Positive pressure air breathing machine adaptive noise reduction method based on air flow spectrum characteristics

Abstract

This invention discloses an adaptive noise reduction method for positive pressure air respirators based on airflow spectrum characteristics, relating to the field of positive pressure air respirator technology. It simultaneously acquires airflow pressure signals from the respirator's airway and ambient noise signals using an airflow sensor and microphone array. Time-frequency analysis is performed on the acquired signals to extract airflow spectrum characteristics. A mapping relationship between the respirator's operating state and the noise spectrum is established based on historical records. This invention accurately identifies useful airflow signals and ambient noise signals by real-time acquisition and analysis of airflow spectrum characteristics, and dynamically adjusts noise reduction parameters to effectively suppress high-frequency noise and low-frequency vibrations, improving noise reduction performance. Compared to traditional fixed-band filtering or passive sound insulation materials, it can better cope with non-stationary noise environments, adjust noise reduction strategies in real time, and ensure clear auditory information in complex noise environments, thereby enhancing firefighters' auditory perception capabilities.

Description

Technical Field

[0001] This invention relates to the field of positive pressure air respirator technology, and more specifically to an adaptive noise reduction method for positive pressure air respirators based on airflow spectrum characteristics. Background Technology

[0002] Firefighters need to wear positive pressure breathing apparatus at fire scenes to ensure breathing safety. Fire scenes are usually accompanied by high-decibel noise, such as the sound of flames burning, the sound of fire-fighting equipment operating, and the sound of buildings collapsing. This noise can seriously interfere with the firefighters' hearing system, affecting their ability to judge the surrounding environment and communicate. Furthermore, the breathing apparatus generates noise such as airflow fluctuations and mechanical vibrations during operation, which further aggravates the noise environment at the scene and reduces the firefighters' comfort and work efficiency. In fire rescue, clear auditory information is crucial for firefighters' command and dispatch, teamwork, and rapid response to emergencies.

[0003] Traditional noise reduction methods typically employ fixed-band filtering or passive sound insulation materials, but they are poorly adaptable to non-stationary noises such as sudden changes in breathing airflow and sudden environmental noise, and cannot adjust noise reduction parameters in real time to cope with dynamically changing noise environments. While traditional spectrum analysis methods can analyze noise to a certain extent, they often cannot simultaneously and effectively suppress high-frequency noise such as motor harmonics and low-frequency vibrations such as valve opening and closing impacts. Furthermore, they do not fully consider the dynamic impact of the ventilator's operating status, including air supply pressure and user breathing frequency, on the noise spectrum. Therefore, this paper proposes an adaptive noise reduction method for positive pressure air ventilators based on airflow spectrum characteristics to solve the above problems. Summary of the Invention

[0004] The purpose of this invention is to provide an adaptive noise reduction method for positive pressure air respirators based on airflow spectrum characteristics, so as to solve the problems mentioned in the background art.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0006] An adaptive noise reduction method for positive pressure air respirators based on airflow spectrum characteristics includes the following steps:

[0007] Step 1: Simultaneously collect ventilator airway pressure signals and ambient noise signals using an airflow sensor and microphone array;

[0008] Step 2: Perform time-frequency analysis (short-time Fourier transform or wavelet transform) on the acquired signal to extract airflow spectral features;

[0009] Step 3: Establish a mapping relationship between the ventilator's operating status and the noise spectrum based on historical records, and perform targeted data matching and differentiated adjustments based on the mapping relationship between the ventilator's operating status and the noise spectrum, combined with the differences between uplink and downlink audio frequencies.

[0010] Step 4: Under the multi-task learning framework of polymorphic transformation, a noise recognition model is constructed using a deep learning model of CNN-LSTM hybrid network to distinguish useful airflow signals from environmental noise signals.

[0011] Step 5: Introduce a multi-module supervision strategy and design a grouping strategy to optimize model deployment, so as to ensure that the algorithm can still achieve leading noise reduction effect under limited computing power;

[0012] Step 6: Introduce a feedback mechanism to analyze user comfort through ear pressure sensor data and optimize model parameters to ensure a balance between noise reduction effect and wearing experience.

[0013] A further improvement to the technical solution of the present invention is that step one specifically includes:

[0014] A high-precision airflow sensor and microphone array are deployed in the positive pressure air respirator. The airflow sensor is installed at the air supply valve to monitor changes in airflow speed and flow rate. The microphone array is arranged inside the mask to collect internal airflow noise and external environmental noise. An ear pressure sensor is installed at the ear to monitor the user's wearing comfort.

[0015] Initialize all sensors, set the sampling frequency (44.1kHz or 48kHz), ensure the sensors are working properly and calibrate them to eliminate system errors, and then synchronously acquire airflow pressure signals and environmental noise signals through the data acquisition card;

[0016] The collected airflow pressure signal and environmental noise signal are preprocessed, including filtering, normalization and feature extraction. Among them, a low-pass filter is used to remove high-frequency noise and retain the main airflow features. The signal amplitude is adjusted to a uniform range through normalization to facilitate subsequent analysis.

[0017] Time synchronization of airflow pressure signals and environmental noise signals is performed to ensure that they are aligned in time. Then, data fusion technology is used to integrate the airflow signals and noise signals into a comprehensive dataset for subsequent noise identification and denoising.

[0018] A further improvement to the technical solution of the present invention is that step two specifically includes:

[0019] Short-time Fourier transform was used to perform time-frequency analysis on the preprocessed airflow pressure signal and environmental noise signal. The short-time Fourier transform is suitable for processing quasi-stationary signals and can provide good frequency resolution.

[0020] The signal is divided into multiple short time periods by using a sliding window, and a Fourier transform is performed on each time period to analyze the change of the signal's spectrum over time, thus obtaining the signal's spectrum.

[0021] After obtaining the time-frequency analysis results, the airflow spectrum characteristics are extracted from the signal spectrum diagram, including the main noise frequency components, noise bandwidth and airflow fluctuation frequency. The airflow spectrum characteristics are used to reflect the characteristics of airflow signals and environmental noise.

[0022] A multi-task learning framework for polymorphic transformation is constructed, with multiple task branches set up to handle different processing tasks, including multiple processing tasks such as airflow spectrum feature extraction, noise identification, and noise reduction parameter adjustment.

[0023] A further improvement to the technical solution of the present invention is that the process of obtaining the spectrum of the signal includes:

[0024] A fixed-length sliding window is used to segment the preprocessed airflow pressure signal and environmental noise signal. The sliding window slides on the signal, moving a fixed step each time, dividing the entire signal into multiple overlapping short signal segments. This approximates the non-stationary signal as a series of short-time stationary signals, which is convenient for subsequent analysis.

[0025] Perform a Fourier transform on the short signal segment within each sliding window to convert it from the time domain to the frequency domain. The Fourier transform yields the spectrum of each short signal segment, including the amplitude spectrum and the phase spectrum. The amplitude spectrum reflects the energy distribution of the signal at different frequencies.

[0026] The amplitude spectrum of each short signal segment is arranged in chronological order to generate the signal's spectrum diagram. In the spectrum diagram, the horizontal axis represents time, the vertical axis represents frequency, and the gray level represents amplitude. The spectrum diagram allows observation of how the signal's spectrum changes over time, clearly showing the appearance and disappearance of different frequency components in the signal, as well as changes in intensity.

[0027] A further improvement to the technical solution of the present invention is that step three specifically includes:

[0028] Historical airflow pressure signals and environmental noise signals of the ventilator under different operating conditions are collected, and the signals are preprocessed and airflow spectrum features are extracted. The data are then integrated to obtain historical records, where the operating conditions include supply pressure, user respiratory rate and respiratory mode (inhalation, exhalation).

[0029] By combining historical records, we analyzed the changes in noise spectrum under different working conditions, constructed a noise database, and established a mapping relationship between the working conditions of the ventilator and the noise spectrum.

[0030] The differences between uplink and downlink audio are analyzed, the spectral characteristics of uplink and downlink audio are compared, and the noise differences caused by different breathing airflow directions are identified. The noise differences include amplitude differences, frequency differences and phase differences. Uplink and downlink audio refer to the audio signals generated by the user's exhaled and inhaled airflow, respectively.

[0031] Based on the established mapping relationship and the differences between uplink and downlink audio, targeted data matching and differential adjustments are performed. For each working state, corresponding noise reduction parameters are matched, including the bandwidth and filtering strength of the filter. By dynamically adjusting the noise reduction parameters, the audio fidelity is optimized. Differential adjustments are made to process different frequency components separately, thereby enhancing low-frequency airflow signals and suppressing high-frequency noise.

[0032] A further improvement to the technical solution of the present invention is that the analysis process of the difference between the uplink and downlink audio includes:

[0033] High-precision airflow sensors and microphone arrays are placed inside the ventilator's air supply valve and mask to collect the up and down audio signals generated by the user's exhaled and inhaled airflows, respectively. The collected signals are preprocessed, including filtering to remove high-frequency noise and normalizing the signal amplitude to a uniform range. The signals are processed in segments through a sliding window to ensure the continuity and stability of the signals.

[0034] Time-frequency analysis is performed on the preprocessed uplink and downlink audio signals. Short-time Fourier transform is used to convert the signal from the time domain to the frequency domain, and the amplitude spectrum, frequency components and phase information of each short signal segment are extracted. The amplitude spectrum reflects the energy distribution of the signal at different frequencies, the frequency components show the main frequency of the signal, and the phase information provides the time characteristics of the signal.

[0035] By comparing the spectral characteristics of uplink and downlink audio, noise differences caused by different breathing airflow directions are identified, including amplitude differences, frequency differences, and phase differences. Specifically, amplitude differences are analyzed by observing the energy changes of uplink and downlink audio at the same frequency and analyzing their relationship with the breathing airflow direction. Frequency differences are analyzed by identifying different frequency components in uplink and downlink audio. Phase differences are analyzed by studying the phase changes of uplink and downlink audio over time. Through difference analysis, the influence of breathing airflow direction on noise characteristics is determined.

[0036] Based on the identified differences in uplink and downlink audio, targeted data matching is performed. For each working state (air supply pressure, breathing rate, breathing mode), corresponding noise reduction parameters are matched. In the inhalation phase, low-frequency airflow signals are enhanced, and in the exhalation phase, high-frequency noise is suppressed.

[0037] By optimizing audio fidelity through differentiated adjustments, the focus is on enhancing low-frequency airflow signals during the inhalation phase to ensure that users can clearly feel the airflow, while the focus is on suppressing high-frequency noise during the exhalation phase.

[0038] A further improvement to the technical solution of the present invention is that step four specifically includes:

[0039] Audio datasets containing useful airflow signals and environmental noise signals are extracted from historical records to ensure that the data covers different working conditions and noise types. The audio datasets are labeled to distinguish between useful airflow signals and environmental noise signals. The datasets are then divided into training sets, validation sets, and test sets to ensure consistent data distribution.

[0040] A deep learning model using a CNN-LSTM hybrid network is employed as the basic architecture to convert signals in the audio dataset into spectrograms, which serve as the hybrid input to the CNN-LSTM network.

[0041] The CNN-LSTM hybrid network is trained using the training set. The cross-entropy loss function is defined to measure the difference between the prediction and the true label. The Adam optimizer is used to adjust the network parameters. The learning rate is dynamically adjusted based on the performance of the validation set. The trained model is then evaluated using the test set. The accuracy, recall, and F1 score are calculated to quantify the model's performance. The misclassification situation is analyzed through the confusion matrix, and the model is optimized accordingly to obtain a well-trained noise recognition model.

[0042] The trained noise recognition model is integrated into the positive pressure air respirator to ensure low-latency inference capability. Real-time airflow pressure signal and environmental noise signal are input, converted into a spectrum, and then input into the noise recognition model. The noise recognition model then outputs classification results, calculates classification scores, analyzes the probability that the signal belongs to the useful airflow signal, and distinguishes the useful airflow signal from the environmental noise signal in real time.

[0043] A further improvement to the technical solution of the present invention is that the calculation process of the classification score includes:

[0044] Real-time acquisition of airflow pressure signals and environmental noise signals; conversion of signals into a spectrum graph using short-time Fourier transform; and calculation of the energy value at each time-frequency point from the spectrum graph.

[0045] The energy values ​​at each time-frequency point are normalized using a baseline energy value, which is obtained through statistical analysis of historical data. The normalized energy values ​​are then weighted and summed.

[0046] A time decay factor is introduced to simulate the decay of the signal over time. The weighted summation result is then mapped to the range [0, 1] using an exponential function and a logistic regression function to obtain the classification score.

[0047] A further improvement to the technical solution of the present invention is that step five specifically includes:

[0048] The noise recognition model is divided into multiple sub-modules, including a feature extraction module, a feature fusion module, and a classification module. Supervision points are set at the output of each sub-module to monitor the output quality and performance of the sub-module. The supervision points use the cross-entropy loss function to measure the difference between the intermediate output and the target value and update the parameters through the backpropagation algorithm. Through the multi-module supervision strategy, it is ensured that each module can be optimized independently, thereby improving the generalization ability and robustness of the overall model.

[0049] Assess the computational limitations of positive pressure air respirators, including available resources for CPU, GPU, and memory. Determine the computational capabilities and memory limitations of different hardware through performance testing. Based on the assessment results, allocate computationally intensive feature extraction modules to GPUs, time-series feature fusion modules to CPUs, and lightweight classification modules to dedicated FPGAs to ensure that each submodule runs on the most suitable hardware and reduce computational latency.

[0050] Based on computing power requirements and grouping strategies, the sub-modules are divided into multiple groups. The feature extraction module and the feature fusion module are grouped together and run on the GPU, while the classification module is grouped into another group and runs on the CPU.

[0051] A further improvement to the technical solution of the present invention is that step six specifically includes:

[0052] Combined with an ear pressure sensor, it collects ear pressure data in real time during the user's wearing process, and preprocesses the collected ear pressure data, including filtering to remove high-frequency noise and normalizing to adjust the signal amplitude to a uniform range.

[0053] Statistical indicators of the mean, variance, and peak of ear pressure were calculated as quantitative indicators of user wearing comfort. The changing trends of the ear pressure statistical indicators were analyzed, and a wearing comfort score was calculated to evaluate the user's wearing comfort under different noise reduction intensities.

[0054] Based on the assessment results of wearing comfort, the noise reduction parameters are dynamically adjusted. If the average ear pressure is too high, the noise reduction intensity is reduced to reduce stimulation to the user's ears. If the peak value is too high, the bandwidth of the filter is adjusted.

[0055] Due to the adoption of the above technical solution, the technical progress achieved by this invention compared to the prior art is as follows:

[0056] 1. This invention provides an adaptive noise reduction method for positive pressure air respirators based on airflow spectrum characteristics. By collecting and analyzing airflow spectrum characteristics in real time, it accurately identifies useful airflow signals and environmental noise signals, and dynamically adjusts noise reduction parameters to effectively suppress high-frequency noise and low-frequency vibrations, thereby improving the noise reduction effect. Compared with traditional fixed-band filtering or passive sound insulation materials, it can better cope with non-stationary noise environments, adjust noise reduction strategies in real time, and ensure clear auditory information in complex noise environments, thereby improving firefighters' auditory perception ability.

[0057] 2. This invention provides an adaptive noise reduction method for positive pressure air respirators based on airflow spectrum characteristics. By introducing a feedback mechanism, the method uses an ear pressure sensor to monitor the user's wearing comfort in real time and dynamically adjusts the noise reduction parameters. According to changes in ear pressure, the noise reduction intensity can be appropriately reduced or the bandwidth of the filter can be adjusted to reduce stimulation to the user's ears. By balancing the noise reduction effect and the wearing experience, the method ensures that the user feels comfortable during long-term use and avoids ear discomfort or fatigue caused by excessive noise reduction. Attached Figure Description

[0058] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0059] Figure 1 This is a flowchart of the present invention;

[0060] Figure 2 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] Example 1, such as Figure 1 , Figure 2 As shown, this invention provides an adaptive noise reduction method for positive pressure air respirators based on airflow spectrum characteristics, comprising the following steps:

[0063] Step 1: Simultaneously acquire ventilator airway pressure signals and ambient noise signals using an airflow sensor and microphone array. Deploy a high-precision airflow sensor and microphone array in the positive pressure air ventilator. The airflow sensor is installed at the air supply valve to monitor airflow velocity and flow rate changes. The microphone array is positioned inside the mask to collect internal airflow noise and external ambient noise. An ear pressure sensor is installed at the ear to monitor user comfort. Initialize all sensors, set the sampling frequency (44.1kHz or 48kHz), ensure the sensors are functioning correctly, and calibrate them to eliminate system errors. Then, airflow pressure signals and environmental noise signals are synchronously acquired through a data acquisition card. The acquired airflow pressure signals and environmental noise signals are preprocessed, including filtering, normalization, and feature extraction. Among them, a low-pass filter is used to remove high-frequency noise and retain the main airflow features. Normalization is used to adjust the signal amplitude to a uniform range to facilitate subsequent analysis. The airflow pressure signals and environmental noise signals are synchronized in time to ensure that the airflow pressure signals and environmental noise signals are aligned in time. Finally, data fusion technology is used to integrate the airflow signals and noise signals into a comprehensive dataset for subsequent noise identification and noise reduction processing.

[0064] Step 2: Perform time-frequency analysis (Short-Time Fourier Transform or Wavelet Transform) on the acquired signals to extract airflow spectral features. Utilize a multi-task learning framework with multi-state transformation for simultaneous lightweight processing, improving model processing efficiency. Short-Time Fourier Transform is used to perform time-frequency analysis on the preprocessed airflow pressure signal and environmental noise signal. Short-Time Fourier Transform is suitable for processing quasi-stationary signals and provides good frequency resolution. The signal is divided into multiple short time intervals using a sliding window, and a Fourier Transform is performed on each interval to analyze the change of the signal's spectrum over time, obtaining the signal's spectrum. After obtaining the time-frequency analysis results, extract the airflow spectral features from the signal's spectrum, including the main noise frequency components, noise bandwidth, and airflow fluctuation frequency. Then, use the... The flow spectrum features reflect the characteristics of airflow signals and environmental noise. A multi-task learning framework with multiple transformations is constructed, with multiple task branches responsible for different processing tasks, including airflow spectrum feature extraction, noise identification, and noise reduction parameter adjustment. Through multi-task learning, the features of airflow signals and noise signals are learned simultaneously, and the processing strategy is dynamically adjusted according to task requirements. In particular, under the multi-task learning framework, pruning and quantization compression techniques are used to lightweight the model. Pruning removes unimportant connections in the model, and quantization converts the model parameters from high-precision representation to low-precision representation, thereby reducing the model's storage space and computational load. Through lightweight processing, it is ensured that the model can still achieve efficient noise identification and noise reduction processing under limited computing power.

[0065] In addition, the process of obtaining the signal's spectrum includes:

[0066] A fixed-length sliding window is used to segment the preprocessed airflow pressure signal and environmental noise signal. The sliding window slides across the signal, moving a fixed step size each time, dividing the entire signal into multiple overlapping short signal segments. This approximates the non-stationary signal as a series of short-time stationary signals, facilitating subsequent analysis. The length of the sliding window is set to 25-50 milliseconds, and the overlap between windows is set to approximately 50% to ensure signal continuity and avoid information loss. The sliding window is used to segment the long signal into multiple short signal segments, each of which can be approximated as a quasi-stationary signal, thus making it suitable for short-time Fourier transform. Through the sliding window, the signal can be gradually... The window position is moved to cover the entire signal, ensuring that the signal in each time period can be analyzed. Fourier transform is performed on the short signal segments within each sliding window to convert them from the time domain to the frequency domain. Through Fourier transform, the spectrum of each short signal segment is obtained, including the amplitude spectrum and phase spectrum. The amplitude spectrum reflects the energy distribution of the signal at different frequencies. The amplitude spectrum of each short signal segment is arranged in chronological order to generate the signal's spectrum graph. In the spectrum graph, the horizontal axis represents time, the vertical axis represents frequency, and the gray level represents the amplitude. The spectrum graph of the signal allows observation of how the signal's spectrum changes over time, clearly showing the appearance and disappearance of different frequency components in the signal, as well as the intensity changes.

[0067] Step 3: Establish a mapping relationship between ventilator operating states and noise spectrum based on historical records. Based on this mapping relationship and considering the differences between uplink and downlink audio frequencies, perform targeted data matching and differential adjustments. Collect historical airflow pressure signals and environmental noise signals from the ventilator under different operating states. Preprocess the signals and extract airflow spectrum features to obtain historical records. Operating states include supply pressure, user respiratory rate, and breathing mode (inspiration, expiration). Combine these historical records to analyze changes in noise spectrum under different operating states, construct a noise database, establish a mapping relationship between ventilator operating states and noise spectrum, analyze the differences between uplink and downlink audio frequencies, and compare... The spectral characteristics of uplink and downlink audio signals are used to identify noise differences caused by different breathing airflow directions. These noise differences include amplitude, frequency, and phase differences. Uplink and downlink audio signals refer to the audio signals generated by the user's exhaled and inhaled airflow, respectively. Based on the established mapping relationship and the differences between uplink and downlink audio signals, targeted data matching and differential adjustments are performed. For each working state, corresponding noise reduction parameters are matched, including the bandwidth and filtering strength of the filter. By dynamically adjusting the noise reduction parameters, the audio fidelity is optimized. Differential adjustments involve the separate processing of different frequency components. By enhancing low-frequency airflow signals and suppressing high-frequency noise, the noise-reduced audio signal retains useful information while minimizing noise interference.

[0068] In addition, the analysis process for the differences between uplink and downlink audio includes:

[0069] High-precision airflow sensors and microphone arrays are placed inside the ventilator's air supply valve and mask to collect the up-and-down audio signals generated by the user's exhaled and inhaled airflows. The collected signals undergo preprocessing, including filtering to remove high-frequency noise and normalizing the signal amplitude to a uniform range. A sliding window is used to process the signal in segments to ensure continuity and stability. Time-frequency analysis is performed on the preprocessed up-and-down audio signals. Short-time Fourier transform is used to convert the signal from the time domain to the frequency domain, extracting the amplitude spectrum, frequency components, and phase information of each short signal segment. The amplitude spectrum reflects the energy distribution of the signal at different frequencies, the frequency components show the main frequencies of the signal, and the phase information provides the signal's time characteristics. By comparing the spectral characteristics of the up-and-down audio, noise differences caused by different airflow directions in amplitude, frequency, and phase are identified. The amplitude differences are analyzed to observe the energy changes of the up-and-down audio at the same frequency. The relationship between the noise level and the direction of breathing airflow was analyzed, frequency differences were analyzed, different frequency components in the uplink and downlink audio were identified, phase differences were analyzed, and the phase changes of the uplink and downlink audio over time were studied. Through difference analysis, the influence of breathing airflow direction on noise characteristics was determined. Based on the identified differences in uplink and downlink audio, targeted data matching was performed. For each working state (air supply pressure, breathing frequency, breathing mode), corresponding noise reduction parameters were matched. In the inhalation phase, low-frequency airflow signals were enhanced, and in the exhalation phase, high-frequency noise was suppressed. By dynamically adjusting the bandwidth and filtering intensity of the filter, it was ensured that noise could be effectively suppressed in different breathing phases while retaining useful information. Audio fidelity was optimized through differential adjustment. In the inhalation phase, the low-frequency airflow signal was enhanced to ensure that the user could clearly feel the airflow. In the exhalation phase, high-frequency noise was suppressed to reduce the interference of environmental noise. The POLQA (Perceptual Objective Listening Quality Assessment) score was used to evaluate the noise-reduced audio to ensure that it was superior to traditional algorithms.

[0070] Step 4: Under the multi-task learning framework of polymorphic transformation, a noise recognition model is constructed using a CNN-LSTM hybrid deep learning model to distinguish between useful airflow signals and environmental noise signals. Audio datasets containing both useful airflow signals and environmental noise signals are extracted from historical records, ensuring data coverage of different operating states and noise types. The audio datasets are labeled to distinguish between useful airflow signals and environmental noise signals. The datasets are then divided into training, validation, and test sets to ensure consistent data distribution. A CNN-LSTM hybrid deep learning model is used as the basic architecture to convert the signals in the audio dataset into spectrograms, which serve as the hybrid input to the CNN-LSTM network. The CNN layers automatically extract local features (frequency distribution, energy changes) from the signal spectrograms, reducing parameter redundancy. The LSTM layers capture long-term dependencies in the feature sequences, modeling the temporal dynamics of the airflow signals. Through a polymorphic transformation mechanism, the network structure (convolutional kernel size, number of LSTM units) is dynamically adjusted to adapt to different operating conditions. For different noise patterns, a CNN-LSTM hybrid network is trained using a training set. A cross-entropy loss function is defined to measure the difference between the prediction and the true label. The Adam optimizer is used to adjust the network parameters, and the learning rate is dynamically adjusted based on the performance of the validation set. The batch size and regularization parameters are adjusted through grid search or random search to improve the model's generalization ability. An early stopping method is introduced to terminate training when the validation set loss does not decrease continuously to prevent overfitting. The trained model is then evaluated using a test set, and the accuracy, recall, and F1 score are calculated to quantify the model's performance. Misclassification is analyzed using a confusion matrix, and the model is optimized accordingly to obtain a trained noise recognition model. The trained noise recognition model is then integrated into a positive pressure air respirator to ensure low-latency inference capability. Real-time airflow pressure signals and environmental noise signals are input, converted into spectrograms, and then input into the noise recognition model. The noise recognition model outputs classification results, calculates classification scores, and analyzes the probability that the signal belongs to a useful airflow signal, thus distinguishing between useful airflow signals and environmental noise signals in real time.

[0071] In addition, the calculation process for categorical scores includes:

[0072] Real-time acquisition of airflow pressure signals and environmental noise signals; conversion of signals into a spectrum using short-time Fourier transform; calculation of energy value at each time-frequency point from the spectrum; normalization of energy value at each time-frequency point using a reference energy value (obtained through historical data statistics); weighted summation of normalized energy values; introduction of a time decay factor to simulate signal decay over time; and mapping of the weighted summation result to the range [0, 1] using an exponential function and a logistic regression function to obtain a classification score.

[0073] The expression for categorical scores is:

[0074]

[0075] E(t, f) = |X(t, f)| 2 ;

[0076]

[0077] In the formula, S(t) is the classification score, representing the probability that the signal belongs to the useful airflow signal at time t, F is the number of frequency points in the spectrum, and w f The weight of frequency f represents the importance of that frequency in the classification. E(t, f) is the energy value at time t and frequency f, representing the energy value at each time-frequency point in the spectrum obtained by short-time Fourier transform (STFT). B(f) is the reference energy value at frequency f, used for normalization. α is the time decay factor, representing the degree of signal decay over time. |X(t, f)| is the amplitude of the spectrum. x(n) is the input signal. w(nt) is the sliding window function. N is the number of points in the Fourier transform. t is the time index. f is the frequency index. j is the imaginary unit. S(t) ranges from 0 to 1. When S(t) > 0.5, the signal is classified as a useful airflow signal. When S(t) ≤ 0.5, the signal is classified as an environmental noise signal.

[0078] Step 5: Introduce a multi-module supervision strategy and design a grouping strategy to optimize model deployment, so as to ensure that the algorithm can still achieve leading noise reduction effect under limited computing power;

[0079] Step 6: Introduce a feedback mechanism to analyze user comfort through ear pressure sensor data and optimize model parameters to ensure a balance between noise reduction effect and wearing experience.

[0080] Example 2, as Figure 1 , Figure 2 As shown, based on Embodiment 1, the present invention provides a technical solution: preferably, step five specifically includes:

[0081] The noise recognition model is divided into multiple sub-modules, including a feature extraction module, a feature fusion module, and a classification module. Supervision points are set at the output of each sub-module to monitor its output quality and performance. The supervision points use the cross-entropy loss function to measure the difference between intermediate outputs and target values, and parameters are updated using the backpropagation algorithm. This multi-module supervision strategy ensures that each module can be optimized independently, improving the overall model's generalization ability and robustness. The computational limitations of positive pressure ventilation machines are evaluated, including available CPU, GPU, and memory resources. Performance testing determines the computational capabilities and memory limitations of different hardware. Based on the evaluation results, the computational... The computationally intensive feature extraction module is allocated to the GPU, the time-series feature fusion module is allocated to the CPU, and the lightweight classification module is allocated to a dedicated FPGA. This ensures that each submodule runs on the most suitable hardware, reducing computational latency. Based on computing power requirements and grouping strategies, the submodules are divided into multiple groups. The feature extraction and feature fusion modules are grouped together and run on the GPU, while the classification module is grouped into another group and runs on the CPU. The communication mechanism between submodules is optimized. Data compression technology and asynchronous communication mechanisms are used to reduce data transmission overhead. Through shared memory and message queues, efficient and low-latency data transmission between modules is ensured.

[0082] Step six specifically includes:

[0083] Combined with an ear pressure sensor, the device collects real-time ear pressure data during the user's wearing process. The collected ear pressure data is preprocessed, including filtering to remove high-frequency noise and normalizing the signal amplitude to a uniform range. Statistical indicators of the mean, variance, and peak value of ear pressure are calculated as quantitative indicators of user wearing comfort. The changing trend of the ear pressure statistical indicators is analyzed to calculate a wearing comfort score, assessing the user's wearing comfort at different noise reduction intensities. Based on the wearing comfort assessment results, the noise reduction parameters are dynamically adjusted. If the mean ear pressure is too high, the noise reduction intensity is reduced to decrease stimulation to the user's ears. If the peak value is too high, the bandwidth of the filter is adjusted to avoid excessive amplification of high-frequency noise. Through a feedback mechanism, a balance between noise reduction effect and wearing experience is ensured.

[0084] The expression for the wearing comfort score is:

[0085]

[0086] In the formula, C(t) is the wearing comfort score, representing the user's wearing comfort at time t, and P... k(t) represents the ear pressure value collected by the k-th ear pressure sensor at time t, in Pascals (Pa). B is the baseline pressure value, obtained through historical data statistics, also in Pascals (Pa). μ is the mean ear pressure, reflecting long-term wearing comfort. σ is the variance of ear pressure, reflecting the stability of pressure changes. K is the number of ear pressure sensors, taking a value of 2, one for each ear. β is the peak weight, representing the degree of influence of the peak value on comfort. k C(t) represents the peak ear pressure at time t, reflecting instantaneous discomfort, and is measured in Pascals (Pa). The value of C(t) ranges from 0 to 1. When C(t) is close to 1, it indicates high wearing comfort, and when C(t) is close to 0, it indicates low wearing comfort.

[0087] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An adaptive noise reduction method for positive pressure air respirators based on airflow spectrum characteristics, characterized in that, Includes the following steps: Step 1: Simultaneously collect ventilator airway pressure signals and ambient noise signals using airflow sensors and microphone arrays; Step 2: Perform time-frequency analysis on the acquired signals to extract airflow spectral characteristics; Step 3: Establish a mapping relationship between ventilator operating status and noise spectrum based on historical records. Then, based on this mapping relationship and the differences between uplink and downlink audio frequencies, perform targeted data matching and differentiated adjustments. Specifically, this includes: Historical airflow pressure signals and environmental noise signals of the ventilator under different operating conditions are collected, and the signals are preprocessed and airflow spectrum features are extracted. The data are then integrated to obtain historical records, where the operating conditions include supply pressure, user respiratory rate, and respiratory mode. By combining historical records, we analyzed the changes in noise spectrum under different working conditions, constructed a noise database, and established a mapping relationship between the working conditions of the ventilator and the noise spectrum. The differences between uplink and downlink audio are analyzed, the spectral characteristics of uplink and downlink audio are compared, and the noise differences caused by different breathing airflow directions are identified. The noise differences include amplitude differences, frequency differences and phase differences. Uplink and downlink audio refer to the audio signals generated by the user's exhaled and inhaled airflow, respectively. Based on the established mapping relationship and the differences between uplink and downlink audio, targeted data matching and differential adjustments are performed. For each working state, corresponding noise reduction parameters are matched, including the bandwidth and filtering strength of the filter. By dynamically adjusting the noise reduction parameters, the audio fidelity is optimized. Differential adjustments are made to process different frequency components separately, by enhancing low-frequency airflow signals and suppressing high-frequency noise. Step 4: Construct a noise recognition model using a CNN-LSTM hybrid deep learning network to distinguish useful airflow signals from environmental noise signals; Step 5: Introduce a multi-module supervision strategy and design a grouping strategy to optimize model deployment; Step 6: Introduce a feedback mechanism to analyze user comfort using ear pressure sensor data and optimize model parameters.

2. The adaptive noise reduction method for positive pressure air respirators based on airflow spectrum characteristics according to claim 1, characterized in that: Step one specifically includes: A high-precision airflow sensor and microphone array are deployed in a positive pressure air respirator. The airflow sensor is installed at the air supply valve, the microphone array is arranged inside the mask, and an ear pressure sensor is installed at the ear. Initialize all sensors, set the sampling frequency, and then synchronously acquire airflow pressure signals and environmental noise signals through the data acquisition card; The collected airflow pressure signals and environmental noise signals are preprocessed, including filtering, normalization and feature extraction; The airflow pressure signal and the environmental noise signal are synchronized in time, and then the airflow signal and noise signal are integrated into a comprehensive dataset using data fusion technology.

3. The adaptive noise reduction method for positive pressure air respirators based on airflow spectrum characteristics according to claim 2, characterized in that: Step two specifically includes: Short-time Fourier transform was used to perform time-frequency analysis on the preprocessed airflow pressure signal and environmental noise signal; The signal is divided into multiple short time periods by using a sliding window, and a Fourier transform is performed on each time period to analyze the change of the signal's spectrum over time, thus obtaining the signal's spectrum. After obtaining the time-frequency analysis results, the airflow spectrum characteristics are extracted from the signal's spectrum diagram, including the main noise frequency components, noise bandwidth, and airflow fluctuation frequency. A multi-task learning framework for polymorphic transformation is constructed, with multiple task branches set up to handle different processing tasks, including multiple processing tasks such as airflow spectrum feature extraction, noise identification, and noise reduction parameter adjustment.

4. The adaptive noise reduction method for positive pressure air respirators based on airflow spectrum characteristics according to claim 3, characterized in that: The process of obtaining the spectrum of the signal includes: A fixed-length sliding window is used to segment the preprocessed airflow pressure signal and environmental noise signal. The sliding window slides on the signal, moving a fixed step each time, dividing the entire signal into multiple overlapping short signal segments, and approximating the non-stationary signal as a series of short-time stationary signals. Perform a Fourier transform on the short signal segment within each sliding window to convert it from the time domain to the frequency domain. The Fourier transform yields the spectrum of each short signal segment, including the amplitude spectrum and the phase spectrum. Arrange the amplitude spectrum of each short signal segment in chronological order to generate a spectrum of the signal. In the spectrum, the horizontal axis represents time, the vertical axis represents frequency, and the gray level represents amplitude.

5. The adaptive noise reduction method for positive pressure air respirators based on airflow spectrum characteristics according to claim 1, characterized in that: The analysis process for the differences between the uplink and downlink audio frequencies includes: High-precision airflow sensors and microphone arrays are placed inside the ventilator's air supply valve and mask to collect the up and down audio signals generated by the user's exhaled and inhaled airflow, and the collected signals are preprocessed. Time-frequency analysis is performed on the preprocessed uplink and downlink audio signals. Short-time Fourier transform is used to convert the signal from the time domain to the frequency domain, and the amplitude spectrum, frequency components and phase information of each short signal segment are extracted. The amplitude spectrum reflects the energy distribution of the signal at different frequencies, the frequency components show the main frequency of the signal, and the phase information provides the time characteristics of the signal. By comparing the spectral characteristics of uplink and downlink audio, noise differences in amplitude, frequency, and phase caused by different breathing airflow directions are identified. Specifically, amplitude differences are analyzed by observing the energy changes of uplink and downlink audio at the same frequency and analyzing their relationship with breathing airflow direction; frequency differences are analyzed by identifying different frequency components in uplink and downlink audio; and phase differences are analyzed by studying the phase changes of uplink and downlink audio over time. Based on the identified differences between uplink and downlink audio, targeted data matching is performed. For each working state, corresponding noise reduction parameters are matched. Specifically, during the inhalation phase, low-frequency airflow signals are enhanced, and during the exhalation phase, high-frequency noise is suppressed. By optimizing audio fidelity through differentiated adjustments, the focus is on enhancing low-frequency airflow signals during inhalation and suppressing high-frequency noise during exhalation.

6. The adaptive noise reduction method for positive pressure air respirators based on airflow spectrum characteristics according to claim 5, characterized in that: Step four specifically includes: Audio datasets containing useful airflow signals and environmental noise signals are extracted from historical records. The audio datasets are labeled to distinguish between useful airflow signals and environmental noise signals. The datasets are then divided into training sets, validation sets, and test sets. A deep learning model using a CNN-LSTM hybrid network is employed as the basic architecture to convert signals in the audio dataset into spectrograms, which serve as the hybrid input to the CNN-LSTM network. The CNN-LSTM hybrid network is trained using the training set. The cross-entropy loss function is defined to measure the difference between the prediction and the true label. The Adam optimizer is used to adjust the network parameters. The learning rate is dynamically adjusted in combination with the performance of the validation set. The trained model is then evaluated using the test set. The accuracy, recall, and F1 score are calculated to quantify the performance of the model, thereby obtaining a trained noise recognition model. The trained noise recognition model is integrated into a positive pressure air respirator. Real-time airflow pressure signal and environmental noise signal are input, converted into a spectrum, and then input into the noise recognition model. The noise recognition model then outputs the classification result, calculates the classification score, analyzes the probability that the signal belongs to the useful airflow signal, and distinguishes the useful airflow signal from the environmental noise signal in real time.

7. The adaptive noise reduction method for positive pressure air respirators based on airflow spectrum characteristics according to claim 6, characterized in that: The calculation process for the classification score includes: Real-time acquisition of airflow pressure signals and environmental noise signals; conversion of signals into a spectrum graph using short-time Fourier transform; and calculation of the energy value at each time-frequency point from the spectrum graph. The energy values ​​at each time-frequency point are normalized using a baseline energy value, which is obtained through statistical analysis of historical data. The normalized energy values ​​are then weighted and summed. A time decay factor is introduced to simulate the decay of the signal over time. The weighted summation result is then mapped to the range [0,1] using an exponential function and a logistic regression function to obtain the classification score.

8. The adaptive noise reduction method for positive pressure air respirators based on airflow spectrum characteristics according to claim 7, characterized in that: Step five specifically includes: The noise recognition model is divided into multiple sub-modules, including a feature extraction module, a feature fusion module, and a classification module. Supervision points are set at the output of each sub-module to monitor the output quality and performance of the sub-module. The supervision points use the cross-entropy loss function to measure the difference between the intermediate output and the target value, and update the parameters through the backpropagation algorithm. The computational limitations of positive pressure air respirators were assessed, including the available resources of CPU, GPU and memory. The computational capabilities and memory limitations of different hardware were determined through performance testing. Based on the assessment results, the computationally intensive feature extraction module was allocated to the GPU, the time-series feature fusion module was allocated to the CPU, and the lightweight classification module was allocated to a dedicated FPGA. Based on computing power requirements and grouping strategies, the sub-modules are divided into multiple groups. The feature extraction module and the feature fusion module are grouped together and run on the GPU, while the classification module is grouped into another group and runs on the CPU.

9. The adaptive noise reduction method for a positive pressure air respirator based on airflow spectrum characteristics according to claim 8, characterized in that: Step six specifically includes: Combined with an ear pressure sensor, it collects ear pressure data in real time during the user's wearing process, and preprocesses the collected ear pressure data, including filtering to remove high-frequency noise and normalizing to adjust the signal amplitude to a uniform range. Statistical indicators of the mean, variance, and peak of ear pressure were calculated as quantitative indicators of user wearing comfort. The changing trends of the ear pressure statistical indicators were analyzed, and a wearing comfort score was calculated to evaluate the user's wearing comfort under different noise reduction intensities. Based on the assessment results of wearing comfort, the noise reduction parameters are dynamically adjusted. If the average ear pressure is too high, the noise reduction intensity is reduced; if the peak value is too high, the bandwidth of the filter is adjusted.

Images