Neural network prediction method and device for corrosion rate of titanium alloy in fluoride ion environment

Abstract

The present invention provides a neural network prediction method and device for the corrosion rate of titanium alloy in a fluoride ion environment, and relates to the technical field of metal material corrosion rate prediction. The method comprises: based on dimensional analysis, selecting the square value of corrosion current density as an intermediate variable, fitting the "time-square value of corrosion current density" relationship; using the complex trapezoidal quadrature formula to numerically integrate the square value of corrosion current density along time; fitting the "square integral value of corrosion current density-yield strength degradation amount" relationship, and obtaining the degradation curve of the yield strength of titanium alloy with time for reference in engineering service; based on the obtained yield strength degradation curve, using a differential matrix to scan the curve to obtain the relative instantaneous corrosion rate v m The prediction model provided by the present invention has good performance and excellent fitting effect, and can be used as a monitoring method for the service safety of titanium alloy materials for submersibles under fluoride ion corrosion conditions in deep sea environments, supplementing the existing prediction method for the corrosion rate of titanium alloy materials for submersibles.

Description

Technical Field

[0001] The invention relates to the technical field of metal material corrosion rate prediction, and in particular to a neural network prediction method and device for the corrosion rate of a titanium alloy in a fluoride ion environment. Background Art

[0002] Cybenko G proved the universal approximation theorem of neural networks in 1989, that is, a single hidden layer feedforward neural network with a sufficient number of neurons can arbitrarily approximate a function. The rise of deep learning shows that compared with a single hidden layer neural network, a relatively large number of neurons may be required to approximate the target function with high precision. By increasing the number of hidden layers of the neural network, that is, deepening the depth of the neural network, the approximation ability of the neural network can be effectively improved by using relatively fewer neurons. Unlike a multi-layer feedforward network, a radial basis neural network has only one hidden layer, and the net input of the hidden layer transfer function is the Euclidean distance between the input vector and the weight matrix, rather than the weighted sum of the weight matrix to the input vector in the feedforward network. A generalized regression neural network is a special radial basis network, in which the number of neurons in the hidden layer is consistent with the total number of input samples, and its linear output layer does not have a bias value.

[0003] However, in the prior art, there are few experimental data on the corrosion of titanium alloy materials used in submersibles under fluoride ion conditions in deep-sea environments, the existing simulated corrosion test cycles are short, and there is a lack of corrosion rate models that effectively fit the experimental results and can predict them in the long term. Summary of the invention

[0004] In view of the problems in the prior art that there are few corrosion experimental data, the existing simulated corrosion experimental cycle is short, and there is a lack of an effective fitting and long-term predictive corrosion rate model for the experimental results, the present invention proposes a neural network prediction method and device for the corrosion rate of titanium alloy in a fluoride ion environment.

[0005] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0006] On the one hand, a neural network prediction method for the corrosion rate of titanium alloy in a fluoride ion environment is provided, which is applied to electronic equipment and includes:

[0007] S1: Collecting titanium alloy corrosion experimental data; preprocessing the experimental data;

[0008] S2: By combining multiple neural networks, the preprocessed data is piecewise fitted; the neural network is trained;

[0009] S3: Arrange the fitting results of the data in the neural network to obtain the yield strength degradation curve of the titanium alloy;

[0010] S4: Based on the yield strength degradation curve, the yield strength degradation curve is scanned using a difference matrix to obtain a relative instantaneous corrosion rate v m , completed the neural network prediction of titanium alloy corrosion rate in fluoride ion environment.

[0011] Optionally, in step S1, titanium alloy corrosion experimental data is collected and preprocessed, including:

[0012] S11: The titanium alloy specimens were divided into two groups, and were placed in immersion solutions with different fluoride ion concentrations for immersion corrosion. Each group was divided into seven time gradients. The same time gradient of the two groups of titanium alloy specimens was a parallel experimental group.

[0013] S12: measuring electrochemical parameters of the titanium alloy sample after each gradient immersion corrosion;

[0014] S13: Calculate the arithmetic mean of the data of the parallel experimental groups; if the corrosion current data is calculated, it is also necessary to calculate the corrosion current density based on the specimen size and further calculate its square value; for the results obtained from the above operations, perform data cleaning.

[0015] Optionally, in S2, multiple neural networks are combined to perform piecewise fitting on the preprocessed data and train the neural network, including:

[0016] S21: By combining and using a variety of neural networks, inputting the pre-processed experimental data of each time gradient, training the neural network, and fitting the time-corrosion current density square value in segments; in the time period with insufficient data, using traditional mathematical methods for pre-fitting to achieve data enhancement, and then using it for neural network training;

[0017] S22: Predict the length of time the user expects Discrete Input it into the trained neural network and output the corresponding corrosion current density square value sequence Numerical integration of the square value sequence of corrosion current density is performed to obtain the square integral value sequence of corrosion current density

[0018] S23: The neural network is trained by combining multiple neural networks; the data is segmentedly fitted with "corrosion current density square integral-yield strength degradation"; in the time period when the data is insufficient, the traditional mathematical method is used for pre-fitting to achieve data enhancement, and then used for neural network training.

[0019] Optionally, data segmentation includes: segmenting the titanium alloy sample data with different fluoride ion concentrations, wherein 0-21 days is the first segment, and after 21 days is the second segment.

[0020] Optionally, in step S21, a plurality of neural networks are combined and used, and the pre-processed experimental data of each time gradient is input to train the neural network, including:

[0021] For the first segment of the two concentration data, BP neural network was selected, the maximum number of training cycles epochs = 8000, and the iteration termination condition was the maximum absolute error MAE of the training sample set <10 -15 , Among them is is the target output value, is the actual output value, k is the sample number;

[0022] For the second segment of data in the first group, a feedforward neural network was selected, and the default algorithm was the Levenberg-Marquardt algorithm;

[0023] For the second segment of data in the second group, a generalized regression neural network was selected.

[0024] Optionally, data enhancement includes: performing data enhancement on the second segment of data, namely supplementing the data to a time period twice the actual number of experimental days through a pre-fitting formula.

[0025] Optionally, in step S21, the user's expected prediction time length is Discrete Input it into the trained neural network and output the corresponding corrosion current density square value sequence include:

[0026] S211: Discrete time series Input into the trained BP network to get the output Only keep the first 21 output results.

[0027] S212: Discrete time series Input into the trained generalized regression neural network net_Time_Jcor2_F8 to get the output

[0028] S213: merging the above two output sequences to obtain a first group of concentration corrosion current density square value sequences;

[0029] S214: Repeat the above steps S211-S213 to obtain a second set of concentrations of the corrosion current density square value sequence.

[0030] Optionally, in S21, numerically integrating the corrosion current density square value sequence includes: numerically integrating the corrosion current density square value along time by using a complex trapezoidal quadrature formula.

[0031] Optionally, in S4, based on the yield strength degradation curve, the yield strength degradation curve is scanned using a difference matrix to obtain a relative instantaneous corrosion rate v m , complete the neural network prediction of titanium alloy corrosion rate in fluoride ion environment, including:

[0032] S41: Single-day degradation of the yield strength of titanium alloy on any day - δσ m As the relative instantaneous corrosion rate v of that day m ,in, δT=1;

[0033] S42: Use the difference matrix to multiply the degradation curve data sequence obtained in step S3 Get the relative instantaneous corrosion rate series Right now Complete the neural network prediction of titanium alloy corrosion rate in fluoride ion environment.

[0034] On the one hand, a neural network prediction device for the corrosion rate of titanium alloy in a fluoride ion environment is provided, and the device is applied to electronic equipment, and the device comprises:

[0035] A data preprocessing module is used to collect titanium alloy corrosion experimental data and preprocess the experimental data;

[0036] A data fitting module is used to perform piecewise fitting on the preprocessed data by combining multiple neural networks;

[0037] The curve integration module organizes the fitting results of the data in the neural network to obtain the yield strength degradation curve of the titanium alloy;

[0038] The corrosion rate prediction module uses a difference matrix to scan the yield strength degradation curve based on the yield strength degradation curve to obtain the relative instantaneous corrosion rate v m , completed the neural network prediction of titanium alloy corrosion rate in fluoride ion environment.

[0039] On the one hand, an electronic device is provided, comprising a processor and a memory, wherein the memory stores at least one instruction, and the at least one instruction is loaded and executed by the processor to implement the neural network prediction method for the corrosion rate of titanium alloy in the fluoride ion environment.

[0040] On the one hand, a computer-readable storage medium is provided, wherein at least one instruction is stored in the storage medium, and the at least one instruction is loaded and executed by a processor to implement the neural network prediction method for the corrosion rate of titanium alloy in the fluoride ion environment.

[0041] The above technical solution of the embodiment of the present invention has at least the following beneficial effects:

[0042] In the above scheme, the prediction model provided by the present invention has good performance and excellent fitting effect. It can be used as a monitoring method for the service safety of titanium alloy materials for submersibles under fluoride ion corrosion conditions in deep-sea environments, and supplements the existing prediction method for the corrosion rate of titanium alloy materials for submersibles. Based on dimensional analysis, the square value of corrosion current density is selected as the intermediate variable, and the "time-corrosion current density square value" relationship is first fitted, and then the complex trapezoidal quadrature formula is used to numerically integrate the square value of corrosion current density along time, and then the "corrosion current density square integral value-yield strength degradation amount" relationship is fitted to obtain the degradation curve of titanium alloy yield strength over time for reference in engineering service. On the basis of the obtained yield strength degradation curve, the curve is scanned using a differential matrix to finally obtain the relative instantaneous corrosion rate of titanium alloy in a fluoride ion environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0044] Figure 1 It is a flow chart of a neural network prediction method for the corrosion rate of titanium alloy in a fluoride ion environment provided by an embodiment of the present invention;

[0045] Figure 2 It is a flow chart of a neural network prediction method for the corrosion rate of titanium alloy in a fluoride ion environment provided by an embodiment of the present invention;

[0046] Figure 3 It is a diagram of the dimensions of experimental samples of a neural network prediction method for the corrosion rate of titanium alloy in a fluoride ion environment provided by an embodiment of the present invention;

[0047] Figure 4 This is a fitting effect diagram of the "time-corrosion current density square value" model under 4mmol / L fluoride ion conditions of a neural network prediction method for the corrosion rate of titanium alloy in a fluoride ion environment provided by an embodiment of the present invention;

[0048] Figure 5 This is a model fitting effect diagram of "corrosion current density square integral value-yield strength" from 0 to 21 days under 4mmol / L fluoride ion conditions of a neural network prediction method for the corrosion rate of titanium alloy in a fluoride ion environment provided by an embodiment of the present invention;

[0049] Figure 6This is a model fitting effect diagram of "corrosion current density square integral value-yield strength" under 4mmol / L fluoride ion conditions for 21 to 600 days of a neural network prediction method for the corrosion rate of titanium alloy in a fluoride ion environment provided by an embodiment of the present invention;

[0050] Figure 7 This is a fitting effect diagram of the "time-corrosion current density square value" model under 8mmol / L fluoride ion conditions of a neural network prediction method for the corrosion rate of titanium alloy in a fluoride ion environment provided by an embodiment of the present invention;

[0051] Figure 8 This is a model fitting effect diagram of "corrosion current density square integral value-yield strength" from 0 to 28 days under 8mmol / L fluoride ion conditions of a neural network prediction method for titanium alloy corrosion rate in a fluoride ion environment provided by an embodiment of the present invention;

[0052] Fig. 9 This is a model fitting effect diagram of "corrosion current density square integral value-yield strength" under 8mmol / L fluoride ion conditions for 28 to 600 days of a neural network prediction method for the corrosion rate of titanium alloy in a fluoride ion environment provided by an embodiment of the present invention;

[0053] Fig.10 A graph of relative instantaneous corrosion rates from 21 to 200 days under 4 mmol / L fluoride ion conditions according to a neural network prediction method for the corrosion rate of a titanium alloy in a fluoride ion environment provided by an embodiment of the present invention;

[0054] Fig.11 It is a device block diagram of a neural network prediction device for titanium alloy corrosion rate in a fluoride ion environment provided by an embodiment of the present invention;

[0055] Fig.12 It is a structural schematic diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0056] In order to make the technical problems, technical solutions and advantages to be solved by the present invention more clear, a detailed description will be given below with reference to the accompanying drawings and specific embodiments.

[0057] The embodiment of the present invention provides a neural network prediction method for the corrosion rate of titanium alloy in a fluoride ion environment, which can be implemented by an electronic device, which can be a terminal or a server. Figure 1 The neural network prediction method flow chart of the corrosion rate of titanium alloy in fluoride ion environment shown in FIG. 1 includes:

[0058] S101: Collecting titanium alloy corrosion experimental data; preprocessing the experimental data;

[0059] S102: performing segmented fitting on the preprocessed data by combining multiple neural networks; and training the neural network;

[0060] S103: sorting out the fitting results of the data in the neural network to obtain a yield strength degradation curve of the titanium alloy;

[0061] S104: Based on the yield strength degradation curve, use a difference matrix to scan the yield strength degradation curve to obtain a relative instantaneous corrosion rate v m , completed the neural network prediction of titanium alloy corrosion rate in fluoride ion environment.

[0062] Optionally, in step S101, titanium alloy corrosion experimental data is collected and preprocessed, including:

[0063] S111: Divide the titanium alloy specimens into two groups, and place them in immersion solutions with different fluoride ion concentrations for immersion corrosion. Each group is divided into seven time gradients. The same time gradient of the two groups of titanium alloy specimens is a parallel experimental group.

[0064] S112: measuring electrochemical parameters of the titanium alloy sample after each gradient immersion corrosion;

[0065] S113: Calculate the arithmetic mean of the data of the parallel experimental groups; if the corrosion current data is calculated, it is also necessary to calculate the corrosion current density based on the specimen size and further calculate its square value; for the results obtained from the above operations, perform data cleaning.

[0066] Optionally, in step S102, multiple neural networks are combined to perform piecewise fitting on the preprocessed data and train the neural network, including:

[0067] S121: By combining and using a variety of neural networks, inputting the pre-processed experimental data of each time gradient, training the neural network, and fitting the time-corrosion current density square value in segments; in the time period with insufficient data, using traditional mathematical methods for pre-fitting to achieve data enhancement, and then using it for neural network training;

[0068] S122: Predict the time length expected by the user Discrete Input it into the trained neural network and output the corresponding corrosion current density square value sequence Numerical integration of the square value sequence of corrosion current density is performed to obtain the square integral value sequence of corrosion current density

[0069] S123: The neural network is trained by combining multiple neural networks; the data is segmentedly fitted with "corrosion current density square integral-yield strength degradation"; in the time period when the data is insufficient, the traditional mathematical method is used for pre-fitting to achieve data enhancement, and then used for neural network training.

[0070] Optionally, data segmentation includes: segmenting the titanium alloy sample data with different fluoride ion concentrations, wherein 0-21 days is the first segment, and after 21 days is the second segment.

[0071] Optionally, in step S121, a plurality of neural networks are combined and used, and the pre-processed experimental data of each time gradient is input to train the neural network, including:

[0072] For the first segment of the two concentration data, BP neural network was selected, the maximum number of training cycles epochs = 8000, and the iteration termination condition was the maximum absolute error MAE of the training sample set <10 -15 , Among them is is the target output value, is the actual output value, k is the sample number;

[0073] For the second segment of data in the first group, a feedforward neural network was selected, and the default algorithm was the Levenberg-Marquardt algorithm;

[0074] For the second segment of data in the second group, a generalized regression neural network was selected.

[0075] Optionally, data enhancement includes: performing data enhancement on the second segment of data, namely supplementing the data to a time period twice the actual number of experimental days through a pre-fitting formula.

[0076] Optionally, in step S121, the time length that the user expects to be predicted Discrete Input it into the trained neural network and output the corresponding corrosion current density square value sequence include:

[0077] S1211: Discrete time series Input into the trained BP network to get the output Only keep the first 21 output results.

[0078] S1212: Discrete time series Input into the trained generalized regression neural network net_Time_Jcor2_F8 to get the output

[0079] S1213: merging the above two output sequences to obtain a first group of concentration corrosion current density square value sequences;

[0080] S1214: Repeat the above steps S1211-S1213 to obtain a second set of concentration sequence of square values ​​of corrosion current density.

[0081] Optionally, in S121, numerically integrating the corrosion current density square value sequence includes: numerically integrating the corrosion current density square value along time by using a complex trapezoidal quadrature formula.

[0082] Optionally, in step S104, based on the yield strength degradation curve, the yield strength degradation curve is scanned using a difference matrix to obtain a relative instantaneous corrosion rate v m , complete the neural network prediction of titanium alloy corrosion rate in fluoride ion environment, including:

[0083] S141: Single-day degradation of the yield strength of titanium alloy on any day - δσ m As the relative instantaneous corrosion rate v of that day m ,in, δT=1;

[0084] S142: Use the difference matrix to multiply the degradation curve and data sequence obtained in step S103 Get the relative instantaneous corrosion rate series Right now Complete the neural network prediction of titanium alloy corrosion rate in fluoride ion environment.

[0085] The prediction model provided by the present invention has good performance and excellent fitting effect. It can be used as a monitoring method for the service safety of titanium alloy materials for submersibles under fluoride ion corrosion conditions in deep sea environments, and supplements the existing prediction methods for the corrosion rate of titanium alloy materials for submersibles.

[0086] The embodiment of the present invention provides a neural network prediction method for the corrosion rate of titanium alloy in a fluoride ion environment, which can be implemented by an electronic device, which can be a terminal or a server. Figure 2 The neural network prediction method flow chart of the corrosion rate of titanium alloy in fluoride ion environment shown in FIG. 1 includes:

[0087] S201: Divide the titanium alloy specimens into two groups, and place them in immersion solutions with different fluoride ion concentrations for immersion corrosion. Each group is divided into seven time gradients. The same time gradient of the two groups of titanium alloy specimens is a parallel experimental group.

[0088] S202: measuring electrochemical parameters of the titanium alloy sample after each gradient immersion corrosion;

[0089] In a feasible implementation manner, the titanium alloy corrosion test method provided by the present invention is specifically as follows: the titanium alloy samples are divided into two groups, 4mmol·L -1 and 8mmol·L -1 Fluoride ion concentration immersion corrosion, each group is divided into 7 time gradient immersion corrosion, with 1 week as the unit, starting at 0 week; add one-fifth of fresh solution to the original electrolyte solution every 40 hours, and heat the immersion corrosion water bath to a constant temperature of 65 degrees Celsius; data acquisition: after each gradient immersion corrosion is completed, the electrochemical parameters are measured. Figure 3 The diagram shows the dimensions of the experimental specimen.

[0090] In a feasible implementation, the titanium alloy samples are divided into two groups, 4 mmol·L -1 and 8mmol·L -1 Fluoride ion concentration immersion corrosion, each group is divided into 7 time gradient immersion corrosions, with 1 week as the unit, starting from 0 week;

[0091] Every 40 hours, add one fifth of the original electrolyte solution to the solution: 0.02L of 3.5% NaCl + 4mmol·L -1 or 0.02L of 3.5% NaCl + 8mmol·L -1 , immerse in a corrosion water bath and heat to a constant temperature of 65 degrees Celsius;

[0092] After the locking time is reached, take out the sample, rinse the immersion solution on the surface of the sample with deionized water, then blow dry with a hair dryer, and use a water bath to maintain a constant temperature of 65 degrees Celsius to measure the electrochemical parameters: self-corrosion current, self-corrosion potential and open circuit, and keep records.

[0093] S203: Calculate the arithmetic mean of the data of the parallel experimental groups; if the corrosion current data is calculated, it is also necessary to calculate the corrosion current density in combination with the specimen size and further calculate its square value; for the results obtained from the above operations, perform data cleaning.

[0094] In a feasible implementation manner, the data preprocessing method provided by the present invention includes: calculating the arithmetic mean of the data of the parallel experimental groups, and if it is corrosion current data, it is also necessary to calculate the corrosion current density in combination with the specimen size and further calculate its square value; for the results obtained from the above operations, remove abnormal data that does not conform to common sense; and simultaneously enlarge or reduce the cleaned data by the same appropriate multiple to remove the order of magnitude of the data.

[0095] In a feasible embodiment, the fluoride ion concentration is 8 mmol·L -1 The data on the 21st day measured under the conditions had only 3 sets of parallel samples (the data measured in other periods had 4 sets of parallel samples), so the arithmetic mean cannot be uniformly divided by 4,8mmol·L -1The data on the 21st day corresponding to the condition should be divided by 3;

[0096] The titanium alloy sample used in the present invention is in the shape of a perforated disc, the disc diameter D=15 mm, the hole diameter d=1 mm, and the surface area A=π / 4(D 2 -d 2 );

[0097] 4mmol·L -1 and 8mmol·L -1 The experimental data corresponding to the two fluoride ion concentration levels respectively discarded the measurement results on the 35th day (including corrosion current density and yield strength);

[0098] 4mmol·L -1 The square value of the corrosion current density corresponding to the fluoride ion concentration needs to be multiplied by 10 4 To remove the original data magnitude, 8mmol·L -1 The square value of the corrosion current density corresponding to the fluoride ion concentration needs to be multiplied by 10 5 ;

[0099] The yield strength data corresponding to the two fluoride ion concentrations are multiplied by 10 -2 Remove the order of magnitude.

[0100] S204: By combining and using a variety of neural networks, the pre-processed experimental data of each time gradient are input to train the neural network, and the time-corrosion current density square value is fitted in segments; in the time period with insufficient data, traditional mathematical methods are used for pre-fitting to achieve data enhancement, and then used for neural network training.

[0101] In a feasible implementation manner, the present invention combines multiple neural networks to perform segmented fitting of "time-corrosion current density square value", with time as input and corrosion current density square value as output: -1 Fluoride ion: The data on days 0, 7, 14, and 21 are the first segment, and a 4-layer (hidden layer × 2) custom BP network is used for fitting; the data on days 21, 28, and 42 are the second segment, and after day 42, the following negative exponential function is used to generate supplementary data: The data were supplemented to twice the actual experimental days, that is, t = 43, 44, ..., 84 were substituted into the above formula. At this time, the data from the 21st day to the 84th day were taken as the second segment and simulated using a 4-layer (hidden layer × 2) feedforward neural network (FeedforwardNet).

[0102] For 8mmol·L -1Fluoride ion: The data of days 0, 7, 14, and 21 are the first segment, and a custom BP neural network with 4 layers (hidden layer × 2) is used for fitting; the data of days 21, 28, and 42 are the second segment, and the following Gaussian function is used for interpolation and extrapolation: The data were supplemented to twice the actual experimental days, that is, t = 21, 22, ..., 42, 43, ..., 84 were substituted into the above formula. At this time, the data from day 21 to day 84 were taken as the second segment and simulated using a generalized regression neural network (GRNN) containing 64 radial basis neurons.

[0103] In a feasible implementation, the square value of the corrosion current density gradually degenerates to a fixed non-negative value over time. To ensure that the neural network learns this change trend, it is necessary to supplement the training data after the 42nd day. Otherwise, the neural network may predict the change law after the 42nd day as a continuous decrease (or even accelerated decrease) to below 0, and fail to learn the true slowing law. Using traditional mathematical methods, the negative exponential function ae -bx +c or rational function After the 42nd day, extrapolation was performed to supplement the training data so that the training set covered the global properties of the function to be fitted.

[0104] In a feasible implementation, if the neural network is trained using only four measured data of "days 0, 7, 14, and 21", even if the degree of fit is high at the measured data points, the shape of the entire fitting curve is not unique. Due to the lack of actual experimental data as a reference, it is impossible to judge which fitting curve is better, and thus it is impossible to guide the training of the neural network. Therefore, traditional mathematical fitting methods are used to achieve data enhancement as a reference value for the neural network.

[0105] S205: Predict the time length expected by the user Discrete Input it into the trained neural network and output the corresponding corrosion current density square value sequence Numerical integration of the square value sequence of corrosion current density is performed to obtain the square integral value sequence of corrosion current density

[0106] In a feasible implementation, in step S205, the user's expected prediction time length Discrete Input it into the trained neural network and output the corresponding corrosion current density square value sequence include:

[0107] S251: Convert discrete time series Input into the trained BP network to get the output Only keep the first 21 output results.

[0108] S252: Discrete time series Input into the trained generalized regression neural network net_Time_Jcor2_F8 to get the output

[0109] S253: merging the above two output sequences to obtain a first group of concentration corrosion current density square value sequences;

[0110] S254: Repeat the above steps S251-S253 to obtain a second set of concentrations of the corrosion current density square value sequence.

[0111] S206: Train the neural network by combining multiple neural networks; perform segmented fitting of the data as “corrosion current density square integral - yield strength degradation amount”; in the time period when data is insufficient, use traditional mathematical methods for pre-fitting to achieve data enhancement, and then use it for neural network training.

[0112] In a feasible implementation, for the first segment of the two concentration data, a BP neural network is selected, the maximum number of training cycles is 8000, and the iteration termination condition is that the maximum absolute error MAE of the training sample set is less than 10 -15 , Among them is is the target output value, is the actual output value, k is the sample number;

[0113] The invention customizes the standard fully connected BP network, and the network structure is 1-5-3-1. Hidden layer neuron transfer function: hyperbolic tangent function The output layer neuron transfer function: linear function y = x; the maximum number of training cycles: epochs = 8000. The training algorithm uses the steepest descent method; the initial learning rate: η = 0.01.

[0114] The learning rate adaptive adjustment rule of the present invention is: Where n is the training cycle number.

[0115] In a feasible implementation, for the second segment of data of the first group, a feedforward neural network is selected, and the preset algorithm is the Levenberg-Marquardt algorithm;

[0116] Among them, the feedforward neural network (FeedforwardNet) built here is recorded as net_Time_Jcor2_F4; the network structure is 1-5-3-1; the training algorithm is Levenberg-Marquardt algorithm:

[0117] net_Time_Jcor2_F4.trainParam.epochs=8000

[0118] net_Time_Jcor2_F4.trainParam.max_fail=20

[0119] net_Time_Jcor2_F4.trainParam.min_grad=10 -20

[0120] net_Time_Jcor2_F4.trainParam.goal=0

[0121] net_Time_Jcor2_F4.trainParam.lr=0.01

[0122] net_Time_Jcor2_F4.trainParam.time=inf

[0123] In a feasible implementation, for the second segment of data of the second group, a generalized regression neural network is selected. Among them, the generalized regression neural network, the neural network constructed here is recorded as net_Time_Jcor2_F8: the network structure is input-radial basis neuron layer-special linear output layer 1-64-1 radial basis neuron layer with bias value, special linear output layer without bias value; the number of neurons in the radial basis neuron layer is consistent with the number of samples, which is 64 radial basis neurons here; the transfer function of the radial basis neuron layer: Gaussian function (a kind of radial basis function) Output layer transfer function: linear function y=x.

[0124] In a feasible implementation manner, the present invention combines multiple neural networks to segmentally fit the "corrosion current density square integral-yield strength degradation amount" to obtain the corrosion current density square integral value I k is the input quantity, and the yield strength degradation amount - Δσ k is the output value (it is emphasized that the degradation value "-Δσ" in the yield strength degradation curve is always based on the yield strength σ of the uncorroded titanium alloy 0 (In the present invention, σ 0 =860MPa) as a reference, which is different from the single-day degradation amount "-δσ" in the relative instantaneous corrosion rate):

[0125] For 4mmol·L -1 Fluoride ion: The data from day 0 to 21 are the first section; the yield strength data from day 0, 7, 14, and 21 are interpolated by cubic smooth spline to obtain Yield strength degradation from 1st to 21st day -Δσ k =σ 0 -σ k ; In step S205, the fluoride ion concentration is 4mmol·L -1 The square integral value sequence of corrosion current density obtained under the condition As input, To output the reference value, a radial basis function neural network (RBF) containing 21 radial basis function neurons was selected for approximation; the data from the 21st to the 600th day was the second segment; the yield strength data from the 21st, 28th, and 42nd days were fitted with a Gaussian function and extrapolated to the 600th day to obtain The selected Gaussian function is as follows:

[0126]

[0127] Yield strength degradation from 21st to 600th day -Δσ k =σ 0 -σ k ; In step S205, the fluoride ion concentration is 4mmol·L -1 The square integral value sequence of corrosion current density obtained under the condition As input, is the output reference value, which is used to train a 5-layer (hidden layer × 3) feedforward neural network.

[0128] In a feasible implementation, the radial basis function network (RBF) network is created by adding one neuron at a time to create the network, that is, iteratively determining the network structure. At each iteration, the input vector that is most effective in reducing the network output error is selected to generate a radial basis neuron, and then the error of the newly generated network is checked. If it does not exceed the expected error, the network creation process is terminated; otherwise, new neurons are added until the mean square error of the network reaches the set error target or the network reaches the maximum number of neurons.

[0129] Network structure: input-radial basis neuron layer (hidden layer)-linear output layer 1-21-1; both the radial basis neuron layer and the linear output layer have deviation values; radial basis neuron layer transfer function: Gaussian function (a type of radial basis function) Output layer transfer function: linear function y=x;

[0130] Feedforward neural network, the neural network built here is recorded as net_Jcor2_Sigma_F4: Network structure: 1-10-8-5-1; Training algorithm: Levenberg-Marquardt algorithm Training parameters are as follows:

[0131] net_Jcor2_Sigma_F4.trainParam.epochs=8000

[0132] net_Jcor2_Sigma_F4.trainParam.max_fail=20

[0133] net_Jcor2_Sigma_F4.trainParam.min_grad=10 -20

[0134] net_Jcor2_Sigma_F4.trainParam.goal=0

[0135] net_Jcor2_Sigma_F4.trainParam.lr=0.01

[0136] net_Jcor2_Sigma_F4.trainParam.time=inf

[0137] For 8mmol·L -1 Fluoride ion: The data from day 0 to 28 is the first section; the yield strength data of day 0, 7, 14, 21, and 28 are interpolated by cubic smooth spline to obtain Yield strength degradation from 1st to 28th day -Δσ k =σ 0 -σ k ; In step S205, the fluoride ion concentration is 8mmol·L -1 The square integral value sequence of corrosion current density obtained under the condition As input, the To output the reference value, a radial basis function neural network (RBF) with 28 radial basis function neurons was selected for approximation; the data from the 28th to the 600th day was the second segment; the yield strength data from the 28th and 42nd day were simulated with adjustment factors. The Gaussian function of is fitted and extrapolated to the 600th day to obtain Selected with adjustment factor The Gaussian function is as follows:

[0138]

[0139] Among them, K=0.015.

[0140] In a feasible implementation, the network creation method of the radial basis neural network is: to create the network by adding one neuron at a time, that is, to iteratively determine the network structure. In each iteration, the input vector that is most effective in reducing the network output error is selected to generate a radial basis neuron, and then the error of the newly generated network is checked. If it does not exceed the expected error, the network creation process is terminated; otherwise, new neurons are added until the mean square error of the network reaches the set error target or the network reaches the maximum number of neurons. Network structure: input-radial basis neuron layer (hidden layer)-linear output layer 1-28-1; both the radial basis neuron layer and the linear output layer have deviation values; radial basis neuron layer transfer function: Gaussian function (a type of radial basis function) Output layer transfer function: linear function y=x.

[0141] Yield strength degradation from 28 to 600 days -Δσ k =σ 0 -σ k ; In step S205, the fluoride ion concentration is 8mmol·L -1 The square integral value sequence of corrosion current density obtained under the condition As input, The output reference value is used to train a 5-layer (hidden layer × 3) feedforward neural network (FeedforwardNet).

[0142] In a feasible implementation, a feedforward neural network, the neural network constructed here is recorded as net_Jcor2_Sigma_F4: network structure: 1-10-8-5-1; training algorithm: Levenberg-Marquardt algorithm training parameters are as follows:

[0143] net_Jcor2_Sigma_F4.trainParam.epochs=8000

[0144] net_Jcor2_Sigma_F4.trainParam.max_fail=20

[0145] net_Jcor2_Sigma_F4.trainParam.min_grad=10 -20

[0146] net_Jcor2_Sigma_F4.trainParam.goal=0

[0147] net_Jcor2_Sigma_F4.trainParam.lr=0.01

[0148] net_Jcor2_Sigma_F4.trainParam.time=inf

[0149] In a feasible implementation, data segmentation includes segmenting the titanium alloy sample data with different fluoride ion concentrations, wherein 0-21 days is the first segment, and after 21 days is the second segment.

[0150] In a feasible implementation manner, data enhancement includes: performing data enhancement on the second segment of data, that is, supplementing the data to twice the actual experimental days through a pre-fitting formula.

[0151] In a feasible implementation manner, numerically integrating the corrosion current density square value sequence includes: numerically integrating the corrosion current density square value along time by using a complex trapezoidal quadrature formula.

[0152] In the embodiment of the present invention, I cor With σ cor There is not necessarily a one-to-one mapping relationship. Even if an approximate relationship is fitted from a pure data perspective, it may not conform to the real physical cause and effect relationship. The corrosion current represents the transfer of charge in the electrochemical reaction and then reflects the increase or decrease in the amount of reactant and product substances. The transformation of the substance will inevitably cause changes in its intrinsic properties, such as changes in mechanical properties. Therefore, the present invention considers fitting the process quantity rather than the state quantity, that is, using the square integral of the corrosion current and the degradation of yield strength in the same period of time - Δσ i =σ i-1 -σ i For the integration and other analytical operations, numerical integration is required to be implemented by computer. The present invention selects the commonly used trapezoidal numerical integration formula to replace

[0153] In a feasible implementation, the Gaussian function has an inflection point (where the second-order derivative is 0), and gradually approaches 0 over time. The training set should cover all the properties of the approximated function (similar to the probability statistical interpretation, the data distribution of the training set needs to be close to the true probability distribution of the model). Therefore, the training set needs to include the inflection point and the slowing trend. The inflection point is In order to include the trend of function value slowing down over time in the training set, it is recommended to select an interval of In the present invention, [21,600] is selected

[0154] In a feasible implementation, different from the experimental data of "fluoride ion concentration 4mmol / L", the experimental data under the condition of fluoride ion concentration 8mmol / L showed that after 21 days, the square value of the corrosion current density still had a short rising process, and then gradually decreased; during the decline, the square value of the corrosion current density gradually degraded to a fixed non-negative value with time evolution. In order to ensure that the neural network learns this kind of change trend of rising first and then falling, it is necessary to interpolate from the 21st day to the 42nd day and extrapolate after the 42nd day to supplement the training data. If the training data is not interpolated and supplemented within 21-42 days, although the neural network can still learn the trend of rising first and then falling, the shape of the obtained fitting curve may be more uncertain. Since there is no reference value, it is impossible to evaluate and choose, so it is necessary to interpolate within 21-42 days to enhance the data and better guide the training of the neural network; if the training data is not extrapolated after the 42nd day, it may cause the neural network to predict the change law after the 42nd day as a continuous decrease (even accelerated decrease) to below 0, and fail to learn the true slowing law. Using traditional mathematical methods, the Gaussian function After the 21st day, interpolation (21-42 days) and extrapolation (after 42 days) are performed to supplement the training data so that the training set covers the global properties of the function to be fitted.

[0155] In a feasible implementation, if the neural network is trained using only five measured data of "days 0, 7, 14, 21, and 28", even if the degree of fit is high at the measured data points, the shape of the entire fitting curve is not unique. Due to the lack of actual experimental data as a reference, it is impossible to judge which fitting curve is better, and thus it is impossible to guide the training of the neural network. Therefore, traditional mathematical fitting methods are used to achieve data enhancement as a reference value for the neural network.

[0156] In a feasible implementation, the Gaussian function has an inflection point (where the second-order derivative is 0) and gradually slows down over time. The training set should cover all the properties of the approximated function (similar to the probability statistical interpretation, the data distribution of the training set needs to be close to the true probability distribution of the model). Therefore, the training set needs to include the inflection point and the slowing trend. The inflection point is In order to include the trend of function value slowing down over time in the training set, it is recommended to select an interval of In the present invention, [28,600] is selected

[0157] S207: Arrange the fitting results of the data in the neural network to obtain the yield strength degradation curve of the titanium alloy.

[0158] In one possible implementation, the discrete time series and yield strength degradation series The corrosion rate (t0 =0) to any day t m The degradation of the yield strength of titanium alloy during this period of time - Δσ m . Use the calculation formula “σ m =σ 0 -(-Δσ m )”, the residual yield strength of the titanium alloy on any day after corrosion can be obtained, and the degradation curve of the yield strength of the titanium alloy over time can be drawn

[0159] S208: Based on the yield strength degradation curve, use a difference matrix to scan the yield strength degradation curve to obtain a relative instantaneous corrosion rate v m , completed the neural network prediction of titanium alloy corrosion rate in fluoride ion environment.

[0160] In a feasible implementation, the daily degradation of the yield strength of the titanium alloy on any day -δσ m As the relative instantaneous corrosion rate v of that day m ,in, δT=1;

[0161] Degradation curve and data series obtained by left multiplication of the difference matrix Get the relative instantaneous corrosion rate series Right now

[0162] The neural network prediction of the corrosion rate of titanium alloy in fluoride ion environment is completed. The fitting effect diagram of the present invention, the yield strength degradation curve and the relative instantaneous corrosion rate curve are shown in the figure. Figure 4-10 shown.

[0163] In a feasible implementation manner, the “corrosion current density square value” is selected as the intermediate variable for “the degradation process of the yield strength of titanium alloy fitting over time according to the present invention” as follows:

[0164] The dimensional analysis is as follows: [σSL] = [W] = [E] = [i 2 RT], where σ: stress; S: area; L: length; W: work; E: energy; i: current; R: resistance; T: time. Both ends are dimensions of work or energy. Where η is the resistivity; i = J·S, where J is the current density, so we have That is, [σ] = [η] [J 2 ][T]=[η][J 2 T]; therefore, the present invention decides to use the "corrosion current density square value" " as an intermediate variable (corresponding to [J 2]), and after numerically integrating the square value of the corrosion current density along time (corresponding to [J 2 T]) and the yield strength after corrosion σ cor Fitting Instead of directly fitting J cor → cor .

[0165] In a feasible implementation manner, the technical problem to be solved by the present invention is: under fluoride ion corrosion conditions, the variation pattern of the corrosion current density and yield strength of the titanium alloy with time (the minimum time scale δT is set to 1 day) is obtained based on the experimental data fitting, and the yield strength degradation amount per unit time (1 day) is used as the predicted relative instantaneous corrosion rate.

[0166] Based on dimensional analysis, the present invention selects the square value of corrosion current density as the intermediate variable, first fits the "time-square value of corrosion current density" relationship, then uses the complex trapezoidal quadrature formula to numerically integrate the square value of corrosion current density along time, and then fits the "square integral value of corrosion current density-yield strength degradation" relationship, and finally obtains the degradation curve of titanium alloy yield strength with time for reference in engineering service.

[0167] When fitting the functional relationship, the present invention fully examines the distribution characteristics of the original experimental data and decides to adopt segmented fitting;

[0168] Taking into account the requirement for the minimum time scale (1 day) in the problem to be solved, the actual experimental sampling period (7 days), and the time trends of the current and yield strength data inspired or constrained by prior knowledge of materials science (as the corrosion time is infinitely extended, the measured data should gradually tend to a certain extremely small fixed value), the present invention first uses traditional mathematical fitting methods to achieve data enhancement and supplement the necessary training data to cover all properties of the function to be fitted, thereby guiding and constraining the next step of neural network training; during segmented fitting, different neural network models are used for different segments of data to ensure a sufficiently accurate fit. The present invention combines a multilayer feedforward neural network, a radial basis neural network, and a generalized regression neural network.

[0169] In a feasible implementation, taking into account the actual service use of the material, the time required for the comprehensive material to undergo obvious and measurable corrosion behavior, as well as the actual experimental conditions and the specific operations of the experimental implementation, similar to the existence of a minimum energy unit in quantum mechanics, the present invention believes that it is reasonable and necessary to use "days" as the minimum time scale for the corrosion process of titanium alloys, that is, there is no need to fit the changes in the mechanical properties of titanium alloys within time scales such as seconds, milliseconds, and microseconds. The reasons are as follows: There is evidence that the service cycle of titanium alloys is based on "years" as the time unit. Therefore, if it is possible to obtain how much the mechanical properties of titanium alloys will degrade every day and how many days it will take for the mechanical properties to drop below the safety line, it will be sufficient to meet engineering needs. That is, the degradation rate of the material mechanical properties obtained in "days" as the time unit is already an "instantaneous rate" relative to the service process itself in "years", which is referred to as the relative instantaneous corrosion rate in the present invention.

[0170] Fig.11 1 is a block diagram of a neural network prediction device for the corrosion rate of titanium alloy in a fluoride ion environment according to an exemplary embodiment. Fig.11 , the device 300 comprises:

[0171] The data preprocessing module 310 is used to collect titanium alloy corrosion experimental data and preprocess the experimental data;

[0172] A data fitting module 320 is used to perform segmented fitting on the preprocessed data by combining multiple neural networks;

[0173] The curve integration module 330 organizes the fitting results of the data in the neural network to obtain the yield strength degradation curve of the titanium alloy;

[0174] The corrosion rate prediction module 340 uses a difference matrix to scan the yield strength degradation curve based on the yield strength degradation curve to obtain a relative instantaneous corrosion rate v k , completed the neural network prediction of titanium alloy corrosion rate in fluoride ion environment.

[0175] Optionally, the data preprocessing module 310 is further used to divide the titanium alloy samples into two groups, respectively place them in immersion solutions with different fluorine ion concentrations for immersion corrosion, and divide each group into seven time gradients; the same time gradient of the two groups of titanium alloy samples is a parallel experimental group;

[0176] Measuring electrochemical parameters of the titanium alloy sample after each gradient immersion corrosion;

[0177] Calculate the arithmetic mean of the data of the parallel experimental groups; if the corrosion current data is calculated, it is also necessary to calculate the corrosion current density based on the specimen size and further calculate its square value; for the results obtained from the above operations, perform data cleaning.

[0178] Optionally, the data fitting module 320 is further used to train the neural network by combining multiple neural networks, inputting the pre-processed experimental data of each time gradient, and fitting the time-corrosion current density square value in segments; in the time period with insufficient data, using traditional mathematical methods for pre-fitting to achieve data enhancement, and then using it for neural network training;

[0179] The length of time that users expect to be predicted Discrete Input it into the trained neural network and output the corresponding corrosion current density square value sequence Numerical integration of the square value sequence of corrosion current density is performed to obtain the square integral value sequence of corrosion current density

[0180] The neural network is trained by combining multiple neural networks. The data is segmentedly fitted with the "square integral of corrosion current density-yield strength degradation amount". In the time period when data is insufficient, traditional mathematical methods are used for pre-fitting to achieve data enhancement, which is then used for neural network training.

[0181] Optionally, data segmentation includes: segmenting the titanium alloy sample data with different fluoride ion concentrations, wherein 0-21 days is the first segment, and after 21 days is the second segment.

[0182] Optionally, the data fitting module 320 is further used to train the neural network by combining multiple neural networks and inputting the preprocessed experimental data of each time gradient, including:

[0183] For the first segment of the two concentration data, BP neural network was selected, the maximum number of training cycles epochs = 8000, and the iteration termination condition was the maximum absolute error MAE of the training sample set <10 -15 , Among them is is the target output value, is the actual output value, k is the sample number;

[0184] For the second segment of data in the first group, a feedforward neural network was selected, and the default algorithm was the Levenberg-Marquardt algorithm;

[0185] For the second segment of data in the second group, a generalized regression neural network was selected.

[0186] Optionally, data enhancement includes: performing data enhancement on the second segment of data, namely supplementing the data to a time period twice the actual number of experimental days through a pre-fitting formula.

[0187] Optionally, the data fitting module 320 is further used to transform the discrete time series Input into the trained BP network to get the output Only keep the first 21 output results.

[0188] Discrete time series Input into the trained generalized regression neural network net_Time_Jcor2_F8 to get the output

[0189] The above two output sequences are combined to obtain the first group of concentrations of the square value sequence of corrosion current density;

[0190] Repeat the above steps to obtain the second set of concentrations of the corrosion current density square value sequence.

[0191] Optionally, numerically integrating the sequence of square values ​​of the corrosion current density includes: numerically integrating the square values ​​of the corrosion current density along time by using a complex trapezoidal quadrature formula.

[0192] Optionally, the corrosion rate prediction module 340 is further used to calculate the daily degradation amount of the yield strength of the titanium alloy on any day -δσ m As the relative instantaneous corrosion rate v of that day m ,in, δT=1;

[0193] Degradation curve and data series obtained by left multiplication of the difference matrix Get the relative instantaneous corrosion rate series Right now

[0194] Complete the neural network prediction of titanium alloy corrosion rate in fluoride ion environment.

[0195] In the embodiment of the present invention, based on dimensional analysis, the square value of corrosion current density is selected as the intermediate variable, and the "time-square value of corrosion current density" relationship is first fitted, and then the complex trapezoidal quadrature formula is used to numerically integrate the square value of corrosion current density along time, and then the "square integral value of corrosion current density-yield strength degradation amount" relationship is fitted to obtain the degradation curve of the yield strength of titanium alloy with time for reference in engineering service, and on the basis of the obtained yield strength degradation curve, the curve is scanned using a differential matrix to finally obtain the relative instantaneous corrosion rate of titanium alloy in a fluoride ion environment.

[0196] Fig.121 is a schematic diagram of the structure of an electronic device 400 provided in an embodiment of the present invention. The electronic device 400 may have relatively large differences due to different configurations or performances, and may include one or more processors (central processing units, CPU) 401 and one or more memories 402, wherein the memory 402 stores at least one instruction, and the at least one instruction is loaded and executed by the processor 401 to implement the following steps of the neural network prediction method for the corrosion rate of titanium alloy in a fluoride ion environment:

[0197] S1: Collecting titanium alloy corrosion experimental data; preprocessing the experimental data;

[0198] S2: By combining multiple neural networks, the preprocessed data is piecewise fitted; the neural network is trained;

[0199] S3: Arrange the fitting results of the data in the neural network to obtain the yield strength degradation curve of the titanium alloy;

[0200] S4: Based on the yield strength degradation curve, the yield strength degradation curve is scanned using a difference matrix to obtain a relative instantaneous corrosion rate v m , completed the neural network prediction of titanium alloy corrosion rate in fluoride ion environment.

[0201] In an exemplary embodiment, a computer-readable storage medium is also provided, such as a memory including instructions, which can be executed by a processor in a terminal to complete the neural network prediction method for the corrosion rate of titanium alloy in a fluoride ion environment. For example, the computer-readable storage medium can be a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, an optical data storage device, etc.

[0202] A person skilled in the art will understand that all or part of the steps to implement the above embodiments may be accomplished by hardware or by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium, and the above-mentioned storage medium may be a read-only memory, a disk or an optical disk, etc.

[0203] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A neural network prediction method for the corrosion rate of titanium alloy in a fluoride ion environment, characterized in that: include: S1: Collecting titanium alloy corrosion experimental data; Preprocessing the experimental data; In the step S1, titanium alloy corrosion experimental data is collected and preprocessed, including: S11: The titanium alloy specimens were divided into two groups, and were placed in immersion solutions with different fluoride ion concentrations for immersion corrosion. Each group was divided into seven time gradients. The same time gradient of the two groups of titanium alloy specimens was a parallel experimental group. S12: measuring electrochemical parameters of the titanium alloy sample after each gradient immersion corrosion; S13: Calculate the arithmetic mean of the data of the parallel experimental groups; if the corrosion current data is calculated, it is also necessary to calculate the corrosion current density in combination with the specimen size and further calculate its square value; for the results obtained from the above operations, perform data cleaning; S2: By combining multiple neural networks, the preprocessed data is piecewise fitted; In S2, a plurality of neural networks are combined to perform segmented fitting on the preprocessed data, including: S21: By combining and using a variety of neural networks, inputting the pre-processed experimental data of each time gradient, training the neural network, and fitting the time-corrosion current density square value in segments; in the time period with insufficient data, using traditional mathematical methods for pre-fitting to achieve data enhancement, and then using it for neural network training; S22: The length of time that the user expects to be predicted , discretized into , input it into the trained neural network, and output the corresponding corrosion current density square value sequence ; Perform numerical integration on the square value sequence of corrosion current density to obtain the square integral value sequence of corrosion current density ; S23: Train the neural network by combining multiple neural networks; segmentally fit the square integral of corrosion current density-yield strength degradation; in the period of insufficient data, use traditional mathematical methods for pre-fitting to achieve data enhancement, and then use it for neural network training; S3: Arrange the fitting results of the data in the neural network to obtain the yield strength degradation curve of the titanium alloy; S4: Based on the yield strength degradation curve, the yield strength degradation curve is scanned using a difference matrix to obtain a relative instantaneous corrosion rate , completed the neural network prediction of titanium alloy corrosion rate in fluoride ion environment.

2. The method according to claim 1, characterized in that The data segmentation includes: segmenting the titanium alloy sample data with different fluoride ion concentrations, wherein 0-21 days is the first segment, and after 21 days is the second segment.

3. The method according to claim 2, characterized in that In step S21, a plurality of neural networks are combined and used, and the pre-processed experimental data of each time gradient are input to train the neural network, including: For the first segment of the two concentration data, BP neural network was selected with a maximum number of training cycles. The iteration termination condition is the maximum absolute error of the training sample set , , where is the target output value, is the actual output value, is the sample serial number; For the second segment of data in the first group, a feedforward neural network was selected, and the default algorithm was the Levenberg-Marquardt algorithm; For the second segment of data in the second group, a generalized regression neural network was selected.

4. The method according to claim 3, characterized in that Data enhancement includes: performing data enhancement on the second segment of data, that is, supplementing the data to twice the actual experimental days through a pre-fitting formula.

5. The method according to claim 4, characterized in that In step S21, the time length that the user expects to predict , discretized into , input it into the trained neural network, and output the corresponding corrosion current density square value sequence ,include: S211: Discrete time series Input into the trained BP network to get the output , only keep the first 21 output results, that is ; S212: Discrete time series Input into the trained generalized regression neural network net_Time_Jcor2_F8 to get the output ; S213: merging the above two output sequences to obtain a first group of concentration corrosion current density square value sequences; S214: Repeat the above steps S211-S213 to obtain a second set of concentrations of corrosion current density square value sequences.

6. The method according to claim 5, characterized in that In the above S21, numerically integrating the corrosion current density square value sequence includes: numerically integrating the corrosion current density square value along time by using a complex trapezoidal quadrature formula.

7. The method according to claim 6, characterized in that In S4, based on the yield strength degradation curve, the yield strength degradation curve is scanned using a difference matrix to obtain a relative instantaneous corrosion rate , complete the neural network prediction of titanium alloy corrosion rate in fluoride ion environment, including: S41: Single-day degradation of the yield strength of titanium alloy on any given day As the relative instantaneous corrosion rate of the day ,in, , ; S42: Use the difference matrix to multiply the degradation curve data sequence obtained in step S3 , the relative instantaneous corrosion rate series is obtained ,Right now Complete the neural network prediction of titanium alloy corrosion rate in fluoride ion environment.

8. A neural network prediction device for the corrosion rate of titanium alloy in a fluoride ion environment, characterized in that: The device is applicable to the neural network prediction method of the titanium alloy corrosion rate in a fluoride ion environment according to any one of claims 1 to 7, and the device comprises: A data preprocessing module is used to collect titanium alloy corrosion experimental data and preprocess the experimental data; A data fitting module is used to perform piecewise fitting on the preprocessed data by combining multiple neural networks; The curve integration module organizes the fitting results of the data in the neural network to obtain the yield strength degradation curve of the titanium alloy; The corrosion rate prediction module uses a differential matrix to scan the yield strength degradation curve based on the yield strength degradation curve to obtain a relative instantaneous corrosion rate. , completed the neural network prediction of titanium alloy corrosion rate in fluoride ion environment.

Images