Turbine transition state blade tip clearance estimation method based on long short-term memory neural network

Abstract

Disclosed in the present invention is a turbine transition state blade tip clearance estimation method based on a long short-term memory neural network. The method comprises: first, using data concatenation and introduction of new parameters to construct multi-transition-state process transient data; then using sensitivity analysis to screen for feature parameters having a great influence on a blade tip clearance, and carrying out dimension reduction on the influence parameters; and on the basis of a long short-term memory neural network, constructing a mapping model between each feature parameter and the blade tip clearance, and building an efficient and high-precision turbine transition state blade tip clearance estimation model. By means of the turbine transition state blade tip clearance estimation model, turbine blade tip clearances of multiple transition state processes can be efficiently and accurately estimated.

Description

Turbine transient tip clearance estimation method based on long short-term memory neural network Technical Field

[0001] The present invention belongs to the technical field of engineering thermophysics, and in particular relates to a method for estimating turbine transition state blade tip clearance based on a long short-term memory neural network. Background Art

[0002] Tip clearance refers to the gap between the high-speed rotating rotor blades and the stationary casing to ensure safe engine operation. Research has shown that for every 0.254mm reduction in tip clearance, exhaust gas temperature can be reduced by 10°C, saving 1% in fuel consumption. This can further extend engine life and reduce significant economic losses. Inappropriate tip clearance can lead to high leakage or blade wear, reducing engine performance or even causing safety accidents. Turbine tip clearance typically occurs in a transient state during the engine's lifecycle. During the design of the tip clearance transient state, the differences in thermal response rates between the casing, impeller, and blades, as well as the loads they are subjected to, can lead to inconsistent deformation of the rotor and stator components. During this transient state, tip clearance is prone to reaching a minimum, which can easily lead to dangerous conditions such as blade wear. Therefore, it is necessary to develop a method for estimating turbine tip clearance during this transient state.

[0003] Currently, scholars are exploring various methods for predicting transitional blade tip clearance. Kypuros et al. established a one-dimensional mathematical model for analyzing turbine tip clearance, which can quickly determine the deformation of each model component and obtain the blade tip clearance value. Furthermore, this model can capture the "pitch point" phenomenon during transitional changes (Kypuros JA, Melcher K JA reduced model for prediction of thermal and rotational effects on turbine tip clearance. 2003). Xu Yijun, based on the structural characteristics and operating conditions of different high- and low-pressure turbine components, established temperature and deformation calculation models for each component using one-dimensional engineering methods and two-dimensional finite element methods, respectively, to achieve the calculation of steady-state and transitional blade tip radial clearance (Xu Yijun. Development of a Radial Clearance Calculation Program for Low-Pollution Civil Aviation Engines [D]. Nanjing University of Aeronautics and Astronautics, 2014).

[0004] However, these common methods currently struggle to achieve efficient and accurate estimation of turbine tip clearance during the transition phase. While one-dimensional mathematical models offer high computational efficiency, they are difficult to capture in a single dimension, significantly limiting their accuracy. Two-dimensional or three-dimensional numerical simulations, on the other hand, achieve high accuracy at the expense of extensive computational effort. These limitations hinder the ability to achieve both efficiency and accuracy, necessitating the development of new approaches for efficient and accurate estimation of turbine tip clearance during the transition phase.

[0005] With the development of machine learning, data-driven agent models provide an efficient and accurate solution for predicting turbine tip clearance in transition state. For the prediction of tip clearance in transition state, Jia Binghui et al. first introduced a rotor vibration model based on a certain-dimensional tip clearance model, provided transition tip clearance data with vibration noise, and used the least squares support vector machine model to learn the dynamic change process of tip clearance (Jia Binghui, Zhang Xiaodong, Hou Yujun. LS-SVM method for dynamic modeling of turbine tip clearance [J]. Computer Simulation, 2012, 29(04).). Huang Xin et al. used a tip clearance sensor to measure the three-dimensional tip clearance transition change characteristics of the blade tip, and diagnosed the blade crack information in the transition tip clearance based on the improved deep belief network model (Huang Xin, Zhang Xiaodong, Zhang Yingjie et al. Three-dimensional tip clearance blade crack diagnosis method based on improved DBNs [J]. Vibration. Testing and Diagnosis, 2022, 42(02).). Nikolay and Tuhfe used the long short-term memory neural network to predict the periodic deformation of the blade tip based on the information of the blade load measurement system (Dimitrov, N, T.Virtual sensors for wind turbines with machine learning-based time series models. Wind Energy. 2022; 25(9):1626-1645.).

[0006] Currently, these surrogate models are still in the early stages of research on transition-state tip clearance estimation. Their tip clearance data is relatively simple and differs significantly from the variability of actual aircraft engine tip clearance. Furthermore, surrogate models must address issues such as high-dimensional tip clearance prediction and the limited sample characteristics within a single process. Key challenges for future research include fully leveraging the influencing factors of the coupling between the timing information of transition-state tip clearance and various parameters, selecting appropriate machine learning methods, and establishing efficient and accurate transition-state tip clearance prediction models.

[0007] Summary of the Invention

[0008] Aiming at the demand for efficient and accurate estimation of turbine blade tip clearance in transient state, the present invention provides a method for estimating turbine blade tip clearance in transient state based on long short-term memory neural network, which can achieve efficient and accurate estimation of blade tip clearance.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: a method for estimating turbine tip clearance in transient state based on long short-term memory neural network, comprising:

[0010] Step 1: Establish a physical model of the turbine components and the air system flow path. Based on the aircraft engine design manual and empirical formulas, obtain the flow rate and temperature of each air system flow path in the turbine physical model, as well as the range of the turbine model speed. Then, use the two-dimensional axisymmetric finite element calculation method and the one-dimensional engineering calculation method to construct a transition state tip clearance calculation model.

[0011] Step 2: Sampling is performed within the variation range in step 1 to obtain multiple sets of variable data, and then the multiple sets of variable data are input into the transition state tip clearance calculation model for calculation to obtain data sets of casing deformation, disk deformation, blade deformation, and tip clearance for different engine histories. The data sets of casing deformation, disk deformation, blade deformation, and tip clearance for different engine histories are then spliced ​​to obtain transition state component deformation and tip clearance data covering multiple engine histories;

[0012] Step 3: Perform sensitivity analysis on the transition state data in step 2 to obtain sensitivity analysis results, and then perform dimensionality reduction processing on the data based on the sensitivity analysis results to obtain dimensionality-reduced data;

[0013] Step 4: Divide the data of multiple engine history records that have undergone dimensionality reduction processing in step 3 into training sets, test sets, and validation sets;

[0014] Step 5: Using a long short-term memory neural network model, the training set and test set in step 4 are used for model training and testing. Through multiple training and testing cycles, a mapping model between each characteristic parameter and the blade tip clearance is constructed to form a turbine transition state blade tip clearance prediction model.

[0015] Step 6: Using the estimation model constructed in step 5, input the data in the validation set in step 4 to verify the accuracy of the estimation model.

[0016] Furthermore, the specific steps for data splicing of the data sets of casing deformation, wheel deformation, blade deformation, and tip clearance of different engine histories in step 2 are as follows:

[0017] By citing a new variable New_cycle, the data sets of casing deformation, disk deformation, blade deformation and tip clearance of different engine histories are spliced, where New_cycle only contains two values, which are 0 and 1.

[0018] Furthermore, the formula for performing sensitivity analysis on multiple sets of variable data in step 3 is: Y=f(X1, X2, ..., X k )

[0019] Where Y is the research model defined by f(X1, X2, ..., X k ) is expressed as, and then expanded into a set of dimension-increasing functions and f, and assuming that f is square integrable, the decomposition function is squared and integrated, and the left side of the integral is defined as the total output variance Var(Y), and the right side is the sum of the variances of each parameter, where V i is the effect of a single parameter on Y, V ij For X i and X j Effect on Y, V 12…d Represented as X1 to X d The impact on Y, and finally, according to the contribution ratio of each variance to the total variance, the sensitivity of each parameter to the output is obtained, that is, the first-order influence index

[0020] Furthermore, in step 4, the ratio of the training set, test set, and validation set is 17:3:10.

[0021] Furthermore, the turbine transition state tip clearance estimation model in step 5 includes:

[0022] An input terminal, for inputting variable data and blade tip clearance into a data processing unit;

[0023] The data processing unit unifies the format of the input data, divides it, and passes it to the LSTM unit;

[0024] Normalization unit, which normalizes the data to improve model accuracy and convergence speed;

[0025] The first LSTM unit reads in sequence data, performs feature extraction, and outputs time step information to the second LSTM unit;

[0026] The second LSTM unit receives the time step information of the first LSTM unit, further extracts features and outputs prediction information;

[0027] Denormalization unit, denormalizes the normalized prediction data and restores it to the true value;

[0028] And an output terminal, used for outputting the predicted value of the blade tip clearance.

[0029] Furthermore, the specific calculation formula of the prediction process of each LSTM unit in the turbine transition state tip clearance estimation model is: t =σ(W f ·[h t-1 ,x t ]+b f ) i t =σ(W i ·[h t-1 ,x t ]+b i ) O t =σ(W o ·[h t-1 ,x t ]+b o ) h t =O t tanh(C t )

[0030] Among them, x t is the feature quantity at the current moment, which is the input of the LSTM model, h t-1 With h t are the hidden state at the previous moment and the hidden state calculated at the current moment, C t-1 with C t are the memory stored at the previous moment and the memory calculated at the current moment, respectively, t 、i t With O t They are the calculation results of forget gate, input gate and output gate respectively. is the updated unit state, W and b are the weight and bias corresponding to each gate respectively, and the subscripts f, i, and O represent the parameters corresponding to the forget gate, input gate, and output gate respectively.

[0031] Beneficial Effects: The present invention constructs a turbine transition tip clearance estimation model based on the relationship between different engine processes, thermodynamic boundary conditions, and tip clearance values, achieving high-efficiency, high-precision estimation of transition tip clearance for different processes. The maximum absolute error in tip clearance prediction is 0.074 mm, and the maximum relative error is only 3.59%. The turbine transition tip clearance estimation model also has very high computational efficiency, completing the calculation for one engine process in just 10 seconds. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 shows the turbine model structure;

[0033] Figure 2 shows the study range of turbine gas thermodynamic parameters;

[0034] Figure 3 shows the turbine transition state tip clearance estimation process;

[0035] Figure 4 shows the result of multi-engine history splicing (in L c for example);

[0036] Figure 5 shows the results of parameter sensitivity analysis (with L c for example);

[0037] Figure 6 shows the LSTM-based turbine transition tip clearance prediction model (with L c-5 Forecasting as an example);

[0038] Figure 7 shows the comparison of prediction results of the intelligent prediction model for turbine transition tip clearance (taking δ5 as an example). DETAILED DESCRIPTION

[0039] The present invention will be further explained below with reference to the accompanying drawings.

[0040] The present invention provides a method for estimating turbine tip clearance in transient state based on a long short-term memory neural network, comprising:

[0041] Step 1: Measure the turbine physical model to obtain the variation range of parameters such as flow rate, temperature, and speed of each air system flow path in the turbine physical model, and then construct a transition state tip clearance calculation model using two-dimensional axisymmetric finite element calculation method and one-dimensional engineering calculation method.

[0042] Step 2: Sampling is performed within the variation range in step 1 to obtain multiple sets of variable data, and then the multiple sets of variable data are substituted into the transition state tip clearance calculation model for calculation to obtain data sets of casing deformation, disk deformation, blade deformation and tip clearance of different engine processes, and the data sets of casing deformation, disk deformation, blade deformation and tip clearance of different engine processes are spliced ​​to obtain multi-transition state transient data.

[0043] Step 3: Perform sensitivity analysis on the multiple sets of variable data in step 2 to obtain sensitivity analysis results. Then, based on the sensitivity analysis results, perform dimensionality reduction processing on the multiple sets of variable data to obtain characteristic data sets of multiple engine processes, and use the transient data of multiple transition state processes as target variables to form target data sets of multiple engine processes.

[0044] Step 4: Divide the feature datasets and target datasets of multiple engine histories in step 3 into training sets, test sets, and validation sets.

[0045] Step 5: Using the long short-term memory neural network model, the feature data sets of the training set and test set in step 4 are used as the input of the neural network model, and the target data sets of the training set and test set in step 4 are used as the output of the neural network model. Through multiple training and testing, a mapping model between each feature parameter and the tip clearance is constructed to form a turbine transition state tip clearance prediction model.

[0046] Step 6: Using the estimation model constructed in step 5, input the data in the validation set in step 4 to verify the accuracy of the estimation model.

[0047] The above method is used to construct a turbine transient tip clearance estimation model with high prediction accuracy, thereby achieving efficient and accurate tip clearance estimation through the estimation model.

[0048] As shown in Figure 1, the turbine physical model in step 1 includes components such as active clearance control cooling pipes, low-pressure turbine casings, low-pressure turbine blades, and low-pressure turbine disks. The turbine air system flow path consists of a mainstream gas flow path and three cooling flow paths, including the casing ACC cooling flow path, the disk left cooling flow path, and the disk right cooling flow path. In the mainstream gas flow path, the mainstream gas heats the blade surface of the low-pressure turbine blades, the gas side of the low-pressure turbine casing, and the upper edge of the low-pressure turbine disk. In the casing ACC cooling flow path, the cooling airflow impinges on the casing surface through the impact holes on the ACC cooling pipes, thereby achieving the purpose of tip clearance control. In the disk left and right cooling flow paths, the cooling airflow enters the low-pressure turbine disk cavity through structures such as grate teeth and air bleed holes, cools the low-pressure turbine disk, and finally flows into the mainstream through structures such as leakage gaps.

[0049] In this embodiment, the turbine physical model is a four-stage turbine model.

[0050] According to the turbine physical model, the parameters related to the blade tip clearance and their variation range during the operation of the aircraft engine can be sorted out as shown in Figure 2. c1 ~Re c6 With T fc1 ~T fc6 The impact Reynolds number and cooling air temperature of the impact holes of the six cooling tubes from front to back are: With T fd1 is the air flow and temperature on the left side of the wheel, With T fd1 is the air flow and temperature on the left side of the wheel, N is the engine speed, Re1~Re4 and T fb1 ~T fb4 is the Reynolds number and gas temperature of the first to fourth stage cascade passages, and K is the number of engine cycles.

[0051] As shown in Figure 3, sensitivity analysis is performed on the offline tip clearance data to reduce the variable dimension, and data splicing is used to expand the data sample characteristics. On this basis, a long short-term memory neural network is used to fit the turbine transition state tip clearance model, and an online turbine transition state tip clearance estimation model is established to achieve tip clearance estimation through the turbine transition state tip clearance estimation model.

[0052] When establishing an offline tip clearance database, sampling is performed within the influencing parameter range to generate multiple sets of parameter data as initial values ​​for tip clearance calculations, resulting in a time series dataset corresponding to casing deformation, disk deformation, and blade deformation. Taking the splicing effect of casing deformation as an example, as shown in Figure 4, based on the LSTM prediction principle, predicting the transitional tip clearance within an engine's history necessarily requires sacrificing information from a period prior to the history to train the model. Therefore, a new variable, New_cycle, is used to splice data from six engine history cycles to achieve prediction of the transitional tip clearance for the entire history.

[0053] In this embodiment, after each process ends, the variable parameters and blade tip clearance are reset, returning to their initial state, and the next engine process begins. To let the neural network model know that the next process is about to begin, the variable parameters and blade tip clearance need to be reset. New_cycle contains only two values, 0 and 1. While within an engine process, New_cycle remains 0. When it enters a process switch, New_cycle changes to 1. Specifically, taking an engine process of 2400s and 4800 time steps as an example, New_cycle is 0 for the first 4799 steps and changes to 1 at the 4800th step. The remaining engine processes are processed in this way.

[0054] To reduce the neural network input dimensionality and improve computational efficiency and accuracy, the obtained dataset was subjected to sensitivity analysis. Taking the sensitivity analysis of casing deformation as an example, the results are shown in Figure 5. Darker points indicate greater sensitivity to the variable. From the numerous variables influencing component deformation, highly sensitive variables were selected as feature variables to form the feature dataset. Component deformation and blade tip clearance formed the target dataset.

[0055] In this embodiment, the Sobol sensitivity analysis method is performed on the variables affecting the tip clearance (i.e., the multiple groups of variable data in step 2), and the characteristic variables with strong sensitivity are screened out and the variable dimension is reduced. The formula for sensitivity analysis is: Y = f(X1, X2, ..., X k )

[0056] Among them, Y represents the research model, f(X) represents the expression of the research model, Var(Y) represents the total variance of the output, V i is the effect of a single parameter on Y, V ij For X i and X j The effect on Y, Indicated as X i First-order impact index, V 12…d Represented as X1 to X d The effect on Y, f represents the expansion of Y into a set of functions with increasing dimensions, f0 represents the 0-dimensional function of Y, f ij Expressed as a 2D function of Y, f 12…k It is expressed as a k-1 dimensional function of Y. The formula obtains the sensitivity of each parameter to the output based on the contribution ratio of each variance to the total variance, and different sensitivities are represented by different depth colors.

[0057] These data sets are divided into training set, validation set, and test set for training and testing the blade tip clearance estimation model. The first 3.4 engine history data are divided into training set, the next 0.6 engine history data are divided into validation set, and the remaining 2 engine history data are divided into test set.

[0058] Based on the above dataset, an LSTM-based model for transient tip clearance estimation was established. Considering that the tip clearance prediction task has been transformed into component deformation prediction, the tip clearance estimation model (Figure 6) was developed, taking deformation prediction at measuring point 5 of the casing as an example. The results, using transient component deformation and tip clearance prediction at measuring point 5 as an example, are shown in Figure 7. When predictions are made for casing deformation, disk deformation, and blade deformation at each measuring point separately (LSTM-data-single), the prediction error for transient component deformation is relatively small across the training, validation, and test sets, essentially coinciding with the real data. The maximum absolute errors for casing deformation, disk deformation, and blade deformation predictions are 0.041 mm, 0.097 mm, and 0.03 mm, respectively; and the maximum relative errors for casing deformation, disk deformation, and blade deformation predictions are 2.15%, 3.19%, and 1.91%, respectively. By predicting the deformation of transitional components, the transitional blade tip clearance can be predicted with high accuracy. The maximum absolute error in tip clearance prediction is 0.074 mm, and the maximum relative error is 3.59%. When simultaneously predicting the deformation of the casing, disk, and blades (LSTM-data-all), the average error between the predicted and true results is 8.77%. In practical applications, different prediction strategies can be selected based on needs. Furthermore, the transitional turbine tip clearance estimation model has very high computational efficiency, completing the calculation of an engine cycle in just 10 seconds, compared to 6 hours using traditional tip clearance calculation methods. This model can meet the needs of engineering calculations.

[0059] In this embodiment, the turbine transient tip clearance estimation model includes an input end, a data processing unit, a first LSTM unit, a normalization unit, a second LSTM unit, a denormalization unit, and an output end. The input end, the data processing unit, the first LSTM unit, the normalization unit, the second LSTM unit, the denormalization unit, and the output end are electrically connected in sequence. The specific calculation formula of the prediction process of each LSTM unit in the turbine transient tip clearance estimation model is: f t =σ(W f ·[h t-1 ,x t ]+b f ) i t =σ(W i ·[h t-1 ,x t ]+b i ) O t =σ(W o ·[h t-1 ,x t ]+b o) h t =O t tanh(C t )

[0060] Among them, x t is the feature quantity at the current moment, which is the input of the LSTM model, h t-1 With h t are the hidden state at the previous moment and the hidden state calculated at the current moment, C t-1 with C t are the memory stored at the previous moment and the memory calculated at the current moment, respectively, t 、i t With O t They are the calculation results of forget gate, input gate and output gate respectively. is the updated unit state, W and b are the weight and bias corresponding to each gate respectively, and the subscripts f, i, and O represent the parameters corresponding to the forget gate, input gate, and output gate respectively.

[0061] The present invention uses a one-dimensional engineering calculation method and a two-dimensional finite element method to establish a transition state tip clearance calculation model, and then substitutes multiple groups of variable data into the transition state tip clearance calculation model for calculation, and obtains data sets of casing deformation, wheel deformation, blade deformation and tip clearance of different engine processes, forming an offline high-precision turbine transition state tip clearance database, and uses data splicing, sensitivity analysis and other methods to enrich the transition state tip clearance sample data and reduce the variable dimension. Based on the long short-term memory neural network method, a mapping relationship between different engine processes, turbine thermodynamic boundary conditions and tip clearance is established, and an online turbine transition state tip clearance estimation model is constructed. The turbine transition state tip clearance estimation model can achieve high-efficiency and high-precision transition state tip clearance estimation of different engine processes.

[0062] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A method for estimating turbine tip clearance in transient state based on long short-term memory neural network, characterized in that: include: Step 1: Establish the physical model of the turbine components and the air system flow path. According to the aircraft engine design manual and empirical formula, obtain the flow rate and temperature of each air system flow path in the turbine physical model and the range of the turbine model speed. Then, construct the transition state blade tip clearance calculation model through the two-dimensional axisymmetric finite element calculation method and the one-dimensional engineering calculation method. Step 2: Sampling is performed within the variation range in step 1 to obtain multiple sets of variable data, and then the multiple sets of variable data are input into the transition state tip clearance calculation model for calculation to obtain data sets of casing deformation, disk deformation, blade deformation and tip clearance of different engine histories, and the casing deformation, disk deformation, blade deformation and tip clearance data of different engine histories are spliced ​​to obtain transition state component deformation and tip clearance data including multiple engine histories; Step 3: Perform sensitivity analysis on the transition state data in step 2 to obtain sensitivity analysis results, and then perform dimensionality reduction processing on the data based on the sensitivity analysis results to obtain dimensionality reduced data; Step 4: Divide the data of multiple engine histories that have undergone dimensionality reduction processing in step 3 into a training set, a test set, and a validation set; Step 5: Using the long short-term memory neural network model, the training set and test set in step 4 are used for model training and testing. Through multiple training and testing, a mapping model between each characteristic parameter and the blade tip clearance is constructed to form a turbine transition state blade tip clearance estimation model. Step 6: Using the estimation model constructed in step 5, input the data in the validation set in step 4 to verify the accuracy of the estimation model.

2. The method for estimating turbine tip clearance in transient state based on neural network according to claim 1, characterized in that: The specific steps for data splicing of the data sets of casing deformation, wheel deformation, blade deformation and blade tip clearance of different engine histories in step 2 are as follows: By citing a new variable New_cycle, the data sets of casing deformation, disk deformation, blade deformation and tip clearance of different engine histories are spliced, where New_cycle only contains two values, which are 0 and 1.

3. The method for estimating turbine tip clearance in transient state based on neural network according to claim 1, characterized in that: The formula for performing sensitivity analysis on multiple sets of variable data in step 3 is: Y = f(X1, X2, ..., X k ) Among them, Y is the research model defined by f(X1,X2,…,X k ) is expressed as, and then expanded into a set of dimensional increasing functions and f, and assuming that f is square integrable, the decomposition function is squared and integrated, and the left side of the integral is defined as the total output variance Var(Y), and the right side is the sum of the variances of each parameter, where V i is the effect of a single parameter on Y, V ij For X i and X j Effect on Y, V 12…d Represented as X1 to X d The impact on Y, and finally, according to the contribution ratio of each variance to the total variance, the sensitivity of each parameter to the output is obtained, that is, the first-order impact index 4. The method for estimating turbine transition state blade tip clearance based on long short-term memory neural network according to claim 1, characterized in that: The ratio of the training set, test set and validation set in step 4 is 17:3:

10.

5. The method for estimating turbine tip clearance in transient state based on long short-term memory neural network according to claim 1, characterized in that: The turbine transition state tip clearance estimation model in step 5 includes: An input terminal, used for inputting variable data and blade tip clearance into a data processing unit; The data processing unit unifies the format of the input data, divides it, and passes it into the LSTM unit; Normalization unit, which normalizes the data to improve the model accuracy and convergence speed; The first LSTM unit reads in the sequence data, performs feature extraction and outputs the time step information to the second LSTM unit; The second LSTM unit receives the time step information of the first LSTM unit, further extracts features and outputs prediction information; A denormalization unit denormalizes the normalized prediction data and restores it to the true value; And an output terminal, used for outputting the predicted value of blade tip clearance.

6. The method for estimating turbine tip clearance in transient state based on long short-term memory neural network according to claim 5, characterized in that: The specific calculation formula of the prediction process of each LSTM unit in the turbine transition state blade tip clearance prediction model is: t =σ(W f ·[h t-1 ,x t ]+b f ) t =σ(W i ·[h t-1 ,x t ]+b i ) O t =σ(W o ·[h t-1 ,x t ]+b o ) t =O t tanh(C t ) Among them, x t is the feature quantity at the current moment, which is the input of the LSTM model, h t-1 With h t are the hidden state at the previous moment and the hidden state calculated at the current moment, C t-1 With C t are the memory stored at the previous moment and the memory calculated at the current moment, respectively, t 、i t With O t They are the calculation results of the forget gate, input gate, and output gate respectively. is the updated unit state, W and b are the weight and bias corresponding to each gate, and the subscripts f, i, and O represent the parameters corresponding to the forget gate, input gate, and output gate, respectively.

Images