Aircraft engine parameter indication detection system and detection method

Abstract

The invention belongs to the field of instrument detection, and provides an aircraft engine parameter indication detection system and method, and the method comprises the steps: generating a first sensor signal for an engine parameter indicator simulation test according to a first parameter; the first sensor signal is output to an engine parameter indicator in an analog signal mode so as to simulate the working state of an engine; collecting a first indication response signal generated after the engine parameter indicator receives the analog signal; calling a first pre-training artificial intelligence model, judging whether the working state of the engine parameter indicator is abnormal or not based on the first sensor signal and the first indication response signal, and outputting a corresponding detection result; calling a second pre-training artificial intelligence model, and combining the first sensor signal and the first indication response signal to generate a second parameter; judging whether to continue to execute the test process based on the second parameter; therefore, the testing precision and efficiency of the aircraft engine can be effectively improved.

Description

Technical Field

[0001] The present invention belongs to the field of instrument detection, and in particular relates to an aircraft engine parameter indication detection system and a detection method. Background Art

[0002] As the core power unit of an aircraft, the operating status of an aircraft engine is directly related to flight safety. To ensure visual monitoring of engine status, modern aircraft are often equipped with an Engine Parameter Indicator (EPI), which displays multiple key operating parameters, including engine speed, temperature, and pressure, in real time. These parameters typically originate from analog signals collected by sensors, which are input into the EPI via an interface circuit and converted into visual indicators. EPI equipment is not only the primary means for pilots to monitor engine operating status, but also a crucial tool for fault analysis during ground maintenance.

[0003] Prior art methods for testing engine parameter indicators primarily involve manually setting standard inputs, simulating physical signals channel by channel, and reading the response values. This type of testing typically relies on analog signal sources, manually adjusting voltage levels or frequencies, and determining the accuracy of each indicator. Some systems utilize specialized test fixtures and testers to complete the testing process. While some advanced testing systems incorporate program-controlled signal generation and response reading modules, most still require testers to pre-set various input parameters and determine the EPI response status based on experience.

[0004] The main problem with existing technologies is that their testing processes rely heavily on manual experience and fixed rules, lacking the ability to deeply learn the behavior of indicators under different engine operating conditions. This results in low testing efficiency, limited detection dimensions, and difficulty in discovering hidden faults. Furthermore, traditional detection methods often fail to automatically adjust input parameters based on EPI responses, lack closed-loop control capabilities, and are unable to dynamically explore the most representative test conditions. They also struggle to effectively diagnose complex, interactive faults. Summary of the Invention In order to solve the problems in the prior art, the present invention provides an aircraft engine parameter indication detection system, comprising: a simulation signal construction module, configured to generate a first sensor signal for an engine parameter indicator simulation test according to a first parameter; an analog signal output module, configured to output the first sensor signal to the engine parameter indicator in the form of an analog signal to simulate the working state of the engine; a response acquisition module, configured to acquire a first indication response signal generated by the engine parameter indicator after receiving the analog signal; a state determination and output module, configured to call a first pre-trained artificial intelligence model, determine whether the working state of the engine parameter indicator is abnormal based on the first sensor signal and the first indication response signal, and output a corresponding detection result; a feedback analysis module, configured to call a second pre-trained artificial intelligence model and generate a second parameter by combining the first sensor signal and the first indication response signal; A loop control module is used to determine whether to continue the test process based on the second parameter. If it is determined to continue the test, the first parameter is replaced by the second parameter, and the analog signal construction module, the analog signal output module, the response acquisition module, the state judgment and output module and the feedback analysis module are called in sequence.

[0005] Furthermore, the first parameter is a preset initialization parameter for starting the first round of testing.

[0006] Furthermore, the first parameter is a random initialization parameter generated within a set range.

[0007] Furthermore, the analog signal construction module includes: A signal configuration submodule, configured to set the voltage value, current value or frequency parameter of the analog signal of each sensor based on the first parameter; A signal conversion submodule is used to convert the configured parameters into an electrical signal format that complies with the target engine parameter indicator interface protocol; The signal stabilization submodule is used to perform amplitude filtering and timing compensation on the generated analog signal to ensure stable signal output.

[0008] Furthermore, the analog signal output module includes: A drive interface submodule, configured to convert the first sensor signal into a hardware signal that can be driven and output by an analog port; The buffer adjustment submodule is used to perform amplitude compensation, anti-interference processing and impedance matching on the signal according to the indicator type; The signal sending submodule is used to send the analog signal to the corresponding port of the target engine parameter indicator according to the preset timing.

[0009] Furthermore, the response collection module includes: Multi-channel acquisition submodule, used to synchronously acquire analog signal outputs corresponding to multiple sensor response channels; The analog-to-digital conversion submodule is used to convert the collected analog response signal into digital data; The buffer storage submodule is used to cache the digital response data into the internal memory for subsequent analysis.

[0010] Furthermore, the first pre-trained artificial intelligence model structure includes: an input feature embedding layer, configured to receive the first sensor signal and the first indication response signal, and map them to a feature space of uniform dimension to form a joint input tensor; A graph construction and graph attention layer is used to construct a feature association graph based on the joint input tensor, use a graph attention mechanism to calculate the attention weight between each feature node and its adjacent nodes, and extract local sensitivity information and the coupling relationship between parameters; The residual connection fusion layer is used to perform a residual connection between the output of the graph attention layer and the output of the input feature embedding layer. It improves the feature expression capability through feature splicing and element-by-element weighting operations to avoid gradient disappearance or feature degradation. The abnormal state discrimination layer is used to perform binary or multi-classification judgment based on the fused high-dimensional feature vector, and output the working state judgment result of the engine parameter indicator under the current input conditions.

[0011] Furthermore, the second pre-trained artificial intelligence model structure includes: An input coding layer, configured to receive the first sensor signal and the first indication response signal, perform feature normalization and channel alignment on each of the first sensor signal and the first indication response signal, and then concatenate the two signals to form an input feature vector; The dual-gated factor generation layer is used to generate a set of weighted parameter adjustment factors based on the current input features, including amplitude factors and direction factors. The amplitude factors control the adjustment degree, and the direction factors control the adjustment direction of each channel. The historical memory fusion layer is used to call the test parameters and feedback status of the previous round, dynamically weighted fuse them with the current feature input, and extract the cross-round feature evolution trend; The parameter update generation layer is used to calculate the second parameter output of the current round based on the fused memory vector.

[0012] Furthermore, the loop control module determines whether to continue executing the test process based on one of the following conditions: The judgment result of the first pre-trained artificial intelligence model is a non-abnormal state, and the error change for two consecutive rounds is lower than a preset threshold; The parameter adjustment direction change angle output by the feedback analysis module is less than a preset angle threshold; The number of test rounds executed exceeds the preset maximum round limit.

[0013] The present invention also provides an aircraft engine parameter indication detection method, which uses any of the aforementioned aircraft engine parameter indication detection systems to perform aircraft engine parameter indication detection.

[0014] This invention provides an aircraft engine parameter indicator detection system based on an artificial intelligence model. It automatically generates simulated sensor signals and performs closed-loop feedback analysis. Using two pre-trained models embedded in the system, it detects anomalies and optimizes parameters, enabling efficient and intelligent detection of EPI devices. By automatically constructing simulated test signals and collecting EPI feedback responses, the system replaces manual test parameter adjustments, effectively improving test accuracy and efficiency.

[0015] The system dynamically updates input parameters based on the feedback from each test round, completing coverage of the multidimensional signal space within a minimum number of test rounds and improving detection coverage. The AI ​​model can discern subtle differences between normal and abnormal responses, enhancing the ability to identify complex or early-stage faults. It is suitable for a variety of scenarios, including mass production, delivery testing, and maintenance inspections. The system also reduces reliance on additional sensors and manual setup processes, lowering maintenance costs and enhancing system versatility and ease of deployment. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0017] Figure 1 is a flow chart of the system execution of the present invention; Figure 2 This is the first pre-trained artificial intelligence model structure diagram; Figure 3 This is the structure diagram of the second pre-trained artificial intelligence model. DETAILED DESCRIPTION

[0018] Below, the invention is preferably described with reference to the accompanying drawings and specific embodiments.

[0019] The parameter indicator in an aircraft engine, also known as the Engine Parameter Indicator (EPI), is an avionics instrument used to monitor and display key engine operating parameters in real time. Its main function is to provide visual information on the engine's operating status to the pilot or maintenance system to ensure the safety and reliability of aircraft flight.

[0020] Aircraft engine parameter indicators generally include but are not limited to the following display modules: Temperature indicators (such as T4 or EGT indicators): Displaying the engine turbine inlet temperature or exhaust temperature is the core basis for judging combustion efficiency and the health status of high-temperature components.

[0021] Speed ​​indicator (N1, N2): They represent the rotation speed of the low-pressure compressor and the high-pressure compressor respectively, and the unit is generally a percentage (%RPM). They are important parameters that reflect the engine thrust output and mechanical rotation state.

[0022] Oil pressure, fuel flow, fan pressure ratio and other indicators: It can be used to monitor fuel supply system, lubrication system and compressor performance.

[0023] These indicators usually receive analog voltage or frequency signals from engine sensors (thermocouples, Hall sensors, pressure transmitters, etc.) and convert them into visual values ​​for display on the cockpit panel, integrated display screen or maintenance terminal.

[0024] Because aircraft engines are complex systems operating at high temperatures, high pressures, high speeds, and long lifespans, while the engines themselves have a high degree of engineering redundancy, inaccurate or ineffective indicators can lead to pilots being unable to correctly judge engine status, posing a potential flight safety hazard. Therefore, pre-delivery inspections, regular maintenance inspections, and post-replacement or repair inspections are essential.

[0025] In order to facilitate detection, the present invention discloses an aircraft engine parameter indication detection system, which is an intelligent system for automatically detecting and determining the status of parameter indicators in aircraft engines. The system can simulate sensor signals such as temperature, speed, voltage, frequency, etc. under various engine operating conditions without manually setting test parameters, and collect EPI response data in real time. The system uses an embedded artificial intelligence model to perform correlation analysis and dynamic adjustment on the simulation signal and response signal, thereby accurately judging whether the working status of the EPI under various input conditions is abnormal. The system is suitable for factory inspection, maintenance calibration and performance verification of core instrument equipment in aircraft engine status indication systems. Its workflow is as follows: Figure 1 shown.

[0026] The system specifically includes the following modules: The simulation signal construction module is used to generate a first sensor signal for an engine parameter indicator simulation test according to a first parameter.

[0027] The simulation signal construction module constructs various sensor signals, including but not limited to temperature, speed, voltage, and frequency, that simulate actual engine operating conditions, based on control parameters used to evaluate EPI performance. Because the physical signals generated by actual aircraft engines during operation are multi-source complex and difficult to fully simulate directly in a ground test environment, this module modeled and analyzed historical operating data to achieve controllable generation of various sensor signals, providing standardized, adjustable, and reproducible input signals for subsequent EPI response testing. This approach avoids the reliance on actual engines or costly sensor arrays, reducing testing costs and improving the adaptability and intelligence of the test system.

[0028] In an optional specific implementation, the analog signal construction module specifically includes the following submodules: The signal parameter input module is used to receive external input or the first parameter generated by the AI ​​model. The first parameter is a set of control quantities used to define the characteristics of the target simulation signal, including temperature setting value, target speed value, simulation time period, voltage amplitude, frequency range, etc.

[0029] In a further implementation, the module can be automatically updated through a touch interface, remote configuration instructions, or control parameters provided by a feedback analysis module within the system.

[0030] In an optional implementation, the first parameter may be a preset initialization parameter used to provide fixed, representative engine operating condition data when the system is first started or performs a standard functional self-test. The preset initialization parameter may be stored in the system's internal non-volatile storage medium and automatically loaded by the main control unit after power-on, ensuring that the system has basic testing and functional verification capabilities and can operate independently even before completing linkage with an external control system or AI feedback module.

[0031] In another optional implementation, the first parameter can be a random initialization parameter, which is used to perform a wide-coverage signal response test on the engine parameter indicator without relying on specific known operating conditions. The random initialization parameter is generated by a random number generator embedded in the system, and the generation range covers multiple signal dimensions such as temperature, voltage, frequency, duty cycle, etc., and its physical rationality and system safety can be controlled according to the set constraint boundaries. Using this method in the early stage of testing helps to collect the response characteristics of EPI under various unknown inputs, enrich the input and output data pairs of the AI ​​model, and thus improve the generalization ability of subsequent state recognition and parameter optimization.

[0032] The temperature signal generation module generates a simulated thermocouple signal or millivolt voltage signal based on the set temperature value. This module integrates a digital-to-analog converter (such as the DAC8565) to generate a corresponding voltage output based on the set input value. The output range is linearly adjustable from 0 to 50 millivolts, simulating the T4 or EGT indication of the engine under different operating conditions. Before output, the simulated temperature signal is scaled by a precision resistor network and filtered by an RC filter to remove high-frequency noise.

[0033] The speed signal generation module generates a simulated speed signal based on the target speed value, typically in the form of a pulse frequency or PWM signal. This module is based on the TIMER peripheral of a microcontroller (such as the STM32). It configures the timer's operating frequency and duty cycle and outputs a PWM waveform through a GPIO port. The frequency linearly corresponds to the N1 / N2 speed. To enhance analog signal stability, the PWM signal undergoes optocoupler isolation and shaping buffering before output, ensuring clear, interference-free waveform edges.

[0034] The frequency signal encoding module generates period-controllable, amplitude-stable sine or square wave signals to simulate the output of certain frequency-encoded pressure or flow sensors. This module is implemented using a digitally controlled oscillator (such as the AD9833). The output frequency and waveform type are configured via the SPI bus. The output signal is then buffered and amplified by an operational amplifier to ensure the signal drive capability meets the requirements of the EPI input interface.

[0035] The signal calibration and output control module performs amplitude compensation, filtering and shaping, and channel switching for various output signals. This module configures the signal path using multiple controllable analog switches (such as the ADG1608) and incorporates an op amp circuit for signal gain adjustment. Before each signal output, the system applies internal calibration coefficients, which are pre-determined through a calibration program and stored in EEPROM, to calibrate the output value. This module also integrates overvoltage protection to prevent interface damage caused by improper operation.

[0036] This analog signal construction module features support for multiple signal types, programmable control, and high-precision output. It can effectively replicate various engine sensor operating conditions and, through its modular design, supports flexible expansion. Compared to traditional systems with manual parameter settings and fixed signal sources, this module supports AI-powered automatic parameter input and feedback optimization, enabling a dynamic signal construction process without human intervention, significantly improving the intelligence of the test system and the credibility of test results.

[0037] In one specific example, the system receives a first set of parameters, including a target T4 temperature of 650°C, an N1 speed of 80%, and an N2 speed of 92%. Based on these parameters, the analog signal construction module generates a thermocouple voltage signal of approximately 22 millivolts and PWM signals with frequencies of 400 Hz and 600 Hz, respectively. These signals are filtered and shaped, and then output to the EPI for testing.

[0038] The analog signal output module is used to output the first sensor signal to the engine parameter indicator in the form of an analog signal to simulate the working state of the engine.

[0039] The analog signal output module outputs the first-class sensor signals generated by the analog signal construction module to the EPI in the form of physically recognizable analog voltages or frequencies, simulating the actual sensor signal input of an aircraft engine under different operating conditions. Because the EPI typically relies on analog interfaces such as temperature sensors and speed sensors for data perception, this module must ensure that the output signal's waveform characteristics, amplitude range, signal stability, and response time meet the EPI input specifications. By constructing a high-precision, low-distortion analog signal channel, it can effectively replace the signal source in a real engine environment, avoiding high-risk field testing and improving the safety, versatility, and automation level of the entire system.

[0040] In an optional specific implementation, the analog signal output module specifically includes the following submodules: The voltage output channel module outputs a DC or low-frequency analog voltage signal corresponding to analog temperature or pressure signals. This module receives the digital signal conversion result from a digital-to-analog converter (such as the DAC8565) and outputs a stable voltage value between 0 and 50 millivolts. Before output, the voltage signal passes through a low-pass filter (an RC network consisting of resistors and capacitors) to remove high-frequency spikes. It is then amplified and buffered by a precision operational amplifier (such as the OPA227) to enhance output drive capability. To prevent voltage reflections or impedance mismatches, a resistor matching network is added to the output of this channel to ensure signal quality meets avionics standards.

[0041] The frequency output channel module outputs a PWM pulse waveform or square wave frequency signal corresponding to the speed signal. This module configures the PWM waveform frequency and duty cycle based on the microcontroller's internal timer. The typical output frequency range is 100 Hz to 5 kHz, corresponding to the range of engine speed signals N1 and N2. The output PWM signal is electrically isolated from the control circuitry and peripheral interfaces via an optoelectronic isolator (such as the PC817). It is then shaped by a Schmitt trigger to ensure sharp signal edges. The shaped PWM signal is then amplified and output by a high-speed buffer (such as the 74HC244) to drive the EPI's digital input channels.

[0042] The signal channel selection module selects the appropriate signal channel based on the interface type of the EPI being tested. This module uses a multi-channel programmable analog switch (such as the ADG1608) that is switched via I / O control signals from the main control unit, enabling dynamic configuration of multiple outputs such as temperature, voltage, and frequency. This module supports simultaneous output of multiple signals to multiple input channels, and can also rotate outputs using time-division multiplexing, ensuring compatibility with different models and versions of EPI devices.

[0043] The overvoltage and short-circuit protection module is used to protect system hardware safety in the event of output anomalies. This module integrates a TVS transient voltage suppressor diode, a current-limiting resistor, and a fuse in the output channel. When the external EPI shorts or the feedback voltage is abnormal, it can quickly cut off the output path to prevent high-voltage backflow or damage to the output terminal. Preferably, this module also includes an output status monitoring subcircuit to collect the actual voltage and current of the output signal and feed it back to the main control unit for abnormality detection and alarm.

[0044] This analog signal output module offers diverse signal types, controllable output parameters, a stable electrical interface, and comprehensive protection mechanisms. Working in conjunction with the analog signal construction module, it accurately simulates multi-dimensional signals such as temperature, speed, and frequency, meeting the testing requirements of various EPI devices. Compared to traditional testing methods that rely on manual knob adjustment of signal sources, this module automatically generates and outputs a variety of signals, improving test automation and efficiency while reducing manual errors and configuration time.

[0045] In a specific example, the analog signal construction module has generated a 22 millivolt analog voltage signal corresponding to the target T4 temperature and a 400Hz PWM pulse signal corresponding to the target N1 speed. In the analog signal output module, the voltage output channel module outputs the voltage signal to the EPI temperature input terminal through the DAC and op amp circuit. The frequency output channel module is configured with a TIMER timer to output a 400Hz PWM wave with a 50% duty cycle. After isolation and shaping, it is output to the EPI's N1 channel input terminal. The channel selection module has set the correct signal path based on the EPI model. The entire output process is stable and noise-free, and an overvoltage protection mechanism is also included to ensure system safety.

[0046] The response acquisition module is used to acquire a first indication response signal generated by the engine parameter indicator after receiving the analog signal.

[0047] The role of the response acquisition module is to collect the output response signal generated by the EPI in real time after it receives the analog signal input, in order to evaluate its actual working status. Since the EPI is a key instrument of the aircraft engine, its response to the input signal directly reflects the health status of its internal perception, calculation and display subsystems. Therefore, collecting its indication response signal is the core link for the system to evaluate its accuracy, stability and reliability. The response acquisition module samples and records the correspondence between the analog input and the actual output of the EPI, providing the necessary data basis for subsequent state judgment, AI modeling, closed-loop adjustment and other modules. In order to ensure acquisition accuracy and system adaptability, the module design needs to be compatible with multiple signal types and amplitude ranges, have electrical protection capabilities such as signal shaping, filtering, isolation, and support multi-channel concurrent acquisition.

[0048] In an optional specific implementation, the response collection module specifically includes the following submodules: The signal access module collects the voltage or frequency signal from the EPI output. This module uses a high-impedance input circuit to prevent load interference on the EPI output. It also incorporates overvoltage clamping protection devices (such as bidirectional TVS diodes) to prevent damage to the acquisition module caused by abnormal voltage feedback from the EPI. Optional implementation options include direct access to the EPI output channels using aviation plugs or connection to an analog test fixture via BNC connectors.

[0049] The signal conditioning module filters, isolates, and adjusts the amplitude of the received raw signal. For voltage signals, a differential input is used to connect to an operational amplifier for primary buffering and amplification, and a low-pass filter is used to filter out high-frequency interference. For frequency signals, a Schmitt trigger is used to shape fuzzy pulses into clear TTL-level square waves to facilitate subsequent frequency counting. Preferably, this module uses an optoelectronic isolator to electrically isolate the EPI from the main system, improving the system's anti-interference capabilities.

[0050] The signal conversion module converts analog voltage signals into digital signals for subsequent processing. This module integrates an analog-to-digital converter (ADC), preferably using a high-precision chip with 12 bits or higher (such as the ADS1115). It supports multiple inputs, programmable gain control, and continuous sampling mode. For frequency signals, this module uses the microcontroller's internal timer to measure time intervals or count cycles, achieving high-precision frequency resolution and outputting digitized frequency data.

[0051] The sampling control and buffering module coordinates the sampling time, sampling frequency, and data storage management for each channel. This module uses the system master's timer interrupt mechanism to control the start and stop, read and write, and read / write of each ADC channel. It also uses a ring buffer to temporarily store sampled data to prevent sampling rate fluctuations from affecting system stability. Preferably, the sampling frequency is dynamically adjusted based on the rate of change of the analog input to avoid resource waste and improve response accuracy.

[0052] Designed with high compatibility, high precision, and high stability, this response acquisition module accurately captures real-time EPI output data under varying analog input conditions. By integrating multi-channel input, digital conversion, noise suppression, and system isolation, it ensures the reliability and security of sampled data, providing a solid data foundation for subsequent AI judgments and parameter iteration. Compared to traditional manual reading of instrument outputs, this module enables automated, multi-channel, real-time response data acquisition, significantly improving test efficiency and evaluation accuracy.

[0053] In one specific example, the analog signal output module outputs a 22 millivolt temperature analog signal and a 400 Hz PWM speed signal to the EPI. The response acquisition module collects the voltage display value and frequency feedback output by the EPI via an air plug connection. The voltage signal is amplified and low-pass filtered via differential input, then fed into the ADS1115 for conversion into 12-bit digital data. The frequency signal is shaped by a Schmitt trigger and fed into a timer for cycle counting. The data from both channels is cached in a ring storage area for real-time analysis and judgment by the AI ​​model. The entire acquisition process is stable and error-free, effectively supporting the operation of the state determination module.

[0054] The state determination and output module is used to call the first pre-trained artificial intelligence model, determine whether the working state of the engine parameter indicator is abnormal based on the first sensor signal and the first indication response signal, and output the corresponding detection result.

[0055] The function of the state judgment and output module is to determine whether there is an abnormality in the current working state of the EPI by calling the first pre-trained artificial intelligence model, combining the first type of sensor signal of the analog input and the first indication response signal generated by the EPI, and then output a clear detection result. Since the performance of the EPI may deviate due to factors such as aging, error drift, and failure of electronic components, it is difficult to cover all types of abnormalities and nonlinear behavior characteristics by relying solely on manually set thresholds or linear comparison methods. Therefore, this module establishes a mapping relationship between input and output based on the artificial intelligence model, which can adapt to a variety of complex, nonlinear, and fuzzy data features, effectively identify a variety of fault conditions including small deviations, sudden responses, distortion mismatches, etc., and output the judgment results in the form of structured data for users to read or for subsequent system processing.

[0056] In an optional specific implementation, the status determination and output module specifically includes the following submodules: A data preprocessing module is used to normalize, filter outliers, and perform feature encoding on the collected first-class sensor signals and first indicator response signals. This module preferably uses a minimum-maximum normalization algorithm to map all signal features to the interval [0, 1]. The conversion function is set to x_norm = (x - x_min) / (x_max - x_min), where x represents the original signal value, and x_min and x_max are the minimum and maximum values ​​of the corresponding feature in the historical training data. Signals with sampling noise or transient spikes are smoothed using a sliding average or median filter to enhance the stability of the input data.

[0057] The model inference module is used to call the first pre-trained artificial intelligence model to perform status judgment on the processed data. Optional implementations of this module include: an embedded inference engine based on a fully connected neural network (FCNN), a convolutional neural network (CNN), or a lightweight graph neural network (GNN-Lite). Preferably, the model structure includes an input layer, two hidden layers, and an output layer. The activation function is ReLU, and the output layer uses a Softmax function to output three types of state label probability values, corresponding to the three judgment results of "normal," "slightly deviated," and "severely abnormal." This module has a built-in model parameter memory for loading pre-trained network weights before the system runs, and performs efficient calculations through a dedicated inference engine (such as CMSIS-NN).

[0058] The judgment output module generates structured judgment information based on the classification probability output by the model and displays or outputs it to a communication interface. This module establishes a threshold judgment mechanism. For example, when the probability of a "serious anomaly" in the Softmax output is greater than 0.8, the output is marked as a "red warning" state. If the probability of a "minor deviation" is between 0.5 and 0.8, the output is a "yellow warning" state. All other conditions are marked as "green normal." Judgment results are displayed on the touchscreen in a JSON structure or visual text format and are simultaneously written to the system log cache, which can be exported or read remotely.

[0059] This state determination and output module, based on a deep learning model that integrates and judges multi-source inputs, boasts strong nonlinear modeling capabilities, a wide error tolerance range, and rich state interpretation. Compared to traditional threshold-based comparison methods, this module can identify potential fault trends, atypical response patterns, and complex fault scenarios, enhancing the system's robustness and intelligence in actual testing, significantly improving the accuracy of fault determination and early warning capabilities.

[0060] Aiming at the problems of heterogeneity, multidimensionality, time series coupling and abnormal sample sparsity in aircraft EPI response data, such as Figure 2 As shown, in a preferred implementation, the structure of the first pre-trained artificial intelligence model includes a residual graph attention fusion neural network, the principle of which is to construct the analog input signal and the response signal into a time series graph structure, and to perform weighted modeling of the heterogeneous correlations between different types of sensor signals by introducing a graph attention mechanism, and to alleviate the gradient vanishing and feature degradation problems of the deep network through the residual connection structure, thereby achieving high-precision judgment of the abnormal state of EPI under complex input conditions.

[0061] The overall structure of the model consists of the following five main hierarchical modules: Input encoding layer, graph construction and graph attention embedding layer, temporal feature fusion layer, residual enhanced multi-layer perceptron layer, and state judgment output layer.

[0062] The input coding layer is used to normalize and encode the analog signal collected by the system and the EPI response signal to form a basic feature vector.

[0063] Assuming that the input signal collected by the system is X_in and the response signal is X_out, the overall feature is vector X, which satisfies: X=[X_in,X_out] Where X_in represents the numerical sequence of input channel signals such as temperature, voltage, and frequency, and X_out represents the corresponding response channel signal.

[0064] The graph construction and graph attention embedding layer is used to treat each dimension of the signal in the vector X as a node in the graph. The edges between nodes are established based on the correlation coefficients between channels in the historical data. The embedded features of each node i are updated by weighted summation of the features of its adjacent nodes.

[0065] The new feature of node i is expressed as: h_i' = activation function (weighted sum (weight α_ij × adjacent node feature h_j)) in: h_i' represents the new feature of node i; h_i represents the original feature of the i-th node; α_ij represents the attention weight of node i to node j; The activation function is usually ReLU; The weight α_ij is calculated by the learnable function a(h_i,h_j) and then normalized.

[0066] The temporal feature fusion layer is used to process continuous signal inputs across multiple time steps and extract temporal correlations. Preferably, this layer has a parallel structure, comprising a one-dimensional convolutional module and a bidirectional gated recurrent unit (Bi-GRU) module.

[0067] The one-dimensional convolution module is used to capture local fluctuations; The Bi-GRU module is used to model long-term dependency change trends.

[0068] The final output is the multi-dimensional feature vector fused at different time steps.

[0069] Residual Enhanced Multilayer Perceptron Layer Used for nonlinear feature transformation and to prevent gradient disappearance.

[0070] The output of each layer is: Current layer output = activation function (weight matrix × previous layer output + bias) + previous layer output in: The weight matrix is ​​a learnable parameter; Bias is a learnable bias term; “+ previous layer output” is the residual connection, which enables the model to learn the sum of original features and high-order features.

[0071] The state judgment output layer is used to map the high-dimensional features after residual enhancement into classification results.

[0072] Output probability vector P = Softmax (output weight matrix × last layer feature vector + bias) in: P is a vector of length 3, representing the predicted probabilities of “normal”, “slightly deviated” and “severely abnormal” states respectively; The label corresponding to the maximum probability is the state judgment result output by the system.

[0073] In this implementation, the graph neural network can model the complex non-Euclidean structural relationship between input signals, which is suitable for the nonlinear coupling characteristics between signals such as temperature, speed, and voltage; the attention mechanism can automatically determine the signal channel that most affects abnormality judgment in the current state; the time series fusion layer takes into account local changes and global trends, and can detect sudden failures and gradual abnormalities; the residual connection structure ensures the stability of deep network training and is suitable for large-scale multi-channel sensor data modeling; the multi-layer perception structure enhances nonlinear expression capabilities and realizes direct mapping from data to state labels.

[0074] This model not only simultaneously models the nonlinear relationship between multi-dimensional, heterogeneous sensor inputs and response outputs, but also automatically focuses on the most critical signal feature combinations for decision-making under different operating conditions, thereby improving the model's generalization and interpretability. By introducing graph-structured modeling, the model can capture the topological dependencies between the internal response characteristics of the EPI and multiple types of input signals, such as complex patterns such as sudden changes in the frequency response curve and drift in the response amplitude. Furthermore, residual connections allow the model to retain the original input features within its deep structure, effectively improving the efficiency and robustness of model training.

[0075] The training process involves the following steps: First, a large amount of EPI response data covering various input conditions of temperature, speed, voltage, and frequency is collected to construct a time-synchronized sample sequence of input signals and output responses. Each sample is then structured as a graph containing multiple feature nodes, such as T4 voltage, N1 frequency, N2 frequency, and response voltage. Edge weights are set based on historical statistical correlations. Second, a graph attention layer (GATLayer) is used to aggregate and weight the features of each node in the graph to generate a high-dimensional state embedding vector. This is then output through a fully connected layer with multiple layers of residual connections for state classification. During training, a cross-entropy loss function is used for supervised learning of three labels: "normal," "slightly deviated," and "severely abnormal." Iterative parameter updates are performed using the Adam optimizer. To prevent overfitting, data augmentation mechanisms are introduced into the training set, including simulated signal perturbations, response signal drift, and interpolation of randomly missing features, to improve model robustness and fault tolerance.

[0076] During the testing phase, the model can be deployed in an embedded AI acceleration chip or a lightweight inference framework to achieve real-time, high-precision determination of the health status of engine parameter indicators.

[0077] In a specific example: An engine maintenance base conducted an in-situ test on an EPI. The test items included the temperature channel (T4) and two speed channels (N1 and N2). The system output the following signals through the analog signal construction module: Temperature input analog signal: 22.00 mV, corresponding to 650 degrees Celsius; N1 speed analog signal: 400 Hz, corresponding to 80% speed; N2 speed analog signal: 600 Hz, corresponding to 92% speed.

[0078] The response signal of EPI is: Temperature display feedback voltage: 20.5 mV; N1 display frequency: 395 Hz; N2 display frequency: 598 Hz.

[0079] Combine the above simulation signal and response signal into a feature vector: X_in=[22.00,400,600] X_out=[20.5,395,598] Eigenvector X = [22.00, 400, 600, 20.5, 395, 598] Normalize it and assume that the historical statistical range is as follows: Temperature signal (millivolt): minimum value 0, maximum value 50 Speed ​​signal (Hz): minimum value 0, maximum value 1000 Then the normalized eigenvector X_norm is: X_norm=[22 / 50,400 / 1000,600 / 1000,20.5 / 50,395 / 1000,598 / 1000] X_norm≈[0.440,0.400,0.600,0.410,0.395,0.598] Create a graph G where each node corresponds to a feature channel: Node v1: Temperature Input Node v2: N1 input Node v3: N2 input Node v4: Temperature Response Node v5: N1 responds Node v6: N2 response The Pearson correlation coefficient (or mutual information) between each node is calculated through historical samples as the edge weight: For example: the weight of temperature input and temperature response is 0.92, and the weight of speed input and response is 0.87 and 0.90 respectively; The adjacent edges in the figure are: e(v1,v4)=0.92 e(v2,v5)=0.87 e(v3,v6)=0.90 Other cross-signal channel edges are set to weak correlation, such as e(v1,v5)=0.15.

[0080] The embedded features of each node are calculated through the graph attention mechanism. For example: The new representation of node v4 (temperature response) is: h_v4 = activation function (0.92 × h_v1 + 0.08 × h_v5 + 0.05 × h_v6) in: h_vi represents the input feature of node vi; The weights are calculated by the attention function and normalized.

[0081] The attention function uses the following form: Original weight calculation: e_ij = W_a × activation function (W × h_i || W × h_j) Normalized weight α_ij=exp(e_ij) / Σ k exp(e_ik) Among them, W represents a learnable linear transformation matrix, which is used to map the node input features from the original dimension to the dimension of the attention space; W_a is a learnable attention weight vector, which acts on the concatenated features to output a scalar; h_i and h_j represent the input feature vectors of node i and node j in the graph respectively; exp represents the exponential function.

[0082] If the input samples are time series (e.g., t=1, 2, 3, 4, 5), a corresponding graph structure is constructed at each moment and fused into a time series feature matrix. A parallel 1D convolution and Bi-GRU are used to extract local changes and long-term trends, and then a residual multi-layer perception network is used for feature transformation.

[0083] For example, at the current time step the model receives the input feature vector H: H=[0.440,0.400,0.600,0.410,0.395,0.598] The output of the first layer residual perceptron is: H1 = activation function (W1×H+b1) + H Among them, W1 is the weight matrix, which is a trainable parameter, and b1 is the bias term.

[0084] The second layer residual output is: H2 = activation function (W2×H1+b2) + H1 Among them, W2 is the weight matrix, which is a trainable parameter, and b2 is the bias term. The Softmax function is used to output the state probability vector: P=Softmax(W_out×H2+b_out) Among them, W_out is the weight matrix of the output layer; b_out is the bias term of the output layer.

[0085] Assume the output is: P=[0.07,0.16,0.77] The corresponding meanings are: The probability of normal state is 7% The probability of a slight deviation is 16%. The probability of severe abnormality is 77% Therefore, the system determines that under the current input and output combination, the EPI working status is "severe abnormality".

[0086] The feedback analysis module is used to call a second pre-trained artificial intelligence model and generate a second parameter by combining the first sensor signal and the first indication response signal.

[0087] The feedback analysis module automatically calls a second pre-trained artificial intelligence model based on the current test data combination after each analog signal output and response signal acquisition. The model comprehensively analyzes the deviation characteristics between the first sensor signal and the first indicator response signal, thereby outputting the control parameters (i.e., the second parameters) used to update the next round of analog signal generation. Because EPI devices may exhibit dynamic nonlinearity, hysteresis, or multiple input coupling characteristics, statically set test signals alone cannot approximate their most sensitive test range. Therefore, this module learns the historical mapping relationship between analog signals and response signals and gradually adjusts the input parameters, forming a closed-loop adaptive iteration of the system testing process, automatically approaching the most discriminative input region to improve detection efficiency and accuracy.

[0088] In a preferred implementation, in view of the multi-input-multi-response-nonlinear coupling characteristics of the first sensor signal and the first indication response signal, this implementation further discloses a deep neural structure with input channel selection capability, error response nonlinear adjustment capability and historical feedback memory capability, so as to realize automatic parameter adjustment of signal input without a clear optimization objective function.

[0089] Specifically, if Figure 3 As shown, the second network model includes the following structure: The input encoding layer is used to fuse the input signal of the current round with the response signal into a learnable vector.

[0090] Input: First sensor signal vector , the first indication response signal vector in, Represents n-dimensional real number space.

[0091] Output: fused vector ,in (Splicing) Using standard normalization method and Mapping to the [0,1] interval eliminates the physical dimension differences between different channels and facilitates subsequent unified modeling.

[0092] The residual compression mapping layer is used to compress high-dimensional redundant features, enhance nonlinear expression capabilities, and preserve the original information structure.

[0093] This layer contains two fully connected sublayers: First sublayer: compression layer enter: Output: , the dimension space is , Represents n-dimensional real number space.

[0094] in is the weight matrix, which is a trainable quantity. For paranoid items.

[0095] Second sublayer: recovery layer enter: , which is the output of the previous layer .

[0096] Output: , the dimension space is in is the weight matrix, which is a trainable quantity. Is the bias term. Residual connection: The residual compression mapping layer introduces a bottleneck structure of "compression-activation-recovery", which improves the model's ability to fit complex nonlinear relationships without significantly increasing the number of parameters. At the same time, the residual skip connection structure is used to retain the initial signal in each layer to prevent gradient disappearance.

[0097] The channel self-modulation gating layer is used to determine whether each signal should be amplified, weakened, or kept unchanged based on the current characteristics on a channel-by-channel basis, and to determine its weight in the next round of parameter adjustment.

[0098] This layer consists of two parallel subnetworks: Positive regulatory gating network Negative regulatory gating network For each channel i∈[1,2n], calculate: in, 、 , represents the positive / negative regulation factor of the i-th channel, ranging from (0,1), indicating the degree to which it should be amplified or suppressed; Sigmoid is the activation function; 、 are the weight vectors of the positive regulation gating network and the negative regulation gating network, respectively, which are learnable parameters; Represents the eigenvalue of the i-th channel in the output vector of the first layer residual perceptron; 、 are the bias terms of the positive and negative gating networks, respectively.

[0099] The final channel modulation factor is: in, It represents the final modulation factor of channel i, with a value range of [-1,1], indicating the intensity of the enhancement (positive) or weakening (negative) direction.

[0100] Apply the modulation factor to the residual representation : in represents the tuning parameter vector, [:n] indicates the corresponding input signal part, is the modulation factor, and ⊙ represents channel-by-channel multiplication.

[0101] The dual-gating structure within the channel's self-modulating gating layer avoids the problem of a single gating layer failing to express "neutral inhibition" or "directional amplification." Positive gating corresponds to enhanced regulation, while negative gating corresponds to regulatory inhibition. The difference between the two can represent a continuous range of channel response dynamics.

[0102] The historical memory fusion layer is used to introduce historical information from multiple rounds of parameter adjustments, giving the model "memory" capabilities and avoiding oscillations or excessive adjustments.

[0103] Cache structure: records the past k rounds of input , model output ΔP, state judgment result S.

[0104] Network structure: Using a single-layer gated recurrent unit GRU, the input is a sequence vector { ^t,S^t}, the output is the historical fusion vector .

[0105] Correct the current pre-tuning result to: in represents the parameter vector, γ∈[0,1] is the historical fusion weight, and preferably γ=0.7.

[0106] This mechanism can prevent the model from repeatedly adjusting at the edge of the state limit and causing oscillations, making parameter changes more continuous and controllable.

[0107] The parameter generation and boundary constraint layer is used to generate the second parameter and ensure security and legality.

[0108] The generation method is: Wherein λ is the adjustment rate, preferably 0.1≤λ≤0.5.

[0109] Then perform physical constraint clipping: like > Upper limit ,but =Upper limit , that is, when Greater than the upper limit When .like <Lower limit ,but =Lower limit , that is, when Less than the lower limit When The constraints are preset based on the engine model, electrical safety standards, and interface specifications. For example, the temperature voltage signal should not exceed 50 millivolts, and the speed frequency should be controlled between 100Hz and 1000Hz.

[0110] The preferred second model uses a dual-driven signal conditioning mechanism, "current error performance" and "historical feedback trends." It internally enhances nonlinear feature expression through a residual structure and automatically identifies the channel direction to be amplified or suppressed through a dual-gating mechanism, avoiding the limitations of manual parameter adjustment strategies. The GRU memory path is also introduced to enable sequential optimization capabilities, preventing oscillation or invalid polling behavior, thereby achieving a "few-round, high-efficiency, and adaptive" signal parameter iteration mechanism.

[0111] This model does not rely on minimizing a standard loss function, but instead focuses on structural adjustments driven by state-response differences. It offers advantages such as strong interpretability, clear adjustment direction, support for online updates, and high historical sensitivity. It is particularly well-suited for complex input parameter adjustment problems under conditions of unstructured feedback, multi-source error responses, and physical boundary constraints. Compared to conventional regression or reinforcement models, this model offers stronger closed-loop parameter adjustment capabilities, faster convergence, and more diagnostically targeted final input parameters.

[0112] During training, given the current round input parameters and corresponding response , output an adjustment vector ΔP so that the next round of input = +ΔP is more likely to trigger anomalies or obtain more significant status feedback.

[0113] Therefore, its training samples do not come directly from the mapping of "input→label", but need to construct sample data based on feedback effect. The training sample triple is represented as: Sample =( , ,ΔP_target).

[0114] in: is the current input parameter vector, dimension is n; is the corresponding response value, with dimension n; ΔP_target is the target adjustment parameter designed manually or derived by the algorithm.

[0115] To obtain ΔP_target, the following generation method is used: During the historical testing process, select an input sequence { ^t, ^t,S^t}, where S^t is the output of the state determination module (such as normal, slightly abnormal, and seriously abnormal labels), and t represents the dimension; For samples that fail to enter the "abnormal" state for multiple consecutive rounds, the target parameter adjustment direction is constructed to increase or decrease a certain amount towards the channel with the maximum error; The formula for determining the value of ΔP_target is: If | - |>ε and S^t≠"abnormal", then = ×δ Where sign represents the sign function, δ is the unit adjustment amplitude (such as 1 mV, 5 Hz), and ε is the response error tolerance threshold.

[0116] In order to balance the regulation direction, amplitude constraint and system stability, SM-RDGNet uses a weighted combination loss function: Total loss L_total = L_dir + α × L_mag + β × L_smooth in: L_dir represents the directional loss used to measure the directional consistency between the model output ΔP_pred and the target adjustment ΔP_target: L_dir=Σ_i[1-cos_sim( , )] Among them, cos_sim represents cosine similarity, and the closer it is to 1, the more consistent the direction.

[0117] L_mag represents the amplitude constraint loss used to punish situations where the adjustment amplitude is too large or too small: L_mag=Σ_i[(| |-|Δ |)²] Ensure that the model output adjustment amount is controlled within the effective range.

[0118] L_smooth represents a smooth regularization term, which is used to limit the jump between the current forecast result and the historical trend: L_smooth=Σ_i[( ^t- ] This prevents the model from producing large, unstable fluctuations in the time series.

[0119] The coefficients α and β are adjustment items, preferably α=0.5 and β=0.1.

[0120] Furthermore, after model training is completed, all parameters are quantized to floating point 16-bit or integer 8-bit format.

[0121] In a specific example: The analog signal channels currently being tested by the EPI of a certain aircraft model include: Temperature voltage (T4_input), in millivolts, range [0,50]; Speed ​​frequency (N1_input, N2_input), unit is Hz, range is [0,1000]; This round of testing is the third round, and the parameters and status of two rounds have been recorded historically.

[0122] The input parameters of this round of system simulation (the first parameter )as follows: T4_input = 22.00 mV N1_input=400Hz N2_input=600Hz The corresponding EPI response signal (first indication response )for: T4_response = 20.50 mV N1_response=395Hz N2_response=598Hz After normalization (maximum values ​​are 50, 1000, and 1000 respectively), we get: =[0.4400,0.4000,0.6000] =[0.4100,0.3950,0.5980] Concatenate and construct the model input feature vector: =[0.4400,0.4000,0.6000,0.4100,0.3950,0.5980] After entering the residual compression mapping layer: =ReLU( × + )→Dimension is compressed to 3 (hypothesis); = × + →Dimension restored to 6; = + →Get the activation feature vector ; Calculate positive gating channel by channel With negative gating , assuming the model is given as follows: =[0.70,0.60,0.80,0.30,0.20,0.10] =[0.10,0.15,0.05,0.05,0.02,0.01] The adjustment factor vector β is: β= -α⁻=[0.60,0.45,0.75,0.25,0.18,0.09] Take the first three channels as the target channels of ΔP_prelim, corresponding to the current input channels: ΔP_prelim=β[:3]⊙ =[0.60×0.4400,0.45×0.4000,0.75×0.6000] =[0.2640,0.1800,0.4500] The system has cached the past two rounds of ΔP respectively: After GRU calculation, the historical memory fusion vector H2 is obtained: =[0.1900,0.0950,0.3600] The fusion weight γ=0.7, then the final ΔP is: ΔP=0.7×ΔP_prelim+0.3× =[0.7×0.2640+0.3×0.1900,0.7×0.1800+0.3×0.0950,0.7×0.4500+0.3×0.3600] =[0.2438,0.1535,0.4230] Current normalized input =[0.4400,0.4000,0.6000] Set λ=0.3 and calculate: ΔP_scaled=λ×ΔP=[0.0731,0.0460,0.1269] = +ΔP_scaled=[0.5131,0.4460,0.7269] Restore physical units: T4_input = 0.5131 × 50 = 25.655 mV N1_input = 0.4460 × 1000 = 446 Hz N2_input = 0.7269 × 1000 = 726.9 Hz Check the bounds and set the upper limit to: T4 ≤ 50 mV, N1 / N2 ≤ 1000 Hz → Legal, no need for cropping.

[0123] The second parameter of this round of output is: [25.655 mV, 446 Hz, 726.9 Hz].

[0124] The result is recorded as the "4th round input parameter" and enters the next round of simulation signal construction; A loop control module is used to determine whether to continue the test process based on the second parameter. If it is determined to continue the test, the first parameter is replaced by the second parameter, and the analog signal construction module, the analog signal output module, the response acquisition module, the state judgment and output module and the feedback analysis module are called in sequence.

[0125] The function of the loop control module is to determine whether the current test has reached a predetermined termination condition, such as state stability, error convergence or maximum test round limit, based on the second parameter output by the feedback analysis module. If the termination condition is not met, the system will replace the first parameter with the second parameter and re-execute the test process. This module is the decision-making center in the closed-loop adaptive detection process of the present invention, ensuring that the system automatically decides whether to continue testing based on the response results of each round without relying on human intervention. Through the dynamic decision-making mechanism, the system can obtain the most effective diagnostic information within the minimum number of rounds, avoid repeated testing and invalid cycles, and improve test efficiency and resource utilization.

[0126] In an optional specific implementation, the loop control module specifically includes the following logical structure: The condition judgment unit is used to judge whether the stop condition is met based on the state label provided by the state judgment and output module and the error change and parameter adjustment amplitude output by the feedback analysis module. Preferably, the stop condition includes one of the following: The current state is "normal", and the error change for two consecutive rounds is less than the set threshold ε; The current status is "severe abnormality", and the change in the parameter adjustment direction for two consecutive rounds is less than the set angle θ; The number of test rounds currently executed exceeds the set upper limit N_max.

[0127] Among them, ε represents the error change convergence threshold, for example, the normalized error is less than 0.01; θ represents the threshold for parameter adjustment direction change, for example, the cosine angle is less than 15 degrees; N_max indicates the maximum number of test rounds, for example, 10 rounds.

[0128] The parameter updating unit is used to update the second parameter when it is judged to be "continue testing" Replace the current first parameter , and sent to the analog signal construction module to start a new round of signal construction and output.

[0129] The module call scheduling unit is used to call the analog signal construction module, analog signal output module, response acquisition module, state judgment and output module and feedback analysis module in sequence according to the set process to form a whole new round of detection process.

[0130] The status recording and termination unit is used to record the termination status, current parameters, and number of test rounds when the test is determined to be "stopped", and output the final results to the user interface or control system. Optional implementation solutions include: displaying prompt information, generating a test report, or issuing remote commands.

[0131] This loop control module, through the use of clear termination conditions and a dynamic feedback mechanism, enables the system to achieve "self-driving and self-convergence." Compared to traditional testing methods that rely on fixed rounds, this module intelligently controls the test iteration process, ensuring detection adequacy while preventing redundant testing. This improves the system's overall response efficiency and energy utilization, while also reducing the risk of equipment loss during testing.

[0132] In a specific example, based on the input parameters of the fourth round =[25.655 mV, 446 Hz, 726.9 Hz]. After the system completes this round of signal output and response acquisition, the state determination module outputs "slight deviation." The feedback analysis module calculates the error change as [0.002, 0.005, 0.001], which is less than the preset threshold ε = 0.01. The angle with the previous round of parameter adjustment is 12 degrees, less than θ = 15 degrees. This indicates that convergence has not yet met the "serious anomaly" threshold, but there is still room for further optimization. Based on this, the loop control module determines "continue testing," uses the second parameter set as the first parameter for the next round, and re-invokes the simulation signal construction module to begin the fifth round of testing. The system also updates the round count and writes all records from the fourth round into the data cache for subsequent backtracking.

[0133] In another embodiment, the present invention further provides an aircraft engine parameter indication detection method, which uses the aforementioned embodiment to perform aircraft engine parameter indication detection.

[0134] It should be noted that the explanations of the aforementioned aircraft engine parameter indication detection system embodiment are also applicable to the method of the embodiment of the present application and will not be repeated here.

[0135] Those skilled in the art will appreciate that the various units and algorithm steps described in the embodiments disclosed herein can be implemented using a combination of electronic hardware, computer software, and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0136] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0137] In the several embodiments provided in this application, if any function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of this application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory; hereinafter referred to as: ROM), random access memory (Random Access Memory; hereinafter referred to as: RAM), magnetic disk or optical disk, and other media that can store program code.

[0138] The above is only a specific embodiment of the present application. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in this application, which should be included in the scope of protection of this application. For some module structures that are not particularly clear in the present invention, the content recorded in the prior art shall prevail. The prior art mentioned in the above background technology section and the specific embodiment section of the present invention can be regarded as part of the present invention and is used to understand the meaning of some technical features or parameters.

Claims

1. An aircraft engine parameter indication and detection system, characterized in that: The system includes the following modules: a simulation signal construction module, configured to generate a first sensor signal for an engine parameter indicator simulation test according to a first parameter; an analog signal output module, configured to output the first sensor signal to the engine parameter indicator in the form of an analog signal to simulate the working state of the engine; a response acquisition module, configured to acquire a first indication response signal generated by the engine parameter indicator after receiving the analog signal; a state determination and output module, configured to call a first pre-trained artificial intelligence model, determine whether the working state of the engine parameter indicator is abnormal based on the first sensor signal and the first indication response signal, and output a corresponding detection result; a feedback analysis module, configured to call a second pre-trained artificial intelligence model and generate a second parameter by combining the first sensor signal and the first indication response signal; A loop control module is used to determine whether to continue the test process based on the second parameter. If it is determined to continue the test, the first parameter is replaced by the second parameter, and the analog signal construction module, the analog signal output module, the response acquisition module, the state judgment and output module and the feedback analysis module are called in sequence.

2. The aircraft engine parameter indication and detection system according to claim 1, characterized in that: The first parameter is a preset initialization parameter used to start the first round of testing.

3. The aircraft engine parameter indication and detection system according to claim 1, characterized in that: The first parameter is a random initialization parameter generated within a set range.

4. The aircraft engine parameter indication and detection system according to claim 1, characterized in that: The analog signal construction module includes: A signal configuration submodule, configured to set the voltage value, current value or frequency parameter of the analog signal of each sensor based on the first parameter; A signal conversion submodule is used to convert the configured parameters into an electrical signal format that complies with the target engine parameter indicator interface protocol; The signal stabilization submodule is used to perform amplitude filtering and timing compensation on the generated analog signal to ensure stable signal output.

5. The aircraft engine parameter indication and detection system according to claim 1, characterized in that: The analog signal output module includes: A drive interface submodule, configured to convert the first sensor signal into a hardware signal that can be driven and output by an analog port; The buffer adjustment submodule is used to perform amplitude compensation, anti-interference processing and impedance matching on the signal according to the indicator type; The signal sending submodule is used to send the analog signal to the corresponding port of the target engine parameter indicator according to the preset timing.

6. The aircraft engine parameter indication and detection system according to claim 1, characterized in that: The response collection module includes: Multi-channel acquisition submodule, used to synchronously acquire analog signal outputs corresponding to multiple sensor response channels; The analog-to-digital conversion submodule is used to convert the collected analog response signal into digital data; The buffer storage submodule is used to cache the digital response data into the internal memory for subsequent analysis.

7. The aircraft engine parameter indication and detection system according to claim 1, characterized in that: The first pre-trained artificial intelligence model structure includes: an input feature embedding layer, configured to receive the first sensor signal and the first indication response signal, and map them to a feature space of uniform dimension to form a joint input tensor; A graph construction and graph attention layer is used to construct a feature association graph based on the joint input tensor, use a graph attention mechanism to calculate the attention weight between each feature node and its adjacent nodes, and extract local sensitivity information and the coupling relationship between parameters; The residual connection fusion layer is used to perform a residual connection between the output of the graph attention layer and the output of the input feature embedding layer. It improves the feature expression capability through feature splicing and element-by-element weighting operations to avoid gradient disappearance or feature degradation. The abnormal state discrimination layer is used to perform binary or multi-classification judgment based on the fused high-dimensional feature vector, and output the working state judgment result of the engine parameter indicator under the current input conditions.

8. The aircraft engine parameter indication and detection system according to claim 1, characterized in that: The second pre-trained artificial intelligence model structure includes: An input coding layer, configured to receive the first sensor signal and the first indication response signal, perform feature normalization and channel alignment on each of the first sensor signal and the first indication response signal, and then concatenate the two signals to form an input feature vector; The dual-gated factor generation layer is used to generate a set of weighted parameter adjustment factors based on the current input features, including amplitude factors and direction factors. The amplitude factors control the adjustment degree, and the direction factors control the adjustment direction of each channel. The historical memory fusion layer is used to call the test parameters and feedback status of the previous round, dynamically weighted fuse them with the current feature input, and extract the cross-round feature evolution trend; The parameter update generation layer is used to calculate the second parameter output of the current round based on the fused memory vector.

9. The aircraft engine parameter indication and detection system according to claim 1, characterized in that: The loop control module determines whether to continue executing the test process based on one of the following conditions: The judgment result of the first pre-trained artificial intelligence model is a non-abnormal state, and the error change for two consecutive rounds is lower than a preset threshold; The parameter adjustment direction change angle output by the feedback analysis module is less than a preset angle threshold; The number of test rounds executed exceeds the preset maximum round limit.

10. A method for detecting aircraft engine parameter indication, characterized in that: An aircraft engine parameter indication detection system according to any one of claims 1 to 9 is used to perform aircraft engine parameter indication detection.

Images