alloy wire according to claim 1, wherein: The performing fast Fourier transform on the echo signal data to extract the amplitude spectrum and phase spectrum of the signal as spectrum characteristic parameters includes: performing truncation processing on the echo signal data to remove the noise segments at the beginning and end of the signal to obtain a valid signal segment; Adding a Hanning window to the valid signal segment to reduce spectrum leakage, thereby obtaining a windowed signal; Performing a fast Fourier transform on the windowed signal to convert the time domain signal into a frequency domain signal; The frequency corresponding to the maximum peak of the amplitude spectrum is extracted from the frequency domain signal as a peak frequency parameter, and the frequency range when the amplitude decays from the peak to half is calculated as an amplitude decay rate parameter.

8. The method for detecting properties of titanium alloy wire according to claim 1, wherein: The collecting of real-time force displacement data of the titanium alloy wire to be tested during the tensile test and echo signal data returned by the ultrasonic testing device includes: The real-time force data and displacement data of the titanium alloy wire during the tensile test are synchronously collected by sensors, wherein the real-time force data is obtained by a tension and pressure sensor, and the displacement data is collected by a laser rangefinder; Controlling the ultrasonic detection device to emit ultrasonic pulses at a preset frequency to different radial cross-sections of the titanium alloy wire, and receiving echo signal data returned from each cross-section, wherein the preset frequency is dynamically adjusted according to the diameter specification of the titanium alloy wire; Timestamp tags are added to the real-time force data, displacement data, and echo signal data to establish a time-correlated index for multi-source data.

9. A titanium alloy wire performance detection system, used to implement a titanium alloy wire performance detection method according to any one of claims 1 to 8, characterized in that: include: An acquisition module is used to collect real-time force and displacement data of the titanium alloy wire to be tested during the tensile test and echo signal data returned by the ultrasonic detection device, and transmit the real-time force and displacement data and the echo signal data to a data processing unit; a preprocessing module, configured to perform denoising processing on the real-time force displacement data by using an adaptive median filtering algorithm through the data processing unit to obtain smoothed force displacement data, and simultaneously perform fast Fourier transform on the echo signal data to extract spectral characteristic parameters; a fusion module, configured to perform feature-layer fusion of the yield strength and tensile strength parameters in the smoothed force-displacement data with the peak frequency and amplitude attenuation rate parameters in the spectrum feature parameters to construct a first multi-dimensional feature matrix; A prediction module is used to input the first multi-dimensional feature matrix into a pre-trained deep belief network model, and predict the elongation and hardness values ​​of the titanium alloy wire through the deep belief network model to obtain a mechanical property prediction result; A report generation module is used to compare the mechanical property prediction results with a preset performance threshold range, mark the mechanical property prediction results as unqualified when they exceed the preset performance threshold range, and generate a titanium alloy wire performance test report containing unqualified items and corresponding parameters.

10. A computing device, characterized in that It comprises a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement a titanium alloy wire performance detection method as described in any one of claims 1 to 8.

Images