CN113050039A - Acoustic fluctuation positioning system used in tunnel - Google Patents
Acoustic fluctuation positioning system used in tunnel Download PDFInfo
- Publication number
- CN113050039A CN113050039A CN202110260239.2A CN202110260239A CN113050039A CN 113050039 A CN113050039 A CN 113050039A CN 202110260239 A CN202110260239 A CN 202110260239A CN 113050039 A CN113050039 A CN 113050039A
- Authority
- CN
- China
- Prior art keywords
- signal
- circuit
- gamma
- tunnel
- sound wave
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Images
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
- G01S5/00—Position-fixing by co-ordinating two or more direction or position line determinations; Position-fixing by co-ordinating two or more distance determinations
- G01S5/18—Position-fixing by co-ordinating two or more direction or position line determinations; Position-fixing by co-ordinating two or more distance determinations using ultrasonic, sonic, or infrasonic waves
- G01S5/24—Position of single direction-finder fixed by determining direction of a plurality of spaced sources of known location
Abstract
The invention relates to the technical field of tunnel positioning, in particular to an acoustic fluctuation positioning system used in a tunnel, which comprises a plurality of equidistant microphone sound wave receivers arranged in the same tunnel and a processor connected with all the microphone sound wave receivers, and a positioning processing method comprises the following steps: the sound wave signal received by the microphone is firstly amplified by a preceding stage signal amplifying circuit and then converted into a differential signal by a single-end to differential signal converting circuit; the differential signal is converted into an analog signal through the differential-to-single-ended signal conversion circuit, the analog signal sequentially passes through the second-order low-pass filter circuit, the band-pass filter circuit and the fixed gain amplification circuit to realize filtering and signal amplification of the signal, and then the signal is converted into a differential signal through the single-ended-to-differential signal conversion circuit II and input into the processor to be processed.
Description
Technical Field
The invention relates to the technical field of tunnel internal positioning, in particular to an acoustic fluctuation positioning system used in a tunnel.
Background
The tunnel generally has the characteristics of space closure, harsh environment and the like, and once a safety accident occurs, rescue workers outside the tunnel can hardly realize the accurate positioning and rescue of trapped people in the tunnel through a traditional communication mode. Therefore, it is important to know information such as the position of the vehicle traveling in the tunnel in real time in order to ensure the safety of the vehicle traveling in the tunnel.
Disclosure of Invention
In view of the shortcomings of the prior art, it is an object of the present invention to provide an acoustic wave localization system for use in tunnels.
In order to achieve the purpose, the invention provides the following technical scheme: the utility model provides an acoustics undulant positioning system for in tunnel, includes that a plurality of sets up the equidistance microphone sound wave receiver in same tunnel and the treater of being connected with all microphone sound wave receivers, and complete sound wave array positioning system is constituteed to all microphone sound wave receivers, microphone sound wave receiver include microphone and processing circuit, processing circuit includes preceding discharge circuit and back discharge circuit, preceding discharge circuit includes preceding signal amplification circuit and single-ended to differential signal conversion circuit one, back discharge circuit includes that the difference changes single-ended signal conversion circuit, second order low pass filter circuit, band-pass filter circuit, fixed gain amplifier circuit and single-ended to differential signal conversion circuit two,
the positioning processing method of the acoustic fluctuation positioning system in the tunnel comprises the following steps:
(1) the sound wave signal received by the microphone is firstly amplified by a preceding stage signal amplifying circuit and then converted into a differential signal by a single-end to differential signal converting circuit;
(2) the converted differential signal is transmitted to a rear discharge circuit through a shielding cable for processing;
(3) the differential signal is converted into an analog signal through a differential-to-single-ended signal conversion circuit, the analog signal sequentially passes through a second-order low-pass filter circuit, a band-pass filter circuit and a fixed gain amplification circuit to realize filtering and signal amplification of the signal, and then the signal is converted into a differential signal through a single-ended-to-differential signal conversion circuit II and is input into a processor for processing;
(4) the processor outputs the positioning information after processing.
Preferably, the processing mode of step (3) in the positioning processing method of the acoustic fluctuation positioning system in the tunnel is as follows:
(1) setting the number of microphones of the acoustic array positioning system to be M, namely, the signal receiving form of the microphones is as follows: y is Ax + e;
wherein y ∈ CMIs a frequency domain signal vector of dimension M, A ═ a (theta)1)a(θ2)…a(θN)]∈CM×NIs a matrix of the perception that is,
for the steering vector, λ is the wavelength of the signal, rmIs the coordinate of the m-th array element, θnFor grazing angle, x ∈ CNIs the signal vector to be estimated, the nth element x of whichnHas a value of thetanThe signal amplitude corresponding to the incident direction, e ∈ CMIs an M-dimensional noise vector;
(2) setting a complex random variable u ∈ CNWhen it satisfies a circularly symmetric complex Gaussian distribution with mean μ and variance Σ, its probability density function is defined as
(3) Setting the noise vector to obey a complex Gaussian distribution with a covariance matrix of
Where I is the identity matrix and the likelihood function is:
(4) setting x to follow a complex gaussian distribution with a mean value of 0, i.e. p (x | γ) ═ CN (x |0, Γ),
where Γ ═ diag { γ }, γ ═ γ1 γ2…γN]T∈CN；
(5) Setting Gamma to obey Gamma priors, i.e.
Wherein a, b ∈ R+By setting the values of a and b, the peak value of x is significantly concentrated on xi0, so most of xiAre both 0;
(6) a real random variable u is set, i.e., Γ (u | a, b) ═ Γ-1(a)baua-1e-buWherein Gamma (a) is a Gamma function and, at the same time, alpha0Subject to Gamma priors, i.e. p (alpha)0|c，d)＝Γ(α0|c,d)，
Wherein c, d ∈ R+In order to make Gamma a priori uninformative, c, d → 0, alpha0And γ is called the hyperparameter;
(7) combining the probability densities of the layered models obtained in the steps (3) to (6) to obtain a combined distribution p (x, y, alpha)0，γ)＝p(y|x，α0)p(x|γ)p(γ)p(α0)；
(8) After the prior is defined, the expression of the posterior probability density of x is obtained according to the Bayes formula
Wherein, the posterior covariance and the posterior average are respectively:
Σ＝(α0AHA+Γ-1)-1，
μ＝α0ΣAHy；
(9) hyperparameter alpha0And gamma is estimated by maximum expectation algorithm, and the signal vector x to be estimated is used as a hidden variable to enable the signal vector x to be estimated
At the maximum, wherein,
expressing an expectation of a posterior probability density relative to x;
(10) neglecting the item irrelevant to the gamma in the probability density function of the step (7) to the gamma, and obtaining the product
(11) By making a pair of gammanN is 1, …, and N is a partial derivative
Wherein
At the same time, to alpha0The method has the advantages that (1) the,
to maximize the value of the above equation, i.e.,
wherein the content of the first and second substances,
(12) in each iteration of the sparse Bayesian learning algorithm, firstly, the hyper-parameter alpha is used0And gamma, then using the parameter x as a hidden variable, and updating the hyperparameter alpha through the algorithm of the step (11) by using the parameter x as a hidden variable0And γ;
(13) processing according to the steps (1) to (12) to obtain the values of gamma elements, mu elements and sigma columns or rows, and performing xnThe anomaly x signal vector is preferably obtained a posteriori.
Preferably, the received sound wave signal is decomposed by wavelet packet to perform tree-shaped decomposition on the sound wave signal to obtain a low-frequency band signal and a high-frequency band signal, a is set as a low-frequency approximate part, a D is set as a high-frequency detail part, the number n of signal decomposition layers is set, and finally, the decomposed signals of the last layer are subjected to sequential combination to obtain an original received signal.
Preferably, the frequency index and scale of wavelet packet decomposition are setIndex and position transformation parameters are i, j and k respectively, then the wavelet packet decomposition calculation function is psii，j，k(t)＝2-j/2ψi(2-jt-k)，
Where i is 0,1, 2 …, wavelet function ψi(t) is obtained by the iterative relationship shown in the following formula
Where h (k) and g (k) are integral mirror filter coefficients associated with the scale function and wavelet function, i.e., h (k) and g (k) are a corresponding low pass filter and high pass filter, respectively.
Preferably, the wavelet packet decomposition and the down-sampling are implemented by using a filter bank, that is, the relation between the down-sampled signal and the original signal is X' (k) ═ 0.5(X (k) + X (N/2-k)),
wherein, X (k) is discrete fourier transform of X (N), N is the number of sampling points, X '(k) is corresponding to fourier transform values of the sequence obtained after down-sampling, and the sampling frequency of X' (k) is N/2, i.e. the detectable frequency band is (0, N/4), and then after one filtering, the detectable frequency band of the signal is halved.
Preferably, m is 0,1, … 2j-1 sequentially from left to right for the node obtained at the last layer in the wavelet packet decomposition tree, and if the effective frequency band of the signal is f, the passband corresponding to the frequency band number m is m (f/2)j)～(m+1)f/2j。
Preferably, in the case of time dispersion, the data is subjected to a discrete fourier transform after being subjected to framing processing, and the short-time fourier transform is formulated as,
where x [ n ] is the signal and wn is a window function, a spectral representation of the power spectral density is obtained by the magnitude squared of the short time fourier transform.
Preferably, the acoustic fluctuation positioning system further comprises a supervised learning module, wherein the supervised learning module is used for dividing the received sound wave signals by using a classifier of logistic regression, setting normal state parameters and abnormal state parameters, and embedding cost functions of regularization parameters and all samples in a training set.
Preferably, in supervised learning of machine learning of the acoustic wave localization system, the training data includes the input vector and the corresponding target value, the corresponding target value is predicted according to the new data x by using the training data, the weighted value estimation is performed by the RVM algorithm, and finally, the training learning of the effective information data is performed.
Compared with the prior art, the invention has the beneficial effects that: on the basis of DOA signal enhancement and direction finding based on SBL, the problem of weak signal detection and identification under the background of strong interference is solved through a detection, tracking and identification integrated framework, the algorithm is further mapped to a high-performance hardware processing platform, a real-time signal acquisition and processing system is constructed, and the effectiveness of the processing algorithm is verified.
Drawings
FIG. 1 is a schematic block diagram of a microphone acoustic wave receiver of the present invention;
FIG. 2 is a wavelet packet decomposition tree diagram according to the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
Referring to fig. 1, the present invention provides a technical solution: the utility model provides an acoustics undulant positioning system for in tunnel, includes that a plurality of sets up the equidistance microphone sound wave receiver in same tunnel and the treater of being connected with all microphone sound wave receivers, and complete sound wave array positioning system is constituteed to all microphone sound wave receivers, microphone sound wave receiver include microphone and processing circuit, processing circuit includes preceding discharge circuit and back discharge circuit, preceding discharge circuit includes preceding signal amplification circuit and single-ended to differential signal conversion circuit one, back discharge circuit includes that the difference changes single-ended signal conversion circuit, second order low pass filter circuit, band-pass filter circuit, fixed gain amplifier circuit and single-ended to differential signal conversion circuit two,
the circuits arranged above are all integrated functional circuits already realized in the prior art, and the microphone acoustic wave receiver in the application is obtained by combining related functional circuits in the prior art.
The positioning processing method of the acoustic fluctuation positioning system in the tunnel comprises the following steps:
(1) the sound wave signal received by the microphone is firstly amplified by a preceding stage signal amplifying circuit and then converted into a differential signal by a single-end to differential signal converting circuit;
(2) the converted differential signal is transmitted to a rear discharge circuit through a shielding cable for processing;
(3) the differential signal is converted into an analog signal through a differential-to-single-ended signal conversion circuit, the analog signal sequentially passes through a second-order low-pass filter circuit, a band-pass filter circuit and a fixed gain amplification circuit to realize filtering and signal amplification of the signal, and then the signal is converted into a differential signal through a single-ended-to-differential signal conversion circuit II and is input into a processor for processing;
(4) the processor outputs the positioning information after processing.
The processing mode of the step (3) in the positioning processing method of the acoustic fluctuation positioning system in the tunnel is as follows: .
Assuming that the array consists of M array elements, then the received signal can be written in the form,
y＝Ax+e (1)
wherein y ∈ CMIs a frequency domain signal vector of dimension M; a ═ a (θ)1)a(θ2)…a(θN)]∈CM×NIs a matrix of the perception that is,
for the steering vector, λ is the wavelength of the signal, rmIs the coordinate of the m-th array element, θnThe angle is grazing angle, 0 degree and 180 degrees are end-fire directions, and 90 degrees are broadside directions; x is formed by CNIs the signal vector to be estimated, the nth element x of whichnHas a value of thetanSignal amplitude corresponding to the incident direction; e is as large as CMIs an M-dimensional noise vector. In a real scenario, only a limited number of target signals are incident, so most of xnIs equal to 0, x is sparse in the spatial dimension. The sparse signal x can then be estimated by bayesian angles.
For a complex random variable u ∈ CNWhen it satisfies a circularly symmetric complex Gaussian distribution with mean μ and variance Σ, its probability density function is defined as
Under the bayesian framework, the solution of the signal is usually represented by a posterior probability density. Considering x as a random variable, assume that the noise vector e obeys a complex Gaussian distribution with a covariance matrix of
Where I is the identity matrix, then the likelihood function is,
in order to obtain the posterior probability density, the sparse prior information of x is required to be given. A widely used sparse prior is Laplace distribution, in a Relevance Vector Machine (RVM), the prior of x is assumed to be a hierarchical prior, whose properties are similar to Laplace prior, but the computation is simpler. Assuming that x obeys a complex gaussian distribution with a mean value of 0,
p(x|γ)＝CN(x|0，Γ) (4)
wherein the covariance Γ ═ diag { γ }, γ ═ γ }, and γ ═ γ }, respectively1 γ2…γN]T∈CN. Further, assuming that γ obeys Gamma prior,
wherein a, b ∈ R+By setting the appropriate values of a and b, the peak value of x can be significantly concentrated on xi0, so most of xiAre all 0 and therefore have a sparse prior. For a real random variable u, Γ (u | a, b) ═ Γ-1(a)baua- 1e-buΓ (a) is a Gamma function. Likewise, α0Also obeying to the Gamma prior,
p(α0|c，d)＝Γ(α0|c,d) (6)
wherein c, d ∈ R+To make Gamma a priori uninformative, c, d → 0. Alpha is alpha0And γ is called the hyperparameter.
The individual probability densities of the hierarchical models are combined, so that the combined distribution is,
p(x,y，α0，γ)＝p(y|x，α0)p(x|γ)p(γ)p(α0) (7)
the probability densities on the right are defined by (3) - (6), respectively.
After defining the prior, the expression of the posterior probability density of x can be obtained according to the Bayesian formula as,
wherein the posterior covariance and the posterior mean are respectively
Σ＝(α0AHA+Г-1)-1 (9)
μ＝α0ΣAHy (10)
Hyperparameter alpha0The sum γ can be estimated by using the maximum Expectation algorithm (EM), and the signal vector x to be estimated is regarded as a hidden variable, so that
At the maximum, wherein,
representing the expectation of the posterior probability density relative to x.
For gamma, ignoring the gamma-independent term in the probability density function (7), there is
By making a pair of gammanN is 1, …, and N is a partial derivative, including,
wherein
Likewise, for α0There is, in some cases,
the method maximizes the formula of the formula, including,
wherein the content of the first and second substances,
in each iteration of the sparse Bayesian learning algorithm, firstly, the hyper-parameter alpha is used0And gamma to estimate the posterior mean (10) and covariance (9) of x, then the parameter x is considered as a hidden variable, and the hyperparameter alpha is updated by (14) and (12)0And gamma. During the iteration, almost all the elements of γ are close to 0; correspondingly, the element of μ and the column (or row) of Σ are both close to 0. Thus, these xnAll of the posteriori of (c) are centered on 0, and it can be considered that these x arenAt 0, the signal vector x is sparse.
Wavelet packet decomposition is provided on the basis of wavelet transformation, and the wavelet packet transformation further divides the low-frequency and high-frequency bands of signals into sub-bands, so that the frequency resolution of the frequency bands is improved. When an abnormality occurs in the tunnel, the frequency spectrum of the collected sound signal changes, the signal is decomposed by using the wavelet packet, so that the details of the frequency spectrum are amplified to observe a frequency band with a fault, and the decomposition tree structure of the signal S is shown in FIG. 2 (taking three-layer wavelet packet decomposition as an example):
where a denotes the approximate part of the low frequency, D denotes the detailed part of the high frequency, and the number denotes the number of decomposition layers from which the original signal can be recovered.
S＝AAA3+AAD3+ADA3+ADD3+DAA3+DAD3+DDA3+DDD3 (30)
A wavelet packet function psii，j，kAnd (t) has three indexes, i, j and k respectively represent a frequency index, a scale index and a position transformation parameter of the wavelet packet function. The concrete expression is as follows
ψi，j，k(t)＝2-j/2ψi(2-jt-k) (31)
Where i is 0,1, 2 …. Wavelet function psii(t) can be obtained by the iterative relationship shown below
The discrete filter systems h (k) and g (k) are integral mirror filter coefficients associated with scale functions and wavelet functions. h (k) and g (k) are the corresponding low pass filter and high pass filter.
The frequency band division is not uniform frequency band division, the key of utilizing a filter bank to realize wavelet packet decomposition is the realization of down sampling, after the DFT analysis signal is subjected to down sampling, the relationship between the sampled signal and the original signal is existed,
X′(k)＝0.5(X(k)+X(N/2-k)) (33)
where X (k) is the discrete fourier transform of X (N), N is the number of sampling points, X '(k) corresponds to the fourier transform value of the sequence obtained after down-sampling, and the sampling frequency of X' (k) is N/2, so the detectable frequency band is (0, N/4). Therefore, after one filtering, the frequency band where the signal can be detected is halved.
The nodes obtained from the last layer of the wavelet packet decomposition tree graph are sequentially coded into m from left to right, wherein m is 0,1 and … 2j-1(j is the number of decomposition levels), called the node number. If the effective frequency band of the signal is f, the pass band corresponding to the frequency band number m is m (f/2)j)～(m+1)f/2j。
The wavelet packet decomposition of the signal is actually to pass the signal through a set of filter combinations (low-pass or high-pass) to the tree nodes at the bottom. Setting 0 to represent low-pass filtering, and setting 1 to represent high-pass filtering, each filter combination can be represented by a binary number, and the binary number is called as an actual path to carry out binary conversion on the frequency band serial number to obtain a frequency band path, namely a binary number; carrying out XOR operation on each digit of the binary number and the left digit of the binary number from left to right in sequence to obtain a node path, namely a new binary number; and performing decimal conversion on the new binary number to obtain the node serial number.
In the case of time dispersion, the data may be subjected to discrete fourier transform after being subjected to framing processing. There is overlap from frame to mitigate the artifacts of the partition boundaries. The short-time fourier transform is formulated as,
where x [ n ] is the signal and w [ n ] is the window function. A spectral representation of the power spectral density can be obtained from the magnitude squared of the short-time fourier transform.
The acoustic fluctuation positioning system further comprises a supervised learning module, wherein the processing mode of the supervised learning module is that a received sound wave signal is subjected to feature vector division through a two-classifier of logistic regression, normal state parameters and abnormal state parameters are set, and cost functions of regularization parameters are embedded and all samples in a training set are added.
A logistic regression-based two-classifier is a supervised learning method for detecting abnormal signals. Due to the nature of the two classifiers, the severity of the damage to the system can only be reflected by the threshold. Human intervention includes visual-based inspection, maintenance to replace blades, or other measures to avoid damage to the wind turbine. Logistic regression classifies features into corresponding categories by optimizing the weight parameter matrix theta. Simplifying m training samples of the received acoustic signal time domain into a matrix formed by n-dimensional characteristic vectors,
the feature matrix x, weighted by the weight parameter θ, defines the hypothesis function for predicting the known output class of the training set. The output y of the two classifiers is 0 to indicate that the blade is not damaged, y is 1 to indicate that the blade is damaged, the sigmoid function used for the assumption is,
for this assumption, θTWhen x is greater than or equal to 0, hθ(x) Not less than 0.5. In the binary classification, the threshold value is used for outputting a predicted value 1 corresponding to the class with damaged blades when the generated value is supposed to exceed the threshold value; on the contrary, when thetaTAnd when x is less than or equal to 0, outputting a predicted value 0 of the undamaged class of the corresponding blade. The computation cost of the output weight parameter θ based on the initial training set is minimal for substituting into the hypothesis to predict future output values. At each moment, when it is assumed thatWhen the output is wrong, the penalty factor is aggravated by the cost function; otherwise, basically no change is made. The cost function is a function of the sum of,
when the cost is calculated, the phenomenon that the algorithm cannot predict the overfitting phenomenon of future samples through learning the relation between input and output of a training set needs to be avoided. The overfitting can be solved by reducing the number of features or using regularization. Regularization adds a term λ to the cost function, keeping θ small to fit each data point. The cost function of the embedded regularization parameter may sum all samples in the training set, if
The gradient descent method is a type of algorithm for minimizing a cost function with respect to a parameter theta, and after the optimization parameter theta is determined, the hypothesis function can define the boundary between classes. As the number of features and output categories increase, the inter-category boundaries may become extremely complex, and at the same time, the value ranges of the features may have significant differences, resulting in a significant increase in the time for algorithm convergence.
In supervised learning of acoustic fluctuation positioning system machine learning, training data comprise input vectors and corresponding target values, the corresponding target values are predicted according to new data x by using the training data, weight value estimation is carried out through an RVM algorithm, and finally training learning of effective information data is carried out.
In a supervised learning problem for machine learning, the training data includes input vectors
And corresponding target value
The goal of machine learning is to use the training data, based on new numbersAnd predicting the corresponding target value t according to x. In the regression problem, t is a continuous variable; in the classification problem, t is a discrete value.
The prediction y (x, w) of the model can be expressed as a basis function φm(x) The linear combination of (a) and (b),
wherein { wmReferred to as weights, the RVM algorithm is used to estimate the weight values.
RVM regression model
In the regression problem, an input-target dataset is given
Assuming that the target value is the prediction of the model plus noise,
tn＝y(xn，w)+∈n (42)
wherein enIs a noise term, which is assumed to be a mean of 0 and a variance of σ2A gaussian distribution of (a). According to the data tnAnd thus the likelihood function can be written as,
the weights based on likelihood function estimation and noise estimation may be over-fitted due to the large number of parameters in the training data. To avoid overfitting, additional forcing conditions may be added to the parameters. The weight apriori distribution is typically chosen to be a zero-mean gaussian distribution,
where α is a hyperparameter. And each weight is independently distributed parameters, so that the complexity of calculation can be reduced.
Variance of noise σ2Also as a hyper-parameter, distribution of hyper-parametersThe Gamma distribution is satisfied,
p(β)＝Γ(β|c，d) (46)
wherein β ═ σ2。
rVM Classification model
The RVM classification model has the same framework as the regression, only the likelihood functions differ slightly. For the class two classification problem, the goal is to predict the posterior probability from the input x, and the sigmoid function σ (y) is used for y (x) as 1/(1+ e) according to the statistical principle-y) Normalized, p (t | x) is a bernoulli distribution, so the likelihood function is,
wherein the value of the objective function tnE {0, 1+, the equation here is free of noise terms.
The iterative approach of RVM is substantially identical to that of SBL.
In addition, for other realizable machine learning algorithms, Support Vector Machines (SVMs) are a class of supervised, classification-based machine learning algorithms, and SVM methods are robust to a greater number of variables and small samples. By controlling the global complexity of the model, the method can process a more complex classification model while better generalizing new samples. Unlike logistic regression, the SVM method is able to generate non-linear boundaries: and after the feature space is converted, a linear boundary is divided in a new space.
The SVM method determines the optimal hyperplane boundary by finding the hyperplane between the maximized output classes. For a super-plane, there is,
wTx＝0 (39)
w and x are vectors, and the expression method of the formula ensures that w is always orthogonal to the hyperplane. In the two-dimensional case, the hyperplane is a straight line that distinguishes the two classes. Similar to logistic regression, the optimal SVM parameters are found by minimizing a cost function that is used to compare the performance of the trained algorithm with the existing training set outputs. Since the SVM is a convex function, any local optimum is a global optimum. SVMs differ from logistic regression in constructing a cost function, but all aim to find the optimal parameters that minimize the corresponding cost function.
Combining the cases of y being 0 and y being 1, the global cost function of the SVM is:
the second term in the above equation is absent λ, introducing a new variable C. The role of C is to avoid overfitting due to the presence of outliers. If the value of C is large, the wild value has a large influence on the determination of the boundary; on the contrary, the SVM algorithm is likely to ignore the existence of the outlier.
Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that changes, modifications, substitutions and alterations can be made in these embodiments without departing from the principles and spirit of the invention, the scope of which is defined in the appended claims and their equivalents.
Claims (9)
1. An acoustic wave localization system for use in a tunnel, comprising: including the equidistance microphone sound wave receiver of a plurality of setting in same tunnel and the treater of being connected with all microphone sound wave receivers, complete sound wave array positioning system is constituteed to all microphone sound wave receivers, microphone sound wave receiver include microphone and processing circuit, processing circuit includes preceding discharge circuit and back discharge circuit, preceding discharge circuit includes preceding stage signal amplification circuit and single-ended changes differential signal conversion circuit one, back discharge circuit includes difference changes single-ended signal conversion circuit, second order low pass filter circuit, band-pass filter circuit, fixed gain amplifier circuit and single-ended changes differential signal conversion circuit two,
the positioning processing method of the acoustic fluctuation positioning system in the tunnel comprises the following steps:
(1) the sound wave signal received by the microphone is firstly amplified by a preceding stage signal amplifying circuit and then converted into a differential signal by a single-end to differential signal converting circuit;
(2) the converted differential signal is transmitted to a rear discharge circuit through a shielding cable for processing;
(3) the differential signal is converted into an analog signal through a differential-to-single-ended signal conversion circuit, the analog signal sequentially passes through a second-order low-pass filter circuit, a band-pass filter circuit and a fixed gain amplification circuit to realize filtering and signal amplification of the signal, and then the signal is converted into a differential signal through a single-ended-to-differential signal conversion circuit II and is input into a processor for processing;
(4) the processor outputs the positioning information after processing.
2. An acoustic wave localization system for use in a tunnel according to claim 1, wherein: the processing mode of the step (3) in the positioning processing method of the acoustic fluctuation positioning system in the tunnel is as follows:
(1) setting the number of microphones of the acoustic array positioning system to be M, namely, the signal receiving form of the microphones is as follows: y is Ax + e;
wherein y ∈ CMIs a frequency domain signal vector of dimension M, A ═ a (theta)1)a(θ2)…a(θN)]∈CM×NIs a matrix of the perception that is,
for the steering vector, λ is the wavelength of the signal, rmIs the coordinate of the m-th array element, θnFor grazing angle, x ∈ CNIs the signal vector to be estimated, the nth element x of whichnHas a value of thetanThe signal amplitude corresponding to the incident direction, e ∈ CMIs an M-dimensional noise vector;
(2) setting a complex random variable u ∈ CNWhen it satisfies a circularly symmetric complex Gaussian distribution with a mean value of μ and a variance of Σ, its probability density function is defined as
(3) Setting the noise vector to obey a complex Gaussian distribution with a covariance matrix of
Where I is the identity matrix and the likelihood function is:
(4) setting x to follow a complex gaussian distribution with a mean value of 0, i.e. p (x | γ) ═ CN (x |0, Γ),
where Γ ═ diag { γ }, γ ═ γ1 γ2...γN]T∈CN；
(5) Setting Gamma to obey Gamma priors, i.e.
Wherein a, b ∈ R+By setting the values of a and b, the peak value of x is significantly concentrated on xi0, so most of xiAre both 0;
(6) a real random variable u is set, i.e., Γ (u | a, b) ═ Γ-1(a)baua-1e-buWherein Gamma (a) is a Gamma function and, at the same time, alpha0Subject to Gamma priors, i.e. p (alpha)0|c，d)＝Γ(α0|c，d)，
Wherein c, d ∈ R+In order to make Gamma a priori uninformative, c, d → 0, alpha0And γ is called the hyperparameter;
(7) combining the probability densities of the layered models obtained in the steps (3) to (6) to obtain a combined distribution p (x, y, alpha)0，γ)＝p(y|x，α0)p(x|γ)p(γ)p(α0)；
(8) After the prior is defined, the expression of the posterior probability density of x is obtained according to the Bayes formula
Wherein, the posterior covariance and the posterior average are respectively:
∑＝(α0AHA+Γ-1)-1，
μ＝α0∑AHy；
(9) hyperparameter alpha0And gamma is estimated by maximum expectation algorithm, and the signal vector x to be estimated is used as a hidden variable to enable the signal vector x to be estimated
At the maximum, wherein,
expressing an expectation of a posterior probability density relative to x;
(10) neglecting the item irrelevant to the gamma in the probability density function of the step (7) to the gamma, and obtaining the product
(11) By making a pair of gammanN is 1, …, and N is a partial derivative
Wherein
At the same time, to alpha0The method has the advantages that (1) the,
to maximize the value of the above equation, i.e.,
wherein the content of the first and second substances,
(12) in each iteration of the sparse Bayesian learning algorithm, firstly, the hyper-parameter alpha is used0And gamma, then using the parameter x as a hidden variable, and updating the hyperparameter alpha through the algorithm of the step (11) by using the parameter x as a hidden variable0And γ;
(13) processing according to the steps (1) to (12) to obtain the values of gamma elements, mu elements and sigma columns or rows, and performing xnThe anomaly x signal vector is preferably obtained a posteriori.
3. An acoustic wave localization system for use in a tunnel according to claim 1 or 2, wherein: performing tree decomposition on the received sound wave signals through wavelet packet decomposition to output the sound wave signals to obtain low-frequency-band signals and high-frequency-band signals, setting A as a low-frequency approximate part and D as a high-frequency detail part, setting the number n of signal decomposition layers, and finally performing serialization combination on the decomposed signals of the last layer to obtain original received signals.
4. An acoustic wave localization system for use in a tunnel according to claim 3, wherein: setting the frequency index, scale index and position transformation parameter of wavelet packet decomposition as i, j and k respectively, then calculating function of wavelet packet decomposition is psii，j，k(t)＝2-j/2ψi(2-jt-k)，
Where i is 0,1, 2i(t) is obtained by the iterative relationship shown in the following formula
Where h (k) and g (k) are integral mirror filter coefficients associated with the scale function and wavelet function, i.e., h (k) and g (k) are a corresponding low pass filter and high pass filter, respectively.
5. An acoustic wave localization system for use in a tunnel according to claim 4, wherein: wavelet packet decomposition is realized by using a filter bank and downsampled, namely the relation between a downsampled signal and an original signal is X' (k) ═ 0.5(X (k) + X (N/2-k)),
wherein, X (k) is discrete fourier transform of X (N), N is the number of sampling points, X '(k) is corresponding to fourier transform values of the sequence obtained after down-sampling, and the sampling frequency of X' (k) is N/2, i.e. the detectable frequency band is (0, N/4), and then after one filtering, the detectable frequency band of the signal is halved.
6. An acoustic wave localization system for use in a tunnel according to claim 5, wherein: sequentially coding m-0, 1, 2 from left to right for nodes obtained from the last layer in the wavelet packet decomposition tree formj-1, setting the effective frequency band of the signal as f, and the passband corresponding to the frequency band number m is m (f/2)j)～(m+1)f/2j。
7. An acoustic wave localization system for use in a tunnel according to claim 6, wherein: in the case of time dispersion, the data is subjected to frame division and then subjected to discrete fourier transform, and the short-time fourier transform has the following formula,
where x [ n ] is the signal and wn is a window function, a spectral representation of the power spectral density is obtained by the magnitude squared of the short time fourier transform.
8. An acoustic wave localization system for use in a tunnel according to claim 7, wherein: the acoustic fluctuation positioning system further comprises a supervised learning module, wherein the processing mode of the supervised learning module is that a received sound wave signal is subjected to feature vector division through a two-classifier of logistic regression, normal state parameters and abnormal state parameters are set, and cost functions of regularization parameters are embedded and all samples in a training set are added.
9. An acoustic wave localization system for use in a tunnel according to claim 8, wherein: in supervised learning of acoustic fluctuation positioning system machine learning, training data comprise input vectors and corresponding target values, the corresponding target values are predicted according to new data x by using the training data, weight value estimation is carried out through an RVM algorithm, and finally training learning of effective information data is carried out.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110260239.2A CN113050039B (en) | 2021-03-10 | 2021-03-10 | Acoustic fluctuation positioning system used in tunnel |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110260239.2A CN113050039B (en) | 2021-03-10 | 2021-03-10 | Acoustic fluctuation positioning system used in tunnel |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113050039A true CN113050039A (en) | 2021-06-29 |
| CN113050039B CN113050039B (en) | 2023-03-07 |
Family
ID=76510905
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110260239.2A Active CN113050039B (en) | 2021-03-10 | 2021-03-10 | Acoustic fluctuation positioning system used in tunnel |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113050039B (en) |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JPS62254083A (en) * | 1986-04-26 | 1987-11-05 | Teru Hayashi | Sound source prober |
| CN103813237A (en) * | 2014-02-18 | 2014-05-21 | 厦门亿联网络技术股份有限公司 | Extensible pick-up with mute function and implementing method thereof |
| CN104714925A (en) * | 2015-02-02 | 2015-06-17 | 北京工业大学 | Gear drive noise analysis method based on fractional order Fourier transform and support vector machine |
| CN107086036A (en) * | 2017-04-19 | 2017-08-22 | 杭州派尼澳电子科技有限公司 | A kind of freeway tunnel method for safety monitoring |
| CN107632288A (en) * | 2017-10-27 | 2018-01-26 | 无锡七百二十度科技有限公司 | A kind of sonic location system |
| CN110045333A (en) * | 2019-04-12 | 2019-07-23 | 上海工程技术大学 | A kind of sound source three-dimensional positioning method based on Kalman filtering |
| WO2020042708A1 (en) * | 2018-08-31 | 2020-03-05 | 大象声科（深圳）科技有限公司 | Time-frequency masking and deep neural network-based sound source direction estimation method |
| CN111540347A (en) * | 2020-05-12 | 2020-08-14 | 山东科华电力技术有限公司 | Cable tunnel monitoring method and system based on audio |
-
2021
- 2021-03-10 CN CN202110260239.2A patent/CN113050039B/en active Active
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JPS62254083A (en) * | 1986-04-26 | 1987-11-05 | Teru Hayashi | Sound source prober |
| CN103813237A (en) * | 2014-02-18 | 2014-05-21 | 厦门亿联网络技术股份有限公司 | Extensible pick-up with mute function and implementing method thereof |
| CN104714925A (en) * | 2015-02-02 | 2015-06-17 | 北京工业大学 | Gear drive noise analysis method based on fractional order Fourier transform and support vector machine |
| CN107086036A (en) * | 2017-04-19 | 2017-08-22 | 杭州派尼澳电子科技有限公司 | A kind of freeway tunnel method for safety monitoring |
| CN107632288A (en) * | 2017-10-27 | 2018-01-26 | 无锡七百二十度科技有限公司 | A kind of sonic location system |
| WO2020042708A1 (en) * | 2018-08-31 | 2020-03-05 | 大象声科（深圳）科技有限公司 | Time-frequency masking and deep neural network-based sound source direction estimation method |
| CN110045333A (en) * | 2019-04-12 | 2019-07-23 | 上海工程技术大学 | A kind of sound source three-dimensional positioning method based on Kalman filtering |
| CN111540347A (en) * | 2020-05-12 | 2020-08-14 | 山东科华电力技术有限公司 | Cable tunnel monitoring method and system based on audio |
Non-Patent Citations (2)
| Title |
|---|
| 卢锋 等: "声光检测在人员搜救中的应用研究", 《西安文理学院学报(自然科学版)》 * |
| 张攀攀: "地震波法隧道地质超前预报关键技术与应用", 《中国优秀博硕士学位论文全文数据库(硕士)·基础科学辑》 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113050039B (en) | 2023-03-07 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Ruff et al. | Deep semi-supervised anomaly detection | |
| CN111971743B (en) | Systems, methods, and computer readable media for improved real-time audio processing | |
| CN110109060B (en) | Radar radiation source signal sorting method and system based on deep learning network | |
| Sugiartawan et al. | Prediction by a hybrid of wavelet transform and long-short-term-memory neural network | |
| Jia et al. | Space object classification using deep neural networks | |
| Ma et al. | Multimodal deep learning for solar radio burst classification | |
| CN114341878A (en) | Information processing device and electronic apparatus including the same | |
| Cai et al. | An early warning model for vegetable pests based on multidimensional data | |
| CN112684012A (en) | Equipment key force-bearing structural part fault diagnosis method based on multi-parameter information fusion | |
| CN114964778A (en) | Bearing fault diagnosis method based on wavelet time-frequency graph and deep learning | |
| Ding et al. | Mine microseismic time series data integrated classification based on improved wavelet decomposition and ELM | |
| CN113050039B (en) | Acoustic fluctuation positioning system used in tunnel | |
| Yao et al. | Identification of underwater targets based on sparse representation | |
| Wang et al. | A hybrid cloud intrusion detection method based on SDAE and SVM | |
| Li et al. | A robust fault diagnosis method for rolling bearings based on deep convolutional neural network | |
| Ren et al. | Semi-supervised classification of PolSAR data with multi-scale weighted graph convolutional network | |
| Stock et al. | Trainable wavelet neural network for non-stationary signals | |
| KR102611399B1 (en) | Abnormal detection device using 1-demension autoencoder | |
| Aslinezhad et al. | A noise-aware deep learning model for automatic modulation recognition in radar signals | |
| Cui et al. | Information theoretical similarity measure for change detection | |
| Al Mudhafar et al. | Image Noise Detection and Classification Based on Combination of Deep Wavelet and Machine Learning: non | |
| Pan et al. | Fast approximate spectral normalization for robust deep neural networks | |
| Chen et al. | Wideband DoA estimation based on joint optimisation of array and spatial sparsity | |
| KR102574605B1 (en) | Method, apparatus and computer program for classifying audio data and measuring noise level using video data and audio data | |
| Jiang et al. | Spectral analysis and recognition using multi-scale features and neural networks |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |