Aero-engine sealing cavity leakage flow detection test correction method based on ant colony

Abstract

The invention provides an aero-engine sealing cavity leakage flow detection test correction method based on ant colonies. The aero-engine sealing cavity leakage flow detection test correction method comprises the following steps of S1, test system establishment and sealing performance verification; s2, designing test working condition parameters and combining a scheme; s3, leakage flow original data acquisition and primary processing; s4, data preprocessing and normalization are carried out; s5, constructing a fusion model of a particle swarm optimization algorithm-BP neural network-ant colony algorithm; s6, the leakage flow detection value is corrected; and S7, verifying a correction effect and evaluating reliability. According to the method, the limitation of a single technology is broken through through a fusion algorithm, error control (from sealing verification to abnormal value processing) is embedded in the whole process, the adaptability is dynamically updated and guaranteed, and finally technical support which is high in precision, reproducible and easy to popularize can be provided for aero-engine sealing cavity leakage flow detection. And the method directly serves engine sealing performance evaluation and safe operation-guarantee.

Description

Technical Field

[0001] This invention relates to the fields of aerospace engineering and key component testing technology for aero-engines, and in particular to a method for correcting leakage flow detection tests in aero-engine sealing cavities based on ant colony. Background Technology

[0002] The sealing cavity of an aero-engine is a core component ensuring the engine's efficient and safe operation. Its core function is to control the leakage of fluids (such as fuel, lubricating oil, and compressed air) to prevent fluid cross-contamination that could lead to decreased engine efficiency, component overheating, or malfunction. Therefore, accurate detection of sealing cavity leakage is a critical step in engine design finalization, production acceptance, and in-service maintenance, directly affecting the engine's performance indicators and operational safety.

[0003] Currently, the core challenges in detecting leakage flow in the sealing cavity of aero-engines are as follows: The detection interference caused by multi-condition coupling is significant: the leakage flow characteristics of the sealing cavity are affected by multiple parameters such as rotational speed (the difference in rotational speed under different engine operating conditions can be more than 10 times), sealing gap (wear causes the gap to change by 0.1 to 0.5 mm), medium temperature (from room temperature to above 300℃), and medium pressure (gradient change from low pressure to high pressure). Traditional detection methods rely on direct readings from a single sensor, which is easily affected by parameter coupling interference, and the detected value deviates greatly from the actual leakage state.

[0004] Limited data processing capabilities of single algorithms: Existing data correction methods mostly use single intelligent algorithms (such as simple ant colony optimization or single neural network). While ant colony optimization has global optimization capabilities, it is not accurate enough in fitting nonlinear mapping relationships. Neural networks are good at nonlinear fitting, but they are prone to getting trapped in local optima. Particle swarm optimization can only optimize a single parameter and cannot take into account the synergy of multiple algorithms, resulting in poor adaptability of the correction model across the entire working range and significant correction errors under extreme working conditions.

[0005] The testing process has a low degree of standardization: existing testing lacks unified system construction specifications (such as sensor installation location, sealing verification standards), operating condition design methods (parameter range determination basis, combination method) and error criteria, resulting in poor consistency of test results from different laboratories and different operators; at the same time, the dynamic adaptation capability for engine model changes and sealing cavity structure improvements is insufficient, requiring the reconstruction of correction models, which is time-consuming and labor-intensive.

[0006] The lack of calibration methods for actual leakage values ​​and verification of correction effects: Traditional detection relies heavily on the accuracy of the sensor itself and does not introduce independent high-precision calibration equipment (such as laser Doppler velocimeters) to obtain actual leakage values, making it impossible to effectively evaluate the reliability of correction methods; some verifications only cover a few intermediate operating conditions and do not cover key operating conditions such as high speed, high temperature, and extreme clearance, which limits the engineering applicability of correction methods. Summary of the Invention

[0007] This invention provides a test correction method for leak flow detection in aero-engine sealing cavities based on ant colony. It constructs a complete technical system of "system construction - operating condition design - data processing - fusion modeling - correction verification - standardization update". Through the coordinated linkage of multiple technical links, it realizes the upgrade of aero-engine sealing cavity leak flow detection from "experience-driven" to "data-driven, algorithm-coordinated, and standard-unified", providing reliable technical support for aero-engine sealing cavity leak control.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: A correction method for detecting leakage flow in the sealing cavity of an aero-engine based on ant colonies includes the following steps: S1: Build a test system, verify the sealing of the test system, and obtain a test system that passes the sealing test. S2: Based on the actual working scenario of the sealing cavity of an aero-engine, multiple sets of working condition combinations are generated using the orthogonal experimental method to form a working condition combination table; S3: Based on the working condition combination table, collect raw data through a sealed and qualified test system, perform preliminary processing on the raw data, and obtain a table of average raw data values; S4: Perform outlier removal and dimension normalization on the data in the original data average table to obtain a normalized data table; S5: Construct a fusion model of Particle Swarm Optimization (PSO), Backpropagation (BP) Neural Network (BSN), and Ant Colony Optimization (ACO). Divide the data in the normalized data table into training and validation sets. Train the fusion model using the training set. The PSO optimizes the initial weights and thresholds of the BP Neural Network and the pheromone evaporation coefficient and heuristic function weights of the ACO algorithm, outputting the optimized parameters. The BP Neural Network is initialized based on the optimized parameters, inputting the normalized condition parameters from the training set, and outputting the deviation correction value for the original predictions from the ACO algorithm. The ACO algorithm constructs a mapping path based on the optimized pheromone evaporation coefficient and heuristic function weights to obtain the original prediction value, and updates its pheromone and fitness functions based on the deviation correction value, forming a closed-loop iteration of PSO parameter optimization, BP correction value output, ACO optimization adjustment, and error feedback optimization. Verify the accuracy of the fusion model using the validation set to obtain a qualified fusion model. S6: Obtain the leakage flow detection value and corresponding operating parameters of the detection condition to be corrected. After outlier removal and normalization preprocessing, input the value into the qualified fusion model to obtain the fusion prediction value. Calculate the correction coefficient and use the correction coefficient to correct the leakage flow detection value to obtain the final value of the corrected leakage flow.

[0009] In this manual, the test correction method for leakage flow detection of aero-engine sealing cavity based on ant colony also includes S7: using a laser Doppler velocimeter with an accuracy of not less than 0.01 kg / s to obtain the actual leakage flow value of the test condition to be corrected, calculating the relative error between the final value of the corrected leakage flow and the actual leakage flow value, and if the relative error is ≤5%, the correction is deemed qualified.

[0010] In this specification, in step S1, the test system includes an aero-engine sealing cavity test piece, a leakage flow detection module, a data acquisition device, and a test control module.

[0011] In this specification, step S5, the parameter optimization process of the particle swarm optimization algorithm includes: initializing the particle position and velocity of the particle swarm, where the particle position includes the weights and threshold of the BP neural network and the pheromone evaporation coefficient and heuristic function weights of the ant colony algorithm; evaluating the particle position by combining the fitness function of the mean square error of the BP neural network and the ant colony algorithm; iteratively updating the individual optimal position and the global optimal position of the particle; and outputting the optimization parameters corresponding to the global optimal position as the initial parameters of the BP neural network and the ant colony algorithm.

[0012] In this specification, step S5, the deviation correction process of the BP neural network includes: initializing the BP neural network based on the optimized parameters output by the particle swarm optimization algorithm; inputting the normalized working condition parameters from the training set into the BP neural network; calculating the deviation correction value through forward propagation; combining the deviation between the original predicted value and the true value of the ant colony algorithm; updating the weights and thresholds of the BP neural network through backpropagation until the network error converges, thus obtaining a trained BP neural network.

[0013] In this specification, step S5, the interaction process between the ant colony algorithm and the BP neural network includes: the ant colony algorithm constructs a mapping path between the operating parameters and the leakage flow based on the optimization parameters output by the particle swarm optimization algorithm, and obtains the original predicted value; the BP neural network outputs a deviation correction value for the original predicted value; the ant colony algorithm incorporates the deviation correction value into the heuristic function and the pheromone update formula, adjusts the state transition probability and the pheromone increment, and optimizes the global optimization path.

[0014] In this specification, step S5, the closed-loop iteration process includes: inputting the optimization parameters output by the particle swarm optimization algorithm into the BP neural network and the ant colony algorithm; inputting the deviation correction value output by the BP neural network into the ant colony algorithm; feeding back the optimization error of the ant colony algorithm to the particle swarm optimization algorithm, and optimizing the parameters twice based on the optimization error; repeating the above process until the comprehensive mean square error of the fusion model meets the preset standard.

[0015] In this manual, in step S2, the operating condition combinations generated by the orthogonal test method must cover the full parameter range of rotational speed, sealing gap, medium temperature and medium pressure, with each parameter containing at least 5 levels, and the number of operating condition combinations not less than 25 groups.

[0016] In this manual, the raw data collected in step S3 includes leakage flow rate, inlet pressure, outlet pressure, and inner wall temperature for each set of operating conditions.

[0017] In this specification, the termination condition for closed-loop iteration is: the overall mean square error of the fusion model on the validation set is ≤0.003, or the number of iterations reaches the preset maximum number of iterations. The overall mean square error is the average of the squared deviations between the predicted values ​​and the true values ​​of the fusion model.

[0018] In summary, the present invention has at least the following beneficial effects: 1. Improve the accuracy and stability of leakage flow detection: Through the three-level synergistic integration of Particle Swarm Optimization (PSO), Backpropagation Neural Network (BP) and Ant Colony Algorithm (ACO), PSO optimization solves the problem of blind initial parameters in BP neural network and empirical core parameters in ACO. BP neural network makes up for the insufficient fitting of nonlinear coupling relationship by ACO, while ACO avoids the local optimum defect of BP neural network. The three work together to achieve accurate mapping of leakage flow characteristics under multiple operating conditions, effectively eliminating system errors caused by sensor drift, installation deviation, parameter coupling and other factors, making the detection results closer to the actual leakage state of the sealing cavity.

[0019] 2. Enhanced adaptability across the entire operating range: Based on the actual working scenarios of aero-engines, orthogonal operating condition combinations are designed to comprehensively cover typical operating conditions such as idling, cruise, and takeoff, as well as boundary operating conditions such as high speed, high temperature, extreme clearance, and high and low pressure. Combined with a dynamic update mechanism, the operating condition range and model parameters can be flexibly adjusted according to changes in engine model and improvements in the sealing cavity structure, without the need to rebuild the entire technical system, significantly improving the technology's adaptability to different testing requirements.

[0020] 3. Achieve standardization and repeatability of the testing process: By clarifying the module selection criteria of the test system (sensor accuracy, equipment connection specifications), sealing verification criteria, data preprocessing methods (outlier removal, normalization operation), fusion algorithm training process (parameter setting, convergence criteria), and correction effect verification criteria, a unified standardized operation manual is formed to ensure that the testing process is consistent for different laboratories and different operators, and that the test results are repeatable and comparable, thereby reducing the impact of human factors on the testing accuracy.

[0021] 4. Enhance the practicality and efficiency of engineering applications: Standardized manuals and automated data processing scripts (such as normalization calculation templates and algorithm training code) simplify the operation process, which can be quickly mastered by ordinary testing personnel; the dynamic update mechanism reduces the technical reconstruction cost after engine model changes and equipment replacements, eliminating the need to design operating conditions and train models from scratch; at the same time, the introduction of independent high-precision calibration equipment verifies the correction effect, ensuring the reliability of the technology, providing direct and usable leakage flow data support for engine design, production, and maintenance, shortening the testing cycle, and lowering the threshold for engineering applications. Attached Figure Description

[0022] Figure 1 This is a schematic flowchart of the test correction method for leakage flow detection in the sealing cavity of an aero-engine based on ant colony technology involved in this invention.

[0023] Figure 2 This is a schematic diagram illustrating the process of constructing the test system and designing operating conditions involved in this invention.

[0024] Figure 3 This is a schematic diagram illustrating the process of constructing the fusion algorithm model involved in this invention.

[0025] Figure 4 This is a schematic diagram of the detection value correction and effect verification process involved in this invention. Detailed Implementation

[0026] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0027] like Figure 1 As shown, this embodiment provides a correction method for detecting leakage flow in the sealing cavity of an aero-engine based on ant colony, including the following: S1: Test system setup and sealing performance verification 1.1 Selection and Configuration of System Core Modules The reliability of the aero-engine sealing cavity leakage flow detection test system directly determines the data quality. It needs to be configured with four core modules according to the principles of "structural equivalence, parameter matching, and accuracy compliance." The test system construction and operating condition design process is as follows: Figure 2As shown. The specific structure and component selection of the test system can be referred to the existing technology. It is only briefly described here for the purpose of full disclosure. The focus of this invention is: the algorithm content built for the existing test system. That is, this invention can be applied to existing conventional test systems and has a wide range of applications.

[0028] The test specimen for the sealing cavity of an aero-engine is a 1:1 replica of the sealing cavity of a certain type of turbofan engine, the FW-10. The material is high-temperature alloy GH4169 (consistent with the actual engine), and it includes three rings of sealing teeth (tooth height 2mm, tooth spacing 5mm). The inner diameter of the cavity is φ150mm, and its dimensional error is verified to be ≤±0.02mm through 3D scanning (ensuring structural equivalence). The test specimen was customized by a component manufacturer in the aviation industry and comes with a material certificate (e.g., number MAT-2024-056) and a dimensional inspection report (e.g., number DIM-2024-112).

[0029] Leakage flow detection module: Includes three types of sensors, all calibrated by the National Institute of Metrology (e.g., calibration certificate numbers JJG-2024-158 to 160): Leakage flow sensor: Model FL-300, range 0-10kg / s, accuracy 0.5%FS, adopts turbine flow meter principle, response time ≤0.1s, installed in the outlet pipe of the sealed cavity (diameter φ50mm), with the sensor probe extending into the center of the pipe (to ensure uniform flow velocity at the measurement cross section).

[0030] Pressure sensor: Model PT-500, range 0-1MPa, accuracy 0.1%FS, installed at the inlet (50mm from the cavity) and outlet (100mm from the cavity) of the sealing cavity respectively, adopting the diffused silicon principle, temperature drift ≤0.02%FS / ℃.

[0031] Temperature sensor: Model TT-800, range -50-400℃, accuracy 0.2℃, is a type K thermocouple. The probe is inserted into the pre-set φ3mm hole in the inner wall of the sealed cavity (10mm depth to ensure good contact) and is protected by a stainless steel sleeve.

[0032] Data acquisition equipment: An NI PXIe-6368 multi-channel data acquisition card is used, with 8 analog input channels and a sampling rate up to 250kHz (actually set to 10Hz to balance data volume and timeliness). It is paired with LabVIEW data acquisition software (version 2023), supporting real-time storage (sampling interval 0.1s) and outlier marking (e.g., automatic red highlighting for over-range measurements). The acquisition card and sensors are connected via shielded twisted-pair cable (length ≤ 5m to reduce electromagnetic interference), with an RS485 interface (supporting Modbus protocol).

[0033] The test control module consists of three types of controllers, which are linked and controlled by a PLC (Siemens S7-1200). Speed ​​controller: Model SC-2000, controls the speed of the sealing cavity drive shaft from 0-20000 r / min, with an accuracy of ±10 r / min. It adopts frequency conversion speed regulation and is connected to the test piece drive shaft through a coupling (coaxiality error ≤0.05mm).

[0034] Temperature controller: Model TC-5000, controls the temperature of the medium (dry air) from 25 to 350℃, with an accuracy of ±1℃. It is heated by an electric heating mantle and is equipped with thermocouple feedback regulation (response time ≤5s).

[0035] Pressure controller: Model PC-3000, controls medium pressure 0.1-0.6MPa, accuracy ±0.001MPa, adopts proportional valve regulation, and is equipped with pressure sensor for closed-loop control.

[0036] 1.2 System Assembly and Sealing Verification Process System assembly must follow the principle of "fixing before connecting, and mechanical before electrical," with specific steps as follows: 1. Fix the sealing cavity test piece on the test bench (using 4 M12 bolts, torque 50 N·m), adjust the levelness (≤0.05 mm / m), and calibrate it with a dial indicator.

[0037] 2. Install the inlet pipe (φ80mm, length 300mm) and outlet pipe (φ50mm, length 200mm) in sequence. Connect them to the test piece with flanges and place fluororubber sealing rings (hardness 70 Shore A, temperature resistance ≥260℃) between the flanges. Tighten the bolts diagonally and evenly (in 3 times to reach 30 N·m).

[0038] 3. Sensor installation: The flow sensor is installed horizontally on the outlet pipe (to avoid the influence of air bubbles), the pressure sensor is installed vertically (to reduce the influence of liquid accumulation), and the temperature sensor is fixed to the inner wall of the cavity via a threaded connection (sealed with sealant to prevent leakage).

[0039] 4. Electrical connection: Connect the sensor signal line to the data acquisition card, connect the controller power line to the regulated power supply (220V±5%), and connect the control signal line to the PLC. After completion, perform an insulation test (insulation resistance ≥10MΩ).

[0040] Sealing verification is a crucial step in ensuring the validity of test data. The procedure is as follows: Close the outlet valve of the sealing chamber, connect the inlet to a compressed air source (dew point ≤ -40℃), slowly increase the pressure to 0.3MPa (hold the pressure for 30 minutes), and record the pressure change through the pressure sensor (sampling interval 1 minute).

[0041] If the pressure drop within 30 minutes is ≤0.005MPa (leakage rate ≤0.017kg / h), the seal is deemed qualified; otherwise, disassemble the flange to check the sealing ring (whether it is damaged or misaligned), replace it with a new sealing ring (from the same batch of products) and re-verify until it is qualified (record the number of verifications and the results, and form a seal verification report).

[0042] S2: Design and Combination Scheme for Test Condition Parameters 2.1 Basis for determining the range of operating parameters The operating parameters need to cover the entire working range of the aero-engine sealing cavity, and their values ​​are based on three types of technical documents: Speed ​​(n): Range 5000-15000 r / min. According to the operating manual of a certain type of engine (such as ENG-2024-001), the idle speed is 5000 r / min, the takeoff speed is 15000 r / min, including transitional conditions (such as cruise at 8000 r / min and climb at 13000 r / min).

[0043] Sealing clearance (g): Range 0.1~0.5mm. Referring to engine component design specifications, the minimum clearance for new parts is 0.1mm, the maximum allowable wear clearance is 0.5mm, and the intermediate values ​​correspond to different wear stages (0.2mm is light wear, 0.4mm is moderate wear).

[0044] Medium temperature (T): Range 25-300℃. Based on engine hot test data (such as TEST-2024-015), 25℃ during ground start-up, 200℃ during cruise, and 300℃ during takeoff, including transition temperatures such as 100℃ (warm-up) and 250℃ (sub-rated).

[0045] Medium pressure (P): Range 0.1-0.5MPa. Refer to compressor outlet pressure data (e.g., DATA-2024-032), low-pressure stage sealing chamber 0.1MPa, high-pressure stage 0.5MPa, intermediate value 0.2-0.4MPa corresponds to different compressor stages.

[0046] 2.2 Orthogonal Test Condition Combination Design Orthogonal experimental design (L25(5)) was adopted 4 The orthogonal array was designed with 25 sets of working conditions to ensure that the parameter combinations are uniform and cover the entire range, with each parameter having 5 levels (equally spaced): Speed ​​levels: n1=5000r / min, n2=8000r / min, n3=11000r / min, n4=13000r / min, n5=15000r / min (interval 2000-3000r / min, matching engine speed gradient).

[0047] Horizontal sealing gap: g1=0.1mm, g2=0.2mm, g3=0.3mm, g4=0.4mm, g5=0.5mm (interval of 0.1mm to facilitate observation of the linear effect of the gap on leakage).

[0048] Medium temperature levels: T1=25℃, T2=100℃, T3=200℃, T4=250℃, T5=300℃ (the intervals between low temperature sections are large, and the intervals between high temperature sections are small, because high temperature has a more significant impact on leakage).

[0049] Medium pressure levels: P1=0.1MPa, P2=0.2MPa, P3=0.3MPa, P4=0.4MPa, P5=0.5MPa (equal interval of 0.1MPa, conforming to pressure gradient design).

[0050] Each row of the orthogonal array represents a set of operating conditions. For example: Operating condition 1: n1, g1, T1, P1 (5000 r / min, 0.1 mm, 25℃, 0.1 MPa); Operating condition 13: n3, g3, T3, P3 (11000 r / min, 0.3 mm, 200℃, 0.3 MPa, intermediate operating condition); Operating condition 25: n5, g5, T5, P5 (15000 r / min, 0.5 mm, 300℃, 0.5 MPa, extreme operating condition). All operating conditions are numbered sequentially (1-25) to form an operating condition combination table, which serves as the basis for S3 data acquisition operations, ensuring that each group of operating conditions is independent and repeatable (the same operating condition is collected 3 times, and the average value is taken to reduce random errors).

[0051] S3: Raw Data Acquisition and Preliminary Processing of Leakage Flow 3.1 System preparation before data acquisition Three preparatory steps must be completed before each set of operating condition data is collected to ensure that the system is in a stable state: 1. Media pretreatment: The incoming dry air must undergo three-stage treatment (filter with a filtration accuracy of 0.1μm, dryer with a dew point ≤-40℃, and pressure regulating valve) to ensure that it is oil-free, water-free, and has stable pressure (fluctuation ≤±0.001MPa). The impurity concentration is monitored by an online particle counter (model PC-200) and is ≤10 particles / mL.

[0052] 2. Parameter setting and preheating: Set the controller parameters according to the working condition combination table (e.g., working condition 5: n=8000r / min, g=0.2mm, T=200℃, P=0.3MPa). After starting the controller, preheat for 30 minutes (for high-temperature working conditions such as 300℃, the time should be extended to 45 minutes) to ensure that the temperature sensor reading is stable (fluctuation ≤±2℃).

[0053] 3. Data Acquisition Equipment Calibration: After the data acquisition software is started, zero-point calibration is performed (the medium source is turned off, the sensor zero-point value is recorded, and subsequent data is automatically deducted) and full-scale calibration (a standard signal of 80% of the range is passed through, and the sensor is recalibrated if the error exceeds 0.5%).

[0054] 3.2 Raw Data Collection and Preliminary Statistics Data collection strictly follows the "stabilization-collection-reproduction" process: Stability determination: When the speed fluctuation is ≤ ±50 r / min, the temperature fluctuation is ≤ ±2℃, and the pressure fluctuation is ≤ ±0.005 MPa, and this condition persists for 5 minutes, the system is considered stable, and the stable moment is recorded (accurate to the second).

[0055] Data Acquisition Process: After stabilization, start data acquisition and continuously acquire 100 sets of data (sampling interval 0.1s, total 10 seconds). Each set of data includes: Flow rate Q (kg / s), inlet pressure (MPa), Export Pressure (MPa), temperature T (°C); Status flags (such as "normal" or "out of range"; the reason for out-of-range data must be recorded separately).

[0056] Reproducibility verification: Under the same operating conditions, the data is collected again after an interval of 2 hours. The average flow deviation between the two collections is ≤1%. Otherwise, the data is collected again (excluding the influence of system drift).

[0057] The purpose of the preliminary processing is to obtain representative values ​​for each set of working conditions. The steps are as follows: 1. For 100 sets of raw data, calculate the arithmetic mean of each parameter (e.g., average flow rate). ).

[0058] 2. Generate a table of raw data averages, including "Operating Condition Number, n, g, T, P, ... , , , ", as input for S4 preprocessing.

[0059] S4: Data Preprocessing and Normalization 4.1 Outlier Removal (Application of the 3σ Criterion) The raw data may contain outliers due to transient sensor interference, which need to be removed according to the 3σ criterion. 1. Calculate the mean of 100 Q values ​​for each working condition. and standard deviation : .

[0060] 2. If a certain satisfy or If an outlier is identified (probability < 0.3%), it will be removed and the remaining data will be used for recalculation. .

[0061] 3. Regarding , The same operation is performed on the T parameter to ensure that all averages are based on valid data.

[0062] For example, among the 100 Q values ​​in operating condition 5, , The outlier range is <2.2kg / s or >2.8kg / s. If there is data with a value of 2.1kg / s, it is removed and the calculation is recalculated using 99 data points. .

[0063] 4.2 Dimensional Normalization (Min-Max Method) To eliminate the impact of parameter dimension differences on the fusion algorithm, the data is normalized to the [0,1] interval: 1. Determine the minimum and maximum values ​​of the parameters for 25 sets of operating conditions: ; ; ; ; (25 groups) (minimum value) (25 groups) (the maximum value).

[0064] 2. Normalization formula: ; For example: normalized value of n=11000r / min .

[0065] 3. Generate a normalized data table, including "operating condition number, , , , , ",in for The normalized values ​​are used as training data for the S5 fusion algorithm (Input: the first 4 normalization parameters; Output: ).

[0066] S5: Construction of Fusion Algorithm Model 5.1 Algorithm Fusion Framework Design and Collaborative Logic In the detection of leakage flow in the sealing cavity of aero-engines, single algorithms often have limitations: ant colony optimization excels at global optimization but has a slow convergence speed; backpropagation neural networks (BP neural networks) excel at nonlinear mapping but are prone to getting trapped in local optima; and particle swarm optimization (PSO) excels at parameter optimization but is less adaptable to complex constraints. Therefore, this scheme designs a three-level collaborative framework: "Particle swarm optimization parameter optimization → BP neural network nonlinear correction → ant colony optimization global optimization," achieving complementary advantages through bidirectional data interaction between algorithms. The process of constructing the fusion algorithm model is as follows: Figure 3 As shown.

[0067] The specific collaboration logic is as follows: First, the Particle Swarm Optimization (PSO) algorithm optimizes the initial weights and thresholds of the Backpropagation (BP) neural network and the core parameters of the Ant Colony Optimization (ACO) algorithm (pheromone evaporation coefficient, heuristic function weights), overcoming the blindness of traditional algorithm parameters set based on experience. Second, the BP neural network, based on the parameters optimized by the PSO algorithm, learns the nonlinear relationship between operating parameters and leakage flow, outputting prediction deviation correction values ​​to compensate for the insufficient mapping accuracy of the ACO algorithm under high-dimensional coupled operating conditions. Finally, the ACO algorithm dynamically adjusts the pheromone update strategy in conjunction with the correction values ​​from the BP neural network, while simultaneously feeding the optimization results back to the PSO algorithm, forming a closed-loop iteration of "parameter optimization - deviation correction - global optimization". The data flow connection between the three is as follows: PSO algorithm optimizes parameters → serves as the initial input for the BP neural network and the ACO algorithm → BP neural network outputs correction values ​​→ ACO algorithm uses the correction values ​​for optimization → ACO algorithm's optimization error is fed back to PSO optimization → PSO algorithm adjusts parameters a second time (looping until convergence).

[0068] 5.2 Particle Swarm Optimization (PSO) Algorithm Model Construction and Parameter Optimization 5.2.1 Definition and Physical Meaning of PSO Core Parameters The PSO algorithm finds the optimal combination of parameters by simulating the foraging behavior of bird flocks, and the following core parameters need to be defined: Particle swarm size (N): 30 particles. Too small a size will lead to insufficient search, while too large a size will increase the computational load. The setting of 30 particles shows the best balance between convergence speed and optimization accuracy in the preliminary experiments.

[0069] Particle dimension (D): 38 dimensions. The corresponding parameters to be optimized include: 4×6=24 weights from the input layer to the hidden layer of the BP neural network (4 neurons in the input layer and 6 neurons in the hidden layer), 6×1=6 weights from the hidden layer to the output layer, 6 thresholds in the hidden layer, 1 threshold in the output layer, pheromone evaporation coefficient of the ant colony algorithm (1), and heuristic function weight of the ant colony algorithm (1), totaling 24+6+6+1+1+1=38 dimensions.

[0070] Particle position ( ):

[0071] in: (j=1,2,3,4; k=1,2,...,6) represents the weights from the j-th neuron in the input layer to the k-th neuron in the hidden layer of the BP neural network; (m=1,2,...,6) represents the weights from the m-th neuron in the hidden layer to the output layer; (m=1,2,...,6) represents the threshold of the m-th neuron in the hidden layer; The threshold for the output layer; The pheromone evaporation coefficient in the ant colony algorithm; The weights are the heuristic functions for the ant colony algorithm.

[0072] Particle velocity ( ): , which represents the optimization step size for each dimension parameter. The initial range is set to [-0.1, 0.1] to avoid the parameters jumping out of the effective range due to excessively large step sizes.

[0073] Individual optimal position ( ): The optimal position of the i-th particle from the initial to the current iteration, recording the best parameter combination for that particle.

[0074] Global optimal position ( ): The optimal position among all particles represents the best combination of parameters found in the current particle swarm.

[0075] Learning factor ( , ): All are set to 2.0. The weights that control the learning of particles towards their historical best positions. The weights that control the particles' learning toward the global optimal position, in setting 2.0, can balance local exploration and global development in pre-experiments.

[0076] Inertia weight (ω): Where t is the current iteration number, This represents the maximum number of iterations. An initial value of 0.9 ensures global search capability, and decreases linearly to 0.4 with each iteration to enhance the accuracy of later local searches.

[0077] 5.2.2 PSO Iteration Formula and Calculation Process PSO optimizes parameters by iteratively updating the particle's velocity and position. The specific formula and calculation steps are as follows: 1. Speed ​​update formula: ; in: t is a random number in the interval [0,1] (generated randomly in each iteration to increase search diversity); t is the current iteration number (1≤t≤50).

[0078] Calculation example (taking the first dimension of the first particle as an example): If t=1, ω=0.9, , , , , , ,but: .

[0079] 2. Position update formula: ; Continuing from the previous example: If the parameter exceeds a reasonable range, it needs to be truncated to the boundary value, such as... If the maximum is 0.2, then any amount exceeding this will be treated as 0.2.

[0080] fitness function ( ): The prediction errors of both the BP neural network and the ant colony algorithm are used to evaluate the quality of particle positions. ; in: The mean squared error of the BP neural network on the training set (17 sets of data); , where is the mean squared error of the ant colony algorithm on the training set; weights 0.4 and 0.6 were determined based on pre-experiments (the global optimization of the ant colony algorithm has a greater impact on the final result).

[0081] 5.2.3 PSO Training Process (Initial Interaction with Other Algorithms) 1. Initialization: Randomly generate the initial positions of 30 particles. (Parameter range: weight ∈ [-1, 1], threshold ∈ [-0.5, 0.5]) ∈[0.05,0.2], (∈[0.5,1.0]) and initial velocity ∈[-0.1,0.1].

[0082] 2. Initial Evaluation: Input the position parameters of each particle into the BP neural network and the ant colony algorithm respectively (at this time, both the BP neural network and the ant colony algorithm are in their initial state), and calculate... and Then obtain the fitness function. .

[0083] 3. Update the optimal position: For each particle, if ,but Global Optimum For all The position with the lowest fitness.

[0084] 4. Iterative optimization: Iterate according to the velocity and position update formula until t=50, and output the globally optimal position. Extract the optimized parameter set: (* indicates optimized values), used as initial parameters for BP neural networks and ant colony algorithms.

[0085] 5.3 BP Neural Network (Backpropagation Neural Network) Model Construction and Bias Correction 5.3.1 BP Neural Network Structure and Parameter Definition A BP neural network is used to learn the nonlinear mapping relationship between operating parameters and leakage flow, and outputs a correction value to compensate for the prediction bias of the ant colony algorithm. Its structure is designed based on the analysis of the influencing factors of leakage flow in the sealing cavity (4 core operating parameters): Input layer: 4 neurons, corresponding to the normalized operating parameters in S4: (Normalized value of rotational speed) (Normalized value of sealing gap) (Normalized value of medium temperature) (Normalized values ​​of medium pressure), all are dimensionless values ​​in the interval [0,1].

[0086] Hidden layer: 6 neurons. Pre-experimentation validated the model's performance (error with 4, 5, 6, and 7 neurons) by showing the lowest error on both the training and validation sets (low risk of overfitting). The sigmoid activation function was used. , used to implement nonlinear mapping.

[0087] Output layer: 1 neuron, outputting the leakage flow prediction bias correction value. The physical meaning is "the amount of compensation for the deviation between the original predicted value and the actual value of the ant colony algorithm", and the range is limited to [-0.1, 0.1] (if it exceeds this range, it will be truncated) to avoid overcorrection.

[0088] Initial weights and thresholds: Optimized using PSO , , , This solves the problems of slow convergence and easy getting trapped in local optima caused by traditional random initialization.

[0089] 5.3.2 Forward Propagation Formula for BP Neural Network (Calculation of Correction Value) Forward propagation is a computational process from input to output used to generate correction values. The specific formula is as follows: 1. Hidden layer input ( ): The input to each hidden layer neuron is a weighted sum of the input layer parameters and their corresponding weights, plus a threshold: (m=1,2,...,6); For the j-th parameter of the input layer ( ).

[0090] Calculation example (taking the first hidden layer neuron as an example): like , , , , Input parameters ,but: .

[0091] 2. Hidden layer output ( ): Output after activation by the Sigmoid function: ; Continuing from the previous example: .

[0092] 3. Output layer input ( ): The input to the output layer neurons is a weighted sum of the hidden layer outputs and their corresponding weights, plus a threshold. ; Calculation example: If , , , , , , Hidden layer output ,but: ,because The range is limited to [-0.1, 0.1], and here it needs to be truncated to 0.1.

[0093] 4. Output layer output ( ): .

[0094] 5.3.3 Backpropagation formula for BP neural network (error correction, interaction with ant colony algorithm) Backpropagation corrects the weights and thresholds using gradient descent, making the corrected values ​​more accurate. Its error calculation is correlated with the original predictions of the ant colony algorithm, thus enabling interaction with the ant colony algorithm.

[0095] 1. Prediction error (E): Error is defined as "the squared deviation between the actual value and (the original prediction value of the ant colony algorithm + the BP correction value)": ; in: The actual normalized leakage flow rate in the training set; This represents the original prediction value from the ant colony algorithm (without correction); the coefficient 0.5 is used to simplify subsequent derivative calculations.

[0096] 2. Weight update formula: Weight updates from hidden layer to output layer: ; in: The learning rate (controls the update step size; preliminary experiments verify that 0.01 can balance convergence speed and stability); partial derivatives (Derived from the chain rule of differentiation).

[0097] Weight updates from input layer to hidden layer: ; Where: partial derivatives .

[0098] 3. Threshold update formula: Hidden layer threshold update: Partial derivative calculations are similar to weight calculations, omitting input parameters. .

[0099] Output layer threshold update: .

[0100] 5.3.4 BP Neural Network Training Process (in conjunction with Ant Colony Algorithm) 1. Initialization: Load the initial weights after PSO optimization. , and threshold , .

[0101] 2. Forward propagation calculation of correction values: Input the operating condition parameters of 17 sets of training data into the network, and calculate... .

[0102] 3. Obtain the original predictions from the ant colony algorithm: Input the same training data into the ant colony algorithm (at this time, the ant colony algorithm uses PSO-optimized parameters) to obtain... .

[0103] 4. Backpropagation to correct parameters: Based on the error E, adjust the network parameters according to the weight and threshold update formula, iterating 500 times (preliminary experiments verify that 500 iterations can bring the error to converge).

[0104] 5. Convergence Criterion: When the average error of the training set... When the time is up, stop training and output the trained BP neural network model (denoted as Model_BP).

[0105] 5.4 Ant Colony Algorithm (ACO) Model Optimization (Integration of PSO and BP Neural Network) 5.4.1 ACO Core Parameter Update (Based on PSO Optimization Results) The core parameters of ACO are determined by PSO optimization, ensuring its global optimization capability: Pheromones volatile coefficient ( ): Optimized using PSO The physical meaning is the proportion of pheromones retained in each iteration. (where evaporation rate is the percentage of pheromones released), the optimized value is usually between [0.08, 0.15], balancing the pheromone update speed and the retention of historical information.

[0106] Heuristic function weights (α): Optimized using PSO The value ranges from [0.6, 0.8], corresponding to the complementary weights of the heuristic function. (β∈[0.2,0.4]), ensuring a reasonable ratio between pheromone concentration and heuristic information.

[0107] Other parameters: Ant colony size = 50 (to cover the search requirements of 25 working conditions), maximum number of iterations = 100 (convergence verified by preliminary experiments), initial pheromone concentration. (To avoid excessively high initial pheromone levels leading to search bias).

[0108] 5.4.2 ACO Pathway Construction and Pheromone Update (Fusing BP Correction Values) ACO searches for the optimal mapping relationship by simulating the behavior of ants releasing pheromones. Its path construction and pheromone update processes require the integration of correction values ​​from a backpropagation neural network. This enables interaction with the BP neural network.

[0109] 1. Path construction (state transition probability): The probability that ant k transitions from operating state i (corresponding to a set of normalized parameters) to leakage flow state j: ; in: Let i be the pheromone concentration from state i to state j; For heuristic function (fusion BP correction value) , (where is the leakage flow rate value corresponding to state j), and its physical meaning is "the smaller the prediction deviation after correction, the higher the probability of transfer"; This is the set of states that ant k has not visited (to avoid duplicate searches).

[0110] 2. Partial pheromone update: After an ant completes path construction, it locally updates its pheromone to avoid premature convergence. ; Where: t is the current iteration number. The initial pheromone is used to ensure that the path retains the basic pheromone after each iteration.

[0111] 3. Global pheromone update (integrated with BP correction value): After the iteration, global pheromone enhancement is performed only on the optimal path (the path with the smallest prediction error), using the following formula: ; Among them: pheromone increment of non-optimal ants Optimal pheromone increment for ants , For pheromone constants, This is the absolute value of the BP correction value (the larger the correction value, the greater the original ACO prediction bias and the more significant the pheromone increment adjustment).

[0112] 5.4.3 ACO Fitness Function (Fused with BP Neural Network Output) The fitness function of ACO is used to evaluate the quality of a path, and a correction value from a BP neural network needs to be introduced. To achieve synergistic optimization between the two: ; in: The correction coefficient (determined through preliminary experiments, balancing the influence of the original ACO error and the BP correction value); if A positive value indicates that the ACO prediction is too low, and the fitness value is reduced to encourage the algorithm to adjust to a higher prediction value; if... If the result is negative, the algorithm will adjust to a lower predicted value.

[0113] 5.4.4 ACO Training Process (Closed-Loop Interaction with PSO and BP Neural Networks) 1. Initialization: Pheromone Matrix Loading PSO optimized and .

[0114] 2. Path construction: 50 ants based on transition probability Construct a mapping path from operating conditions to leakage flow rate to obtain the original predicted value. .

[0115] 3. BP Correction: Input the operating condition parameters into Model_BP to obtain... Calculate the corrected predicted value .

[0116] 4. Fitness assessment: according to Calculate the fitness of each path and record the optimal path (the path with the lowest fitness).

[0117] 5. Pheromone Update: Adjust the pheromone matrix according to local and global update formulas. .

[0118] 6. Feedback and Iteration: The optimal path... Feedback is sent to PSO, which optimizes the parameters twice based on this (if the ACO error is large, PSO will adjust ρ and α); steps 2-5 are repeated for up to 100 iterations to output the fusion model (denoted as Model_ACO-BP-PSO).

[0119] 5.5 Fusion Model Validation and Parameter Tuning Mechanism To ensure the reliability of the fusion model, it is necessary to evaluate it through a validation set and establish a parameter adjustment mechanism to connect with the subsequent leakage flow detection value correction steps.

[0120] 5.5.1 Validation Set Evaluation Metrics and Calculation The model accuracy was evaluated using 8 sets of validation data (randomly selected from 25 operating conditions, covering high, medium, and low speeds / clearance / temperature / pressure). The core metric was the comprehensive mean square error. ; in: These are the predicted values ​​from the fusion model; The actual normalized leakage flow rate for the validation set.

[0121] Qualification Standard: (More stringent than the 0.005 of a single ACO, ensuring correction accuracy).

[0122] 5.5.2 Parameter Adjustment Mechanism (for verification failures) like Adjustments should be made according to the following priority (based on error source diagnosis): 1. If (Insufficient accuracy of BP correction): Increase the number of hidden layer neurons to 8 (to enhance nonlinear fitting ability); Reduce the learning rate η to 0.005 (to avoid oscillations in weight updates); Retrain the BP neural network until .

[0123] 2. If (Insufficient ACO optimization accuracy): Increase the ant colony size to 60 (to expand the search area); Reduce the pheromone volatility coefficient ρ to (Retain more historical information); Retrain ACO until .

[0124] 3. If PSO optimization fails ( (Slow convergence) Increase the particle swarm size to 40. The formula for adjusting the inertia weight decay is as follows: (Enhanced global search); Rerun PSO to optimize parameters, then retrain the BP neural network and ACO.

[0125] After adjustments, verification is required until... This ensures that the model can provide a reliable predictive benchmark for correcting the leakage flow detection values ​​of S6.

[0126] 5.6 Core Contributions and Synergistic Advantages of the Fusion of Three Algorithms This fusion algorithm overcomes the limitations of single algorithms in detecting leakage flow in sealed cavities through a three-level collaborative approach. Its core contributions are as follows: 1. The role of PSO in parameter optimization: It upgrades the parameters of BP neural network and ACO from "empirical setting" to "data-driven optimization", which improves the convergence speed of the model by 40% (in the preliminary test, a single ACO required 150 iterations to converge, while the fusion model only required 90 iterations).

[0127] 2. Nonlinear correction effect of BP neural network: It provides accurate deviation compensation for the strong nonlinear characteristics of leakage flow in the sealing cavity affected by the coupling of "speed-gap-temperature-pressure". Under high temperature (300℃) and high speed (15000r / min) conditions, the prediction deviation is reduced by 35%.

[0128] 3. Global optimization role of ACO: Combined with BP correction value, the pheromone update strategy is dynamically adjusted to avoid getting trapped in local optima and ensure the prediction stability in the entire working condition range (such as extreme gaps of 0.1mm or 0.5mm). The working condition coverage is increased from 80% of the single algorithm to 100%.

[0129] The synergistic advantage of the three is ultimately reflected in the average prediction accuracy of the fusion model. This method improves upon the single ant colony algorithm by more than 50%, providing a more reliable theoretical basis for correcting the leakage flow detection value in the subsequent S6 step, and ensuring that the corrected detection results can accurately reflect the actual leakage state of the aero-engine sealing cavity.

[0130] S6: Leakage Flow Detection Value Correction Process (Detection value correction and effect verification process as follows) Figure 4 (As shown) 6.1 Acquisition and Preprocessing of Detection Data to be Corrected The working condition to be corrected must meet the requirement that "the parameters are within the range of S2" (to ensure the model is effective). For example, take a certain actual detection working condition (number D1: n=10000r / min, g=0.3mm, T=220℃, P=0.35MPa). Data Acquisition Process: Start the system using method S3, and collect 100 sets of data after it stabilizes. (Leakage flow detection value), synchronously record the real-time values ​​of n, g, T, and P (ensure the deviation from the set value is ≤5%).

[0131] 2. Pretreatment: Outlier removal: for 100 Outliers were removed using the 3σ criterion (same as S4.1), and the average detection value was calculated. (m is the amount of data after removal, usually 95-100).

[0132] Normalization: determined using S4.2 and ,calculate: ; ; ; ; .

[0133] 6.2 Calculation of Prediction and Correction Coefficients for Fusion Model The physical meaning of the correction coefficient k is "the ratio of the model's predicted value to the detected value", which is used to eliminate systematic errors.

[0134] 1. Fusion model prediction: , , , Inputting a successfully trained Model_ACO-BP-PSO from S5, we obtain: the raw ACO prediction values. BP correction value ; Fusion prediction value .

[0135] 2. Calculation of correction factor: ; For example: if , Then k=1.2 (the detection value is too low and needs to be amplified by 1.2 times).

[0136] 6.3 Detection Value Correction and Result Output The corrected leakage flow rate needs to reflect the actual physical quantity. Steps: 1. Individual value correction: for 100 groups (Original detection values, unnormalized), applying the correction formula: (i=1,2,...,100).

[0137] 2. Corrected average: for 100 groups Remove outliers (3σ criterion), calculate (m' is the amount of data after removal).

[0138] 3. Result Recording: Generate a correction result record table, including "Operating Condition D1 parameters, ,k, ",in This is the final correction value, used to evaluate the leakage performance of the sealing cavity (e.g., compared with the design allowable value of 0.8 kg / s to determine whether it is qualified).

[0139] S7: Verification of Correction Effect and Reliability Assessment 7.1 Obtaining the actual leakage flow rate (high-precision measurement) To verify the effectiveness of the correction, the actual values ​​need to be obtained using high-precision equipment. ; Measuring equipment: Laser Doppler velocimeter (LDV-5000), accuracy 0.01 kg / s, spatial resolution 0.1 mm. 3 The measurement principle, calibrated by the National Institute of Metrology, is to calculate the fluid velocity through the laser Doppler effect and then calculate the flow rate by combining the cross-sectional area of ​​the pipe.

[0140] Measurement Procedure: Under operating condition D1, align the LDV probe with the outlet pipe of the sealed cavity (50mm from the sensor), adjust the laser focus to the center of the pipe (where the flow velocity is highest), collect 100 sets of velocity data, and convert them into flow rate. Calculate after removing outliers ,Right now .

[0141] 7.2 Correction Error Calculation and Judgment Criteria The effect of the correction is evaluated using the relative error ε: ; Acceptance criteria: According to the test specifications for the sealing performance of aero-engines, the relative error of leakage flow detection must be ≤5%, i.e., ε≤5%.

[0142] Example: If , If ε ≈ 4.3% (qualified); if ε = 6.2% (unqualified), the cause needs to be investigated: like Large fluctuations (standard deviation > 0.05 kg / s): Check if the sensor installation is loose and re-collect data.

[0143] If the prediction bias of the fusion model is large ( ): Return to S5, supplement the training model with the D1 data until the prediction error is reduced.

[0144] 7.3 Batch Validation and Method Validity Confirmation To ensure the method's universality, five typical operating conditions (covering the parameter range) need to be validated: Low speed and low pressure (D2: 5000 r / min, 0.1 MPa). Temperature in high gap (D3: 0.5mm, 200℃); Intermediate full parameters (D4: 11000r / min, 0.3mm, 200℃, 0.3MPa). High speed and high temperature (D5: 15000r / min, 300℃); Low gap, high pressure (D6: 0.1mm, 0.5MPa); If ε for all 5 working conditions is ≤5%, the correction method is deemed effective overall; if any one condition is unqualified, the following needs to be addressed: 1. Add training data for this working condition (collect 5 additional sets of repeated data). 2. Adjust the parameters of the fusion algorithm in S5 (e.g., increase the number of PSO particles to 40 and the number of BP hidden layer neurons to 8). 3. Retrain the model and validate it until all operating conditions are satisfactory.

[0145] S8: Standardization and Dynamic Updating of Correction Methods 8.1 Compilation of Standardized Operation Manual The processes, parameters, and criteria for S1-S7 have been compiled into a standardized operation manual (version V1.0), which includes: System configuration list: sensor model, accuracy, calibration cycle (once every 6 months), equipment connection diagram (including wiring definition).

[0146] Operating Condition Design Guidelines: Parameter Range Determination Methods, Orthogonal Array Selection Principles (e.g., L36(6) for extending to 36 operating conditions) 4 )).

[0147] Data processing details: Example of calculating the 3σ criterion for outlier removal, and a table of normalized parameters (attached). and ).

[0148] Fusion algorithm operation: software script for model training (MATLAB code), screenshot of parameter setting interface, convergence criteria ( ).

[0149] Correction and verification process: Correction coefficient calculation template (Excel spreadsheet), LDV measurement operation steps.

[0150] The manual must be reviewed by three engine testing experts (who must hold senior engineer titles and have more than 5 years of experience). The review will focus on the rationality of the process, the operability of the parameters, and the effectiveness of error control. The manual will be released and implemented after it passes the review.

[0151] 8.2 Method Dynamic Update Mechanism To adapt to changes in engine models or testing conditions, an update mechanism will be established every 6 months; Triggering conditions: Engine model change (e.g., upgrade from FW-10 to FW-11, resulting in changes to the sealing cavity structure). Sensor replacement (e.g., changing the model of the flow sensor to improve accuracy); In batch verification, ε>5% for 3 consecutive operating conditions (method failure warning).

[0152] Update process: 1. Data Supplementation: Collect new operating condition data (such as the sealing gap range of FW-11 from 0.08 to 0.4 mm), process it according to S3-S4, and form a new normalized data table.

[0153] 2. Model Iteration: Merge the new data with the historical data (total sample size ≥ 30 groups), and retrain the fusion model according to S5 (increase the number of PSO iterations to 60 to ensure convergence).

[0154] 3. Effect verification: Test 5 new working conditions. If ε is ≤5% for all of them, the test is passed.

[0155] 4. Manual revision: Update parameter ranges, normalization coefficients, and model parameters; release a new version (e.g., V1.1); and provide training to operators.

[0156] Through dynamic updates, the correction method is ensured to always adapt to actual testing needs, providing continuous and reliable technical support for the detection of leakage flow in the sealing cavity of aero-engines.

[0157] In some embodiments, a surface fitting algorithm (SFA) is introduced to form a closed-loop system of "data preprocessing - parameter optimization - nonlinear mapping - global optimization". The core function of the surface fitting algorithm is to establish a continuous and smooth mapping relationship between operating parameters and leakage flow, solve the interpolation error problem between discrete data points, and provide a high-quality basic model for other algorithms.

[0158] The collaborative logic of the four algorithms is as follows: Surface Fitting (SFA) first fits a continuous function to the normalized data of S4, outputting smoothed baseline predictions. and fitting residuals ; Particle Swarm Optimization (PSO) simultaneously optimizes the initial parameters of the backpropagation neural network, the core parameters of the ant colony algorithm, and the polynomial coefficients of the surface fitting to fit the residuals. The prediction error is the optimization objective; The BP neural network optimizes parameters based on PSO, learns the relationship between the surface fitting residual and operating parameters, and outputs the residual correction value. , so that the corrected residual minimize; Ant Colony Algorithm (ACO) combined with surface fitting for basic prediction values The residual correction value of BP is used to dynamically adjust the pheromone update strategy, search for the globally optimal mapping relationship, and finally output the fused prediction value. ; The errors of each algorithm are fed back to PSO, forming an iterative closed loop of "fitting-optimization-correction-optimization" until the overall error converges.

[0159] 1. Surface Fitting Algorithm (SFA) Model Construction and Basic Mapping 1.1 Form and Parameter Definition of Surface Fitting Function For leakage in the sealing cavity, four operating parameters are considered ( To determine the characteristics of the influence, a four-dimensional surface fitting function extended by a bivariate high-order polynomial is selected, with the following form: ; i,j,k,l are the polynomial orders (0≤i,j,k,l≤3, total order≤3, to avoid overfitting); These are the surface fitting coefficients (unknown parameters to be solved), totaling... Number (the number of coefficients of a four-dimensional third-order polynomial calculated by the number of combinations). This is the normalized baseline prediction of leakage flow rate output from the surface fitting.

[0160] The fitting residual is defined as the deviation between the actual value and the fitted value: (p=1,2,...,25, corresponding to 25 sets of working condition data).

[0161] 1.2 Solving for Fitting Coefficients Using the Least Squares Method The coefficients are solved by minimizing the sum of squared residuals using the least squares method. : Objective function: ; in Let be the combination of independent variables for the p-th data set.

[0162] make This yields a system of linear equations: ; Written in matrix form: ,in It is a 25×35 matrix of independent variables. The coefficient vector is 35×1. It is a 25×1 vector of actual values.

[0163] Solving for the given information, we get: (Calculate the coefficients by inverting the matrix).

[0164] 1.3 Surface Fitting Training Process (Initial Interaction with PSO) Initial fitting: 25 sets of normalized data from S4 ( Substitute into the above formula to calculate the initial coefficients. and initial fitted values .

[0165] Residual calculation: The initial residual vector is obtained. .

[0166] Parameter feedback: sum of squared residuals Feedback is sent to PSO as one of the objectives of PSO in optimizing the surface fitting coefficients (PSO's fitness function will include an S term).

[0167] 2. Particle Swarm Optimization (PSO) Algorithm Model Extension (Fusing Surface Fitting Parameters) 2.1 PSO Parameter Dimension Expansion and Optimization Target Adjustment Based on the original 38-dimensional parameters, surface fitting coefficients are added. The 35 dimensions expand to a total of 73 dimensions; Particle position: ; The fitness function is adjusted to a weighted sum of the errors of the four algorithms:

[0168] S is the sum of squared residuals from the surface fitting; This represents the residual correction error of the BP neural network. This represents the optimization error of the ant colony algorithm. This is the sum of squared deviations between the surface fitting and the ant colony prediction (measuring the consistency between the two).

[0169] 2.2 Interaction between PSO and surface fitting During PSO initialization, the surface fitting coefficients The initial range is set to [-10, 10] (based on the coefficient distribution of the preliminary experiment).

[0170] In each iteration, the particle position As the current candidate coefficients, they are substituted into the surface fitting function for calculation. And S.

[0171] According to the new fitness function Evaluate the quality of particles and update and .

[0172] After the iteration ends, from Extract the optimized surface fitting coefficients This is used for subsequent high-precision fitting.

[0173] 3. BP Neural Network Model Extension (Fusing Surface Fitting Residuals) 3.1 Input and Output Adjustment of BP Neural Network The input layer of a BP neural network adds surface fitting residuals. As the fifth neuron, it is used to learn the nonlinear characteristics of the residual; Input layer: ; Output layer: Residual correction value (The goal is to) (Approaching 0).

[0174] 3.2 Forward Propagation and Error Calculation of BP Neural Network (Interaction with SFA) Hidden layer input: (j=1-4 are operating parameters, j=5 is...) ); Output layer output: (o is the weighted sum of the hidden layers, and the activation function is the same as before); Error function: .

[0175] Update the weights through backpropagation, so that Minimize to achieve accurate correction of the surface fitting residual.

[0176] 4. Ant Colony Algorithm (ACO) Model Extension (Fusing Surface Fitting Base Values) 4.1 ACO state transition probability and pheromone update (interaction with SFA) Heuristic function fusion of surface fitting values: ;in This is the deviation of the ACO search (the adjustment relative to the baseline value of the surface fitting).

[0177] The global pheromone update introduces the fitting residual: ; The residual is the result of BP correction. The smaller the residual, the greater the pheromone increment.

[0178] Final prediction fusion formula: ; 4.2 Collaborative Optimization Process of ACO and SFA ACO uses surface fitting values Based on the baseline, the search deviation (Range [-0.1, 0.1]).

[0179] In each iteration, Residual after BP correction Combined, the final prediction error is calculated.

[0180] Error feedback is fed back to pheromone updates, causing the algorithm to converge toward the direction of "surface fitting base value + optimal deviation".

[0181] 5. Training Process and Validation of Four-Algorithm Fusion 5.1 Iterative Training Steps initialization: Surface fitting: Calculating initial coefficients ,get PSO: Randomly generates 73-dimensional particles and initializes them. and ; BP neural network: Initialize weights and thresholds using PSO initial parameters; ACO: Initialize the pheromone matrix and set... and The initial value.

[0182] First iteration: PSO Optimization Output BP parameters, ACO parameters; SFA uses calculate and ; BP neural network learning Output ,get ACO is based on search ,calculate Calculate the error of each algorithm and feed it back to the PSO fitness function.

[0183] Iterative loop: Repeat the steps until the number of iterations reaches 100 or... (More stringent than the fusion of three algorithms).

[0184] 5.2 Validation Indicators and Synergistic Advantages Overall mean square error of the validation set: .

[0185] The core contribution of the four-algorithm fusion: Surface fitting provides a continuous and smooth basic model, reducing the interpolation error of discrete data points by 40% and solving the prediction jump problem of ACO in sparse operating conditions.

[0186] PSO optimizes four algorithm parameters simultaneously, improving the overall convergence speed by 25% and avoiding the limitations of optimizing a single algorithm parameter.

[0187] BP neural networks specifically correct surface fitting residuals, reducing residuals by 50% under highly nonlinear conditions (such as high temperature and high pressure).

[0188] ACO searches for local optimal deviations based on surface fitting, taking into account both global trends and local details, and improves prediction accuracy by 15% compared to the fusion of three algorithms.

[0189] Through deep fusion of four algorithms, the final model achieves an average prediction error across the entire operating condition range. This provides a more accurate theoretical benchmark for correcting leakage flow detection values.

Claims

1. An ant colony-based method for modifying a test of a turbine engine seal cavity leakage flow, the method comprising: The method comprises the following steps: ​ S1: building a test system, sealing verification is performed on the test system, and a sealed test system is obtained; S2: based on the actual working scene of the aero-engine sealing cavity, a plurality of working condition combinations are generated by using an orthogonal test method to form a working condition combination table; S3: according to the working condition combination table, original data are collected by the sealed test system, and the original data are preliminarily processed to obtain an average value table of the original data; S4: the data in the average value table of the original data are subjected to outlier rejection and dimensionless normalization processing to obtain a normalized data table; S5: a fusion model of a particle swarm optimization algorithm-BP neural network-ant colony algorithm is constructed, the data in the normalized data table are divided into a training set and a verification set, and the fusion model is trained by using the training set; the particle swarm optimization algorithm is used to optimize the initial weight, the initial threshold value of the BP neural network, and the pheromone evaporation coefficient and the heuristic function weight of the ant colony algorithm, and the optimized parameters are output; the BP neural network is initialized based on the optimized parameters, the normalized working condition parameters in the training set are input, and a deviation correction value of an original prediction value of the ant colony algorithm is output; the ant colony algorithm is used to construct a mapping path based on the optimized pheromone evaporation coefficient and the heuristic function weight to obtain the original prediction value, the pheromone and the fitness function of the ant colony algorithm are updated in combination with the deviation correction value, and a closed-loop iteration of particle swarm parameter optimization-BP neural network correction value output-ant colony optimization adjustment-error feedback optimization is formed; the precision of the fusion model is verified by using the verification set, and a qualified fusion model is obtained; S6: the leakage flow detection value and the corresponding working condition parameters of a to-be-corrected detection working condition are obtained, preprocessed by using the outlier rejection and normalization, input into the qualified fusion model, a fusion prediction value is obtained, a correction coefficient is calculated, the leakage flow detection value is corrected by using the correction coefficient, and a final value of the corrected leakage flow is obtained.

2. The ant colony based aeroengine seal cavity leakage flow test correction method of claim 1, wherein, S7: a laser Doppler velocimeter with a precision of not less than 0.01 kg / s is used to obtain a real leakage flow value of the to-be-corrected detection working condition, a relative error between the final value of the corrected leakage flow and the real leakage flow value is calculated, and if the relative error is less than or equal to 5%, it is determined that the correction is qualified.

3. The ant colony based aeroengine seal cavity leak flow test correction method of claim 1, wherein, In step S1, the test system comprises an aero-engine sealing cavity test piece, a leakage flow detection module, a data acquisition device and a test control module.

4. The ant colony based aeroengine seal cavity leak flow test correction method of claim 1, wherein, In step S5, the parameter optimization process of the particle swarm optimization algorithm comprises: initializing the particle position and the speed of the particle swarm, the particle position comprises the weight and the threshold value of the BP neural network and the pheromone evaporation coefficient and the heuristic function weight of the ant colony algorithm; the particle position is evaluated by using the fitness function of the mean square error of the BP neural network and the ant colony algorithm, the individual optimal position and the global optimal position of the particle are iteratively updated, the optimized parameters corresponding to the global optimal position are output, and the optimized parameters are used as the initial parameters of the BP neural network and the ant colony algorithm.

5. The ant colony based aeroengine seal cavity leak flow test correction method in accordance with claim 1, wherein, In step S5, the bias correction process of the BP neural network includes: initializing the BP neural network based on the optimization parameters output by the particle swarm optimization algorithm, inputting the normalized working condition parameters in the training set into the BP neural network, and calculating the bias correction value through forward propagation; combining the bias between the original prediction value and the real value of the ant colony algorithm, updating the weights and thresholds of the BP neural network through back propagation until the network error converges, and obtaining the trained BP neural network.

6. The ant colony based aeroengine seal cavity leak flow test correction method of claim 1, wherein, In step S5, the interaction process of the ant colony algorithm and the BP neural network includes: the ant colony algorithm constructs a mapping path of working condition parameters and leakage flow based on the optimization parameters output by the particle swarm optimization algorithm to obtain an original prediction value; the BP neural network outputs a bias correction value for the original prediction value; the ant colony algorithm integrates the bias correction value into the heuristic function and the pheromone update formula, adjusts the state transition probability and the pheromone increment, and optimizes the global optimization path.

7. The ant colony based aeroengine seal cavity leak flow test correction method in accordance with claim 1, wherein, In step S5, the process of closed-loop iteration includes: the optimization parameters output by the particle swarm optimization algorithm are input into the BP neural network and the ant colony algorithm; the bias correction value output by the BP neural network is input into the ant colony algorithm; the optimization error of the ant colony algorithm is fed back to the particle swarm optimization algorithm, and the particle swarm optimization algorithm optimizes the parameters based on the optimization error; the above process is repeated until the comprehensive mean square error of the fusion model meets the preset standard.

8. The ant colony based aeroengine seal cavity leak flow test correction method of claim 1, wherein, In step S2, the working condition combinations generated by the orthogonal test method cover the full parameter range of the rotation speed, the sealing gap, the medium temperature and the medium pressure, each parameter contains at least 5 levels, and the number of working condition combinations is not less than 25.

9. The ant colony based turbine shroud leak flow test correction method of claim 1, wherein, In step S3, the collected original data include the leakage flow, the inlet pressure, the outlet pressure and the inner wall temperature under each working condition.

10. The ant colony-based aeroengine seal cavity leak flow test correction method in accordance with Claim 7, wherein, The end condition of the closed-loop iteration is that the comprehensive mean square error of the fusion model on the verification set is ≤0.003, or the iteration number reaches the preset maximum iteration number, and the comprehensive mean square error is the average value of the square of the bias between the prediction value and the real value of the fusion model.

Images