Sound emission amplitude time sequence prediction method during titanium alloy tension-torsion combined loading

Abstract

The invention provides an acoustic emission amplitude time sequence prediction method during titanium alloy tension-torsion combined loading. The method comprises the following steps: carrying out sound velocity calibration on a test piece by utilizing an acoustic emission system; performing a stretching-torsion combined test on the test piece according to a set loading scheme; collecting acoustic emission data of each test in real time; and based on the acquired acoustic emission data, constructing an acoustic emission amplitude prediction model at different tensile and torsional loading rates, and predicting time sequence data of the acoustic emission amplitude at the corresponding loading rate. According to the method, the dynamic time sequence rule of material damage can be accurately captured, a quantitative basis is provided for damage monitoring, service life evaluation and structural safety analysis of the Ti6Al7Nb titanium alloy, and the method can meet the acoustic emission amplitude prediction requirements of different metal materials and loading conditions.

Description

Technical Field

[0001] This invention relates to a method for predicting the timing of acoustic emission amplitude under combined tension and torsion loading on Ti6Al7Nb titanium alloy, belonging to the field of nondestructive testing of materials and acoustic emission signal analysis. Background Technology

[0002] Ti6Al7Nb titanium alloys are widely used in medical devices such as artificial joints and orthopedic implants due to their excellent mechanical properties, biocompatibility, and corrosion resistance. Implants typically endure multiaxial coupled loads during their service life in the human body, with tensile-torsional combined loads being the most common and representative of the actual stress state. Under such complex stress conditions, the material undergoes the entire process from micro-damage initiation and crack propagation to eventual fracture. The internal damage evolution exhibits significant temporal and uncertainties, making it difficult to achieve real-time monitoring and reliable prediction of damage development using traditional mechanical testing methods alone.

[0003] Existing research largely focuses on the mechanical property analysis of Ti6Al7Nb titanium alloy under uniaxial loading or the static interpretation of acoustic emission characteristics at a single loading rate. Research on the correlation between "loading rate-damage sequence-acoustic emission response" under combined tensile and torsional loading conditions is relatively scarce. Acoustic emission (AE) technology, as a core non-destructive testing method for real-time monitoring of material micro-damage evolution, has its core parameter—acoustic emission amplitude—which directly quantifies the energy release level of internal material failure and is a key indicator characterizing the degree of damage. However, current technology cannot accurately predict the acoustic emission amplitude under different tensile and torsional loading rates, and particularly lacks the ability to proactively predict dynamic damage processes.

[0004] The dynamic changes in tensile and torsional loading rates directly alter the material damage evolution path: low-speed loading results in slow damage accumulation, requiring a complete test to last tens of hours or even longer, leading to long test cycles and high equipment energy consumption; high-speed loading causes concentrated damage outbreaks, making signals prone to distortion. Furthermore, the coupling effect of different tensile and torsional rates causes acoustic emission amplitudes to exhibit complex temporal fluctuations. Establishing patterns solely through numerous repeated tests is not only costly but also fails to cover all operating conditions. More importantly, existing technologies largely rely on traditional statistical fitting or shallow machine learning, failing to fully explore the bidirectional temporal evolution patterns of acoustic emission signals and lacking technical solutions for constructing high-precision prediction models using deep learning techniques.

[0005] For critical scenarios such as medical implants, engineering practice requires both accurate damage warning data and a pressing need to reduce the time and economic costs of low-speed loading tests. Existing technologies, unable to balance "prediction accuracy" and "test efficiency," fail to meet practical application requirements. Therefore, there is an urgent need to develop a deep learning prediction method based on acoustic emission timing characteristics to dynamically predict the acoustic emission amplitude of Ti6Al7Nb titanium alloy under different tensile and torsional loading rates. In particular, this method can predict amplitude changes under long-term low-speed conditions using short-term test data, providing technical support for implant service safety assessment while addressing the pain points of low efficiency and high cost in traditional testing. Summary of the Invention

[0006] The technical problem this invention aims to solve is the lack of an effective method for predicting the acoustic emission amplitude of Ti6Al7Nb titanium alloy under combined tensile and torsional loading conditions in existing technologies. Furthermore, traditional tensile and torsional tests suffer from problems such as long cycles under low-speed loading conditions, high equipment energy consumption, and high costs for full-condition coverage. This invention overcomes these problems by providing a dynamic prediction method for acoustic emission amplitude based on tensile and torsional loading rate grouping and acoustic emission timing characteristics. The method utilizes limited combined tensile and torsional loading test data, combined with real-time acoustic emission monitoring technology and Bidirectional Long Short-Term Memory (BiLSTM) deep learning modeling, to construct an acoustic emission amplitude prediction model applicable to different coupled tensile and torsional rates. This model can achieve early warning and amplitude trend prediction of material micro-damage evolution, and can also predict acoustic emission amplitude changes under long-term low-speed loading conditions using short-term test data, balancing prediction accuracy and test efficiency. This provides core technical support for damage assessment in key scenarios such as Ti6Al7Nb titanium alloy medical implants.

[0007] This invention provides a method for predicting the timing of acoustic emission amplitude under combined tension and torsion loading on titanium alloys, comprising the following steps:

[0008] S1, use an acoustic emission system to calibrate the sound velocity of the processed tension-torsion test piece;

[0009] S2, Design a set of tensile-torsion combined deformation loading schemes, and conduct tensile-torsion combined tests on the specimens according to the set loading schemes;

[0010] S3, During the test, load, deformation and acoustic emission data for each test are collected in real time. The acoustic emission data includes time-series data of acoustic emission amplitude, acoustic emission energy and ring count.

[0011] S4. Based on the collected acoustic emission data, a deep learning model is used for time series prediction. The time series evolution characteristics of acoustic emission amplitude under different loading rates are learned, and an acoustic emission amplitude prediction model under different tensile and torsional loading rates is constructed to predict the time series data of acoustic emission amplitude under the corresponding loading rates.

[0012] This invention, based on acoustic emission amplitude as the most intuitive indicator of material damage under load, provides a method for predicting the temporal sequence of acoustic emission amplitude under combined tensile-torsional loading of Ti6Al7Nb titanium alloy using a deep learning model. Specifically, under combined tensile-torsional loading conditions, based on the acoustic emission temporal sequence characteristics corresponding to different tensile-torsional loading rates, a method is constructed using deep learning technology to build a dynamic prediction model of acoustic emission amplitude. This enables the prediction of the material's damage evolution process, providing technical support for the safety assessment of metallic materials and is applicable to damage early warning scenarios for Ti6Al7Nb titanium alloys in key fields such as medical implants.

[0013] The following is a further optimized technical solution of the present invention:

[0014] Preferably, an acoustic emission signal acquisition system is established to acquire the acoustic emission signals of the specimens and calculate the sound velocity value of each specimen. The acoustic emission signal acquisition system includes an amplifier, an acoustic emission instrument, a computer, and two acoustic emission piezoelectric ceramic sensors. The tension-torsion specimens are positioned linearly to determine the position coordinates of the acoustic emission source in one-dimensional space.

[0015] Preferably, the specific positioning method of the tension-torsion test specimen is as follows: a sensor is placed on each of the two ends of the tension-torsion test specimen, and the position of the sound velocity calibration point is 30mm away from the position of the sensor, that is, it is set at 30mm and 180mm from the left end of the specimen gauge, respectively.

[0016] Preferably, a tensile-torsional combined deformation loading test scheme is designed, wherein the loading rate of the test design includes at least 8 levels, and the load, deformation signal and acoustic emission signal of the tensile-torsional specimen are collected in real time during the test. The load includes axial force and torque, the deformation signal includes axial deformation and torsional angle, and the acoustic emission signal includes acoustic emission amplitude, acoustic emission energy and ringing count.

[0017] Preferably, the tensile-torsion test specimens are standard specimens made of the same material, processed in the same batch, and matched with the tensile-torsion loading chuck of the tensile-torsion electronic testing machine.

[0018] Preferably, the deep learning model is a bidirectional long short-term memory (BiLSTM) network model, which, through training on experimental data and verification, can efficiently predict the acoustic emission amplitude under different tensile and torsional loading rates.

[0019] This invention establishes a time-series prediction model for acoustic emission amplitude using a BiLSTM model. This model can predict the changes in acoustic emission amplitude of Ti6Al7Nb titanium alloy under different tensile and torsional loading rates based on acoustic emission signal data from experiments. Through training and validation with data, the model can effectively capture the dynamic evolution of micro-damage in the material from initiation to propagation, exhibiting high prediction accuracy. During validation, the effectiveness of the model is further verified by calculating correlation coefficients and other statistical indicators. Based on this model, the evolution of acoustic emission amplitude over time under different loading conditions can be accurately predicted, thus providing technical support for the service life assessment and structural safety analysis of implants.

[0020] Acoustic emission amplitude reflects the intensity of energy release during material damage and is the most intuitive indicator of damage. Therefore, this invention selects acoustic emission amplitude as the prediction object. Considering the strong temporal correlation of data in the temporal prediction of acoustic emission amplitude, a bidirectional long short-term memory (BiLSTM) network model is selected to obtain the predicted values ​​of acoustic emission amplitude at different times. The acoustic emission amplitude signal exhibits dynamic temporal characteristics with changes in material state—there is both a positive correlation of "early amplitude fluctuations influencing later changes" and an inverse correlation of "later peaks inferring earlier subtle trends." BiLSTM, as an extension of the long short-term memory (LSTM) network, has the core advantage of accurately capturing these bidirectional correlation features by integrating "positive temporal dependence" and "inverse temporal dependence," making it a preferred deep learning model for this task and providing support for accurate prediction of acoustic emission amplitude.

[0021] Traditional recurrent neural networks (RNNs) struggle with the long temporal sequences often found in acoustic emission data due to the "vanishing / exploding gradient" problem, easily losing crucial long-term temporal information. LSTM, however, effectively overcomes the limitations of long-term processing by dynamically preserving effective features of acoustic emission amplitude and forgetting environmental noise through a gating mechanism consisting of a forget gate, input gate, and output gate. Building on this, BiLSTM further utilizes parallel "forward LSTM" and "reverse LSTM": the forward LSTM mines the evolutionary patterns of acoustic emission amplitude in chronological order, while the reverse LSTM traces the source features of amplitude changes in reverse chronological order. The final representation formed by concatenating the output features of both covers the historical dependence and future correlation of acoustic emission amplitude, significantly improving the accuracy of temporal prediction. The gating mechanism of BiLSTM consists of a forget gate, input gate, cell state, and output gate.

[0022] In summary, the acoustic emission amplitude prediction method of the present invention utilizes the bidirectional time series analysis capability of the BiLSTM model, which can analyze the changing trend of acoustic emission amplitude in chronological order, and also trace the source characteristics of amplitude fluctuations backward, thereby fully covering the dynamic signal changes of the material from the appearance of micro-damage to crack propagation.

[0023] Preferably, the method for constructing a prediction model for acoustic emission amplitude under different tensile and torsional loading rates using a deep learning model for time-series prediction includes:

[0024] (1) Construct a BiLSTM model architecture. The BiLSTM model consists of a forward LSTM layer, a backward LSTM layer and a fully connected layer. The forward LSTM layer is used to mine the evolution law of acoustic emission amplitude in time sequence. The backward LSTM layer is used to trace the source features of amplitude change in reverse time sequence. The output features of the forward LSTM layer and the backward LSTM layer are concatenated and mapped to the acoustic emission amplitude prediction value through the fully connected layer.

[0025] (2) Define the LSTM gating mechanism, which includes a forgetting gate, an input gate, a cell state gate, and an output gate, respectively calculated using the following formulas:

[0026] Forgotten Gate: ,

[0027] in For the Gate of Oblivion It is the sigmoid activation function. Here is the forget gate weight matrix. The state was hidden in the previous moment. The timing input is the acoustic emission amplitude at the current moment. To offset the forget gate;

[0028] Input gate: Generate candidate cell states And calculate the information retention ratio. ,

[0029] in Candidate cell state, To control the proportion of new information retained, tanh is used to limit the range of candidate cell state values. , These are the candidate cell state weight matrix and the bias term, respectively. , These are the input gate weight matrix and the bias term, respectively;

[0030] Cell status update: ,

[0031] in For the renewal of cell state, Represents element-wise multiplication. This represents the cell state at the previous moment.

[0032] Output gate: The current state is hidden. ,

[0033] in For output gate, , These are the output gate weight matrix and the bias term, respectively. Hide the current state;

[0034] The final output of the BiLSTM model is the concatenation of the forward and backward hidden states: Then, the predicted acoustic emission amplitude is obtained through a fully connected layer: ,

[0035] in This is a concatenation of the forward and reverse hidden states. This is the hidden state of the forward LSTM. This is the hidden state of the inverse LSTM. This is the predicted value for acoustic emission amplitude;

[0036] (3) Use the coefficient of determination The mean absolute percentage error (MAPE) quantifies the model prediction accuracy:

[0037] The formula for the coefficient of determination is: ,

[0038] in These are experimental values. For predicted values, The average of the test values;

[0039] The formula for mean absolute percentage error is: ,

[0040] Where MAPE is the mean absolute percentage error, and n is the number of time-series data points for acoustic emission amplitude.

[0041] (4) Divide the experimental dataset into a training set and a test set. The training set is used for iterative optimization of model parameters, and the test set is used to evaluate the generalization ability of the model.

[0042] This invention presents a method for predicting tensile and torsional failure stresses in materials based on acoustic emission data. It requires only a small number of tensile and torsional tests to establish a model relating acoustic emission characteristics to loading rates, enabling accurate prediction of material failure stresses under different tensile and torsional loading rates. This method is highly flexible and versatile, adaptable to various types of metallic materials; it is also simple to operate, significantly reducing experimental costs and time. Through a data-driven prediction model, this invention avoids the drawbacks of traditional methods that rely on numerous complex experiments and calculations, significantly improving prediction efficiency and providing reliable technical support for material performance evaluation and optimization in engineering practice.

[0043] In summary, this invention can accurately capture the dynamic temporal pattern of material damage, providing a quantitative basis for damage monitoring, service life assessment and structural safety analysis of Ti6Al7Nb titanium alloys, and can be adapted to the acoustic emission amplitude prediction needs of different metal materials and loading conditions. Attached Figure Description

[0044] Figure 1 This is a schematic diagram illustrating the working principle of the material tensile-torsion testing system in this invention.

[0045] Figure 2 This is a schematic diagram illustrating the principle of sound velocity measurement for standard tensile-torsional specimens in this invention.

[0046] Figure 3 This is a dimensional diagram of the standard titanium alloy tensile-torsional specimen used in this invention.

[0047] Figure 4 This is a photograph of the physical specimen used in the tensile-torsion test of this invention.

[0048] Figure 5 This is a diagram of the tensile-torsional acoustic emission test of the sample in this invention.

[0049] Figure 6 This is a comparison chart of the predicted and measured acoustic emission amplitude curves based on the BiLSTM model in this invention.

[0050] Figure 7 These are the failure diagrams of titanium alloy specimens under eight different loading rates in this invention. Detailed Implementation

[0051] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings: This embodiment is implemented under the premise of the technical solution of the present invention, and provides detailed implementation methods and specific operation processes, but the protection scope of the present invention is not limited to the following embodiments.

[0052] A method for predicting the timing of acoustic emission amplitude under combined tension and torsion loading on titanium alloys includes the following steps:

[0053] S1, use an acoustic emission system to calibrate the sound velocity of the processed tension-torsion test piece;

[0054] S2, Design a set of tensile-torsional combined deformation loading schemes, and conduct tensile-torsional combined tests on the specimens according to the set loading scheme (i.e. loading rate);

[0055] S3, During the test, load, deformation and acoustic emission data for each test are collected in real time. The acoustic emission data includes time-series data such as acoustic emission amplitude, acoustic emission energy, and ringing count.

[0056] S4. Based on the collected acoustic emission data, a deep learning BiLSTM model is used for time series prediction. The time series evolution characteristics of acoustic emission amplitude under different loading rates are learned, and an acoustic emission amplitude prediction model under different tensile and torsional loading rates is constructed to predict the time series data of acoustic emission amplitude under the corresponding loading rates.

[0057] The tensile-torsion test specimens used in this invention are standard specimens made of the same material, processed in the same batch, and matched with the special clamps for tensile-torsion loading of the electronic tensile-torsion testing machine. When processing the standard specimens for the combined tensile-torsion test, the same metal material is... Figure 3 The specified dimensions are used to machine a set of specimens for tensile and torsion tests, with at least 8 specimens of the same material. See the physical specimens for details. Figure 4 .

[0058] This invention establishes a full-information acoustic emission signal acquisition system, which acquires acoustic emission signals from a specimen and calculates the velocity of sound in the specimen. The full-information acoustic emission signal acquisition system includes an amplifier, an acoustic emission instrument, a computer, and two acoustic emission piezoelectric ceramic sensors.

[0059] Based on the full-information acoustic emission signal acquisition system, an acoustic emission lead-breaking test was conducted on a group of dedicated tensile-torsion specimens made of the same material before the tensile-torsion test. The specific sound velocity value of each specimen was calculated and recorded in sequence. The specific operation process is as follows: First, two acoustic emission piezoelectric ceramic sensors were glued to the two end faces of the specimen with coupling agent. Then, an amplifier was connected with a signal line, and then an acoustic emission instrument was connected with a cable. The acoustic emission instrument transmitted the signal to the dedicated acoustic emission signal processing software on the computer. Then, lead was broken 5 times at 30mm (i.e., lead-breaking point 1) and 180mm (i.e., lead-breaking point 2) from the left end of the gauge length of the specimen. After 10 lead-breaking tests, the waveform file was played back. Finally, the sound velocity value of the specimen material was calculated and recorded using the sound velocity calibration function of the acoustic emission software. During the experiment, the specimen was positioned linearly to determine the position coordinates of the acoustic emission source in one-dimensional space. Specifically, a sensor was placed on each of the two ends of the tension-torsion specimen, with the distance between the sound velocity calibration point and the sensor position 30mm, i.e., set at 30mm and 180mm from the left end of the specimen's gauge length, respectively (see...). Figure 2 ).

[0060] This invention designs a tensile-torsional combined deformation loading test scheme for acoustic emission testing, the principle of which is as follows: Figure 1 As shown, the experiment is as follows Figure 5 As shown. Based on the tensile-torsional electronic material testing machine and its supporting signal acquisition system, a tensile-torsional combined deformation loading scheme was designed for a group of special tensile-torsional specimens processed from the same material. The specimens were subjected to tensile-torsional combined deformation tests at a set loading rate. At least eight different loading rates were designed, and tensile-torsional tests were conducted at the set loading rates. After the tests, the specimens showed... Figure 7 As shown in the figure. During the experiment, the following signals were collected in real time: mechanical signals such as axial force, axial deformation, torque, and torsional angle, as well as acoustic emission signals such as acoustic emission amplitude, acoustic emission energy, and ring count. These signals can reflect the damage evolution of the specimen during the loading process and provide necessary data support for subsequent acoustic emission amplitude prediction.

[0061] This invention establishes a time-series prediction model for acoustic emission amplitude and calculates the correlation coefficient. The specific method includes:

[0062] (1) Construct a BiLSTM model architecture. The BiLSTM model consists of a forward LSTM layer, a backward LSTM layer and a fully connected layer. The forward LSTM layer is used to mine the evolution law of acoustic emission amplitude in time sequence. The backward LSTM layer is used to trace the source features of amplitude change in reverse time sequence. The output features of the forward LSTM layer and the backward LSTM layer are concatenated and mapped to the acoustic emission amplitude prediction value through the fully connected layer.

[0063] (2) Define the LSTM gating mechanism, which includes a forgetting gate, an input gate, a cell state gate, and an output gate, respectively calculated using the following formulas:

[0064] Forgotten Gate The formula for calculating the information to be discarded from a cell state is as follows:

[0065] ,

[0066] in This is the sigmoid activation function (output range 0-1, used to control the information retention ratio). Here is the forget gate weight matrix. The state was hidden in the previous moment. The input is the timing sequence of the acoustic emission amplitude at the current moment, i.e., the time sequence of the acoustic emission amplitude. To offset the forget gate;

[0067] The input gate is responsible for filtering and storing new information, and generates candidate cell states using the following formulas. and control the retention ratio of new information ,

[0068] ,

[0069] ,

[0070] Where tanh is used to limit the range of values ​​for candidate cell states. This is the weight matrix for the candidate cell states. This is a bias term for the candidate cell state. Here is the weight matrix of the input gate. This is the bias term for the input gate;

[0071] Cell state update The result of combining the forget gate and the input gate is:

[0072] ,

[0073] in Represents element-wise multiplication. This represents the cell state at the previous moment.

[0074] Output gate The final decision on the hidden state at the current moment The calculation formula is as follows:

[0075] ,

[0076] ,

[0077] in Here is the weight matrix of the output gate. This is the bias term for the output gate. Hide the current state;

[0078] For the BiLSTM model, the final output is the concatenation of the forward and backward hidden states. ,Right now:

[0079] ,

[0080] In the formula This is the hidden state of the forward LSTM. The hidden state is obtained by inverting the LSTM, and finally mapped to the final emission amplitude prediction value through a fully connected layer. ,Right now:

[0081] ;

[0082] (3) Determine the reliability of the prediction model

[0083] The BiLSTM model uses the coefficient of determination. The mean absolute percentage error (MASE) quantifies the predictive accuracy of a research model, while the coefficient of determination measures the model's ability to explain data variation. Its calculation formula is as follows:

[0084] ,

[0085] in These are experimental values. For predicted values, The average of the test values. The closer the value is to 1, the better the fit.

[0086] Mean Absolute Percentage Error (MAPE) reflects relative error; the smaller the value, the higher the relative accuracy. Its calculation formula is:

[0087] ,

[0088] Where n is the number of time-series data points of acoustic emission amplitude, and the smaller the MAPE value, the higher the relative accuracy;

[0089] (4) Verify the generalization performance of the model

[0090] The experimental dataset was divided into a 60% training set and a 40% test set. The training set was used for iterative optimization of model parameters, and the test set was used to evaluate the model's generalization ability. If the value is close to 1 and the MAPE value is small, the model is reliable and can be used to predict the time series of acoustic emission amplitude under different tensile and torsional loading rates.

[0091] This invention employs a full-information acoustic emission signal acquisition system to simultaneously acquire acoustic emission time-series data of Ti6Al7Nb titanium alloy specimens during tensile-torsional tests. Based on eight pre-set combinations of different tensile-torsional loading rates, group loading tests are conducted on standard tensile-torsional specimens processed in the same batch, ensuring a one-to-one correspondence between each rate condition and acoustic emission data. Using the acoustic emission amplitude recorded during the test as the core feature parameter, after data preprocessing (denoising and normalization) and bidirectional time-series feature extraction, a BiLSTM-based acoustic emission amplitude prediction model is constructed. This model uses forward LSTM to mine the "history-current" evolution law of acoustic emission amplitude and reverse LSTM to trace the "current-future" source characteristics. By integrating bidirectional time-series dependencies, it achieves accurate prediction of the AE amplitude change trend with the loading process under different loading rates. It can effectively capture the acoustic emission response law of materials from crack initiation to propagation stages, and the prediction accuracy meets engineering requirements.

[0092] Compared to traditional static analysis methods, the BiLSTM model solves the gradient vanishing problem in long-term acoustic emission data processing through a gating mechanism (forget gate, input gate, output gate). Its bidirectional modeling characteristics are better suited to the dual temporal characteristics of acoustic emission amplitude, namely "forward evolution + reverse tracing." Compared to the full-condition repeated test scheme, the method of this invention only needs to train the model with 60% of the short-term acoustic emission data to predict the amplitude change of the last 40% under the same loading rate. Especially for low-speed conditions, it can shorten the test cycle and significantly reduce equipment energy consumption and labor costs. The model constructed by this invention has excellent generalization ability after being verified by multiple sets of rate conditions. It can provide quantitative basis for implant service life assessment, early identification of internal damage, and structural safety judgment, filling the technical gap in dynamic prediction of acoustic emission amplitude of Ti6Al7Nb titanium alloy under combined tension and torsion loading.

[0093] The prediction method of this invention is based on acoustic emission signal data. By collecting characteristic parameters such as acoustic emission amplitude under different tensile and torsional loading rates and combining them with a BiLSTM deep learning model, it can predict the change trend of material acoustic emission amplitude with loading time relatively well. This method breaks away from the reliance on traditional material mechanics theory calculations and directly constructs a correlation model between acoustic emission signal and loading rate through experimental data, effectively capturing the dynamic characteristics of material damage evolution and achieving high prediction accuracy (coefficient of determination R). 2 ≥0.85, Mean Absolute Percentage Error (MAPE) ≤0.03).

[0094] Meanwhile, for low-speed loading conditions, only short-term acoustic emission data needs to be collected to predict subsequent amplitude changes, shortening the test cycle by more than 50% and significantly reducing test costs and equipment energy consumption. Furthermore, this method possesses a universal prediction framework, adaptable to various metallic materials and loading conditions, providing a reliable basis for material damage monitoring, service life assessment, and structural safety analysis, thus driving the upgrade of acoustic emission technology from "post-event analysis" to "pre-event prediction."

[0095] Example 1

[0096] Medical-grade titanium alloy Ti6Al7Nb was selected as the test material. Before conducting the tensile-torsion test, the sound velocity of each specimen was calibrated. The specimens were numbered T1 to T8, and the corresponding sound velocity values ​​were measured. x Specific data are shown in Table 1. The sound velocity values ​​in Table 1 were then sequentially entered into the acoustic emission software. Subsequently, eight sets of tensile-torsional tests were conducted on the specimen at different tensile-torsional loading rates. Acoustic emission signals were simultaneously acquired during the tests. Specific loading rates and loading path schemes are detailed in Table 2. Where v is the torsional loading rate and γ is the axial tensile loading rate.

[0097] Table 1. Sound velocity values ​​of medical titanium alloy Ti6Al7Nb

[0098] Specimen <![CDATA[T1]]> <![CDATA[T2]]> <![CDATA[T3]]> <![CDATA[T4]]> <![CDATA[T5]]> <![CDATA[T6]]> <![CDATA[T7]]> <![CDATA[T8]]> <![CDATA[ x / (m·s -1 )]]> 4336 4352 3945 4095 4244 4773 4453 4624

[0099] Table 2. Tensile and Torsional Loading Rate Schemes and Sound Velocity Values ​​for Ti6Al7Nb

[0100] Specimen <![CDATA[v / (mm·min -1 )]]> <![CDATA[γ / (°·min -1 )]]> <![CDATA[ x / (m·s -1 )]]> <![CDATA[T1]]> 1.00 1.00 4336 <![CDATA[T2]]> 1.25 1.00 4352 <![CDATA[T3]]> 1.50 1.00 3945 <![CDATA[T4]]> 1.75 1.00 4095 <![CDATA[T5]]> 2.00 1.00 4244 <![CDATA[T6]]> 2.25 1.00 4773 <![CDATA[T7]]> 2.50 1.00 4453 <![CDATA[T8]]> 2.75 1.00 4624

[0101] According to the experimental scheme in Table 2, eight sets of experiments were conducted on Ti6Al7Nb titanium alloy at different tensile-torsional loading rates. Based on the acoustic emission data collected from the experiments and the BiLSTM model, program 1 in Matlab software was run to obtain the predicted acoustic emission amplitude values ​​corresponding to the eight sets of different tensile-torsional loading rates. The coefficient of determination (R²) was used to evaluate the prediction accuracy of the model. 2 The predicted acoustic emission amplitude and mean absolute percentage error (MAPE) are shown in Table 3. A comparison chart of the predicted and measured values ​​is plotted using the results, as detailed in the attached figure. Figure 6 As shown.

[0102] Table 3. Time-series prediction performance of the BiLSTM model for 8 groups of acoustic emission amplitude samples.

[0103]

[0104] As can be seen from the results in Table 3, the R of the test set... 2 The MAPE value is above 0.8, and the average MAPE value on the test set is 0.33%, indicating a very small relative error. This demonstrates that the model can efficiently fit the temporal correlation characteristics of acoustic emission amplitude. Temporal prediction of amplitude can provide technical support for early damage in medical Ti6Al7Nb implants and has positive reference value for optimizing their clinical service safety. Therefore, the reliability of the prediction model is high, indicating that the prediction method is feasible.

[0105] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any transformations or substitutions that can be conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

[0106] The MATLAB software fitting programming described in this invention consists of Program 1, as shown below:

[0107] Program 1: Calculate the undetermined parameters of the prediction model

[0108] warning off

[0109] close all

[0110] clear

[0111] clc

[0112] Import eight sets of data (single-column time series data), run eight times.

[0113] result = xlsread('D:\Desktop\1026.xlsx'); %% Ensures the Excel file has no header and contains only plain data.

[0114] %% Data Analysis

[0115] num_samples = length(result); % Number of samples

[0116] kim = 15; % Keep the delay step size at 15 to avoid changes in the comparison baseline.

[0117] zim = 1; % Single-step prediction remains unchanged

[0118] %% Split the dataset

[0119] for i = 1: num_samples - kim - zim + 1

[0120] res(i, :) = [reshape(result(i: i + kim - 1), 1, kim), result(i + kim+ zim - 1)];

[0121] end

[0122] %% Dataset Analysis

[0123] outdim = 1;

[0124] num_size = 0.6; % Keep the training set size at 0.6

[0125] num_train_s = round(num_size * num_samples);

[0126] f_ = size(res, 2) - outdim;

[0127] %% Split the dataset into training and test sets

[0128] P_train = res(1: num_train_s, 1: f_)';

[0129] T_train = res(1: num_train_s, f_ + 1: end)';

[0130] M = size(P_train, 2);

[0131] P_test = res(num_train_s + 1: end, 1: f_)';

[0132] T_test = res(num_train_s + 1: end, f_ + 1: end)';

[0133] N = size(P_test, 2);

[0134] %% Data normalization

[0135] [P_train, ps_input] = mapminmax(P_train, 0, 1);

[0136] P_test = mapminmax('apply', P_test, ps_input);

[0137] [t_train, ps_output] = mapminmax(T_train, 0, 1);

[0138] t_test = mapminmax('apply', T_test, ps_input);

[0139] %% Data Tile

[0140] P_train = double(reshape(P_train, f_, 1, 1, M));

[0141] P_test = double(reshape(P_test , f_, 1, 1, N));

[0142] t_train = t_train';

[0143] t_test = t_test';

[0144] %% Data format conversion

[0145] for i = 1 : M

[0146] p_train{i, 1} = P_train(:, :, 1, i);

[0147] end

[0148] for i = 1 : N

[0149] p_test{i, 1} = P_test( :, :, 1, i);

[0150] end

[0151] %% Create Model

[0152] layers = [

[0153] sequenceInputLayer(f_) % Creates the input layer

[0154] bilstmLayer(8, 'OutputMode', 'last') % BiLSTM layer

[0155] reluLayer % Relu activation layer

[0156] fullyConnectedLayer(1) % Fully Connected Layer

[0157] regressionLayer]; % Regression layer

[0158] %% Parameter Settings

[0159] options = trainingOptions('adam', ...

[0160] 'MaxEpochs', 200, ... % Training epochs

[0161] 'GradientThreshold', 1, ...

[0162] 'InitialLearnRate', 3e-3, ... % Initial learning rate

[0163] 'LearnRateSchedule', 'piecewise', ...

[0164] 'LearnRateDropPeriod', 150, ... % Learning rate decay timing

[0165] 'LearnRateDropFactor',0.1, ...

[0166] 'L2Regularization', 1e-3, ... % Regularization coefficients

[0167] 'ExecutionEnvironment', 'auto',...

[0168] 'Verbose', false, ...

[0169] 'Plots', 'training-progress');

[0170] %% Training Model

[0171] net = trainNetwork(p_train, t_train, layers, options);

[0172] %% Simulation Prediction

[0173] t_sim1 = predict(net, p_train);

[0174] t_sim2 = predict(net, p_test );

[0175] %% Data denormalization

[0176] T_sim1 = mapminmax('reverse', t_sim1, ps_output);

[0177] T_sim2 = mapminmax('reverse', t_sim2, ps_output);

[0178] %% Root Mean Square Error

[0179] error1 = sqrt(sum((T_sim1' - T_train).^2) . / M);

[0180] error2 = sqrt(sum((T_sim2' - T_test ).^2) . / N);

[0181] %% View network structure

[0182] analyzeNetwork(net)

[0183] %% Drawing

[0184] figure

[0185] plot(1: M, T_train, 'r-', 1: M, T_sim1, 'b-', 'LineWidth', 1)

[0186] legend('actual value', 'predicted value')

[0187] xlabel('predicted sample')

[0188] ylabel('prediction result')

[0189] string = {'Comparison of training set prediction results'; ['RMSE=' num2str(error1)]};

[0190] title(string)

[0191] xlim([1, M])

[0192] grid

[0193] figure

[0194] plot(1: N, T_test, 'r-', 1: N, T_sim2, 'b-', 'LineWidth', 1)

[0195] legend('actual value', 'predicted value')

[0196] xlabel('predicted sample')

[0197] ylabel('prediction result')

[0198] string = {'Comparison of test set prediction results'; ['RMSE=' num2str(error2)]};

[0199] title(string)

[0200] xlim([1, N])

[0201] grid

[0202] %% Calculation of relevant indicators

[0203] % R2

[0204] R1 = 1 - norm(T_train - T_sim1')^2 / norm(T_train - mean(T_train))^2;

[0205] R2 = 1 - norm(T_test - T_sim2')^2 / norm(T_test - mean(T_test ))^2;

[0206] disp(['R2 of the training set data is:', num2str(R1)])

[0207] disp(['R2 of the test set data is:', num2str(R2)])

[0208] % MAPE

[0209] mape1 = sum(abs((T_sim1' - T_train). / T_train)) . / M ;

[0210] mape2 = sum(abs((T_sim2' - T_test ). / T_test )) . / N ;

[0211] disp(['MAPE of training set data is:', num2str(mape1)])

[0212] disp(['MAPE of the test set data is:', num2str(mape2)])

[0213] %% Draw a scatter plot (format as before)

[0214] sz = 25;

[0215] c = 'b';

[0216] figure

[0217] scatter(T_train, T_sim1, sz, c)

[0218] hold on

[0219] plot(xlim, ylim, '--k')

[0220] xlabel('True value in training set');

[0221] ylabel('predicted value from training set');

[0222] xlim([min(T_train) max(T_train)])

[0223] ylim([min(T_sim1) max(T_sim1)])5

[0224] title('Training set predicted values ​​vs. training set actual values')

[0225] figure

[0226] scatter(T_test, T_sim2, sz, c)

[0227] hold on

[0228] plot(xlim, ylim, '--k')

[0229] xlabel('True value of test set');

[0230] ylabel('test set predicted value');

[0231] xlim([min(T_test) max(T_test)])

[0232] ylim([min(T_sim2) max(T_sim2)])

[0233] title('Test set predicted values ​​vs. test set actual values')

[0234] Export test set results to Excel

[0235] test_result = [1:N; T_test; T_sim2']';

[0236] xlswrite('D:\Desktop\Results.xlsx', {'Sample Points','True Values','Predicted Values'}, 'Test Set Data', 'A1');

[0237] xlswrite('D:\Desktop\Results.xlsx', test_result, 'Test Set Data', 'A2');

[0238] disp('Test set results have been exported to "D:\Desktop\Results.xlsx", please check!').

Claims

1. A method for predicting the timing of acoustic emission amplitude under combined tension and torsion loading on titanium alloys, characterized in that, Includes the following steps: S1, use an acoustic emission system to calibrate the sound velocity of the processed tension-torsion test piece; S2, Design a set of tensile-torsion combined deformation loading schemes, and conduct tensile-torsion combined tests on the specimens according to the set loading schemes; S3, During the test, load, deformation and acoustic emission data for each test are collected in real time. The acoustic emission data includes time-series data of acoustic emission amplitude, acoustic emission energy and ring count. S4. Based on the collected acoustic emission data, a deep learning model is used for time series prediction. The time series evolution characteristics of acoustic emission amplitude under different loading rates are learned, and an acoustic emission amplitude prediction model under different tensile and torsional loading rates is constructed to predict the time series data of acoustic emission amplitude under the corresponding loading rates.

2. The method for predicting the acoustic emission amplitude timing under combined tension and torsion loading of titanium alloys according to claim 1, characterized in that, An acoustic emission signal acquisition system is established to acquire acoustic emission signals from the specimens and calculate the sound velocity value of each specimen. The acoustic emission signal acquisition system includes an amplifier, an acoustic emission instrument, a computer, and two acoustic emission piezoelectric ceramic sensors. The tension-torsion specimens are positioned linearly to determine the position coordinates of the acoustic emission source in one-dimensional space.

3. The method for predicting the acoustic emission amplitude timing under combined tension and torsion loading of titanium alloys according to claim 2, characterized in that, The specific positioning method for the tension-torsion test specimen is as follows: a sensor is placed on each of the two ends of the tension-torsion test specimen, and the position of the sound velocity calibration point is 30mm away from the position of the sensor, that is, it is set at 30mm and 180mm from the left end of the specimen gauge, respectively.

4. The method for predicting the acoustic emission amplitude timing under combined tension and torsion loading of titanium alloys according to claim 1, characterized in that, Design a tensile-torsional combined deformation loading test scheme. The loading rate of the test design includes at least 8 levels. During the test, the load, deformation signal and acoustic emission signal of the tensile-torsional specimen are collected in real time. The load includes axial force and torque. The deformation signal includes axial deformation and torsional angle. The acoustic emission signal includes acoustic emission amplitude, energy and ring count.

5. The method for predicting the acoustic emission amplitude timing under combined tension-torsion loading of titanium alloys according to claim 1, characterized in that, The tensile-torsion test specimens are standard specimens made of the same material, processed in the same batch, and matched with the tensile-torsion loading chuck of the tensile-torsion electronic testing machine.

6. The method for predicting the acoustic emission amplitude timing under combined tension-torsion loading of titanium alloys according to claim 1, characterized in that, The deep learning model is a bidirectional long short-term memory network model.

7. The method for predicting the acoustic emission amplitude timing under combined tension-torsion loading of titanium alloys according to claim 1, characterized in that, Methods for constructing acoustic emission amplitude prediction models under different tensile and torsional loading rates using deep learning models for time-series prediction include: (1) Construct a BiLSTM model architecture. The BiLSTM model consists of a forward LSTM layer, a backward LSTM layer and a fully connected layer. The forward LSTM layer is used to mine the evolution law of acoustic emission amplitude in time sequence. The backward LSTM layer is used to trace the source features of amplitude change in reverse time sequence. The output features of the forward LSTM layer and the backward LSTM layer are concatenated and mapped to the acoustic emission amplitude prediction value through the fully connected layer. (2) Define the LSTM gating mechanism, which includes a forgetting gate, an input gate, a cell state gate, and an output gate, respectively calculated using the following formulas: Forgotten Gate: , in For the Gate of Oblivion It is the sigmoid activation function. Here is the forget gate weight matrix. The state was hidden in the previous moment. The timing input is the acoustic emission amplitude at the current moment. To offset the forget gate; Input gate: Generate candidate cell states And calculate the information retention ratio. , in Candidate cell state, To control the proportion of new information retained, tanh is used to limit the range of candidate cell state values. , These are the candidate cell state weight matrix and the bias term, respectively. , These are the input gate weight matrix and the bias term, respectively; Cell status update: , in For the renewal of cell state, Represents element-wise multiplication. This represents the cell state at the previous moment. Output gate: The current state is hidden. , in For output gate, , These are the output gate weight matrix and the bias term, respectively. Hide the current state; The final output of the BiLSTM model is the concatenation of the forward and backward hidden states: Then, the predicted acoustic emission amplitude is obtained through a fully connected layer: , in This is a concatenation of the forward and reverse hidden states. This is the hidden state of the forward LSTM. This is the hidden state of the inverse LSTM. This is the predicted value for acoustic emission amplitude; (3) Use the coefficient of determination And the mean absolute percentage error quantifies the model prediction accuracy: The formula for the coefficient of determination is: , in These are experimental values. For predicted values, The average of the test values; The formula for mean absolute percentage error is: , Where MAPE is the mean absolute percentage error, and n is the number of time-series data points for acoustic emission amplitude. (4) Divide the experimental dataset into a training set and a test set. The training set is used for iterative optimization of model parameters, and the test set is used to evaluate the generalization ability of the model.

Images