Flight phase self-adaptive cockpit speech enhancement system based on multi-parameter fusion

Abstract

The invention discloses a flight stage adaptive cockpit speech enhancement system based on multi-parameter fusion, and the system comprises a data collection module which collects key flight parameters in real time from an airborne data network of an aircraft; the flight stage recognition module is used for analyzing the input parameters of the data acquisition module, deducing the current flight stage in real time and judging whether a special situation is entered or not; outputting the stage code and the confidence coefficient to an adaptive speech enhancement module; and the adaptive speech enhancement module is used for selecting a proper noise reduction model to suppress the current noise according to the stage code and the confidence coefficient, and carrying out optimal scheduling through model hot switching and double-model mixed execution. According to the flight stage self-adaptive cockpit voice enhancement system based on multi-parameter fusion, the flight stage is automatically recognized by analyzing the flight parameters in real time, and a targeted noise reduction algorithm is dynamically called, so that the voice recording quality can be remarkably improved, and noise environments of different stages in the flight process can be well adapted.

Description

TECHNICAL FIELD

[0001] The present application relates to a cockpit voice system, in particular to a flight phase adaptive cockpit voice enhancement system based on multi-parameter fusion. BACKGROUND

[0002] There are several related technical solutions for cockpit voice enhancement or aviation voice noise reduction in the market and literature at present, but most of them have the following limitations: 1. Fixed parameter noise reduction system In traditional digital voice recorders (CVR) or enhancement modules, fixed band-pass filters, spectral subtraction algorithms, etc. are often used for basic processing of cockpit voice (such as some early Honeywell, L3 CVR series products): • Technical mechanism: set a fixed frequency bandwidth to filter out low-frequency engine noise and high-frequency electromagnetic interference; • Main defects: Cannot be adjusted in real time according to changes in flight state; No adaptive mechanism for sudden abnormalities (such as reverse thrust, tire burst, lightning strike); When the voice and noise spectrum overlap, it is easy to cause voice distortion or information loss.

[0003] 2. Deep voice enhancement technology of single model In recent years, some research and products have proposed voice enhancement algorithms based on neural networks (such as DNN, U-Net), such as the literature “Research on Speech Enhancement Based on Deep Neural Network”, and commercial products such as some embedded DSP modules (such as TI Jacinto) voice preprocessing IP. But the limitations of these methods are: • Use a single neural network model to process audio throughout the flight phase, which cannot specifically suppress phase-specific noise; • No flight state recognition mechanism, model selection and parameter configuration are static and non-adaptive; • Weak processing capability for extreme weather, mechanical failure, and impact non-stationary noise under sudden events.

[0004] 3. Insufficient noise label-driven training mechanism Some algorithms attempt to train deep models based on offline labeled noise types, but in actual flight, it is difficult to accurately perceive whether the current airborne environment belongs to a specific state such as “takeoff”, “taxi” or “SP1 engine failure”; Lack of state discrimination mechanism leads to insufficient algorithm generalization ability, making it difficult to achieve robustness improvement in real scenarios.

[0005] The present application aims to solve the following key technical problems in the prior art: 1. Lack of flight state recognition mechanism leads to mismatch of voice enhancement strategy The traditional cockpit voice recording system uses fixed parameters or static models for voice noise reduction processing, without considering that the noise types are significantly different in different stages of flight (such as taxiing, climbing, cruising, descending, etc.), thus leading to unsatisfactory noise reduction effect, decline in voice clarity, and even possible weakening of important voice segments.

[0006] 2. Unable to identify and respond to non-stationary noise caused by sudden special situations In sudden events such as engine failure, cabin pressure abnormalities, and extreme weather, strong impact broadband noise (such as shaft frequency resonance, lightning pulse sound, etc.) often accompanies. The existing system lacks real-time state recognition and adaptive algorithm switching mechanism, and cannot effectively suppress sudden noise without affecting voice recognition quality.

[0007] 3. Existing voice enhancement algorithm model has poor generalization ability Some existing deep learning voice enhancement methods rely on a unified model to process all flight stages or noise environments, lack of specificity, leading to phenomena such as excessive voice filtering or large residual noise in specific stages (such as takeoff thrust enhancement, landing reverse thrust), limiting the post-analysis value of CVR voice. SUMMARY

[0008] The technical problem to be solved by the present application is to provide a flight stage adaptive cockpit voice enhancement system based on multi-parameter fusion, which can significantly improve the quality of voice recording and adapt well to different noise environments in the flight process.

[0009] To solve the above technical problems, the present application provides a flight stage adaptive cockpit voice enhancement system based on multi-parameter fusion, which includes a data acquisition module that acquires key flight parameters in real time from the aircraft's onboard data network; a flight stage recognition module that analyzes the input parameters of the data acquisition module, infers the current flight stage in real time, and determines whether to enter a special situation; outputs the stage code and confidence to the adaptive voice enhancement module; the adaptive voice enhancement module selects the appropriate noise reduction model to suppress the current noise according to the stage code and confidence, and performs optimization scheduling through model hot switching and double model mixing.

[0010] Further, the data acquisition module acquires key flight parameters in real time from the onboard bus, with a sampling period of 50-500ms, the key flight parameters including kinetic parameters, spatial state parameters, engine state parameters and configuration state parameters, and a time reference parameter is obtained to synchronize the time axis of all parameters.

[0011] Further, the flight phase recognition module divides the flight phase into 9 regular phases and 3 special phases and outputs phase codes, and uses fuzzy logic to process the parameter crossing problem of transition phase and outputs confidence.

[0012] Further, the phase code setting and determination of the regular phase are as follows: the phase code F1 corresponds to ground taxi, the determination parameter threshold is airspeed < 40 knots, and engine N1 < 50%; the corresponding noise is ground broadcast and APU high-frequency screech; the phase code F2 corresponds to takeoff roll, the determination parameter threshold is airspeed 40-80 knots, and throttle lever > 90%; the corresponding noise is tire vibration and engine 3rd order harmonic; the phase code F3 corresponds to initial climb, the determination parameter threshold is height < 1000 ft, and climb rate > 800 ft / min; the corresponding noise is landing gear vortex shedding noise and flap slot screech; the phase code F4 corresponds to climb cruise, the determination parameter threshold is height 1000-cruise altitude, and acceleration is 0.05-0.2g; the corresponding noise is engine rotor noise and fuselage skin vibration; the phase code F5 corresponds to level cruise, the determination parameter threshold is height variation rate ± 50 ft / min, and Mach number is 0.78-0.85; the corresponding noise is broadband turbulent flow noise and air conditioning system airflow sound; the phase code F6 corresponds to initial descent, the determination parameter threshold is descent rate 500-1000 ft / min, and throttle lever is 30-60%; the corresponding noise is passenger cabin pressure regulating valve periodic opening and closing; the phase code F7 corresponds to approach preparation, the determination parameter threshold is height < 3000 ft, flap > 15°, and airspeed < 180 knots; the corresponding noise is configuration warning sound and slat turbulent flow; the phase code F8 corresponds to final approach, the determination parameter threshold is glide path capture, and radio altitude < 1000 ft; the corresponding noise is GPWS voice alarm and raindrop impact spectrum; the phase code F9 corresponds to landing roll, the determination parameter threshold is airspeed < 80 knots, and reverse thrust is activated; the corresponding noise is brake disc friction and reverse thrust turbulent flow.

[0013] Further, the phase code setting and determination of the special phase are as follows: the phase code SP1 corresponds to engine failure, the determination parameter threshold is engine speed difference > 15%; the corresponding noise is rotor unbalanced vibration; the phase code SP2 corresponds to passenger cabin emergency, the determination parameter threshold is cabin pressure altitude > 8000 ft; the corresponding noise is oxygen mask release mechanical sound; the phase code SP3 corresponds to extreme weather, the determination parameter threshold is vertical acceleration > 0.5g; the corresponding noise is full-band random impulse noise.

[0014] Further, the adaptive speech enhancement module receives phase code and confidence from the flight phase recognition module; if it is judged that the phase is switched or in a special phase, and the confidence exceeds the preset threshold, the dispatching strategy table retrieval is triggered, the default enhancement strategy combination or alternative strategy suitable for the current flight phase is adopted to perform fundamental frequency / resonant frequency protection processing on the human voice part in the input speech frame, and the enhanced speech frame is output to the CVR recording buffer, and the current signal-to-noise ratio improvement index is recorded.

[0015] Further, in the take-off running stage, when the flight phase recognition module identifies that the air speed is between 40-80 knots, the throttle lever position exceeds 90%, and the 3rd harmonic of 180Hz is detected in the engine noise, it is determined that the current is the take-off running stage, and the phase code F2 is output; the adaptive speech enhancement module calls the following noise reduction strategy: the main strategy adopts a lightweight U-Net model to infer the separation of noise and speech components, and focuses on suppressing tire vibration and engine harmonic; the alternative strategy is a multi-channel spectral subtraction, which is used to process sudden high-frequency splash sound in a wet runway scene.

[0016] Further, in the extreme lightning weather, when the flight phase recognition module identifies that the vertical acceleration exceeds the threshold, the radar echo intensity is out of limit, the 0.1-10kHz full-band random pulse noise and electromagnetic pulse noise appear, and the cabin pressure height is stable, it is determined that the current is the extreme weather, and the phase code SP3 is output; the adaptive speech enhancement module calls the following noise reduction strategy: the main strategy adopts an IRS-Mapper network to dynamically suppress the pulse component, and specifically eliminates the transient pulse noise caused by lightning; the auxiliary strategy is high-frequency clipping combined with zero-crossing limiting, which protects the speech frequency band of 300-3400Hz.

[0017] Further, in the cruise phase transition to the initial descent phase, when the flight phase recognition module identifies that the descent rate exceeds 500ft / min for 3 seconds, the throttle lever position decreases to 45%, and the 1Hz periodic pressure regulating sound is detected, it is determined that the current is the transition state from phase code F5 to phase code F6; the adaptive speech enhancement module calls the following noise reduction strategy: the transition strategy is to run the Gated-Conv-DNN model left by phase code F5 and the spectral subtraction model preloaded by phase code F6 in parallel, and the Gated-Conv-DNN weight gradually changes from 70% to 0% and the spectral subtraction weight gradually increases from 30% to 100% during the transition process; the stable strategy is to switch to LSTM plus filtering fusion, and optimize the periodic noise suppression.

[0018] Furthermore, when the flight phase identification module detects that the engine speed difference exceeds the threshold and the vibration sensor detects shaft frequency harmonics of 120Hz and its harmonics, it determines that the current situation is an engine fault and outputs phase code SP2; the adaptive voice enhancement module calls the following noise reduction strategies: the core strategy uses a notch filter with a center frequency of 120Hz±Δ, combined with the Redundant-Net general enhancement network to repair the speech spectrum loss caused by the notch filter; the emergency strategy is to forcibly increase the gain of the 500-2500Hz speech band by 3dB to ensure command priority; Δ is the frequency tracking tolerance, with a value range of 5~15Hz.

[0019] Compared with the prior art, the present invention has the following beneficial effects: The flight phase adaptive cockpit voice enhancement system based on multi-parameter fusion provided by the present invention can automatically identify the flight phase by analyzing flight parameters in real time and dynamically calling targeted noise reduction algorithms, thereby significantly improving the voice recording quality and adapting well to the noise environment at different stages of flight. Attached Figure Description

[0020] Figure 1 This is a diagram of the adaptive cockpit voice enhancement system architecture for the flight phase of the present invention. Figure 2 This is a flight data acquisition diagram of the present invention; Figure 3 This is a flowchart of the model scheduling and adaptive speech enhancement module of the present invention. Detailed Implementation

[0021] The present invention will now be further described with reference to the accompanying drawings and embodiments.

[0022] Figure 1 This is a schematic diagram of the adaptive cockpit voice enhancement system for the flight phase of the present invention.

[0023] Please see Figure 1 The adaptive cockpit voice enhancement system based on multi-parameter fusion provided by this invention consists of three core modules: The data acquisition module acquires 14 types of key flight data, such as airspeed, altitude, and engine parameters, in real time from the airborne bus (ARINC 429 / CAN, etc.), with a sampling period of 50-500ms; it supports multi-protocol parsing to ensure data and voice timestamp synchronization.

[0024] The flight phase identification module uses fuzzy logic to determine the current flight phase (9 regular phases + 3 special situations) and solves the problem of parameter overlap during transition phases; it outputs phase codes (such as F2 takeoff roll, SP1 engine failure) and confidence scores to the scheduling module.

[0025] Adaptive speech enhancement module, according to the stage characteristics to call the optimal noise reduction strategy (such as U-Net, spectral subtraction, harmonic trap wave, etc.), support model hot switching and mixed execution.

[0026] The following is a detailed process step introduction of each module: 1. Data acquisition module This module serves as the front-end perception layer of the entire speech enhancement system, responsible for real-time acquisition of key flight parameters from the aircraft's onboard data network, and outputs in a unified format for use by the "flight phase recognition module". It is the input basis for subsequent state judgment of flight state and adaptive noise reduction strategy call. The flight data acquisition module uses a modular structure to display each parameter category and its relevance, as shown in Figure 2 Kinetic parameters: directly affect the motion of the aircraft (such as acceleration determines speed change, pitch / roll controls flight trajectory); coupled with spatial state (such as altitude, airspeed) through aerodynamic equations.

[0027] Spatial state parameters: reflect real-time position and heading, dependent on kinetic parameters (such as attitude changes cause altitude changes); airspeed is strongly related to engine state (N1 / N2).

[0028] Engine state parameters: thrust output (N1 / N2) directly changes kinetic parameters (acceleration) and spatial state (climb rate); throttle lever position is the core input command of the pilot.

[0029] Configuration state parameters: mechanical configuration (flaps, landing gear) changes aerodynamic characteristics, affecting dynamics and spatial state; cabin pressure height is related to spatial state (true altitude) through the pressurization system.

[0030] Time reference parameters: synchronize the time axis of all parameters to ensure data consistency (such as aligning acceleration and flap state when recording the black box).

[0031] The enhancement system of the present invention needs to collect specific flight data types and sources, as shown in Table 1 below: Table 1 Flight data types and sources

[0032] FADEC is Full Authority Digital Engine Control, a core digital electronic system in the aviation field for precise control of engine operating state, providing engine speed N1 / N2. It realizes dynamic control of the engine from start-up, acceleration, cruise to shutdown through full authority (no mechanical backup or mechanical backup only for emergency) digital logic.

[0033] ​The data bus protocol of the application supports interface protocols such as ARINC 429, ARINC 664 (AFDX, Avionics Full Duplex Switched Ethernet), CAN and the like; each parameter corresponds to a unique LRU label or data word, and the sampling period is 50ms-500ms; a parsing dictionary table is maintained inside the module to realize protocol parsing and parameter mapping.

[0034] 2. Flight phase identification module (1) The flight phase identification module is mainly used for analyzing the input parameters of the flight data acquisition module, inferring the current flight phase (such as take-off, climb, cruise, descent, etc.) in real time; determining whether to enter a special situation (such as engine abnormality, cabin pressure out of control, lightning strike / extreme weather); outputting a standardized phase identification code for voice enhancement scheduling.

[0035] The application divides the flight phase into the following 9 kinds of regular flight phases and 3 kinds of special situations, as shown in Table 2 below: Table 2 Flight phase types and characteristics

[0036] APU is the abbreviation of Auxiliary Power Unit, which is a small power system independent of the main engine on the aircraft. Its core function is to provide auxiliary energy (electricity, compressed air) for the aircraft during ground parking and taxiing, as well as before take-off and after landing, while avoiding frequent starting of the main engine in non-flying state (high cost and high fuel consumption of main engine starting, and greater noise and pollutant emissions during ground starting).

[0037] GPWS (Ground Proximity Warning System) is a core safety device on civil aviation passenger aircraft and military aircraft to prevent "controlled flight into terrain (CFIT)" - a dangerous scenario where the aircraft accidentally approaches the ground or obstacles due to pilot's judgment error or height perception deviation, etc. under normal flight conditions (pilot controllable). GPWS can remind the pilot to avoid risks in time through sound and light alarms, and provide noise characteristics of the flight phase.

[0038] (2) Fuzzy logic is used to handle the parameter cross problem of the transition phase The application realizes flexible judgment and transition smoothing control of the flight phase boundary state through fuzzy logic reasoning, effectively avoids the misidentification problem of "hard threshold + jump", provides more continuous, accurate and robust state control signals for the voice enhancement engine, and improves the adaptability and intelligence level of the whole system.

[0039] Fuzzy logic allows a state to belong to multiple categories at the same time, just to different degrees. The calculation of confidence is based on this concept of "membership". The calculation process can be broken down into the following steps: Step 1: Fuzzyfication Convert the collected precise flight parameters (such as airspeed, altitude, throttle lever position) into membership degrees of the corresponding fuzzy sets. For example, for the "takeoff roll (F2)" phase, one of its judgment conditions is "airspeed 40-80 knots". A measured airspeed value of 60 knots, through a membership function (such as a trapezoidal or triangular function), can be calculated as follows: the membership degree of "roll speed" set is 0.95; the membership degree of "taxi speed" set is 0.1; the membership degree of "climb speed" set is 0.05.

[0040] Step 2: Fuzzy Inference Combine the membership degrees of multiple parameters, apply fuzzy rules (IF-THEN rules), and obtain the membership degree of each flight phase.

[0041] Rule example: IF airspeed is "roll speed" AND throttle lever is "takeoff thrust" THEN phase is "takeoff roll (F2)". The system will calculate the firepower (activation strength) of this rule, usually using the AND operator to take the minimum value, or the OR operator to take the maximum value.

[0042] Continue the example: airspeed membership degree = 0.95, throttle lever position membership degree = 0.90, so the activation strength of this rule for F2 phase is min(0.95, 0.90) = 0.90.

[0043] Step 3: Synthesis and Defuzzification (convert to confidence) The system performs the above calculations for all rules (covering F1-F9, SP1-SP3) to obtain a total membership degree score for each flight phase. At this time, you may get a set like this: {F1: 0.1, F2: 0.9, F3: 0.2,...}. The highest membership degree at this time, after normalization or direct output, is the "confidence". In the above example, the system determines the F2 phase with a confidence of 90%.

[0044] 3. Adaptive Speech Enhancement Module (1) Model Scheduler The model scheduler is located in the core of the speech processing link in the system architecture, and its main responsibilities are: to determine the current flight phase according to the output of the flight phase identification module (F1-F9 or SP1-SP3); to select the most suitable noise reduction model or processing strategy in the flight phase speech enhancement strategy scheduling table; to control the speech processing engine to execute different algorithms (U-Net, spectral subtraction, harmonic notch, etc.); to support smooth transition or parallel fusion between models to ensure stable voice quality.

[0045] The system configures the optimal model structure for each type of phase according to the typical phase noise characteristics, and the flight phase speech enhancement strategy scheduling table is shown in Table 3 below: Table 3 Flight phase speech enhancement strategy scheduling table

[0046] (2) Use confidence to make more intelligent and robust decisions The usage is as follows: i. Determine the "aggressiveness" of the enhancement strategy High confidence (e.g. > 85%): The system is confident in the current phase. The scheduler will fully enable the "default enhancement strategy combination" (such as the U-Net dedicated model) corresponding to the phase, which will provide the strongest targeted noise reduction.

[0047] Low confidence (e.g. 60%-85%): The system is in a transitional or uncertain state. The scheduler will adopt a conservative or mixed strategy: Start dual-model hybrid execution: Run the models of the previous and next two phases in parallel, with the weights gradually changing according to the confidence or time.

[0048] Enable more general alternative algorithms: temporarily do not call complex specialized neural networks, but use less computationally intensive and more stable classic algorithms (such as spectral subtraction) to avoid voice distortion due to phase misjudgment.

[0049] ii. Trigger the timing and smoothness of model "hot switching" The scheduler will monitor the trend of confidence. When the confidence is steadily and at a high level (e.g. > 90%) from one phase to another, perform a decisive hot switch.

[0050] When the confidence oscillates or slowly transitions between two phases, a smooth transition mechanism (such as a hybrid model) is started, with the switching speed proportional to the confidence change rate, ensuring that there is no perceptible jump in voice output.

[0051] iii. As an input for system health status and degradation strategy If the system is in a very low confidence state (e.g. < 50%) for a long time, it may mean that the sensor is malfunctioning or encountering an unexpected flight mode.

[0052] At this time, the dispatcher can trigger an alarm to prompt maintenance personnel to check the system. A safe system-wide default noise reduction mode (such as a lightweight general DNN or a set of fixed filters) is enabled to ensure basic system functionality at the expense of part of the optimization performance in exchange for robustness.

[0053] iv. Priority judgment in special situation handling For the SP special situation stage, high confidence means that the situation is urgent and certain. The dispatcher will not only call the corresponding noise reduction strategy, but also may trigger other emergency logic at the same time. For example, when the confidence of SP2 (cabin emergency) is high, in addition to noise reduction, it may also forcibly increase the gain of voice recording to ensure that the emergency instructions are clearly recorded.

[0054] (3) Dispatcher workflow Figure 3 The specific workflow of the model scheduling and adaptive speech enhancement module is as follows: Step 1: Obtain the phase state Receive the phase code (such as F3, SP1) and confidence from the flight phase identification module; Determine whether it is in phase switching, transition zone or special situation state.

[0055] Step 2: Model decision logic If phase switching occurs, trigger the dispatching strategy table retrieval; Check if it is necessary: model hot switching (old model unloading, new model loading); double model mixing (such as F3→F4, mixing U-Net+spectral subtraction); Check resource availability (CPU load, memory window); Step 3: Model loading / activation Call the compiled lightweight network in the local model cache (ONNX / TFLite); Or enable classic DSP strategies (such as spectral subtraction, notch); All models are pre-compressed to <1MB, supporting NPU or NEON SIMD optimization.

[0056] Step 4: Audio processing execution Input the speech frame (20ms), perform: preprocessing (normalization, frame window division); enhancement model inference / algorithm calculation; fundamental frequency / resonance frequency protection processing (only for human voice part); signal reconstruction and resampling; Step 5: Output and feedback Output the enhanced speech frame to the CVR recording cache; record the current model running information and signal-to-noise ratio improvement index.

[0057] Example 1: Application of Intelligent Cockpit Voice Recording System in Takeoff Roll Phase Step 1: Data Collection The system collects the following flight parameters in real time through the onboard bus (ARINC 429): • Airspeed: 60 knots; • Engine N1 speed: 95%; • Throttle lever position: 92%; • Altitude: 50 feet; • Flap status: retracted; • Timestamp: synchronized to UTC standard time;

[0058] Step 2: Flight Phase Identification The flight phase identification module analyzes the collected data based on fuzzy logic and determines that the current phase is "F2 takeoff roll phase" with a confidence level of 92%. The system identifies the following characteristics: • Airspeed between 40-80 knots; • Throttle lever position exceeds 90%; • 3rd harmonic of 180Hz detected in engine noise;

[0059] Step 3: Adaptive Voice Enhancement The model scheduler calls the following noise reduction strategies according to the phase code F2: Primary strategy: lightweight U-Net (Lite-F2) model, focusing on suppressing tire vibration (20-60Hz) and engine harmonics (180Hz).

[0060] Alternative strategy: multi-channel spectral subtraction (MC-Spectral Sub), used to handle sudden high-frequency splash noise (such as wet runway scenarios).

[0061] Step 4: Voice Processing Execution 1. Input 20ms voice frames, perform normalization and frame preprocessing.

[0062] 2. Inference through the U-Net model to separate noise and voice components.

[0063] 3. Use a tracking filter to further suppress engine harmonics.

[0064] 4. Output the enhanced voice frames to the cockpit voice recorder (CVR).

[0065] Effect Verification After testing, the signal-to-noise ratio (SNR) is improved from the original 15dB to 28dB, the tire vibration noise attenuation exceeds 90%, and the pilot's command clarity is significantly improved.

[0066] Example 2: Emergency handling under extreme weather (lightning strike) Step 1: Data acquisition The system detects the following abnormal parameters: • Vertical acceleration: 0.6g (exceeds threshold 0.5g); • Radar echo intensity: off-scale; • Cabin pressure altitude: stable (SP2 not triggered); Step 2: Flight phase identification The fuzzy logic module integrates the parameters and determines "SP3 extreme weather phase" with a confidence of 88%. The characteristics include: • Full-band random impulse noise (0.1-10kHz); • Electromagnetic pulse noise (characteristic of lightning strike); Step 3: Adaptive speech enhancement The dispatcher enables emergency noise reduction strategies: Primary strategy: IRS-Mapper impact suppression network, targeted to eliminate transient impulse noise caused by lightning strikes.

[0067] Secondary strategy: high-frequency clipping + zero-crossing limiting, protecting the speech frequency band (300-3400Hz).

[0068] Step 4: Speech processing execution 1. Real-time detection of impulse interference in speech frames (through short-time energy analysis).

[0069] 2. IRS-Mapper network dynamically suppresses impulse components, preserving speech fundamental frequency.

[0070] 3. Limiting processing on remaining high-frequency noise.

[0071] Effect verification The impulse noise caused by lightning strikes is completely suppressed, and the speech intelligibility is maintained at the level of 4 or above required by the International Civil Aviation Organization (ICAO).

[0072] Example 3: Smooth transition from cruise phase to initial descent Step 1: Data acquisition The system monitors parameter changes: • Altitude rate of change: from ±50ft / min (F5 flat cruise) to -800ft / min; • Throttle lever position: drops to 45%; Step 2: Flight phase identification The module identifies "F5→F6 transition state" with a confidence of 76%. The characteristics include: • Descent rate exceeds 500ft / min for 3 seconds; • Cabin pressure regulating valve periodic noise (1 Hz modulation) Step 3: Adaptive speech enhancement The scheduler executes a hybrid strategy: Transition period strategy: Run Gated-Conv-DNN (F5 legacy model) and Spectral subtraction (F6 preloaded model) in parallel, with weights gradually changing over time (70% -> 30%). For example, during the transition period from F5 (cruise) to F6 (initial descent), the system does not abruptly switch models, but instead runs both models in parallel and mixes their output results; the mixing weights are dynamically adjusted over time.

[0073] Starting moment (t0): The system just enters the transition state, Gated-Conv-DNN (F5 model) weight: 70%, Spectral subtraction (F6 model) weight: 30%; final output speech frame = (Gated-Conv-DNN output x 0.7) + (spectral subtraction output x 0.3).

[0074] Transition process (t0...t n ): Linear (or nonlinear) weight change, Gated-Conv-DNN weight gradually decreases from 70%, spectral subtraction weight gradually increases from 30%; for example, at the midpoint of the transition period: the weights may be 50% / 50%.

[0075] End moment (t n ): End of transition period, completely enter F6 stage; Gated-Conv-DNN (F5 model) weight: 0%, spectral subtraction (F6 model) weight: 100%; at this time, the system has actually completely switched to the model of the F6 stage, and the Gated-Conv-DNN is safely unloaded.

[0076] Stable strategy: Switch to LSTM + filter fusion, optimize periodic noise suppression.

[0077] Step 4: Speech processing execution 1. Dual model output weighted fusion, avoiding switching transient distortion.

[0078] 2. LSTM network learns the periodic characteristics of pressure regulating noise.

[0079] Effect verification The speech quality fluctuation in the transition stage is less than 1 dB, with no perceptual jump, meeting the DO-214 aviation speech recording standard.

[0080] Example 4: Special processing of engine failure (SP1) Step 1: Data collection Flight control system alarm: • N1-N2 speed difference: 18% (exceeds 15% threshold); • Vibration sensor detects shaft frequency harmonics (120Hz and its multiples); Step 2: Flight phase identification The module determines that "SP1 engine failure" with a confidence level of 95%. The noise characteristics include: • Rotor imbalance vibration (120Hz narrowband noise); • Impact oscillation (caused by fan blade shedding); Step 3: Adaptive speech enhancement Scheduler calls: Core strategy: Notch filter (center frequency 120Hz ± Δ, Δ is the frequency tracking tolerance, preferably 5-15Hz), combined with Redundant-Net general enhancement network.

[0081] Emergency strategy: Forcefully increase the gain of the speech band (500-2500Hz) by 3dB to ensure priority of commands.

[0082] Step 4: Speech processing execution 1. Dynamically adjust the notch bandwidth to track shaft frequency drift.

[0083] 2. Redundant-Net repairs speech spectrum loss caused by notch.

[0084] Effect verification Rotor noise is reduced by 20dB, and the recognition accuracy of key commands such as "Mayday" calls reaches 100%.

[0085] As can be seen from the above, the flight phase adaptive cockpit speech enhancement system based on multi-parameter fusion provided by the present application has the following advantages: 1. Strong system versatility, adaptable to different types of civil aircraft platforms Supports mainstream aviation bus protocols such as ARINC 429, ARINC 664, CAN, RS422, etc., with good scalability and platform compatibility, facilitating integration into various civil aviation cabin recording systems for rapid deployment and landing application.

[0086] 2. Automatically identify flight phases to improve the intelligence and relevance of speech processing Unlike existing systems that generally use uniform noise reduction processing, the present application first introduces a "fuzzy logic-based flight phase identification module" that can determine the aircraft's 9 regular flight phases and 3 special situations in real time, implement targeted speech processing strategy scheduling, and significantly improve the system's environmental adaptability and response speed.

[0087] 3. Multi-model dynamic calling, significantly improving voice clarity and fidelity The application adopts "flight phase-noise reduction model" one-to-one mapping design, matches the optimal model (such as U-Net, spectral subtraction, Gated-TCN, etc.) for each type of flight phase, and supports model hot switching and parallel fusion, effectively suppresses complex noises such as tire vibration, wind shear, GPWS warning, etc., and realizes optimal voice enhancement under different flight phases.

[0088] Although the present application has been disclosed with the preferred embodiments as above, it is not intended to limit the present application, and any person skilled in the art can make some modifications and improvements without departing from the spirit and scope of the present application, so the protection scope of the present application is defined by the claims.

Claims

1. A multi-parameter fusion based flight phase adaptive cockpit speech enhancement system, characterized in that, Comprise: Data acquisition module, real-time acquisition of key flight parameters from the aircraft's onboard data network; Flight phase identification module, analyze the input parameters of the data acquisition module, real-time inference of the current flight phase, and determine whether to enter special situations; Output phase code and confidence to the adaptive speech enhancement module; Adaptive speech enhancement module, according to the phase code and confidence, select the appropriate noise reduction model to suppress the current noise, and through model hot switching and double model mixing execution to optimize scheduling.

2. The multi-parameter fusion based flight phase adaptive cockpit speech enhancement system of claim 1, wherein, The data acquisition module acquires key flight parameters from the onboard bus in real time, with a sampling period of 50-500ms, and the key flight parameters include kinetic parameters, spatial state parameters, engine state parameters and configuration state parameters, and obtains time reference parameters to synchronize the time axis of all parameters.

3. The multi-parameter fusion based flight phase adaptive cockpit speech enhancement system according to claim 1, wherein, The flight phase identification module divides the flight phase into 9 regular phases and 3 special situation phases and outputs the phase code, uses fuzzy logic to handle the parameter crossing problem of the transition phase and outputs the confidence.

4. The multi-parameter fusion based flight phase adaptive cockpit speech enhancement system of claim 3, wherein, The phase code setting and determination of the regular phase are as follows: Phase code F1 corresponds to ground taxi, the determination parameter threshold is airspeed <40 knots, and engine N1 <50%; The corresponding noise is ground broadcast and APU high-frequency howling; Phase code F2 corresponds to take-off roll, the determination parameter threshold is airspeed 40-80 knots, and throttle lever >90%; The corresponding noise is tire vibration and engine 3rd harmonic; Phase code F3 corresponds to initial climb, the determination parameter threshold is height <1000ft, and climb rate >800ft / min; The corresponding noise is landing gear vortex shedding noise and flap gap howling; Phase code F4 corresponds to climb cruise, the determination parameter threshold is height 1000-cruise altitude, and acceleration is 0.05-0.2g; The corresponding noise is engine rotor noise and fuselage skin vibration; Phase code F5 corresponds to level cruise, the determination parameter threshold is height variation rate ±50ft / min, and Mach number is 0.78-0.85; The corresponding noise is broadband turbulent flow noise and air conditioning system airflow sound; Phase code F6 corresponds to initial descent, the determination parameter threshold is descent rate 500-1000ft / min, and throttle lever is 30-60%; The corresponding noise is passenger cabin pressure regulating valve periodic opening and closing; Phase code F7 corresponds to approach preparation, the determination parameter threshold is height <3000ft, flap >15°, and airspeed <180 knots; The corresponding noise is configuration warning sound and slat turbulent flow; Phase code F8 corresponds to final approach, the determination parameter threshold is glide path capture, and radio altitude <1000ft; The corresponding noise is GPWS voice alarm and raindrop impact spectrum; Phase code F9 corresponds to landing roll, the determination parameter threshold is airspeed <80 knots, and reverse thrust is activated; The corresponding noise is brake disc friction and reverse thrust turbulent flow.

5. The multi-parameter fusion based flight phase adaptive cockpit speech enhancement system of claim 3, wherein, The phase code setting and determination of the special situation phase are as follows: Phase code SP1 corresponds to engine failure, the determination parameter threshold is engine speed difference >15%; The corresponding noise is rotor unbalanced vibration; Phase code SP2 corresponds to cabin emergency, the threshold of the determination parameter is cabin pressure altitude > 8000 ft; the corresponding noise is the mechanical sound of oxygen mask release; Phase code SP3 corresponds to extreme weather, the threshold of the determination parameter is vertical acceleration > 0.5g; the corresponding noise is full-band random impulse noise.

6. The multi-parameter fusion based flight phase adaptive cockpit speech enhancement system of claim 1, wherein, The adaptive speech enhancement module receives phase code and confidence from the flight phase identification module; if it is determined that the phase is switched or in a special phase, and the confidence exceeds the preset threshold, the scheduling strategy table retrieval is triggered, the default enhancement strategy combination or alternative strategy suitable for the current flight phase is adopted to perform fundamental frequency / resonant frequency protection processing on the human voice part in the input speech frame, and the enhanced speech frame is output to the CVR recording buffer, and the current signal-to-noise ratio improvement index is recorded.

7. The multi-parameter fusion based flight phase adaptive cockpit speech enhancement system of claim 6, wherein, In the take-off running stage, when the flight phase identification module identifies that the airspeed is between 40-80 knots, the throttle lever position exceeds 90%, and the 3rd harmonic of 180Hz is detected in the engine noise, it is determined that the current is the take-off running stage, and the phase code F2 is output; the adaptive speech enhancement module calls the following noise reduction strategy: the main strategy uses a lightweight U-Net model to infer the separation of noise and speech components, focusing on suppressing tire vibration and engine harmonics, and the alternative strategy is multi-channel spectral subtraction, which is used to process sudden high-frequency splash sound in wet runway scene.

8. The multi-parameter fusion based flight phase adaptive cockpit speech enhancement system of claim 6, wherein, In extreme lightning weather, when the flight phase identification module identifies that the vertical acceleration exceeds the threshold, the radar echo intensity is out of limit, 0.1-10kHz full-band random impulse noise and electromagnetic pulse noise appear, and the cabin pressure altitude is stable, it is determined that the current is extreme weather, and the phase code SP3 is output; the adaptive speech enhancement module calls the following noise reduction strategy: the main strategy uses IRS-Mapper network to dynamically suppress impulse components, and specifically eliminates transient impulse noise caused by lightning, and the auxiliary strategy is high-frequency clipping combined with zero-crossing limiting, which protects the speech frequency band of 300-3400Hz.

9. The multi-parameter fusion based flight phase adaptive cockpit speech enhancement system of claim 6, wherein, In the cruise phase transition to the initial descent phase, when the flight phase identification module identifies that the descent rate exceeds 500ft / min for 3 seconds, the throttle lever position is reduced to 45%, and the 1Hz periodic pressure regulating sound is detected, it is determined that the current is the transition state of phase code F5 to phase code F6; the adaptive speech enhancement module calls the following noise reduction strategy: the transition strategy is to run the Gated-Conv-DNN model left by phase code F5 and the spectral subtraction model preloaded by phase code F6 in parallel, and the Gated-Conv-DNN weight gradually changes from 70% to 0% and the spectral subtraction weight gradually increases from 30% to 100% during the transition process; the stable strategy is to switch to LSTM plus filtering fusion to optimize periodic noise suppression.

10. The multi-parameter fusion based flight phase adaptive cockpit speech enhancement system of claim 6, wherein, When the flight phase recognition module recognizes that the engine speed difference exceeds the threshold value, and the vibration sensor detects the shaft frequency harmonic of 120Hz and its multiple, it is determined that the current is an engine failure, and the phase code SP2 is output; the adaptive speech enhancement module calls the following noise reduction strategy: the core strategy adopts a notch filter with a center frequency of 120Hz±Δ, and cooperates with a Redundant-Net general enhancement network to repair the speech spectrum damage caused by the notch; the emergency strategy: forcibly increases the gain of the 500-2500Hz speech frequency band by 3dB to ensure that the instruction is prioritized; Δ is a frequency tracking tolerance, and the value range is 5-15Hz.

Images