Isolation section intelligent suction control method for shock wave string position feedback

Abstract

The invention relates to the technical field of supersonic speed isolation section active flow control, and discloses an isolation section intelligent suction control method for shock wave string position feedback, and the method comprises the steps: obtaining isolation section wall surface pressure data through a synchronous data collection device, and transmitting the data to an edge calculation platform for preprocessing; inputting the data into a lightweight flow field reconstruction model to generate a flow field schlieren image; utilizing a target detection algorithm to identify wave system features, and calculating the front edge position of a shock train through a confidence weighting algorithm; a feedback control instruction is generated according to the relative relation between the shock wave string front edge position and the suction groove position, the electromagnetic valve is driven to be opened and closed through the serial port relay, and intelligent suction regulation and control are achieved. Based on an edge computing architecture, through reconstruction and weighted positioning from sparse pressure to a dense flow field, accurate capture and real-time closed-loop control of the shock wave string position are realized, the kick of the shock wave string towards the upstream is effectively inhibited, and the stability of the flow field of the isolation section is maintained.

Description

Technical Field

[0001] This invention relates to the field of active flow control technology for supersonic isolation sections, specifically to an intelligent suction control method for isolation sections based on shock train position feedback. Background Technology

[0002] Currently, in the operation of air-breathing hypersonic propulsion systems, the isolator, as a key aerodynamic component connecting the inlet and combustion chamber, plays a crucial role in containing the shock train and isolating combustion back pressure. The flow direction and position of the shock train structure within the isolator are not fixed but dynamically shift with fluctuations in flight Mach number and combustion chamber back pressure. Maintaining a stable position of the shock train within the isolator, preventing excessive forward movement that could cause the inlet to malfunction, is a core element in ensuring the safe operation of the propulsion system. To address the risks posed by the forward movement of the shock train, flow control is typically implemented on the isolator wall, enhancing the flow field's resistance to back pressure by removing a low-energy boundary layer.

[0003] To address the monitoring and control requirements of the flow field in the aforementioned isolated section, existing technical solutions mostly employ detection methods based on discrete wall pressure sensors combined with passive or simple open-loop active control. In practical applications, pressure sensor arrays are arranged along the flow direction to obtain the wall pressure distribution. Specific pressure thresholds are set using the pressure surge method or standard deviation method. When the pressure at a certain measuring point exceeds the threshold, it is determined that the leading edge of the shock wave has reached that position. Based on this determination, some solutions use a passive suction chamber with a fixed orifice ratio for continuous suction; or, based on a preset Mach number operating condition table, simply control the opening and closing of valves, without involving real-time analysis of the detailed flow field morphology.

[0004] However, the aforementioned existing technologies have limitations when dealing with complex flow fields with highly dynamic changes. Discretely distributed pressure measurement points are spatially sparse, making it difficult to reconstruct a complete flow field topology that includes details of shock wave / boundary layer interference, resulting in blind spots in the understanding of the flow field state. Judgment logic relying solely on pressure thresholds is highly sensitive to signal noise and cannot accurately define the blurred physical boundaries of the shock wave lead edge, easily leading to drastic jumps in positioning results and false alarms. Furthermore, control methods lacking real-time perception of the flow field morphology often exhibit lag in response. Passive suction, while simple, causes continuous flow loss, reducing engine thrust performance; while simple on / off control struggles to accurately anchor the shock wave position, and in the event of sudden high back pressure, it cannot effectively suppress various unsteady oscillations and abrupt movements of the shock wave through rapid and accurate flow field feedback.

[0005] Therefore, this invention provides an intelligent suction control method for the isolation section with shock train position feedback to address the shortcomings of the prior art. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides an intelligent suction control method for the isolation section with shock train position feedback. This method solves the problems of low shock train positioning accuracy due to sparse discrete pressure monitoring information in the supersonic isolation section, and the difficulty in effectively suppressing the upstream jump of the shock train in real time, which can lead to the inability of the intake to start.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a smart suction control method for an isolation section with shock train position feedback, comprising the following steps: The wall pressure data output by the wall pressure sensor of the isolation section is collected in real time using synchronous data acquisition equipment, and the wall pressure data is transmitted to the edge computing platform for preprocessing. The preprocessed wall pressure data is input into a lightweight flow field reconstruction model deployed on the edge computing platform, and the output is a flow field schlieren image reflecting the density gradient change inside the isolation section. The wave system features in the flow field schlieren image are identified using a target detection algorithm, and the position of the shock train leading edge is calculated using a confidence weighting algorithm. Based on the relative relationship between the leading edge position of the shock wave train and the preset flow direction position range of the suction groove, a feedback control command is generated. The feedback control command is sent through the serial port relay to drive the solenoid valve to open and close, thereby realizing intelligent suction control of the flow field in the isolation section.

[0008] By employing the above technical solutions, the problem of obtaining overall flow field structure information by limited sensors is solved by establishing a nonlinear mapping from sparse wall pressure to high-resolution flow field schlieren images using a lightweight flow field reconstruction model. Combined with target detection and confidence-weighted position correction algorithms, high-confidence region boundary information is used to correct positioning deviations caused by image noise, achieving accurate capture of the physical coordinates of the shock wave leading edge. The real-time closed-loop feedback mechanism based on an edge computing architecture can trigger boundary layer suction in milliseconds when the shock wave approaches the danger zone. Therefore, the technical effect of effectively suppressing the upstream jump of the shock wave, maintaining the stability of the flow field in the isolation section, and preventing the inlet from failing to start is achieved.

[0009] Preferably, the process of using a synchronous data acquisition device to collect wall pressure data output by the wall pressure sensor of the isolation section in real time and transmitting the wall pressure data to the edge computing platform for preprocessing includes: the synchronous data acquisition device performing analog-to-digital conversion on the analog voltage signal output by the wall pressure sensor according to a preset sampling frequency, encapsulating the converted digital pressure data and sending it to the edge computing platform via TCP / IP protocol; the edge computing platform establishing a first-in-first-out buffer pool for the data channel of each wall pressure sensor, performing a moving average filtering algorithm on the received raw pressure data sequence, and calculating the arithmetic mean within the filtering window as the effective pressure data.

[0010] By adopting the above technical solution, the timing consistency of data transmission is ensured by using high-speed network protocols, and high-frequency random noise caused by the electromagnetic environment in the industrial field is effectively filtered out by moving average filtering, thereby improving the signal-to-noise ratio of subsequent model inputs.

[0011] Preferably, the preprocessing process of the edge computing platform further includes: using the real-time inlet static pressure data collected by the wall pressure sensors arranged at the inlet of the isolation section as a reference quantity, dividing the effective pressure value of each wall pressure sensor after moving average filtering by the real-time inlet static pressure data at the same time to obtain the dimensionless wall pressure ratio; using the maximum-minimum normalization algorithm to linearly map the dimensionless wall pressure ratio to the numerical range of zero to one, generating a normalized pressure feature vector as the input of the lightweight flow field reconstruction model.

[0012] By adopting the above technical solution, the interference of fluctuations in the absolute pressure of the incoming flow on feature extraction is eliminated, the data distribution range is unified, and the generalization ability and numerical stability of the reconstruction model under different Mach numbers and total pressure conditions are enhanced.

[0013] Preferably, the process of inputting the preprocessed wall pressure data into a lightweight flow field reconstruction model deployed on the edge computing platform and outputting a flow field schlieren image reflecting the density gradient changes within the isolation section includes: inputting the wall pressure data into a convolutional neural network branch of the lightweight flow field reconstruction model, performing upsampling operations through fully connected layers and transposed convolutional layers to output two-dimensional spatial structure features of the flow field; inputting the wall pressure data into a fully connected neural network branch of the lightweight flow field reconstruction model, extracting global nonlinear correlation features between the wall pressure data through multiple fully connected layers; and concatenating the two-dimensional spatial structure features with the global nonlinear correlation features in the fusion layer of the lightweight flow field reconstruction model, and decoding the output convolutional layer to generate the flow field schlieren image.

[0014] By adopting the above technical solution, the dual-branch architecture can simultaneously capture the local spatial details of the flow field and the global pressure evolution trend, so that the reconstructed image retains both clear shock wave edge features and maintains an accurate overall flow field intensity distribution.

[0015] Preferably, the lightweight flow field reconstruction model undergoes model lightweighting processing before deployment on the edge computing platform. The model lightweighting processing includes: structural pruning optimization, which analyzes the sparsity of weights in the model's convolutional layers, identifies and removes redundant feature channels whose contribution to the output results is lower than a preset threshold, and reduces the number of floating-point operations in the model's forward inference; and half-precision quantization, which converts the weight parameters and bias parameters in the model from 32-bit floating-point format to 16-bit floating-point format and compiles them into a file format suitable for mobile deployment, thereby reducing the memory throughput latency of the edge computing platform.

[0016] By adopting the above technical solutions, the computational load and memory usage of deep learning models are reduced, enabling complex flow field reconstruction algorithms to achieve high frame rate real-time inference that meets control requirements on embedded edge devices.

[0017] Preferably, the process of identifying wave system features in the flow field schlieren image using a target detection algorithm includes: running a target detection algorithm based on the YOLOv8 architecture to perform dual-class target recognition on the flow field schlieren image; identifying a first detection category, labeled as a background wave system region, wherein the background wave system region is the upstream region within the isolation segment that is not affected by the main shock wave disturbance or a conventional background wave system structure region; identifying a second detection category, labeled as a shock wave dominant region, wherein the shock wave dominant region is the region containing the core shock wave structure and its induced boundary layer separation region; and outputting the predicted bounding box coordinates and corresponding classification confidence values ​​for each region.

[0018] By adopting the above technical solution, the background flow field and the shock wave interference area were clearly distinguished, providing semantically clear geometric boundary information for subsequent precise positioning.

[0019] Preferably, the process of calculating the shock wave leading edge position using a confidence-weighted algorithm includes: extracting the right edge coordinates of the predicted bounding box of the background wave system region on the flow direction coordinate axis, and the classification confidence score corresponding to the background wave system region; extracting the left edge coordinates of the predicted bounding box of the dominant shock wave region on the flow direction coordinate axis, and the classification confidence score corresponding to the dominant shock wave region; constructing a weighted average correction model, using the classification confidence score as a weighting factor to perform weighted calculation on the right edge coordinates and the left edge coordinates, to obtain the corrected shock wave leading edge position.

[0020] By adopting the above technical solution, the jitter problem that may exist in the transition area of ​​a single detection frame edge is overcome. By using probabilistic statistical characteristics to fuse the position information of adjacent areas, the robustness and accuracy of shock train leading edge localization are improved.

[0021] Preferably, the process of generating a feedback control command based on the relative relationship between the leading edge position of the shock wave train and the preset flow direction position range of the suction channel includes: comparing the numerical values ​​of the leading edge position of the shock wave train and the flow direction position range of the suction channel; when the leading edge position of the shock wave train is less than or equal to the flow direction position range of the suction channel, determining that the flow field of the isolation section is in a state of anti-back pressure requirement, generating an opening command and setting the control flag to 1; when the leading edge position of the shock wave train is greater than the flow direction position range of the suction channel, determining that the flow field of the isolation section is in a safe area state, generating a closing command and setting the control flag to 0.

[0022] By adopting the above technical solution, the continuously changing flow field position state is transformed into a clear binary control logic, and an intelligent judgment mechanism with the position of the suction groove as the physical threshold is established.

[0023] Preferably, the process of sending the feedback control command to drive the solenoid valve to open and close via the serial port relay includes: the edge computing platform sending a hexadecimal control message to the serial port relay via the Modbus RTU protocol; when the control flag is 1, the serial port relay responds to the opening command by driving its internal physical contacts to close, connecting the power supply circuit of the solenoid valve, and opening the pneumatic pipeline channel connected to the suction groove of the isolation section; when the control flag is 0, the serial port relay responds to the closing command by driving its internal physical contacts to open, cutting off the power supply circuit of the solenoid valve, and the solenoid valve resets and closes.

[0024] By adopting the above technical solution, a reliable conversion from digital control signals to physical circuit switching is achieved, directly driving the actuator to change the flow field boundary conditions.

[0025] Preferably, the arrangement of the wall pressure sensors includes: arranging the wall pressure sensors at the inlet of the isolation section to acquire the incoming static pressure data as a reference quantity; arranging the wall pressure sensors at several equally spaced measuring points on the lower wall of the isolation section to acquire the flow pressure distribution data inside the isolation section; and arranging the wall pressure sensors at the outlet of the isolation section to monitor the outlet back pressure; the wall pressure sensors at the above locations together constitute a sparse pressure input array and are physically connected to the synchronous data acquisition device through signal cables.

[0026] By adopting the above technical solution, a standardized sparse sensor array was constructed, which covers the key feature areas of the flow field with the fewest physical measurement points, reducing the complexity of the hardware system.

[0027] This invention provides an intelligent pumping control method for an isolation section based on shock train position feedback. It offers the following advantages: 1. This invention utilizes a lightweight fusion neural network model to establish a nonlinear mapping relationship from sparse wall pressure distribution to dense flow field schlieren images. This method overcomes the limitations of traditional discrete sensors in capturing the overall shape of shock trains and complex wave system structures, achieving visualized reconstruction of the flow field within a closed flow channel without the need for optical windows, providing an intuitive and high-resolution foundation for flow field condition monitoring.

[0028] 2. This invention employs a target detection and confidence-weighted correction algorithm based on YOLOv8. By fusing boundary information and classification confidence of the background wave system region and the shock train dominant region, it achieves accurate calculation of the shock train leading edge position. This technique effectively suppresses positioning jitter errors caused by blurred edges or noise in the reconstructed image, improves the robustness and accuracy of the shock train position feedback signal, and ensures the accuracy and reliability of the control system's judgment benchmark.

[0029] 3. This invention constructs a real-time closed-loop feedback control loop based on an edge computing architecture, integrating flow field reconstruction, feature extraction, and logic determination into an embedded computing platform, and directly driving the actuator to perform boundary layer suction. Through a millisecond-level fast response mechanism, the low-energy boundary layer is removed in time when the shock train approaches the danger zone, effectively anchoring the shock train position and suppressing its upstream jump movement, thereby maintaining the anti-back pressure capability of the flow field in the isolation section and preventing the inlet from failing to start. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the implementation environment and hardware system architecture of an embodiment of the present invention; Figure 2 This is a flowchart of the intelligent suction control method for the isolation section with shock train position feedback according to an embodiment of the present invention.

[0031] Among them, 10 is a wall pressure sensor; 20 is a synchronous data acquisition device; 30 is an edge computing platform; 40 is a serial port relay; and 50 is a solenoid valve. Detailed Implementation

[0032] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] See attached document Figure 1This invention provides an intelligent suction control method for an isolation section with shock train position feedback. The physical carriers on which this method operates mainly include an isolation section test bench, a sensing array composed of wall pressure sensors 10, and a matching active flow control structure.

[0034] The isolation section ground test was conducted in a supersonic direct-drive wind tunnel configured for downdraft operation. High-pressure air was supplied from an air storage tank upstream of the wind tunnel, with a volume of 10 cubic meters and a total pressure of approximately 13 MPa. The air flowed through pneumatic control valves, a stabilization section, and nozzles before entering the isolation section test area. The isolation section test area had dimensions of 50 mm wide, 30 mm high, and 320 mm long. Fused silica observation windows, 260 mm long and 30 mm high, were installed on both side walls of the isolation section test area for optical observation of the internal flow field during the test to obtain comparative verification data. The air within the isolation section test area was ultimately discharged into the atmosphere through a pipe downstream of the test area. By installing different types of nozzles upstream of the wind tunnel, different inlet Mach numbers for the isolation section could be obtained; this embodiment primarily involved inlet Mach numbers of 1.8 and 2.4.

[0035] To simulate downstream congestion caused by combustion chamber back pressure, a flap device is installed at the outlet of the isolation section passage. This flap device is mounted on top of the isolation section outlet and creates a geometric throat by changing the flap angle, thereby generating a throttling effect. As the flap angle increases, the outlet area of ​​the test section decreases, leading to a continuous increase in outlet back pressure, which in turn induces the formation of a shock train structure within the test section. Under different back pressure conditions, the shock train is located at different flow directions within the isolation section.

[0036] To acquire flow field state data in real time, high-frequency dynamic wall pressure sensors 10 are arranged on the wall of the isolation section. The wall pressure sensor 10 has a measurement range of 0 to 1 MPa, an effective response frequency of approximately 20 kHz, and a comprehensive accuracy of ±0.1% of full scale. In this embodiment, to meet the specific requirements of the edge computing flow field reconstruction model for sparse pressure input, specific pressure measurement points are selected to form a pressure input array. Specifically, an inlet static pressure measurement point T0 is arranged at the inlet of the isolation section. The sensor at this location is used to acquire the incoming flow static pressure data as a reference value for dimensionless processing of the subsequently acquired wall pressure data. On the lower wall of the isolation section, pressure measurement points are selected at equal intervals, specifically including measurement points T12, T14, T16, T18, and T20 shown in the figure. In addition, a back pressure measurement point T21 is arranged at the bottom of the isolation section outlet, 30 mm from the top. The wall pressure sensors 10 distributed at different locations are used to convert the pressure physical quantities at each measurement point into analog voltage signals.

[0037] This embodiment features a suction flow control structure on the upper wall of the isolation section for active flow control. The suction channel is located in the center of the visible area on the upper wall of the isolation section, specifically within a flow direction distance of 150 to 160 millimeters from the inlet of the isolation section. The spanwise width of the suction channel is set to 40 millimeters, which is less than the 50-millimeter width of the flow channel in the isolation section to ensure the airtightness of the suction pipeline. The suction channel is connected to the outside atmosphere via a suction pipeline, which contains a valve base. The valve base, designed based on the Vitósinski formula, ensures smooth airflow within the channel. Preferably, this embodiment uses a valve base configuration with an airflow outlet diameter of 16 millimeters. With this configuration, when the suction passage is open, the pressure difference between the high pressure inside the isolation section and the external atmospheric pressure creates a passive suction effect, which can control the shock wave train structure passing through the suction channel.

[0038] The system in this embodiment employs an edge computing architecture as its core processing unit to meet the high computing power requirements of online flow field reconstruction and real-time control. The system is equipped with an edge computing platform 30, which is an industrial-grade edge computer developed based on embedded modules, specifically model TW-T906. This edge computing platform 30 integrates an NVIDIA JETSON AGXORIN module and is equipped with a high-performance graphics processor. It possesses a floating-point operation capability of 200 TOPS, enabling real-time inference computation for deep learning models. In terms of software environment, the edge computing platform 30 is pre-installed with the Ubuntu 20.04 operating system and configured with the TensorFlow deep learning framework runtime environment for deploying and running flow field reconstruction models and object detection algorithms. The edge computing platform 30 is designed with ultra-strong, lightweight aluminum alloy materials and employs conductive passive heat dissipation to adapt to the physical environment of the test site. Furthermore, the edge computing platform 30 supports multiple input / output interfaces, including USB, CAN, RS232, GPIO, and synchronization signal interfaces, for signal interaction with external devices.

[0039] At the data acquisition end, the system is equipped with a synchronous data acquisition device 20, specifically model MCCUSB1808. This synchronous data acquisition device 20 has a high-precision resolution of 18 bits and is configured with 8 synchronous analog input channels. It supports differential signal input and can synchronously acquire voltage signals from all channels at a set sampling frequency. The synchronous data acquisition device 20 connects to the edge computing platform 30 via a communication interface and is responsible for the physical connection to the front-end wall pressure sensor 10. Running on the open-source driver framework uldaq based on the Linux system, the synchronous data acquisition device 20 is responsible for converting the analog electrical signals output by the wall pressure sensor 10 into digital signals and interacting with the operating system of the edge computing platform 30 through a driver program, providing real-time pressure input data for flow field reconstruction.

[0040] At the control end, the system is equipped with a serial port relay 40 and a solenoid valve 50 to implement physical-level suction control. The serial port relay 40, specifically model LH-IO204, serves as a control signal transfer unit. Its input is connected to the edge computing platform 30, receiving digital commands from it. The output is connected to the power supply circuit of the solenoid valve 50, controlling the load's operation by controlling the on / off state of the control circuit. The solenoid valve 50 is a direct-acting type with a rated operating voltage of 24 volts DC and a rated power of 13 watts. It is configured as a normally closed valve; in the de-energized state, an internal spring presses the closing element against the valve seat, keeping the valve closed and isolating the internal isolation section from the outside. When energized, the generated electromagnetic force overcomes the spring force, opening the valve and allowing the internal isolation section to connect to the outside atmosphere through the suction pipeline. The solenoid valve 50 is installed on the suction pipeline, directly responding to the control system's commands to perform active flow control.

[0041] This embodiment constructs a complete hardware communication and connection topology from the perception layer to the control execution layer, and establishes the physical link and data interaction logic between the wall pressure sensor 10, the synchronous data acquisition device 20, the edge computing platform 30, the serial port relay 40, and the solenoid valve 50.

[0042] In the sensing and acquisition link, the wall pressure sensor 10 is physically connected to the analog input channel of the synchronous data acquisition device 20 via a shielded signal cable. The wall pressure sensor 10 converts the sensed airflow pressure into an analog voltage signal and transmits it to the synchronous data acquisition device 20 in real time. A high-speed data communication link is established between the synchronous data acquisition device 20 and the edge computing platform 30. At the communication protocol level, the synchronous data acquisition device 20 and the edge computing platform 30 use the TCP / IP communication protocol for data interaction. In specific implementation, the edge computing platform 30, acting as a client, initiates a connection request to the synchronous data acquisition device 20, configured as a server, by creating a connection-oriented Socket. After the two parties establish a connection through a handshake, the synchronous data acquisition device 20 encapsulates the digital pressure signal, which has undergone analog-to-digital conversion, into a data packet and sends it to the edge computing platform 30 through an Ethernet interface. The edge computing platform 30 reads the data stream in real time by calling a receive function, and utilizes the retransmission control and sequence control mechanisms of the TCP protocol to ensure the integrity and timing consistency of the wall pressure data during transmission, thereby providing a reliable input source for subsequent flow field reconstruction algorithms.

[0043] In the control and execution link, the edge computing platform 30 is connected to the serial relay 40 via a physical interface. Specifically, the edge computing platform 30 uses a USB to RS485 converter to convert the USB communication interface into an RS485 serial communication interface and connects it to the communication terminal of the serial relay 40. At the command transmission level, the edge computing platform 30 and the serial relay 40 use the Modbus RTU industrial communication protocol. The edge computing platform 30 encodes control commands into hexadecimal data frames conforming to the Modbus RTU format according to the control strategy generated by its internal algorithm, and sends them to the serial relay 40 via the serial port. The output terminal of the serial relay 40 is connected in series to the power supply circuit of the solenoid valve 50. This power supply circuit is equipped with a 24-volt DC power supply. When the serial relay 40 receives an open command from the edge computing platform 30, its internal contacts close, connecting the power supply circuit of the solenoid valve 50; conversely, when it receives a close command, the contacts open, cutting off the power supply circuit. The solenoid valve 50 performs corresponding mechanical actions according to the on / off state of the power supply circuit, thereby realizing the physical opening and closing control of the airflow channel of the suction pipeline.

[0044] See attached document Figure 2 This invention provides an intelligent suction control method for the isolation section based on shock train position feedback. The method is executed continuously based on the aforementioned implementation environment and hardware system architecture, and specifically includes the following steps.

[0045] Step S1: Wall pressure data acquisition and moving average preprocessing, edge computing platform 30 / synchronous data acquisition device 20. The synchronous data acquisition device 20 acquires wall pressure data of the isolation section in real time, and performs moving average preprocessing on the edge computing platform 30. The synchronous data acquisition device 20 performs analog-to-digital conversion on the analog voltage signal output by the wall pressure sensor 10 arranged on the wall of the isolation section according to a set sampling frequency. The edge computing platform 30 receives the converted digital pressure signal stream in real time through a communication interface. To eliminate high-frequency noise and signal glitches caused by electromagnetic interference in the industrial test environment, the edge computing platform 30 performs a moving average filtering algorithm on the received raw pressure data sequence, calculating the arithmetic mean within a certain time window as the effective pressure data, thereby providing high signal-to-noise ratio input features for subsequent model inference.

[0046] Step S2: Real-time flow field reconstruction based on a lightweight fusion neural network, outputting a flow field schlieren image. Preprocessed pressure data is input into the lightweight flow field reconstruction model deployed on the edge computing platform 30, outputting a flow field schlieren image. The edge computing platform 30 calls its internally stored flow field reconstruction model, which is a lightweight neural network built based on a deep learning algorithm and optimized by quantization pruning. This model establishes a nonlinear mapping relationship from sparse wall pressure distribution to dense flow field structure. The edge computing platform 30 feeds the filtered pressure data output in step S1 as an input vector into this model, generating a flow field schlieren image reflecting the density gradient changes within the isolated section in real-time.

[0047] Step S3: Based on YOLOv8 and confidence-weighted shock train detection, the position of the shock train leading edge is calculated. The wave system features in the reconstructed image are identified using a target detection algorithm, and the position of the shock train leading edge is calculated using a confidence-weighted algorithm. The edge computing platform 30 performs image processing on the flow field schlieren image generated in step S2. The system uses a target detection algorithm to identify key flow field feature regions in the image, specifically distinguishing between background wave system regions and shock train dominant regions, and obtaining the predicted bounding boxes and corresponding confidence values ​​for each region. Based on this, the system employs a confidence-weighted correction algorithm, comprehensively utilizing the right boundary information of the background wave system regions and the left boundary information of the shock train dominant regions, to calculate the precise physical coordinates of the shock train leading edge in the flow direction of the isolation section.

[0048] Step S4: Compare and determine the position of the shock wave leading edge with the suction channel position. Based on the relative relationship between the shock wave leading edge position and the suction channel position, a feedback control command is generated. The edge computing platform 30 compares the shock wave leading edge position coordinates calculated in step S3 with the pre-stored suction channel flow direction position range. The system evaluates whether the current flow field state meets the back pressure resistance control requirements according to preset judgment logic. When the judgment result shows that the shock wave leading edge position reaches or crosses the suction channel position, the system generates a logic control command to start suction; when the judgment result shows that the shock wave is located in the safe area downstream of the suction channel, the system generates a logic control command to stop suction.

[0049] Step S5: The serial port relay 40 drives the solenoid valve 50 to perform intelligent suction control. The serial port relay 40 sends control commands to drive the solenoid valve 50 to open and close, achieving intelligent suction control of the flow field in the isolation section. The edge computing platform 30 converts the logic control commands generated in step S4 into a communication protocol format and sends them to the serial port relay 40 through the communication interface. The serial port relay 40 controls the on / off state of the output circuit according to the received commands, thereby driving the solenoid valve 50 connected in the circuit to perform mechanical opening or closing actions. After the solenoid valve 50 opens, the high-pressure area inside the isolation section is connected to the low-pressure atmosphere outside through the suction pipeline. The pressure difference is used to remove the low-energy boundary layer, thereby suppressing the abrupt upstream movement of the shock wave train and achieving closed-loop steady-state control of the flow field state in the isolation section.

[0050] See attached document Figure 2 This invention provides an intelligent suction control method for an isolated section with shock train position feedback. In step S1, real-time data acquisition and transmission rely on a high-speed network communication link established between an edge computing platform 30 and a synchronous data acquisition device 20. To ensure the real-time performance and reliability of pressure data transmission, this embodiment uses the Transmission Control Protocol (TCP) as the underlying communication standard. In terms of network architecture configuration, the edge computing platform 30 is configured as a TCP client, and the synchronous data acquisition device 20 is configured as a TCP server. This master-slave communication architecture enables the edge computing platform 30 to proactively initiate data requests based on the load status of the computing task, thereby effectively coordinating the data stream reception rhythm.

[0051] Edge computing platform 30 establishes a communication connection with synchronous data acquisition device 20 by calling socket interface functions. During system initialization, edge computing platform 30 initiates a connection request to synchronous data acquisition device 20, which points to the Internet Protocol address and port number pre-set by synchronous data acquisition device 20. After both parties complete the handshake protocol, a full-duplex communication channel is established. Synchronous data acquisition device 20 then enters the data transmission state, while edge computing platform 30 starts a receiving thread, reading the data stream from the network buffer by repeatedly calling the receiving function. This process utilizes the inherent error control and retransmission mechanism of the TCP protocol, which can automatically correct bit errors or packet loss that may occur during transmission, ensuring the integrity and strict timing consistency of the wall pressure data sequence during transmission, providing accurate time series input for subsequent flow field reconstruction.

[0052] The synchronous data acquisition device 20 digitizes the analog signal output by the wall pressure sensor 10 according to the set sampling parameters. In this embodiment, the sampling frequency is set to 10 kHz, which meets the requirements for capturing the high-frequency oscillation characteristics of the shock train within the isolation section. The analog-to-digital converter inside the synchronous data acquisition device 20 synchronously acquires the voltage values ​​of all channels in each sampling cycle and converts these voltage values ​​into digital quantities. The converted digital pressure data is not transmitted discretely in a single-point manner, but is encapsulated according to a predefined data packet format. Each data packet contains a frame header, pressure data payloads for each channel, and a check bit. The synchronous data acquisition device 20 continuously sends the encapsulated data packets to the edge computing platform 30 through the Ethernet physical interface. After receiving the data packets, the edge computing platform 30 parses them according to the protocol format, extracts the real-time pressure values ​​corresponding to each measuring point, and stores them in a memory queue for the next step of filtering processing.

[0053] This invention provides an intelligent suction control method for the isolation section with shock train position feedback. In step S1, in order to ensure the accuracy of the subsequent flow field reconstruction model input, the edge computing platform 30 immediately performs moving average filtering processing after receiving the original pressure data.

[0054] Because the industrial environment where the isolation section test rig is located typically has complex electromagnetic fields, and the operation of high-power wind tunnel equipment generates continuous electromagnetic interference, high-frequency random noise and signal spikes inevitably get mixed into the analog signals collected by the wall pressure sensor 10. If the raw data with noise is used directly for flow field reconstruction, artifacts will appear in the reconstructed flow field schlieren image, thus affecting the positioning accuracy of the shock train leading edge. Therefore, this embodiment uses software filtering to eliminate interference. The edge computing platform 30 independently establishes a first-in-first-out buffer pool for the data channels of each wall pressure sensor 10 in its memory space. The capacity of this buffer pool is determined by the filtering window length parameter. In this embodiment, in order to balance the denoising effect and the dynamic response delay of the system, the filtering window length parameter is set to 10.

[0055] When the edge computing platform 30 receives a new sampling data packet, the system pushes the latest pressure data into the corresponding buffer pool, while removing the oldest data point from the buffer pool, always maintaining the 10 most recently sampled data points in the buffer pool. The system calculates the arithmetic mean of all data points currently stored in the buffer pool and uses the calculated average as the effective pressure value at the current moment. For any given pressure measurement point, the filtered pressure value is calculated according to the following formula: ; in, Representing the The effective pressure value of a wall pressure sensor at the current moment after filtering; This represents the length of the sliding filter window; Representing the The wall pressure sensor at the current moment Push forward The system collects raw pressure data at each sampling period. Through the above calculation process, the system can effectively smooth out instantaneous pressure fluctuations caused by environmental interference, improve the signal-to-noise ratio of the input data, and thus provide a stable and accurate data foundation for the flow field reconstruction algorithm that reflects the physical state of the flow field.

[0056] In step S1, after the edge computing platform 30 completes the moving average filtering process, it continues to perform dimensionless and normalization processing on the pressure data.

[0057] Because the isolation section test bench may be under different incoming Mach number conditions or different total pressure conditions during actual operation, the absolute values ​​of the wall pressure collected by the wall pressure sensor 10 will have significant differences in magnitude. In order to eliminate the influence of the incoming absolute pressure value on the flow field feature extraction and improve the generalization ability of the flow field reconstruction model under different conditions, this embodiment adopts a dimensionless processing strategy. The edge computing platform 30 uses the real-time inlet static pressure data collected by the inlet static pressure measuring point T0 arranged at the inlet of the isolation section as a reference quantity. The system divides the effective pressure value of each wall pressure sensor 10 after moving average filtering by the inlet static pressure value at the same time to calculate the wall pressure ratio sequence that eliminates the influence of dimensions. This sequence can purely reflect the pressure change characteristics caused by the shock train structure inside the isolation section, and is independent of the total pressure level of the incoming flow.

[0058] To adapt to the input layer data specifications of the lightweight flow field reconstruction model deployed in the edge computing platform 30, and to prevent gradient vanishing or gradient explosion during neural network training and inference due to excessive differences in the range of input data values, the system further employs a max-min normalization algorithm to linearly map the dimensionless wall pressure ratio to a closed interval between zero and one. The edge computing platform 30 performs a linear transformation on the dimensionless wall pressure ratio based on pre-defined pressure ratio boundary parameters, generating normalized pressure feature values. These normalized pressure feature values ​​form a standardized feature vector, which serves as the direct input to the flow field reconstruction model in step S2. The normalization calculation process follows the core formula below: ; in, Representing the The input characteristic value of each wall pressure sensor after normalization at any time; Representing the The effective pressure value of a wall pressure sensor after being filtered by a moving average at any time; The inlet static pressure value collected at inlet static pressure measuring point T0 at time T0; This represents the maximum boundary value of the preset dimensionless pressure ratio; This represents the minimum boundary value of the preset dimensionless pressure ratio. Through the above processing, the system ensures that the data distribution of the input flow field reconstruction model is numerically stable and uniform.

[0059] See attached document Figure 2In step S2, the edge computing platform 30 uses the constructed fusion neural network model to reconstruct the preprocessed sparse wall pressure data into a high-resolution flow field schlieren image. This fusion neural network model employs a dual-branch parallel processing architecture, specifically including a convolutional neural network branch and a fully connected neural network branch. These two branches extract flow field features from different dimensions, and finally, feature stitching and decoding are performed in the fusion layer for output.

[0060] The convolutional neural network branch is primarily used to recover the two-dimensional spatial structure features of the flow field from discrete pressure data. The normalized pressure feature vector from the input layer first enters the initial fully connected layer of this branch, where matrix multiplication maps the low-dimensional pressure input to a high-dimensional latent feature space. Subsequently, the system performs a reshaping operation, transforming the one-dimensional high-dimensional feature vector into an initial two-dimensional feature matrix with a specific number of channels. This initial two-dimensional feature matrix then enters a cascaded transposed convolutional layer. The transposed convolutional layer performs an upsampling operation, progressively enlarging the spatial resolution of the feature matrix through learnable convolutional kernel parameters, making it approximate the size of the target output image. After each upsampling operation, the model is configured with a standard convolutional layer to spatially filter the enlarged feature map, sharpening local details such as shock wave edges and shear layers in the flow field, and eliminating the checkerboard effect generated during the upsampling process.

[0061] The fully connected neural network branch is primarily used to extract global nonlinear correlation features between wall pressure data. This branch contains multiple stacked fully connected layers, with nonlinear activation functions inserted between each layer. The input pressure vector undergoes layer-by-layer linear transformations and nonlinear mappings within this branch, uncovering the intrinsic coupling relationships between pressure values ​​at different spatial locations. The output of this branch is a global feature vector encoding the overall flow field intensity and macroscopic evolution trend.

[0062] In the fusion layer of the model, the system concatenates the spatial feature map output by the convolutional neural network branch with the global feature vector output by the fully connected neural network branch. Specifically, the system first broadcasts and copies the global feature vector in the spatial dimension to match its size with the spatial feature map, and then performs the concatenation operation in the channel dimension. The concatenated fused feature tensor contains both local detail information and global intensity information. This fused feature tensor is finally passed through the output convolutional layer for dimensionality reduction and regression to generate a single-channel pixel matrix, which is the final reconstructed flow field schlieren image. The end-to-end mapping relationship of this fusion neural network model is defined as follows: ; To optimize model parameters, mean squared error is used as the loss function during the training phase, and its definition is as follows: ; In the above formula, The normalized wall pressure feature vector represents the input; It represents the set of all learnable weights and bias parameters in a fusion neural network model; This represents the non-linear mapping function from input to output as defined in a neural network. The reconstructed flow field schlieren image matrix represents the output of the model; Representing the The labels of the real flow field schlieren images corresponding to each training sample; Representative model for the first The predicted image output from each training sample; This represents the size of the training batch; This represents the mean squared error loss function value, and the goal of model training is to minimize this loss function value through the backpropagation algorithm.

[0063] In step S2, in order to ensure that the fused neural network model can achieve an inference speed that meets the real-time control requirements on the edge computing platform 30, the present invention implements model lightweighting and edge deployment processing after the model training is completed.

[0064] Considering the computational resource and power consumption limitations of the edge computing platform 30, this embodiment first performs structural pruning optimization on the trained fusion neural network model. The system analyzes the sparsity of weights in each convolutional layer of the model, identifies and removes redundant feature channels whose contribution to the output result is lower than a preset threshold. Specifically, the system reduces the number of channels in the deep convolutional layers of the model's backbone network, for example, reducing the number of feature channels from the original 256 to 128. Through this structural pruning operation, the number of floating-point operations in the model's forward inference process is reduced, alleviating the computational load while maintaining the accuracy of flow field reconstruction.

[0065] Building upon structural pruning, the system further performs numerical quantization on the model parameters. While 32-bit single-precision floating-point format is typically used during model training to ensure the accuracy of gradient descent, excessively high data precision can consume significant GPU memory bandwidth during inference. Therefore, this embodiment employs a half-precision quantization strategy, converting all weight and bias parameters in the model from 32-bit floating-point format to 16-bit floating-point format. This quantization process converts the model file into a TensorFlow Lite format suitable for mobile deployment. This conversion not only halves the storage size of the model file but also effectively reduces the memory throughput latency of the edge computing platform 30 when loading model parameters.

[0066] After lightweight processing, the model file is finally deployed to the runtime environment of the edge computing platform 30. The edge computing platform 30 calls its built-in high-performance deep learning inference compiler to perform hardware-level compilation optimization on the model computation graph. This compiler fuses and pipelines convolution operators for the graphics processing unit architecture inside the edge computing platform 30, eliminating redundant nodes in the computation graph. After the above series of processes including pruning, quantization, and compilation optimization, the inference time for single-frame flow field schlieren image reconstruction on the edge computing platform 30 is controlled to around 20 milliseconds, thereby ensuring that the response frequency of the entire control loop can match the high dynamic change characteristics of the flow field in the isolation section.

[0067] See attached document Figure 2 In step S3, to accurately extract the location information of the shock wave train from the reconstructed flow field schlieren image, the edge computing platform 30 runs a target detection algorithm based on the YOLOv8 architecture. This algorithm is configured to perform a specific dual-class target recognition task, aiming to separate and locate different flow structures in the flow field image.

[0068] This embodiment predefines two detection target categories to adapt to the physical characteristics of the flow field in the isolation section. The first detection category, labeled Class0, is defined as the background wave system region. This region physically corresponds to the upstream region within the isolation section unaffected by the main shock train disturbance or a conventional background wave system structure region, which typically exhibits relatively regular or weak density gradient changes in the reconstructed image. The second detection category, labeled Class1, is defined as the shock train-dominated region. This region physically includes the core shock structure of the shock train and its induced boundary layer separation region, which manifests as strong alternating bright and dark stripes and complex turbulent structure features in the reconstructed image. By distinguishing between these two target categories, the system can effectively isolate the shock train body from the complex background flow field.

[0069] Before the actual deployment of the algorithm, this invention performed supervised training on the YOLOv8 model using a pre-constructed image dataset. This dataset consists of a large number of historical flow field schlieren images generated by the flow field reconstruction algorithm, covering flow field morphology under different Mach numbers and backpressure conditions. To adapt to the standard input layer resolution of the YOLOv8 network model and improve inference efficiency at the edge, the reconstructed image dataset used for training underwent downsampling. During the training phase, the neural network iteratively optimized the network weights through backpropagation, learning to extract the morphological and textural features of the Class 0 background wave system region and the Class 1 shock wave dominant region under different flow field conditions. After training, the target detection model can receive the real-time flow field schlieren image output from step S2 and output prediction results containing the target bounding box coordinates and corresponding classification confidence, providing quantified geometric boundary information for subsequent shock wave leading edge position calculation.

[0070] In step S3, after obtaining preliminary detection box information using the target detection algorithm, the edge computing platform 30 further executes a position weighted correction algorithm based on the detection box confidence to accurately calculate the physical position of the shock train leading edge.

[0071] Since the flow field schlieren image reconstructed from sparse pressure data is essentially an approximate mapping of the physical properties of the real flow field, its sharpness and edge sharpness are limited by the spatial resolution of the input data and the inference performance of the reconstruction model. In some transition regions, the pixel boundaries between the background wave system and the shock train structure may appear blurred or jittery. This uncertainty in image features can cause slight localization deviations in the detection box boundaries output by the target detection model. To eliminate this localization error caused by image reconstruction noise, the system does not rely solely on the left edge of the shock train target box as the leading edge position, but instead introduces the position information of the adjacent background wave system target box and the detection confidence of both for joint calculation.

[0072] The edge computing platform 30 analyzes the detection result tensor output by the YOLOv8 model, extracting two key geometric coordinate parameters and two corresponding probability parameters. The system first reads the right boundary coordinates of the detection boxes classified as background wave systems on the flow axis, defining this as the termination position of the background wave system; simultaneously, it reads the left boundary coordinates of the detection boxes classified as shock wave dominant regions on the flow axis, defining this as the starting position of the shock wave. Physically, these two coordinates should theoretically coincide at the same point, i.e., the leading edge of the shock wave. Furthermore, the system simultaneously extracts the classification confidence scores of the background wave system detection boxes and the shock wave dominant region detection boxes. These two confidence scores, output by the Softmax layer of the detection model, range from zero to one, objectively reflecting the degree of certainty the neural network has in recognizing the features of the current detection box.

[0073] The system constructs a weighted average correction model using the four extracted parameters. The basic logic of this model is to use high-confidence location information to correct low-confidence location information. For example, when the shock wave region features are not obvious, leading to low detection confidence, while the background wave system features are clear, leading to high detection confidence, the final position calculated by the system will be more inclined towards the right boundary of the background wave system, thereby suppressing the interference caused by shock wave detection frame fluctuations. The weighted calculation formula for the shock wave leading edge position is as follows: ; in, This represents the corrected position of the shock train leading edge; The coordinates of the right edge of the detection box representing the background wave system region; The classification confidence level represents the detection box corresponding to the background wave system region; The coordinates of the left edge of the detection box representing the dominant region of the shock train; This represents the classification confidence score corresponding to the detection box in the dominant region of the shock train. It is calculated using this algorithm. The value integrates the characteristic information of two adjacent flow field regions, providing a highly robust position feedback signal for subsequent feedback control.

[0074] See attached document Figure 2 In step S4, the edge computing platform 30, based on the relative position of the shock wave leading edge and the physical suction channel calculated in step S3, runs the shock wave position feedback control law to generate specific control commands. This control law aims to determine whether to initiate boundary layer suction by judging whether the shock wave crosses a specific flow direction control boundary, thereby achieving closed-loop control of the flow field state inside the isolation section.

[0075] The edge computing platform 30 pre-stores fixed position parameters of the suction channel in the flow direction coordinate system of the isolation section. These position parameters are defined as the flow direction position interval of the suction channel, denoted by the symbol... This position parameter corresponds to the physical center coordinates or starting coordinates of the suction groove opened on the lower wall of the isolation section, and is a key geometric threshold for determining the flow field's resistance to back pressure. Within each control cycle, the system will output the shock train leading edge position from step S3. The flow direction of the suction channel is within the range of positions. Perform numerical comparisons.

[0076] When the edge computing platform 30 determines the position of the leading edge of the shock wave... Less than or equal to the flow direction position range of the suction channel At this point, it indicates that the leading edge structure of the shock train has moved upstream along the flow direction and reached or crossed the physical position of the suction channel. The system determines that the flow field in the isolation section is in a state requiring resistance to back pressure, meaning the shock train is being driven by the high back pressure downstream and is approaching the inlet throat, posing a risk of the inlet not starting. Based on this determination, the edge computing platform 30 generates an activation command, setting the internal control flag to 1. This state indicates that the system needs to immediately activate the suction mechanism, using the suction channel to remove low-energy boundary layer fluid to enhance the wall boundary layer's resistance to back pressure near the leading edge of the shock train, thereby suppressing further forward movement of the shock train.

[0077] Conversely, when the edge computing platform 30 determines the position of the shock train leading edge... Larger than the flow direction range of the suction channel When the shock wave train's leading edge structure is located downstream of the suction channel, the system determines that the flow field in the isolation section is in a safe region. This means the shock wave train is relatively far back, has sufficient stability margin from the intake throat, and does not cause significant interference to the upstream flow field. Based on this determination, the edge computing platform 30 generates a shutdown command, setting the internal control flag to 0. This state indicates that the system does not need to actively intervene and should cut off the suction path to avoid unnecessary mass flow loss and maintain system efficiency. Through the aforementioned binary logic determination process, the system converts the continuously changing shock wave train position signal into a discrete switching control signal.

[0078] See attached document Figure 2 In step S5, the edge computing platform 30 converts the generated feedback control command into a standard industrial control signal through a physical communication link and sends it to the actuator to drive the hardware device to complete the corresponding air path switching action.

[0079] The edge computing platform 30 and the serial relay 40 use the Modbus RTU serial communication protocol for data exchange to ensure the reliability and anti-interference capability of command transmission. At the physical connection level, the edge computing platform 30 uses its built-in universal serial bus interface to connect to a USB-to-RS485 converter module, transmitting differential level signals to the communication terminals of the serial relay 40 via twisted-pair cable. The serial relay 40 is pre-configured as a slave device on the bus and assigned a unique communication address to match the master station's addressing request. The edge computing platform 30, acting as the control master station, assembles hexadecimal control messages conforming to the Modbus protocol frame format in real time based on the logical state of the control flag bit (Flag) output in step S4.

[0080] When the control flag generated in step S4 is 1, indicating that the system determines a suction operation is required, the edge computing platform 30 sends a relay activation command message to the serial port relay 40. The serial port relay 40 receives and parses the message, and after passing cyclic redundancy check, drives its corresponding internal physical contact to close. This physical contact is connected in series in the 24V DC external power supply circuit of the solenoid valve 50. The closing action of the contact creates a closed circuit between the external power supply and the electromagnetic coil of the solenoid valve 50. Under the electromagnetic force generated by the coil, the solenoid valve 50 overcomes the resistance of the internal mechanical spring and quickly lifts the valve core, thereby opening the pneumatic pipeline channel connected to the suction groove of the isolation section and initiating the boundary layer suction process.

[0081] When the control flag generated in step S4 is 0, indicating that the system determines that suction is unnecessary or needs to be stopped, the edge computing platform 30 sends a relay disconnect command message to the serial port relay 40. The serial port relay 40 responds to this command by releasing its internal physical contacts, physically cutting off the power supply circuit to the solenoid valve 50. The solenoid valve 50 then loses its electromagnetic attraction and, under the mechanical force of its built-in return spring, pushes the valve core back to its normally closed state, thereby blocking the airflow through the suction pipeline. Through this conversion process from digital signals to analog actions, the system achieves real-time closed-loop intervention of the physical boundary conditions of the flow field in the isolation section.

[0082] See attached document Figure 2 After step S5 is completed and the solenoid valve 50 is opened, the isolation section suction control system produces a substantial physical regulation effect, which achieves passive constraint on the position of the shock wave train by changing the boundary conditions of the flow field.

[0083] Because the flow field inside the isolation section has a high static pressure level in the region where the shock train exists, while the external environment or vacuum container connected to the outlet of solenoid valve 50 is under relatively low pressure, a pressure gradient pointing outwards from the suction tank is established inside and outside the suction tank. At the instant solenoid valve 50 is turned on, this pressure gradient drives the low-energy boundary layer fluid accumulated near the suction tank region to be rapidly discharged from the isolation section flow channel through the suction pipe. The 16 mm diameter suction pipe selected in this embodiment provides sufficient flow area to ensure that continuous mass flow removal effectively weakens the shock train-induced boundary layer separation.

[0084] Removing the boundary layer weakens the interaction between the shock wave and the boundary layer, creating a hydrodynamically stable node at the suction chute location. This physical mechanism anchors the shock train structure near the flow direction of the suction chute, limiting the tendency for the shock train to move violently upstream during front back pressure fluctuations or increases. By suppressing the abrupt jumps of the shock train towards the inlet throat, this control method maintains stable aerodynamic parameters at the inlet of the isolation section, effectively preventing inlet start-up and ensuring the operational safety of the propulsion system under off-design conditions.

Claims

1. A method for intelligent suction control of an isolation section with shock train position feedback, characterized in that, Includes the following steps: The wall pressure data output by the isolation section wall pressure sensor (10) is collected in real time using the synchronous data acquisition device (20), and the wall pressure data is transmitted to the edge computing platform (30) for preprocessing. The preprocessed wall pressure data is input into the lightweight flow field reconstruction model deployed on the edge computing platform (30), and the flow field schlieren image reflecting the density gradient change inside the isolation section is output. The wave system features in the flow field schlieren image are identified using a target detection algorithm, and the position of the shock train leading edge is calculated using a confidence weighting algorithm. Based on the relative relationship between the leading edge position of the shock wave train and the preset flow direction position range of the suction groove, a feedback control command is generated. The feedback control command is sent through the serial port relay (40) to drive the solenoid valve (50) to open and close, thereby realizing intelligent suction control of the flow field in the isolation section.

2. The intelligent suction control method for the isolation section with shock train position feedback according to claim 1, characterized in that, The process of using a synchronous data acquisition device (20) to collect wall pressure data output by the isolation section wall pressure sensor (10) in real time, and transmitting the wall pressure data to the edge computing platform (30) for preprocessing includes: The synchronous data acquisition device (20) performs analog-to-digital conversion on the analog voltage signal output by the wall pressure sensor (10) according to a preset sampling frequency, encapsulates the converted digital pressure data, and sends it to the edge computing platform (30) via TCP / IP protocol. The edge computing platform (30) establishes a first-in-first-out buffer pool for the data channel of each wall pressure sensor (10), performs a moving average filtering algorithm on the received raw pressure data sequence, and calculates the arithmetic mean within the filtering window as the effective pressure data.

3. The intelligent suction control method for the isolation section with shock train position feedback according to claim 2, characterized in that, The preprocessing process performed by the edge computing platform (30) also includes: Using the real-time inlet static pressure data collected by the wall pressure sensor (10) arranged at the entrance of the isolation section as a reference quantity, the effective pressure value of each wall pressure sensor (10) after sliding average filtering is divided by the real-time inlet static pressure data at the same time to obtain the dimensionless wall pressure ratio. The dimensionless wall pressure ratio is linearly mapped to a numerical range of zero to one using the max-min normalization algorithm, generating a normalized pressure feature vector as the input to the lightweight flow field reconstruction model.

4. The intelligent suction control method for the isolation section with shock train position feedback according to claim 1, characterized in that, The process of inputting the preprocessed wall pressure data into the lightweight flow field reconstruction model deployed on the edge computing platform (30) and outputting a flow field schlieren image reflecting the density gradient change inside the isolation section includes: The wall pressure data is input into the convolutional neural network branch of the lightweight flow field reconstruction model. Upsampling is performed through fully connected layers and transposed convolutional layers to output the two-dimensional spatial structure features of the flow field. The wall pressure data is input into the fully connected neural network branch of the lightweight flow field reconstruction model, and the global nonlinear correlation features between the wall pressure data are extracted through multiple fully connected layers. In the fusion layer of the lightweight flow field reconstruction model, the two-dimensional spatial structural features are stitched together with the global nonlinear correlation features, and the flow field schlieren image is generated by decoding through the output convolutional layer.

5. The intelligent suction control method for the isolation section with shock train position feedback according to claim 4, characterized in that, The lightweight flow field reconstruction model underwent model lightweighting processing before being deployed on the edge computing platform (30). The model lightweighting processing included: Structural pruning optimization: Analyze the sparsity of weights in the model's convolutional layers, identify and remove redundant feature channels whose contribution to the output is lower than a preset threshold, and reduce the number of floating-point operations in the model's forward inference. Half-precision quantization: Convert the weight parameters and bias parameters in the model from 32-bit floating-point format to 16-bit floating-point format, and compile them into a file format suitable for mobile deployment, thereby reducing the memory throughput latency of the edge computing platform (30).

6. The intelligent suction control method for the isolation section with shock train position feedback according to claim 1, characterized in that, The process of identifying wave system features in the flow field schlieren image using a target detection algorithm includes: A target detection algorithm based on the YOLOv8 architecture is used to perform dual-class target recognition on the flow field schlieren image; Identify the first detection category and mark it as the background wave system region. The background wave system region is the upstream region within the isolation section that is not affected by the main disturbance of the shock train or the conventional background wave system structure region. The second detection category is identified and marked as the shock train dominance region, which is a region containing the core shock structure and the induced boundary layer separation region; Output the predicted bounding box coordinates and corresponding classification confidence scores for each region.

7. The intelligent suction control method for the isolation section with shock train position feedback according to claim 6, characterized in that, The process of calculating the position of the shock train leading edge using the confidence-weighted algorithm includes: Extract the right edge coordinates of the predicted bounding box of the background spectral region on the flow direction coordinate axis, and the classification confidence corresponding to the background spectral region; Extract the left edge coordinates of the predicted bounding box of the shock wave dominant region on the flow direction coordinate axis, and the classification confidence corresponding to the shock wave dominant region; A weighted average correction model is constructed, and the classification confidence score is used as a weighting factor to perform weighted calculations on the right edge coordinates and the left edge coordinates to obtain the corrected shock train leading edge position.

8. The intelligent suction control method for the isolation section with shock train position feedback according to claim 1, characterized in that, The process of generating feedback control commands based on the relative relationship between the leading edge position of the shock wave train and the preset flow direction position range of the suction channel includes: The position of the leading edge of the shock wave train is numerically compared with the flow direction position range of the suction groove; When the position of the leading edge of the shock wave train is less than or equal to the position range of the flow direction of the suction groove, it is determined that the flow field of the isolation section is in a state of anti-back pressure requirement, an opening command is generated and the control flag is set to 1. When the position of the leading edge of the shock wave is greater than the flow direction range of the suction channel, it is determined that the flow field of the isolation section is in a safe area, a shutdown command is generated, and the control flag is set to 0.

9. The intelligent suction control method for the isolation section with shock train position feedback according to claim 8, characterized in that, The process of sending the feedback control command through the serial port relay (40) to drive the solenoid valve (50) to open and close includes: The edge computing platform (30) sends hexadecimal control messages to the serial port relay (40) via the Modbus RTU protocol; When the control flag is 1, the serial port relay (40) responds to the opening command to drive the internal physical contacts to close, connect the power supply circuit of the solenoid valve (50), and open the pneumatic pipeline channel connected to the suction groove of the isolation section. When the control flag is 0, the serial port relay (40) responds to the shutdown command and drives the internal physical contacts to disconnect, cutting off the power supply circuit of the solenoid valve (50), and the solenoid valve (50) is reset and closed.

10. The intelligent suction control method for the isolation section with shock train position feedback according to claim 1, characterized in that, The arrangement of the wall pressure sensor (10) includes: The wall pressure sensor (10) is arranged at the entrance of the isolation section to obtain the static pressure data of the incoming flow as a reference quantity; Several equally spaced measuring points are selected on the lower wall of the isolation section to arrange the wall pressure sensor (10) to obtain the flow pressure distribution data inside the isolation section; The wall pressure sensor (10) is installed at the outlet of the isolation section to monitor the outlet back pressure; The wall pressure sensors (10) at the above locations together form a sparse pressure input array and are physically connected to the synchronous data acquisition device (20) via signal cables.

Images