Single crystal turbine blade solidification interface regulation and control method based on LSTM

Abstract

The invention relates to the technical field of single-crystal high-temperature alloy directional solidification control, in particular to a single-crystal turbine blade solidification interface regulation and control method based on LSTM (Long Short Term Memory), which comprises the following steps: collecting time sequence data of a multi-source sensor in real time, constructing a multi-dimensional spatial-temporal feature vector after synchronization preprocessing, and inputting a pre-trained LSTM model to predict the morphology and position of a solidification interface. Deviation is calculated in combination with an ideal process target curve, a drawing rate and heating power adjusting instruction is generated through a self-adaptive fuzzy PID controller, and an incremental learning mechanism is introduced to continuously optimize a model. According to the application, high-precision closed-loop dynamic regulation and control of the solidification interface can be realized, and the tissue quality and the process stability of the single crystal blade are improved.

Description

Technical Field

[0001] This invention belongs to the field of advanced manufacturing and intelligent control technology, specifically a method for controlling the solidification interface of single-crystal turbine blades based on LSTM. Background Technology

[0002] Single-crystal turbine blades are core hot-end components of aero-engines and gas turbines, operating in environments characterized by high temperature, high pressure, and high corrosion. This places extremely high demands on the crystal quality and uniformity of the microstructure of the blades. Directional solidification technology is the core technology for preparing high-performance single-crystal turbine blades. Its core objective is to maintain the flatness of the solid-liquid interface during solidification, avoiding defects such as small-angle grain boundaries, eutectic phase precipitation, and interface depressions. The stability of the solid-liquid interface directly depends on the precise control of process parameters such as the furnace temperature field and the pulling rate.

[0003] Currently, interface control in the directional solidification process of single-crystal turbine blades mostly adopts the traditional proportional-integral-derivative (PID) control algorithm. This algorithm is simple in structure and easy to implement, but it has obvious limitations: On the one hand, the single-crystal solidification process is a complex time-series process with strong nonlinearity, large lag, and multivariable coupling. The temperature field changes in the furnace have a significant long-range dependence, and abrupt changes in the cross-sectional dimensions of the turbine blade tenon, blade body, etc., will cause thermal field disturbances. The fixed parameters of traditional PID control cannot adapt to the fluctuation of operating conditions and equipment state drift, and are prone to problems such as control lag, large overshoot, and insufficient accuracy. On the other hand, the evolution state of the solidification interface is difficult to measure directly in real time. Existing control methods mostly rely on experience to set process parameters, lacking the ability to accurately predict the interface evolution and cannot predict the occurrence of interface defects in advance, resulting in poor control effect.

[0004] To address these issues, some studies have introduced neural network technology for solidification interface prediction. However, existing neural network models mostly employ shallow structures, failing to effectively capture the long-range dependencies of temperature field changes. Furthermore, overfitting is prone to occur during model training, resulting in prediction accuracy that is insufficient to meet actual production requirements. Simultaneously, existing control methods lack effective data feedback and model update mechanisms, making it impossible to dynamically optimize control parameters based on actual solidification microstructure quality. This hinders the continuous improvement of control accuracy, restricting the fabrication quality and production efficiency of single-crystal turbine blades.

[0005] Therefore, in view of the problems of low control precision, poor adaptability and insufficient prediction ability in the existing technology, there is an urgent need for a method that can accurately predict the evolution of solidification interface, dynamically adapt to the fluctuation of operating conditions and realize closed-loop optimization control, so as to solve the interface defect problem in the directional solidification process of single crystal turbine blades and improve the preparation quality of single crystal turbine blades. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of existing technologies in the control of solidification interface of single-crystal turbine blades, such as low precision, poor adaptability, lack of accurate prediction and closed-loop optimization capabilities. This invention provides a method for controlling the solidification interface of single-crystal turbine blades based on LSTM. The method accurately predicts the evolution state of the solidification interface through an LSTM model, combines fuzzy PID adaptive control to achieve dynamic parameter correction, and uses an incremental learning mechanism to achieve continuous model optimization. Ultimately, this method achieves high-precision closed-loop dynamic control of the solidification interface, reduces interface defects, and improves the fabrication quality of single-crystal turbine blades.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following solution: A method for controlling the solidification interface of a single-crystal turbine blade based on LSTM includes the following steps: S1. Real-time acquisition of multi-source sensor timing data during the directional solidification process of single crystal turbine blades in a single crystal growth furnace; S2. The collected multi-source sensor time-series data is preprocessed for synchronization, and then a multi-dimensional spatiotemporal feature vector that can reflect the dynamic evolution law of the solidification process is constructed based on the sliding window technique. S3. Input the multidimensional spatiotemporal feature vector into the pre-trained solidification interface evolution prediction model. The model uses a long short-term memory neural network as its core architecture to output the morphological feature parameters of the solidification interface and the absolute position of the solidification interface within a future preset time period. S4. Compare the predicted solidification interface morphology parameters and the absolute position of the solidification interface with the preset ideal process target curve, calculate the deviation value and the deviation change rate, and then generate the pulling rate correction command and heating power compensation command through the adaptive controller. S5. The pulling rate correction command and heating power compensation command are output to the single crystal growth furnace actuator. The single crystal growth furnace actuator adjusts the temperature gradient and cooling rate at the solid-liquid interface in real time, and finally realizes the closed-loop dynamic control of the solidification interface of the single crystal turbine blade. S6. During the closed-loop dynamic control process, an online-offline combined incremental learning and update mechanism is established. Control data and solidification quality results during actual operation are collected regularly. Then, the weight parameters of the long short-term memory neural network model are incrementally learned using the stochastic gradient descent algorithm to achieve continuous optimization of the model and adapt to the equipment state drift and operating condition fluctuation of the single crystal growth furnace.

[0008] This method addresses the challenge of precisely controlling the solidification interface during the directional solidification of single-crystal turbine blades. It achieves dynamic and precise control of the solidification interface through six steps: multi-source sensor time-series data acquisition and preprocessing, multi-dimensional spatiotemporal feature construction, LSTM model prediction, adaptive closed-loop control, and incremental learning model optimization. The core of this method leverages the time-series prediction advantages of LSTM neural networks to capture the dynamic evolution of the solidification process. Combined with adaptive control strategies, it corrects process parameters in real time and adapts to equipment state drift and operating condition fluctuations through incremental learning, ultimately ensuring the stability of the solidification microstructure quality of the single-crystal turbine blades. Compared to existing methods for controlling the solidification interface of single-crystal turbine blades, which often employ fixed process parameters or simple feedback control, making it difficult to cope with the dynamic changes in multi-source time-series data, equipment state drift, and operating condition fluctuations during solidification, and suffering from limitations such as control lag, insufficient accuracy, and poor adaptability, the innovation of this method is mainly reflected in three aspects: First, it adopts an LSTM neural network as the core prediction architecture, combined with sliding window technology to construct multi-dimensional spatiotemporal feature vectors, breaking through the limitation of traditional methods in capturing the temporal correlation of the solidification process, and can accurately predict the solidification interface parameters within a preset time period, providing support for advance control; Second, it constructs a closed-loop control mechanism of "prediction-comparison-correction", combined with an adaptive controller to generate targeted process correction commands, realizing real-time dynamic adjustment of temperature gradient and cooling rate, solving the problems of traditional control lag and insufficient accuracy; Third, it introduces an online-offline combined incremental learning mechanism, continuously optimizing model weights through actual control data, effectively adapting to equipment state drift and operating condition fluctuations in single-crystal growth furnaces, improving the stability and applicability of the method, and distinguishing it from the limitations of existing fixed models that are difficult to adapt to complex operating conditions.

[0009] Preferably, in step S1, the multi-source sensor time-series data includes the output power of the single crystal growth furnace heater, the current pulling rate of the crucible, the multi-point temperature sampling value of the melting chamber, and the real-time liquid metal level. The liquid metal refers to the high-temperature alloy liquid in a molten state used to prepare single crystal turbine blades.

[0010] Preferably, in step S2, the synchronization preprocessing of the multi-source sensor time-series data specifically includes the following: using Gaussian filtering to reduce noise in the sampled data, wherein the kernel function of the Gaussian filter is 5×5 and the standard deviation is 1.2; using the Min-Max normalization method to normalize the data of each dimension to the [0,1] interval to eliminate the influence of data dimensions; the step size of the sliding window is reasonably set according to the sampling frequency of the multi-source sensor, and its window length is set as the characteristic period of the thermal hysteresis effect in single crystal growth.

[0011] Preferably, in step S3, the long short-term memory neural network includes an input layer, a hidden layer, and an output layer; wherein the hidden layer is configured as two layers of long short-term memory neural network units, each layer of long short-term memory neural network units has 64 neurons, and each layer of long short-term memory neural network units includes a forget gate, an input gate, and an output gate. This structure can effectively capture the long-range dependence of temperature field changes during solidification; the loss function of the solidification interface evolution prediction model adopts a weighted combination of root mean square error and mean absolute percentage error, with a weight ratio of 7:3. An L2 regularization term is introduced to prevent overfitting when predicting nonlinear interface fluctuations, wherein the coefficient of the L2 regularization term is 1e. -5 .

[0012] Preferably, the solidification interface morphology features in step S4 include the solidification interface curvature, radial temperature gradient, and solid-liquid coexistence zone height; the absolute position of the solidification interface is the vertical distance between the center point of the solid-liquid interface and the bottom of the induction coil, which is measured using a laser rangefinder, and the position of the bottom of the induction coil is clearly defined with reference to the reference surface of the single crystal growth furnace body.

[0013] Preferably, the specific implementation process of step S4 is as follows: First, the predicted solidification interface morphology characteristic parameters and the absolute position of the solidification interface are compared and calculated with the preset ideal process target curve. The comparison adopts the point-by-point corresponding comparison method, that is, the predicted solidification interface curvature, radial temperature gradient, solid-liquid coexistence zone height, and absolute position of the solidification interface are compared with the standard values ​​of the corresponding parameters in the ideal curve one by one. After the above point-by-point comparison is completed, the corresponding deviation value is calculated for each parameter using the absolute deviation method. The unified calculation formula is deviation value = |predicted value - standard value|. At the same time, the deviation change rate of each parameter is calculated, that is, deviation change rate = (current deviation value - previous deviation value) / sampling time interval. The solidification interface curvature threshold is then set to ±5%, which is the allowable fluctuation range for a flat interface. The calculated solidification interface curvature deviation value is compared with this ±5% curvature threshold. If the curvature deviation value exceeds the threshold, the interface is determined to deviate from a flat state. If a convex liquid phase is detected at the interface, a linkage command is generated to increase the pulling rate and decrease the upper heating power. Here, "upper" refers to the upper heating area of ​​the single crystal growth furnace. The pulling rate adjustment range is 5%-10%, and the upper heating power adjustment range is 3%-8%. If a concave liquid phase is detected at the interface, a linkage command is generated to decrease the pulling rate and increase the upper heating power. The adjustment range is the same as above.

[0014] This method specifically optimizes step S4: it clarifies the morphological characteristics of the solidification interface, including curvature, radial temperature gradient, and the height of the solid-liquid coexistence zone; it defines the absolute position of the solidification interface as the vertical distance between the center point of the solid-liquid interface and the bottom of the induction coil with the furnace body reference surface as a reference, and this distance is measured using a laser rangefinder; in the specific implementation of step S4, the predicted parameters are compared with the ideal process target curve using a point-to-point comparison method, the deviation value and deviation change rate are calculated using the absolute deviation method, and a ±5% allowable fluctuation threshold for the curvature of the solidification interface is set. Based on the interface deviation (convex or concave liquid phase), a linkage adjustment command is generated for the corresponding pulling rate (adjustment range 5%-10%) and the upper heating power (adjustment range 3%-8%), thereby improving the accuracy of solidification interface control and ensuring the solidification structure quality of the single-crystal turbine blade.

[0015] Preferably, the adaptive controller in step S4 employs a fuzzy proportional-integral-derivative (PI-DI) control algorithm. The fuzzy rule table of this algorithm contains five fuzzy subsets: negative large, negative small, zero, positive small, and positive large, which correspond to the fuzzy ranges of the deviation values ​​and deviation change rates calculated in step S4, respectively. The proportional gain correction range is set to 0.5-1.0. The adaptive controller can dynamically correct the proportional-integral-derivative parameters online based on the deviation trend predicted by the long short-term memory neural network, i.e., the trend of the deviation values ​​and deviation change rates of each parameter calculated in step S4. This allows for the precise generation of pulling rate correction commands and heating power compensation commands, thereby effectively compensating for thermal field disturbances caused by abrupt changes in the cross-sectional dimensions of the single-crystal turbine blade tenon, blade body, etc., ensuring control accuracy.

[0016] Preferably, the control data in step S6 is the pulling rate correction command and heating power compensation command generated in S4, and the actual adjustment parameters in S5. The solidification structure quality result is the metallographic detection result after the solidification of the single crystal turbine blade. The specific period of the period is for each batch of single crystal turbine blades produced. The solidification structure quality result is judged based on the metallographic detection data. The judgment criteria for the solidification structure quality result specifically include four categories: qualified, small-angle grain boundary, eutectic phase precipitation, and severe interface depression. This is used to provide clear quality feedback basis for incremental learning and assist in the optimization of model weight parameters.

[0017] The beneficial effects of this invention are as follows: 1. High prediction accuracy and effective capture of long-range temperature field dependence: This invention uses a two-layer LSTM neural network to construct a solidification interface evolution prediction model. Each layer has 64 neurons and includes a three-gate mechanism, which can accurately capture the long-range dependence of temperature field changes during single crystal solidification and adapt to the temporal effects of thermal hysteresis. At the same time, it uses a weighted combination loss function of root mean square error and mean absolute percentage error plus L2 regularization to effectively avoid model overfitting, improve the prediction accuracy of interface evolution parameters, and provide a reliable basis for subsequent precise control.

[0018] 2. High control precision and strong adaptability: The fuzzy PID adaptive control algorithm is adopted. By using the fuzzy intervals of the deviation value and the rate of change of deviation corresponding to the five fuzzy subsets, the PID parameters are dynamically corrected online. This can effectively compensate for the thermal field disturbance caused by the sudden change in the blade cross-sectional size, and solve the problems of fixed parameters, control lag and poor adaptability of traditional PID control. At the same time, only the curvature deviation value is used as the core of the interface deviation judgment, and the other parameters are used as auxiliary decision-making basis, so as to balance the control precision and efficiency.

[0019] 3. Achieve closed-loop optimization and continuously improve control effect: Establish an online-offline combined incremental learning and update mechanism, collect control data and metallographic inspection results on a cycle of each furnace production, optimize the LSTM model weights through stochastic gradient descent algorithm, realize dynamic adaptation of model with equipment status and operating condition fluctuations, continuously improve prediction and control accuracy, and reduce interface defects such as small-angle grain boundaries and eutectic phase precipitation.

[0020] 4. High feasibility and fits the actual production scenario: The multi-source sensor data acquisition, preprocessing method and parameter setting (such as sliding window period of 833 milliseconds, threshold ±5%) of the present invention are all in line with the actual production conditions of single crystal turbine blades. The steps are clear and the parameters are well-defined. No complicated additional equipment is required. It can be directly adapted to existing single crystal growth furnaces and is easy to promote and apply in industrial applications. Attached Figure Description

[0021] Figure 1 This is a flowchart of a method for controlling the solidification interface of a single-crystal turbine blade based on LSTM according to the present invention.

[0022] Figure 2 This is a time-series data diagram of the multi-source sensor of the present invention.

[0023] Figure 3 This is a diagram of the long short-term memory neural network of the present invention. Detailed Implementation

[0024] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0025] like Figure 1 As shown, Figure 1 This is a flowchart illustrating a method for controlling the solidification interface of a single-crystal turbine blade based on LSTM, according to the present invention. The method includes the following steps: S1. Real-time acquisition of multi-source sensor timing data during the directional solidification process of single crystal turbine blades in a single crystal growth furnace; S2. The collected multi-source sensor time-series data is preprocessed for synchronization, and then a multi-dimensional spatiotemporal feature vector that can reflect the dynamic evolution law of the solidification process is constructed based on the sliding window technique. S3. Input the multidimensional spatiotemporal feature vector into the pre-trained solidification interface evolution prediction model. The model uses a long short-term memory neural network as its core architecture to output the morphological feature parameters of the solidification interface and the absolute position of the solidification interface within a future preset time period. S4. Compare the predicted solidification interface morphology parameters and the absolute position of the solidification interface with the preset ideal process target curve, calculate the deviation value and the deviation change rate, and then generate the pulling rate correction command and heating power compensation command through the adaptive controller. S5. The pulling rate correction command and heating power compensation command are output to the single crystal growth furnace actuator. The single crystal growth furnace actuator adjusts the temperature gradient and cooling rate at the solid-liquid interface in real time, and finally realizes the closed-loop dynamic control of the solidification interface of the single crystal turbine blade. S6. During the closed-loop dynamic control process, an online-offline combined incremental learning and update mechanism is established. Control data and solidification quality results during actual operation are collected regularly. Then, the weight parameters of the long short-term memory neural network model are incrementally learned using the stochastic gradient descent algorithm to achieve continuous optimization of the model and adapt to the equipment state drift and operating condition fluctuation of the single crystal growth furnace.

[0026] like Figure 2 As shown, Figure 2 This is a time-series data diagram of the multi-source sensors of the present invention. The time-series data of the multi-source sensors includes the output power of the heater of the single crystal growth furnace, the current pulling rate of the crucible, the multi-point temperature sampling values ​​of the melting chamber, and the real-time liquid metal level. The liquid metal refers to the high-temperature alloy liquid in a molten state used to prepare single crystal turbine blades.

[0027] Preferably, in step S2, the synchronization preprocessing of the multi-source sensor time-series data specifically includes the following: using Gaussian filtering to reduce noise in the sampled data, wherein the kernel function of the Gaussian filter is 5×5 and the standard deviation is 1.2; using the Min-Max normalization method to normalize the data of each dimension to the [0,1] interval to eliminate the influence of data dimensions; the step size of the sliding window is reasonably set according to the sampling frequency of the multi-source sensor, and its window length is set as the characteristic period of the thermal hysteresis effect in single crystal growth.

[0028] like Figure 3 As shown, Figure 3 This is a diagram of the Long Short-Term Memory (LSTM) neural network of the present invention. The LSM neural network includes an input layer, a hidden layer, and an output layer. The hidden layer consists of two LSM neural network units, each with 64 neurons. Each LSM neural network unit includes a forget gate, an input gate, and an output gate. This structure effectively captures the long-range dependence of temperature field changes during solidification. The loss function of the solidification interface evolution prediction model uses a weighted combination of root mean square error and mean absolute percentage error, with a weight ratio of 7:3. An L2 regularization term is introduced to prevent overfitting when predicting nonlinear interface fluctuations. The coefficient of the L2 regularization term is 1e. -5 .

[0029] Preferably, the solidification interface morphology features in step S4 include the solidification interface curvature, radial temperature gradient, and solid-liquid coexistence zone height; the absolute position of the solidification interface is the vertical distance between the center point of the solid-liquid interface and the bottom of the induction coil, which is measured using a laser rangefinder, and the position of the bottom of the induction coil is clearly defined with reference to the reference surface of the single crystal growth furnace body.

[0030] Preferably, the specific implementation process of step S4 is as follows: First, the predicted solidification interface morphology characteristic parameters and the absolute position of the solidification interface are compared and calculated with the preset ideal process target curve. The comparison adopts the point-by-point corresponding comparison method, that is, the predicted solidification interface curvature, radial temperature gradient, solid-liquid coexistence zone height, and absolute position of the solidification interface are compared with the standard values ​​of the corresponding parameters in the ideal curve one by one. After the above point-by-point comparison is completed, the corresponding deviation value is calculated for each parameter using the absolute deviation method. The unified calculation formula is deviation value = |predicted value - standard value|. At the same time, the deviation change rate of each parameter is calculated, that is, deviation change rate = (current deviation value - previous deviation value) / sampling time interval. The solidification interface curvature threshold is then set to ±5%, which is the allowable fluctuation range for a flat interface. The calculated solidification interface curvature deviation value is compared with this ±5% curvature threshold. If the curvature deviation value exceeds the threshold, the interface is determined to deviate from a flat state. If a convex liquid phase is detected at the interface, a linkage command is generated to increase the pulling rate and decrease the upper heating power. Here, "upper" refers to the upper heating area of ​​the single crystal growth furnace. The pulling rate adjustment range is 5%-10%, and the upper heating power adjustment range is 3%-8%. If a concave liquid phase is detected at the interface, a linkage command is generated to decrease the pulling rate and increase the upper heating power. The adjustment range is the same as above.

[0031] Preferably, the adaptive controller in step S4 employs a fuzzy proportional-integral-derivative (PI-DI) control algorithm. The fuzzy rule table of this algorithm contains five fuzzy subsets: negative large, negative small, zero, positive small, and positive large, which correspond to the fuzzy ranges of the deviation values ​​and deviation change rates calculated in step S4, respectively. The proportional gain correction range is set to 0.5-1.0. The adaptive controller can dynamically correct the proportional-integral-derivative parameters online based on the deviation trend predicted by the long short-term memory neural network, i.e., the trend of the deviation values ​​and deviation change rates of each parameter calculated in step S4. This allows for the precise generation of pulling rate correction commands and heating power compensation commands, thereby effectively compensating for thermal field disturbances caused by abrupt changes in the cross-sectional dimensions of the single-crystal turbine blade tenon, blade body, etc., ensuring control accuracy.

[0032] Preferably, the control data in step S6 is the pulling rate correction command and heating power compensation command generated in S4, and the actual adjustment parameters in S5. The solidification structure quality result is the metallographic detection result after the solidification of the single crystal turbine blade. The specific period of the period is for each batch of single crystal turbine blades produced. The solidification structure quality result is judged based on the metallographic detection data. The judgment criteria for the solidification structure quality result specifically include four categories: qualified, small-angle grain boundary, eutectic phase precipitation, and severe interface depression. This is used to provide clear quality feedback basis for incremental learning and assist in the optimization of model weight parameters.

[0033] The present invention will be further described in detail below with reference to specific embodiments. This embodiment takes the directional solidification control of a single-crystal turbine blade (made of nickel-based superalloy) of an aero-engine as an example, and uses the method described in the present invention to achieve closed-loop dynamic control of the solidification interface. The specific implementation process is as follows.

[0034] Step S1: Multi-source sensor time-series data acquisition Multi-source sensors were installed inside the single crystal growth furnace to collect real-time time-series data during the directional solidification process. The sampling frequency was set to 60Hz. The collected multi-source sensor time-series data included: the output power of the single crystal growth furnace heater (collection range 0-100kW, accuracy ±0.1kW), the current crucible pulling speed (collection range 0.1-1.0mm / min, accuracy ±0.01mm / min), the multi-point temperature sampling values ​​of the melting chamber (6 infrared temperature probes were installed at equal angles along the circumference of the inner wall of the melting chamber, with a temperature range of 1300-1600℃ and an accuracy better than ±1.5℃), and the real-time liquid metal level (collection range 50-200mm, accuracy ±0.5mm). The liquid metal was a nickel-based superalloy liquid (molten state, temperature maintained at around 1500℃) used to prepare single crystal turbine blades.

[0035] Step S2: Data preprocessing and construction of multidimensional spatiotemporal feature vectors The time-series data from the acquired multi-source sensors undergoes synchronization preprocessing, specifically as follows: 1. Noise Reduction Processing: Gaussian filtering is used to reduce noise in all sampled data. The kernel size of the Gaussian filter is set to 5×5 and the standard deviation is set to 1.2, which effectively filters out random noise in the sensor sampling process. 2. Normalization: The Min-Max normalization method is used to normalize the data of each dimension (heater power, pumping speed, temperature, liquid level) to the [0,1] interval, eliminating the influence of different data units and ensuring data comparability; 3. Feature Vector Construction: Multidimensional spatiotemporal feature vectors are constructed based on the sliding window technique. The step size of the sliding window is set to 16.7ms according to the 60Hz sampling frequency (consistent with the sampling period to ensure no omission or redundancy of time series data). The window length is set to 833ms, which is the characteristic period of thermal hysteresis effect in single crystal growth (i.e., the delay time for the solid-liquid interface to respond after the temperature field changes). By sliding the window, the temporal features within each window are extracted to construct a multidimensional spatiotemporal feature vector for subsequent model input.

[0036] Step S3: Training and Prediction of Solidification Interface Evolution Prediction Model 1. Model Construction: A solidification interface evolution prediction model with LSTM neural network as the core architecture is constructed. The model includes an input layer, a hidden layer, and an output layer. The number of neurons in the input layer is consistent with the dimension of the multidimensional spatiotemporal feature vector. The hidden layer consists of two LSTM units, each with 64 neurons. Each LSTM unit contains a forget gate, an input gate, and an output gate to capture the long-range dependence of temperature field changes. The output layer outputs the solidification interface morphology features (curvature, radial temperature gradient, and solid-liquid coexistence zone height) and the absolute position of the solidification interface within the next 10 seconds (preset time period).

[0037] 2. Model Training: Historical production data (directional solidification data of 100 furnaces of single-crystal turbine blades) were collected and divided into a training set (80%) and a test set (20%). The training set was used for model training, and the test set was used for model validation. The loss function adopted a weighted combination of root mean square error and mean absolute percentage error, with a weight ratio of 7:3. An L2 regularization term (with a coefficient of 1e) was also introduced. -5 To prevent overfitting, the model is trained using the stochastic gradient descent algorithm with 1000 training iterations and a learning rate of 0.001 until the model loss function converges, thus obtaining a pre-trained solidification interface evolution prediction model.

[0038] 3. Interface prediction: Input the multidimensional spatiotemporal feature vector constructed in step S2 into the trained model. The model outputs the solidification interface curvature, radial temperature gradient, solid-liquid coexistence zone height (solidification interface morphology feature parameters) and the absolute position of the solidification interface (the vertical distance between the center point of the solid-liquid interface and the bottom of the induction coil, with the bottom of the induction coil defined by the furnace body reference plane) within the next 10 seconds.

[0039] Step S4: Deviation Calculation and Control Command Generation 1. Comparison and Deviation Calculation: The solidification interface morphology parameters and absolute positions predicted by the model are compared point by point with the preset ideal process target curve. The point-by-point comparison method is as follows: the predicted curvature, radial temperature gradient, solid-liquid coexistence zone height, and absolute position are compared with the standard values ​​of the corresponding parameters in the ideal curve (ideal curvature is 0, radial temperature gradient standard value is 5℃ / mm, solid-liquid coexistence zone height standard value is 10mm, and absolute position standard value is 80mm). For each parameter, the deviation value is calculated using the absolute deviation method (deviation value = |predicted value - standard value|), and the deviation change rate of each parameter is calculated (the ratio of the difference between two adjacent deviation values ​​to the 16.7ms sampling time interval).

[0040] 2. Interface Deviation Judgment: The solidification interface curvature threshold is set to ±5% (the allowable fluctuation range of the flat state). Only the curvature deviation value is compared with this threshold. If the curvature deviation value exceeds ±5%, the interface is determined to deviate from the flat state and adjustment is required. The deviation values ​​and deviation change rates of other parameters (radial temperature gradient, solid-liquid coexistence zone height, absolute position) are used as auxiliary decision-making basis to determine the cause of the curvature deviation and the adjustment command range.

[0041] 3. Control Command Generation: The adaptive controller adopts a fuzzy PID control algorithm. The fuzzy rule table contains 5 fuzzy subsets (negative large, negative small, zero, positive small, positive large), which correspond to the fuzzy range of deviation value and deviation change rate, respectively. The proportional gain correction range is set to 0.5-1.0. The controller dynamically corrects the PID parameters online based on the deviation trend of the output (the change trend of each parameter deviation value and deviation change rate) predicted by the LSTM model. If a convex liquid phase is detected at the interface (the curvature deviation value is positive and exceeds the threshold), a linkage command is generated to increase the pulling speed (adjustment range 5%-10%, 8% in this embodiment) and decrease the upper heating power (adjustment range 3%-8%, 5% in this embodiment). If a concave liquid phase is detected at the interface (the curvature deviation value is negative and exceeds the threshold), a linkage command is generated to decrease the pulling speed (8%) and increase the upper heating power (5%). Here, "upper" refers to the upper heating area of ​​the single crystal growth furnace.

[0042] Step S5: Execution of closed-loop dynamic control The pulling rate correction command and heating power compensation command generated in step S4 are output to the single crystal growth furnace actuator. The actuator adjusts in real time according to the command: the pulling rate is adjusted by the crucible drive device, and the upper heating power is adjusted by the heater power adjustment module. In turn, the temperature gradient and cooling rate at the solid-liquid interface are adjusted in real time, and the curvature of the solidification interface is controlled within the allowable range of ±5%, so as to realize the closed-loop dynamic control of the solidification interface.

[0043] Step S6: Incremental Model Learning and Continuous Optimization The control data and solidification quality results during actual operation are collected periodically, with each batch of single-crystal turbine blades produced as a cycle: the control data includes the pulling rate correction command and heating power compensation command generated by S4 and the actual adjustment parameters (actual pulling rate and actual heating power) of S5; the solidification quality results are the metallographic test results after the blade solidifies, and are judged according to four categories: qualified, small-angle grain boundary, eutectic phase precipitation, and severe interface depression.

[0044] The collected control data and quality results are used as feedback information. The stochastic gradient descent algorithm is used to incrementally learn the weight parameters of the LSTM model, update the model parameters, and achieve continuous optimization of the model. For example, if the metallographic test results show small-angle grain boundaries, it indicates that the model's predicted curvature is biased. The model weights are adjusted through incremental learning to improve the prediction and control accuracy of the next batch and adapt to equipment state drift and operating condition fluctuations.

[0045] (III) Verification of Implementation Results Using the method described in this embodiment, 50 batches of single-crystal turbine blades were continuously produced. Compared with the traditional PID control method, the results are as follows: 1. Interface defect rate: The blade solidification interface defect rate (small angle grain boundary, eutectic phase precipitation, severe interface depression) of the method of the present invention is 2%, while that of the traditional method is 12%, and the defect rate is significantly reduced. 2. Control precision: The fluctuation range of the solidification interface curvature is controlled within ±4.5% (better than the ±5% threshold), while the fluctuation range of the traditional method is ±7.8%, thus improving the control precision by 42%. 3. Model adaptability: After incremental learning through 50 furnaces, the model prediction error has been reduced to less than 3%, which can effectively adapt to equipment state drift and operating condition fluctuations, and significantly improve the stability of regulation.

[0046] Verification results show that the method of the present invention can achieve high-precision closed-loop dynamic control of the solidification interface of single-crystal turbine blades, effectively reduce interface defects, improve the quality of blade preparation, and meet the performance requirements of core components of aero-engines.

[0047] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Based on the technical essence of the present invention, any simple modifications, equivalent substitutions, and improvements made to the above embodiments within the spirit and principles of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for controlling the solidification interface of a single-crystal turbine blade based on LSTM, characterized in that, Includes the following steps: S1. Real-time acquisition of multi-source sensor timing data during the directional solidification process of single crystal turbine blades in a single crystal growth furnace; S2. The collected multi-source sensor time-series data is preprocessed for synchronization, and then a multi-dimensional spatiotemporal feature vector that can reflect the dynamic evolution law of the solidification process is constructed based on the sliding window technique. S3. Input the multidimensional spatiotemporal feature vector into the pre-trained solidification interface evolution prediction model. The model uses a long short-term memory neural network as its core architecture to output the morphological feature parameters of the solidification interface and the absolute position of the solidification interface within a future preset time period. S4. Compare the predicted solidification interface morphology parameters and the absolute position of the solidification interface with the preset ideal process target curve, calculate the deviation value and the deviation change rate, and then generate the pulling rate correction command and heating power compensation command through the adaptive controller. S5. The pulling rate correction command and heating power compensation command are output to the single crystal growth furnace actuator. The single crystal growth furnace actuator adjusts the temperature gradient and cooling rate at the solid-liquid interface in real time, and finally realizes closed-loop dynamic control of the solidification interface of the single crystal turbine blade.

2. The method for controlling the solidification interface of a single-crystal turbine blade based on LSTM according to claim 1, characterized in that, In step S1, the multi-source sensor time-series data includes the output power of the single crystal growth furnace heater, the current pulling rate of the crucible, the multi-point temperature sampling values ​​of the melting chamber, and the real-time liquid level height of the liquid metal.

3. The method for controlling the solidification interface of a single-crystal turbine blade based on LSTM according to claim 1, characterized in that, In step S2, the synchronization preprocessing of the multi-source sensor time-series data specifically includes the following: using Gaussian filtering to reduce noise in the sampled data, wherein the kernel function of the Gaussian filter is 5×5 and the standard deviation is 1.2; using the Min-Max normalization method to normalize the data of each dimension to the [0,1] interval to eliminate the influence of data dimensions; the step size of the sliding window is reasonably set according to the sampling frequency of the multi-source sensor, and its window length is set as the characteristic period of the thermal hysteresis effect in single crystal growth.

4. The method for controlling the solidification interface of a single-crystal turbine blade based on LSTM according to claim 1, characterized in that, In step S3, the long short-term memory neural network includes an input layer, a hidden layer, and an output layer; wherein the hidden layer is configured as two layers of long short-term memory neural network units, each layer of long short-term memory neural network units has 64 neurons, and each layer of long short-term memory neural network units includes a forget gate, an input gate, and an output gate; the loss function of the solidification interface evolution prediction model adopts a weighted combination of root mean square error and mean absolute percentage error, with the weight ratio of the two set at 7:3, and an L2 regularization term is introduced to prevent overfitting when predicting nonlinear interface fluctuations, wherein the coefficient of the L2 regularization term is 1e -5 .

5. The method for controlling the solidification interface of a single-crystal turbine blade based on LSTM according to claim 1, characterized in that, The solidification interface morphology parameters mentioned in step S4 include the solidification interface curvature, radial temperature gradient, and solid-liquid coexistence zone height; the absolute position of the solidification interface is the vertical distance between the center point of the solid-liquid interface and the bottom of the induction coil. This vertical distance is measured using a laser rangefinder, and the position of the bottom of the induction coil is clearly defined with reference to the reference surface of the single crystal growth furnace body.

6. The method for controlling the solidification interface of a single-crystal turbine blade based on LSTM according to claim 5, characterized in that, The specific implementation process of step S4 is as follows: First, the predicted solidification interface morphology characteristic parameters and the absolute position of the solidification interface are compared and calculated with the preset ideal process target curve. The comparison adopts the point-by-point corresponding comparison method, that is, the predicted solidification interface curvature, radial temperature gradient, solid-liquid coexistence zone height, and absolute position of the solidification interface are compared with the standard values ​​of the corresponding parameters in the ideal curve one by one. After the point-by-point comparison is completed, the corresponding deviation value is calculated for each parameter using the absolute deviation method. The unified calculation formula is deviation value = |predicted value - standard value|. At the same time, the deviation change rate of each parameter is calculated, that is, deviation change rate = (current deviation value - previous deviation value) / sample. The sampling time interval is then set to ±5% for the solidification interface curvature threshold, which is the allowable fluctuation range for a flat interface. The calculated solidification interface curvature deviation value is compared with the ±5% curvature threshold. If the curvature deviation value exceeds the threshold, the interface is determined to deviate from a flat state. If a convex liquid phase is detected at the interface, a linkage command is generated to increase the pulling rate and decrease the upper heating power. The upper part refers to the upper heating area of ​​the single crystal growth furnace. The pulling rate adjustment range is 5%-10%, and the upper heating power adjustment range is 3%-8%. If a concave liquid phase is detected at the interface, a linkage command is generated to decrease the pulling rate and increase the upper heating power.

7. The method for controlling the solidification interface of a single-crystal turbine blade based on LSTM according to claim 6, characterized in that, The adaptive controller described in step S4 adopts a fuzzy proportional-integral-derivative control algorithm. The fuzzy rule table of this fuzzy proportional-integral-derivative control algorithm contains 5 fuzzy subsets, specifically negative large, negative small, zero, positive small, and positive large, which correspond to the fuzzy intervals of the deviation value and the rate of change of deviation calculated in step S4, respectively. The proportional gain correction range is set to 0.5-1.

0. The adaptive controller can dynamically correct the proportional-integral-derivative parameters online based on the deviation trend of the output predicted by the long short-term memory neural network, that is, the deviation value of each parameter and the change trend of the deviation rate calculated in step S4. In this way, it can accurately generate the pulling rate correction command and the heating power compensation command, thereby effectively compensating for the thermal field disturbance caused by the sudden change in the size of the single crystal turbine blade tenon and blade cross-section, and ensuring the control accuracy.

8. The method for controlling the solidification interface of a single-crystal turbine blade based on LSTM according to claim 1, characterized in that, The process also includes step S6, which involves establishing an online-offline combined incremental learning and update mechanism during the closed-loop dynamic control process. This mechanism periodically collects control data and solidification quality results from the actual operation process, and then uses a stochastic gradient descent algorithm to incrementally learn the weight parameters of the long short-term memory neural network model. This achieves continuous model optimization and adapts to equipment state drift and operating condition fluctuations in the single crystal growth furnace. The control data in step S6 consists of the pulling rate correction command and heating power compensation command generated in S4, and the actual adjustment parameters in S5. The solidification quality results are the metallographic detection results after the solidification of the single crystal turbine blades. The specific period is for each batch of single crystal turbine blades produced. The solidification quality results are judged based on the metallographic detection data. The specific criteria for judging the solidification quality results include four categories: qualified, small-angle grain boundaries, eutectic phase precipitation, and severe interface depressions. These criteria provide clear quality feedback for incremental learning and assist in optimizing the model weight parameters.

Images