CN115618201A

CN115618201A - Breathing machine signal processing method based on compressed sensing and breathing machine

Info

Publication number: CN115618201A
Application number: CN202211229071.XA
Authority: CN
Inventors: 黄絮; 徐德祥; 杜春玲
Original assignee: Hunan Ventmed Medical Technology Co Ltd
Current assignee: Hunan Ventmed Medical Technology Co Ltd
Priority date: 2022-10-09
Filing date: 2022-10-09
Publication date: 2023-01-17

Abstract

The invention discloses a respirator signal processing method based on compressed sensing, which comprises the following steps: the method comprises the steps of collecting respiratory signals including a flow sensor and a pressure sensor of a breathing machine, performing multiple filtering pretreatment on the signals, obtaining a sparse matrix of the signals of the breathing machine by combining the sparse transformation of wavelet transformation and Fourier transformation, obtaining an optimal observation matrix based on an improved bee colony algorithm, and realizing a compressive sensing reconstruction model based on a one-dimensional residual convolution network model; the global optimization seeking and the optimization speed of the observation matrix are enhanced; the accuracy of the signal reconstruction is improved. The invention also provides a breathing machine. The invention realizes the optimized compression storage of the data of the breathing machine, and leads the breathing machine to store more data.

Description

Breathing machine signal processing method based on compressed sensing and breathing machine

Technical Field

The invention relates to the field of medical instruments, in particular to a breathing machine signal processing method based on compressed sensing and a breathing machine.

Background

The respirator is a device which can replace, control or change the normal physiological respiration of a person, increase the ventilation capacity of the lung, improve the respiratory function, reduce the consumption of the respiratory function and save the reserve capacity of the heart. The pressure of the airway is detected by a pressure sensor near the mask, the microprocessor receives an actual airway pressure signal fed by the pressure sensor, compares the actual airway pressure signal with the set treatment pressure, and continuously adjusts the rotating speed of the motor according to the difference value of the actual airway pressure signal and the set treatment pressure to maintain the stable pressure during treatment.

When the existing breathing machine stores data, a common method is to directly store the acquired data, so that the capacity of storage equipment is limited, and the data volume occupies a large memory, so that the breathing signal data volume capable of being stored is limited, and the breathing signal data for a long time cannot be stored.

Disclosure of Invention

Technical problem to be solved

In order to solve the technical problem, the invention provides a method for processing a breathing machine signal based on compressed sensing, which compresses a breathing signal obtained by a breathing machine, stores the breathing signal in a storage unit, and reconstructs the compressed signal based on a signal reconstruction method when data reading is needed, so that a large amount of storage space can be saved, and long-time storage of breathing machine data can be realized.

(II) technical scheme

In order to solve the technical problems and achieve the purpose of the invention, the invention is realized by the following technical scheme:

the method comprises the following steps: collecting respiratory signals, wherein the collected signals comprise flow sensor data and pressure sensor data;

step two: signal preprocessing, namely filtering a flow sensor signal by adopting three filters and filtering a pressure sensor signal by adopting a low-pass filter;

step three: sparse expression of the breathing machine signals, namely, obtaining a sparse matrix of the breathing machine signals by combining the sparse transformation of wavelet transformation and Fourier transformation on the acquired breathing machine signals;

step four: selecting an observation matrix, and combining the observation matrix with a random observation matrix based on an improved bee colony algorithm to realize optimization of the observation matrix;

step five: and (5) respiratory signal reconstruction, and realizing a compressed sensing reconstruction model based on a one-dimensional residual convolution network model.

Further, the three filters in the second step include two low-pass filters and one high-pass filter;

further, the improved bee colony algorithm described in step four comprises the following steps:

step 1, initializing parameters, and setting the maximum iteration times of an artificial bee colony algorithm, the maximum search times of honey source stay and the bee population, wherein half of each of the collected bees and the observed bees.

And 2, initializing the population and randomly generating an initial solution according to an initialization formula.

The initialization formula is as follows:

x _ij ＝x _minj +rand[0,1](x _maxj -x _minj )

wherein x is _ij Representing the value of j in the sequence of the ith honey source, i is more than or equal to 1 and less than or equal to SN and is a natural number, j is more than or equal to 1 and less than or equal to D and is a natural number, D is the number of the number sequence to be solved, x _maxj 、x _minj The maximum and minimum values of all solutions for the j dimension. And randomly selecting the assignment of the honey source between the extreme values, and then calculating the fitness of the honey source.

And 3, searching the honey source neighborhood by adopting the bees to generate a new position, performing probability crossing on the global optimal solution of the algorithm and the solution obtained after the bee neighborhood search to obtain a new solution, comparing the original solution with the new solution according to a greedy criterion, and selecting the honey source with less task completion time in task scheduling.

The search was performed using the following formula:

wherein x is _ij Is honey source location, x' _j The j dimension value of the global optimal solution, a and b are random values, and the value ranges are respectively-1<a<1,0<b<1.5, R is an equilibrium coefficient threshold value, and r is a random number between 0 and 1.

And 4, observing bees to obtain honey source information transmitted by the swinging dance of the collected bees, calculating honey source pheromones, and selecting honey sources according to the relation between the sensitivity and the honey source pheromones.

Setting a variable m (i), wherein n (i) is used for recording the number of times of continuous updating of the ith honey source, and updating according to the following formula:

and (5) representing the updating times of the honey source through m (i), n (i), thereby obtaining the search activity of the honey source. Calculating pheromone H (i) of the ith honey source as:

wherein A is a reference value of the honey source, lim is the maximum number of local search, and fit (i) is the fitness of the ith honey source.

The fitness fit (i) is calculated as follows:

where f (i) is the objective function of the algorithm.

The observation bees select the honey sources according to the pheromones and the sensitivities of the honey sources, and the probability of dynamic selection is as follows:

where H (i) is the pheromone of the ith employing bee source and N is the number of employing bees.

Randomly generating sensitivity S (i) of the ith observation bee, if the sensitivity S (i) of the ith observation bee is less than or equal to P (i), performing neighborhood search, and selecting a better honey source; if S (i) > P (i), the honey source position is unchanged.

And 5, observing the bee to be changed into a honey bee neighborhood, searching the current solution, crossing the global optimal solution with the probability of the current solution to obtain another solution, and selecting a new solution according to a greedy strategy.

And 6, if the honey source reaches the exploitation limit, converting the honey bee collection into a reconnaissance bee, and generating a new honey source.

And 7, if the iteration times of the algorithm are ended, returning the optimal solutions found by all the current bee colonies and the minimum task scheduling completion time, and ending the algorithm.

And 8, calculating to obtain an optimal observation matrix according to the optimal solution of the bee colony algorithm.

Furthermore, the network structure in the fifth step mainly comprises an input layer, a convolution layer, a residual block, a pooling layer, a normalization layer, a full-link layer and an activation function.

The normalization mode is as follows:

wherein x is ^j Data for layer j neurons, E mean and Var variance.

The convolution kernel formats in the 3 one-dimensional convolution layers in the residual block unit are 64, 64 and 256, respectively, and the sizes of the corresponding convolution kernels are 1 × 1, 3 × 3 and 1 × 1, respectively.

Further comprising: network training and signal reconstruction

And training the network model by adopting an optimizer, setting learning rate and termination conditions, adopting truncated Gaussian distribution as model initialization weight, and randomly selecting 80% of the acquired compressed sensing signals by adopting a training set.

And performing signal reconstruction on the compressed signal by adopting the trained network model to obtain reconstructed data. To evaluate reconstruction accuracy, signal reconstruction is considered successful when the reconstruction error satisfies the following equation:

wherein d is the original signal and d is the original signal,

is the reconstructed signal.

The present invention also provides a ventilator, comprising:

the signal acquisition module: the system is used for acquiring signals of a flow sensor and a pressure sensor of the respirator;

a signal processing module: the acquired signals are processed based on a digital filter, and the respiratory flow signals and the air leakage are obtained by combining two low-pass filters and one high-pass filter.

A thinning module of the breathing machine signals: the method is used for sparse expression of the breathing machine signals, and particularly, a sparse matrix of the breathing machine signals is obtained by combining the sparse transformation of wavelet transformation and Fourier transformation;

selecting an observation matrix module: the method is used for selecting the optimal observation matrix, and specifically, the optimal observation matrix is obtained through calculation based on an improved bee colony algorithm;

a ventilator signal reconstruction module: the method is used for reconstructing the breathing machine signals based on the deep learning network. The method mainly comprises two parts of establishing a deep learning network and training the deep learning network and reconstructing signals.

(III) advantageous effects

Compared with the prior art, the invention has the following beneficial effects:

(1) The respiration signal is compressed and stored through the induction compression technology, so that the storage space is saved, and the time for storing data is prolonged;

(2) The optimal observation matrix is selected based on the improved bee colony algorithm, so that the overall optimization seeking and the optimization speed of the observation matrix are enhanced;

(3) The invention realizes a compressed sensing reconstruction model based on a one-dimensional residual convolution network model, and improves the accuracy of signal reconstruction.

Drawings

The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this application, illustrate embodiment(s) of the application and together with the description serve to explain the application and not to limit the application. In the drawings:

FIG. 1 is a flow chart of a method for processing a ventilator signal according to an embodiment of the present application

FIG. 2 is a schematic view of a flow chart of an observation matrix optimization algorithm according to an embodiment of the present application

FIG. 3 is a schematic diagram of a network model structure according to an embodiment of the present application

FIG. 4 is a schematic diagram of a residual block according to an embodiment of the present application

Detailed Description

The embodiments of the present disclosure are described in detail below with reference to the accompanying drawings.

The embodiments of the present disclosure are described below with specific examples, and other advantages and effects of the present disclosure will be readily apparent to those skilled in the art from the disclosure in the specification. It is to be understood that the embodiments described are only a few embodiments of the present disclosure, and not all embodiments. The disclosure may be embodied or carried out in various other specific embodiments, and various modifications and changes may be made in the details within the description without departing from the spirit of the disclosure. It should be noted that the features in the following embodiments and examples may be combined with each other without conflict. All other embodiments, which can be derived by a person skilled in the art from the embodiments disclosed herein without inventive step, are intended to be within the scope of the present disclosure.

It should be noted that the drawings provided in the following embodiments are only for illustrating the basic idea of the present disclosure, and the drawings only show the components related to the present disclosure rather than the number, shape and size of the components in actual implementation, and the type, amount and ratio of the components in actual implementation may be changed arbitrarily, and the layout of the components may be more complicated.

Referring to fig. 1, a method for processing a ventilator signal based on compressed sensing includes:

s1: collecting respiratory signals

The breathing detection is mainly that flow signals of a breathing machine pipeline and output ventilation pressure are acquired by corresponding sensors, filtered, amplified and finally sent into a microprocessor, and basic data are provided for processing of a breathing detection algorithm. The collected data includes flow sensor data and pressure sensor data.

S2 Signal preprocessing

The function of the filter is to reduce or eliminate the interference of noise to the effective signal according to different characteristics of the effective signal and the noise. The filter includes both a digital filter and an analog filter. The input and output of the digital filter are discrete time sequences, and the input sequences are subjected to filtering processing through software, so that the components in an effective frequency range are extracted, and the noise in an invalid frequency range is filtered. Compared with an analog filter, the digital filter has the advantages of higher reliability, precision and flexibility, and digital waveform storage can be realized.

The respiratory flow signal obtained from the flow sensor contains the normal respiratory flow, the air leakage and the interference high-frequency signal of the flow waveform. The magnitude of the leak flow rate is related to the patient mask usage.

The invention adopts 3 filters for filtering, comprising two low-pass filters 1 and 2 and a high-pass filter.

The normal range of the human respiratory frequency is 16-20 times per minute, and the respiratory frequency is a low-frequency signal, so a digital low-pass filter 1 is adopted to filter high-frequency interference signals in the respiratory flow signals. Compared with the respiratory signal of the patient, the air leakage fluctuates in a certain time, so that the respiratory flow is separated from the air leakage, the respiratory flow waveform in the waveform is filtered by the flow signal through the low-pass filter 2 to obtain the air leakage, and the air leakage is filtered by the high-pass filter to obtain the respiratory flow.

And for the pressure signal acquired by the pressure sensor, a low-pass filter is adopted for filtering processing, and a high-frequency interference signal is filtered.

S3: sparse representation of ventilator signals;

the compressed sensing requires that signals are sparse or sparse in a certain transform domain, most signals in the nature are not sparse, and signals of a breathing machine are also non-sparse in a time domain, so that a proper sparse transform matrix needs to be designed to obtain sparse representation of the breathing machine signals projected onto the sparse matrix.

The mathematical criterion for signal sparsity is defined as X ∈ R for the signal ^N×1 A sparse representation S of the signal is obtained under the sparse matrix Ψ:

S＝Ψ×X

obtaining a sparse matrix of the breathing machine signals by combining the sparse transformation of wavelet transformation and Fourier transformation on the acquired breathing machine signals;

s4: selecting an observation matrix;

the characteristics of the observation matrix determine whether the compressed sensing theory is feasible or not, and also determine whether high-probability reconstruction of signals can be realized or not, and the observation matrix is an important part of the compressed sensing theory. Three types of observation matrices commonly found in the prior art are random matrices, deterministic matrices, and partially orthogonal matrices.

N-dimensional signal passing observation matrix phi (phi epsilon R) ^M×N M.ltoreq.N) can be reduced to dimension M as shown in the following formula:

Y＝Φ×X＝Φ×Ψ×θ＝A×θ

wherein A is a perception matrix;

at present, most observation matrixes have some defects, such as poor reconstruction performance of the Toeplitz observation matrix. The observation matrix needs to satisfy a finite equidistant criterion, which indicates that the performance of the sensing matrix is inversely proportional to the cross-correlation between the observation matrix and the sparse matrix, i.e. the smaller the cross-correlation between the matrices, the better the performance of this sensing matrix. The random measurement matrix has the advantages of less measurement quantity and good reconstruction performance. However, due to the complexity of the calculation, the random measurement matrix has a larger storage capacity than the deterministic measurement matrix, and the matrix reconstruction has higher uncertainty.

In order to solve the defects of the random observation matrix, the optimization of the observation matrix is realized by combining the improved swarm algorithm with the random observation matrix.

The algorithm flow chart is shown in fig. 2:

in the basic bee colony algorithm in the prior art, according to a roulette strategy, observers can select honey sources with better fitness, and poor honey sources are abandoned, so that the diversity of the colony is reduced, the optimizing capability of the algorithm is influenced, and the algorithm is early mature and falls into local optimization. The improved algorithm adopted by the invention can enlarge the range of selecting the honey source by matching with the pheromone of the honey source to select the area. The pheromone of the search area is matched with the sensitivity of the search area, so that the algorithm has guidance. Through the selection mode, honey sources with any fitness can be selected, the population diversity is richer, the algorithm can be prevented from falling into local optimum, the excellent honey source selection probability is higher, and the direction of observing the honey source selected by the bees is ensured.

The improved bee colony algorithm comprises the following steps:

The initialization formula is as follows:

x _ij ＝x _minj +rand[0,1](x _maxj -x _minj )

And 3, searching the honey source neighborhood by adopting the honeybees to generate a new position, performing probability crossing on the global optimal solution of the algorithm and the solution obtained after the honeybee neighborhood search to obtain a new solution, comparing the original solution with the new solution according to a greedy criterion, and selecting the honey source with less task completion time in task scheduling.

The search strategy of the bee colony algorithm is improved, in the artificial bee colony algorithm, the observation bees select the honey sources to mine through the honey source information shared by the honey-gathering bees, the greater the fitness value of the honey sources is, the higher the probability of selection is, the greater the number of observation bees for neighborhood search of the honey sources is, and the algorithm has stronger neighborhood search capability. The prior art uses the following formula for location update:

v _ij ＝x _ij +ε(x _ij -x _kj )；

wherein x is _k For other honey sources than i, k ≠ i, ε is an arbitrary decimal number, which is the coefficient of magnitude of the difference from the original value, and ranges from [ -1,1]And selecting and calculating a value with a larger value by adopting a greedy algorithm.

However, the position updated by the above formula is stronger in the ability of exploring honey sources and weaker in the ability of developing honey sources, and in order to make up for the defect of weaker ability of developing honey sources, the invention provides a new search mode, and the following formula is adopted for searching:

The exploration capacity and the development capacity of the honey sources are adjusted through the position updating formula of the formula, and the strength of the development capacity can be adjusted through adjusting the value of R.

And 4, observing the bee to obtain honey source information transmitted by the swinging dance of the bee, calculating honey source pheromones, and selecting the honey source according to the relation between the sensitivity and the honey source pheromones.

In the traditional algorithm, the observers determine the following probability according to the employing honey source fitness value, and the method is favorable for algorithm convergence but is easy to fall into local optimization. The method is considered to have more search potential and search value near the continuously optimized honey source. After the non-updated honey source is searched for many times, the neighborhood searching significance is not great, and the local optimum is already involved. The honey source neighborhood which is updated for many times is more active in searching and has greater significance in algorithm optimization guidance, so that the observation bees are more probabilistically distributed to the region.

The specific implementation mode is as follows:

The fitness fit (i) is calculated as follows:

where f (i) is the objective function of the algorithm.

where H (i) is the pheromone of the ith employed honey source and N is the number of employed bees. Therefore, when the observation bees select the employed bees, the fitness and the search activity of the honey source are considered, the honey source with higher activity is selected, the algorithm is guaranteed to jump out of local optimality in the early stage, the search is meticulous in the later stage, and finally the improved algorithm is high in convergence speed and high in convergence precision.

According to the sensitivity S (i) of the ith observation bee randomly generated, if the sensitivity S (i) of the ith observation bee is less than or equal to P (i), neighborhood searching is carried out, and a better honey source is selected; if S (i) > P (i), the honey source position is unchanged.

Based on the optimal observation matrix calculated by the improved bee colony algorithm, the convergence rate of the algorithm is improved, the running time is shortened, and meanwhile, the defect that the original algorithm is easy to fall into local optimal is overcome.

S5: reconstructing a respiratory signal;

reconstruction is the most important and critical part in compressed sensing research, and the reconstruction problem of the sparse signals aims to reconstruct original high-dimensional signals to the maximum extent through low-dimensional observation data, so that continuous optimization of reconstruction of the sparse signals has research value and significance.

The invention realizes a compressed sensing reconstruction model based on a one-dimensional residual convolution network model, and the signal reconstruction process comprises the following steps:

s51: establishment of network model

(1) Network model structure establishment

The network model is shown in fig. 3:

the network structure mainly comprises an input layer, a convolution layer, a residual block, a pooling layer, a normalization layer, a full-connection layer and an activation function.

In a residual network, the input is not only passed to the next layer in turn, but is also superimposed into the output of the next layer. Compared with the traditional convolutional neural network, the residual error network directly transfers the data stream to the following layers through the constant shortcut connection, so that gradient attenuation caused by the nonlinear transformation of multiple stacking can be reduced. Therefore, the residual network can build a deeper network model and training will be faster.

The normalization layer BN is added between the convolution layer and the pooling layer, the input of the neural network is easy to cause the change of data distribution after passing through the convolution layer and the activation layer, if the magnitude of the data is greatly different, the change of a large number can cause the change of a small number to be very unobvious when the neural network processes the data, so the data needs to be normalized in the network, and the normalization is also beneficial to gradient reduction and convergence speed acceleration. A simple normalization is usually:

wherein x is ^j Data for layer j neurons, E mean and Var variance.

(2) Residual block establishment

A residual block is created as shown in fig. 4, and the residual block unit proposed herein contains 3 one-dimensional convolutional layers, and in the residual block unit, in order to ensure that the input and output can be directly added, the dimensions and sizes of the input and output must be kept consistent. The convolution kernel formats in the 3 one-dimensional convolution layers in the residual block unit are 64, 64 and 256, respectively, and the sizes of the corresponding convolution kernels are 1 × 1, 3 × 3 and 1 × 1, respectively. In order to keep the feature size constant, appropriate padding values are set in each layer of convolution according to the size of the current convolution kernel. A band leakage corrected linear unit function is used as the activation function behind each convolutional layer.

Firstly, a 1x1 convolution block is used for reducing the number of channels, a common convolution block is used in the middle to enable the number of output channels to be equal to the number of input channels, and finally, the number of channels is restored by the 1x1 convolution block, and the number of output channels of the residual block is equal to the number of input channels in the whole view. These two 1x1 volume blocks are flexible to reduce the number of parameters.

S52: network training and signal reconstruction

wherein d is the original signal and d is the original signal,

is the reconstructed signal.

In the embodiment, the respiratory signals are compressed and stored by the induction compression technology, so that the storage space is saved, and the time for storing data is prolonged; an optimal observation matrix is selected based on an improved bee colony algorithm, so that the global optimization seeking and the optimization seeking speed of the observation matrix are enhanced; and a compressed sensing reconstruction model is realized based on a one-dimensional residual convolution network model, and the accuracy of signal reconstruction is improved.

An embodiment of the present invention further provides a ventilator, which includes:

a ventilator signal reconstruction module: the method is used for reconstructing the breathing machine signal based on the deep learning network. The method mainly comprises two parts of establishing a deep learning network and training the deep learning network and reconstructing signals.

The above-mentioned embodiments are merely illustrative of the preferred embodiments of the present invention, and do not limit the scope of the present invention, and various modifications and improvements made to the technical solution of the present invention by those skilled in the art without departing from the spirit of the present invention shall fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for processing a ventilator signal based on compressed sensing is characterized by comprising the following steps:

step two: signal preprocessing, namely filtering the signals of the flow sensor by adopting three filters and filtering the signals of the pressure sensor by adopting a low-pass filter;

2. The method according to claim 1, wherein the three filters in step two comprise two low-pass filters and one high-pass filter.

3. The method according to claim 1, wherein the fourth step further comprises:

the improved bee colony algorithm comprises the following steps:

(1) Initializing parameters, and setting the maximum iteration times of an artificial bee colony algorithm, the maximum search times of honey source stay and the number of bee populations, wherein half of each of the collected bees and half of each of the observed bees;

(2) Initializing a population, and randomly generating an initial solution according to an initialization formula;

the initialization formula is as follows:

x _ij ＝x _minj +rand[0,1](x _maxj -x _minj )

wherein x is _ij Representing the value of j in the sequence of the ith honey source, i is more than or equal to 1 and less than or equal to SN and is a natural number, j is more than or equal to 1 and less than or equal to D and is a natural number, D is the number of the number sequence to be solved, x _maxj 、x _minj Maximum and minimum values for all solutions in the j dimension; randomly selecting the assignment of the honey source between the extreme values, and then calculating the fitness of the honey source;

(3) Adopting bees to search a honey source neighborhood to generate a new position, performing probability crossing on the global optimal solution of the algorithm and the solution obtained after the bee neighborhood search to obtain a new solution, comparing the original solution with the new solution according to a greedy criterion, and selecting a honey source with less task completion time in task scheduling;

(4) Observing bees to obtain honey source information transmitted by the swinging dance of the collected bees, calculating honey source pheromones, and selecting honey sources according to the relation between sensitivity and the honey source pheromones;

randomly generating sensitivity S (i) of the ith observation bee, selecting probability P (i) of honey sources, and if the sensitivity S (i) of the ith observation bee is less than or equal to P (i), performing neighborhood search and selecting a better honey source; if S (i) > P (i), the position of the honey source is unchanged;

(5) The observation bee becomes a honey bee neighborhood to search a current solution, the global optimal solution is crossed with the probability of the global optimal solution to obtain another solution, and a new solution is selected according to a greedy strategy;

(6) If the honey source reaches the exploitation limit, the honey bees are converted into reconnaissance bees to generate a new honey source;

(7) If the iteration times of the algorithm is ended, returning the optimal solutions and the minimum task scheduling completion time found by all the current bee colonies, and ending the algorithm;

(8) And calculating to obtain an optimal observation matrix according to the optimal solution of the bee colony algorithm.

4. The method of claim 3, wherein the bee sampling generates a new position for honey source neighborhood search by using the following formula:

5. The method of compressed sensing-based ventilator signal processing according to claim 4, wherein the step (4) further comprises:

setting a variable m (i), wherein n (i) is used for recording the number of times of continuous updating of the ith honey source and updating according to the following formula:

and (5) representing the updating times of the honey source through m (i), n (i), thereby obtaining the search activity of the honey source.

6. The method of claim 5, wherein the step (4) further comprises:

calculating pheromone H (i) of the ith honey source as:

wherein A is a reference value of the honey source, lim is the maximum number of local search, and fit (i) is the fitness of the ith honey source;

the fitness fit (i) is calculated as follows:

where f (i) is the objective function of the algorithm.

7. The method of claim 6, wherein the step (4) further comprises:

the observation bees select the honey sources according to pheromones and sensitivities of the honey sources, and the probability of dynamic selection is P (i):

8. The method according to claim 1, wherein the step five further comprises: the network structure comprises an input layer, a convolution layer, a residual block, a pooling layer, a normalization layer, a full-connection layer and an activation function; the convolution kernel formats in the 3 one-dimensional convolution layers in the residual block unit are 64, 64 and 256, respectively, and the sizes of the corresponding convolution kernels are 1 × 1, 3 × 3 and 1 × 1, respectively.

9. The method according to claim 8, wherein the normalization is:

wherein x is ^j Data for layer j neurons, E mean and Var variance.

10. A ventilator based on compressed sensing ventilator signal processing method according to any of claims 1-9, characterized in that the ventilator comprises:

the signal processing module: processing the acquired signals based on a digital filter, and combining two low-pass filters and one high-pass filter to obtain respiratory flow signals and air leakage;

a thinning module of the breathing machine signals: the method is used for sparse representation of the breathing machine signals, and particularly is used for obtaining a sparse matrix of the breathing machine signals by combining the sparse transformation of wavelet transformation and Fourier transformation;

selecting an observation matrix module: the method is used for selecting the optimal observation matrix, and particularly, the optimal observation matrix is obtained through calculation based on an improved bee colony algorithm;

a ventilator signal reconstruction module: the system is used for reconstructing a breathing machine signal based on a deep learning network; the method mainly comprises two parts of establishing a deep learning network and training the deep learning network and reconstructing signals.