A cochlear dead zone edge detection method based on PTC
By constructing the subject response probability model and introducing attention mechanism, the PTC detection process is optimized, and the problem of excessive time-consuming cochlear dead zone detection in the prior art is solved, and fast and accurate cochlear dead zone edge detection is achieved.
Patent Information
- Application Number
- CN202211341410.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-10-31
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2042-10-31
AI Technical Summary
The method of detecting cochlear dead zones based on PTC in the prior art takes a long time, resulting in limited application in actual production work.
Using the PTC-based cochlear dead-zone edge detection method, the subject's response probability model is constructed as a V-shaped curve, the attention mechanism and active learning are introduced, and the Bayesian conditional probability formula and mutual information calculation are used to optimize the test process and reduce the number of tests. and time.
It significantly shortens the detection time, improves the accuracy and efficiency of the detection results, and can complete the same test within 6 minutes, saving more than 30% of the time, while ensuring the accuracy of the results.
Smart Images

Figure CN115644860B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of cochlear dead zone detection, and in particular to a cochlear dead zone edge detection method based on PTC. Background Art
[0002] The cochlea is an organ in the human body that converts the mechanical energy of external sound waves into nerve impulses. The round window of the cochlea receives the sound vibrations from the ossicles and causes the basilar membrane to vibrate. Sounds of different frequencies cause the basilar membrane to vibrate in different parts, and the vibration of the basilar membrane drives the hair cells to vibrate and produce hearing. Auditory hair cells are divided into outer hair cells and inner hair cells. The outer hair cells are like amplifiers, which are mainly responsible for amplifying the vibrations caused by sound on the basilar membrane. Their damage will reduce the gain of the cochlear amplifier; while the inner hair cells are like transducers, which mainly convert the vibrations of the basilar membrane into the excitement of the auditory nerve. Their damage will reduce the number of action potentials of the auditory nerve. Usually, sensorineural hearing loss is mainly caused by damage to the outer hair cells. At this time, amplifying the input through hearing aids can enable the patient to hear the sound again. However, in some cases, the inner hair cells or neurons of the auditory nerve in a certain frequency area of the cochlea are severely damaged, resulting in the inability of the patient to hear in this part of the frequency domain no matter how the sound is amplified. This part of the frequency area is called the cochlear dead zone, or simply the dead zone.
[0003] The cochlear dead zone is an area in the cochlea where the inner hair cells or auditory neurons have lost or almost lost their functions. This area has difficulty converting sound signals into electrical signals no matter how strong the stimulation is. Approximately 36% to 43% of hearing aid wearers or cochlear implant recipients have cochlear dead zones to varying degrees, which can lead to inaccurate hearing aid fitting results or accurate cochlear implantation. The detection of cochlear dead zones is of great significance in both hearing aid fitting and cochlear implantation. Most existing methods for detecting cochlear dead zones are based on psychophysical tuning curves (PTC). In the PTC test, the subject's task is to determine whether a pure tone signal exists in the presence of narrow-band masking noise. During the entire test, the frequency of the pure tone signal (f sig ) and strength (L sig ) is fixed, and each test will use different center frequencies and intensities (f mask , L mask The PTC test result is the masking sound intensity L required for the subject to just not be able to hear the pure tone. mask Values with respect to masking sound frequency f mask For subjects with normal hearing and subjects with hearing loss but no dead zone, the minimum value of PTC is at f sigWhen the signal frequency is within the dead zone, the subject actually perceives the pure tone through the inner hair cells outside the dead zone, or through the edge frequency of the dead zone (f e ) to feel the pure tone, so when f mask ≈f e The PTC method in the prior art can obtain a more accurate estimate of the dead zone edge frequency, but it also has many shortcomings: the PTC test in the prior art needs to measure the masking threshold at multiple frequency points, which usually takes a lot of time (usually 30 minutes to an hour), resulting in limited application scenarios in actual production work. Summary of the invention
[0004] In order to solve the problem that the method of detecting the cochlear dead zone based on PTC in the prior art is time-consuming and leads to limited application in actual production work, the present invention provides a cochlear dead zone edge detection method based on PTC, which not only has accurate detection results but also can greatly shorten the detection time.
[0005] The technical solution of the present invention is as follows: a method for detecting the edge of a cochlear dead zone based on PTC, characterized in that it comprises the following steps:
[0006] S1: Construct a subject response probability model based on the PTC principle;
[0007] The subject response probability model is a V-shaped curve based on the parameter θ = [f e ,L e ,k 1 ,k 2 ] represents the model parameters of the subject response probability model;
[0008] Among them, f e is the frequency at the tip of the V-shaped curve, L e is the sound intensity at the tip of the V-shaped curve; (f e ,L e ) is the vertex coordinate of the subject response probability model curve; k 1 represents the forward slope of the subject response probability model curve, k 2 represents the backward slope of the subject response probability model curve;
[0009] S2: Based on the subject response probability model, the probability P(y=1|x,θ) that the subject hears a pure tone is expressed as:
[0010]
[0011] Where Φ(x) is the cumulative distribution function of the standard normal distribution; f maskis the masking sound frequency;
[0012] x=[f mask ,L mask ] represents the stimulus parameter, f mask is the frequency of the masking noise, L mask is the intensity of the masking noise; f e is the frequency at the tip of the V-shaped curve, L e is the sound intensity at the tip of the V-shaped curve; y = 1 means that the subject can hear the pure tone under the condition of stimulation parameter x; y = 0 means that the subject cannot hear the pure tone under the condition of stimulation parameter x; σ is a hyperparameter;
[0013] S3: Assume that there are already samples Then we have:
[0014]
[0015] in, Indicates that under a certain set of parameters θ, the sample is obtained by sampling The probability of , N is the number of sampling points, i represents the i-th iteration;
[0016] S4: Based on the Bayesian conditional probability formula, we have:
[0017]
[0018] in, Indicates that in the existing observation samples p(θ) represents the prior distribution of the parameter θ.
[0019] S5: Introduce the attention mechanism into the subject response probability model, the i-th frequency f i The calculation method of the attention weight A at is:
[0020]
[0021] in, Indicates that the mean is lnf e , a normal distribution with a variance of 0.01 is lnf i probability;
[0022] Represents the current observation sample; Indicates the current f e The marginal distribution of σ 2 represents variance;
[0023] S6: Stimulus parameter x for the next test of the model next The calculation method is:
[0024]
[0025] Where I is the identity matrix; is the mutual information of the sampling point x to the parameter θ under the current sampling result;
[0026] S7: Before each test, set the pure tone signal parameter f sig and L sig and narrowband noise test parameter f e ,L e ,k 1 ,k 2 ,f mask ,l mask Implement initialization testing after testing the scope;
[0027] The initialization test includes:
[0028] Take 10dB as L mask The step size is, when testing the pure tone signal frequency f sig , pure tone signal frequency and f e The mean of the minimum values in the test range Pure tone signal frequency and f e The mean of the maximum values in the test range The test responses at three frequencies. Each time a frequency is tested, the measurement of this frequency point is stopped when the test response changes once. This way, the initial series of sampling results can be obtained.
[0029] The sampling results obtained by initialization Send to middle;
[0030] S8: Pass Calculate the y corresponding to x; calculate the current posterior parameter distribution After that, we calculate
[0031] S9: Calculate the next stimulation parameter x using the next stimulation parameter formula next ;
[0032]
[0033] S10: Set the stimulus parameter x next As input x, steps S8 to S9 are executed repeatedly until the following stop condition is met, and the test stops;
[0034] The stopping conditions are:
[0035] Three consecutive tests is greater than the preset stopping confidence threshold, and the three f obtained in these three tests e The difference between the values is less than f e 1 / 20 of the octave;
[0036] S11: construct the basic diagram of PTC curve;
[0037] The abscissa of the PTC curve basic diagram is frequency, and the ordinate is intensity;
[0038] During the test, when y=0, the corresponding x=[f mask ,L mask ] are marked on the PTC curve base diagram, and all points are fitted to obtain the subject response probability model curve as the test result.
[0039] It is further characterized by:
[0040] In step S6, the mutual information The calculation method is:
[0041]
[0042] in, is Shannon entropy, which is used to measure the uncertainty of a random variable;
[0043] is the mutual information of the sampling point x to the parameter θ under the current sampling result; it is specifically calculated using the following formula:
[0044]
[0045]
[0046] The stopping confidence threshold is set to 0.5;
[0047] Step S8 can also be implemented by the following steps:
[0048] a1: Get the f in this test respectively e 、f mask , L e , L mask , k 1 and k 2 The scope of the test;
[0049] a2: In the test range, the intensity parameter L e and L mask According to the equal interval, the frequency parameter f e and f mask According to the exponential equal intervals, 25 data points were taken respectively;
[0050] Use f e 、f mask , L e , L mask All data points are used to construct a 6-dimensional parameter matrix P;
[0051] In the parameter matrix P, each group of elements [f e ,L e ,k 1 ,k 2 ,f mask ,L mask ] all conform to the calculation relationship of P(y=1|x,θ);
[0052] a3: When step S8 needs to be executed, the stimulation parameter x=[f mask ,L mask ] Query in the parameter matrix P and get the corresponding
[0053] The test scope includes:
[0054] f e and f mask The test range is set at the pure tone signal frequency f sig A location one octave nearby;
[0055] L e and L mask The test range is set to pure tone signal strength L sig Below 15dB to 90dB;
[0056] k 1 and k 2 The test range is set to 10 -3 Up to 10 -0.5 .
[0057] The present invention provides a method for detecting the edge of a cochlear dead zone based on PTC, which constructs a subject response probability model, the subject response probability model is a V-shaped curve, and the test result can be obtained without testing all the frequencies of narrow-band masking noise during the test process, thereby improving the test speed; the present application introduces an attention mechanism into the subject response probability model, and constructs an active-PTC detection method based on active learning. The active-PTC detection method based on active learning can quickly and accurately find the stimulation parameter x of the model for the next test next , which not only greatly speeds up the test speed, but also solves the problem of the PTC method taking too long; at the same time, because x next The calculation method is based on mutual information, where the mutual information reflects the uncertainty of the sampling point x to the parameter θ through Shannon entropy. The x calculated in this application isnext The information amount of θ is maximized, ensuring that the amount of information obtained in each sampling is as much as possible, further accelerating the PTC test, and improving the accuracy of the test results. In the specific implementation, based on the subject response probability model P (y = 1 | x, θ), the parameter matrix P is pre-constructed. For each stimulus parameter x, there is no need to calculate, but to obtain the corresponding Further improved the test speed. BRIEF DESCRIPTION OF THE DRAWINGS
[0058] Figure 1 Schematic diagram of the probability model used in active-PTC in this application;
[0059] Figure 2 It is a schematic diagram of the cumulative distribution function of the standard normal distribution with a mean of 60dB;
[0060] Figure 3 Example of test results saved for active-PTC;
[0061] Figure 4 A schematic diagram showing the comparison of the test time in the experimental results of the technical solution in the present application and the prior art solution;
[0062] Figure 5 It is a schematic diagram comparing the tip frequency of the PTC curve in the experimental results of the technical solution in this application and the prior art solution. DETAILED DESCRIPTION
[0063] The present invention comprises a PTC-based cochlear dead zone edge detection method, which comprises the following steps.
[0064] S1: Construct a subject response probability model based on the PTC principle;
[0065] The subject response probability model is a V-shaped curve based on the parameter θ = [f e ,L e ,k 1 ,k 2 ] represents the model parameters of the subject response probability model;
[0066] Among them, f e is the frequency at the tip of the V-shaped curve (corresponding to the dead zone edge frequency), L e is the sound intensity at the tip of the V-shaped curve (corresponding to f e The minimum masking strength required at e ,L e ) is the vertex coordinate of the subject response probability model curve; k 1 represents the forward slope of the subject response probability model curve, k 2 Represents the backward slope of the subject response probability model curve.
[0067] S2: Based on the subject response probability model, if the masking parameter is x and the model parameter is θ, the probability that the subject hears the pure tone is P(y=1|x,θ), which is expressed as:
[0068]
[0069] Where Φ(x) is the cumulative distribution function of the standard normal distribution; f mask is the masking sound frequency;
[0070] x=[f mask ,L mask ] represents the stimulus parameter, f mask is the frequency of the masking noise, L mask is the intensity of the masking noise; f e is the frequency at the tip of the V-shaped curve, L e is the sound intensity at the tip of the V-shaped curve; y=1 indicates that the subject can hear the pure tone under the condition of stimulation parameter x; y=0 indicates that the subject cannot hear the pure tone under the condition of stimulation parameter x; σ is a hyperparameter. A larger σ enables the model to tolerate a larger error, but at the same time the convergence will also be slower. In this embodiment, setting σ between 5 and 10 can achieve better test results.
[0071] is the cumulative distribution function of the standard normal distribution, and Φ(x) converts the distance between the test point and the hearing threshold into a probability, such as Figure 1 As shown, the horizontal axis is frequency, the vertical axis is intensity, the PTC curve in the prior art (marked as PTC result in the figure) is the curve shown by the solid line, and in the technical solution of the present application, the curve corresponding to the constructed subject response probability model is the curve shown by the dotted line (marked as active-PTC result in the figure), the active-PTCresult curve and its parameter θ=[f e ,L e ,k 1 ,k 2 ] At 1250Hz, the value corresponding to the yellow line is 60dB. The relationship between the masking sound intensity and the probability of the subject hearing the pure tone at 1250Hz is as follows: Figure 2 As shown, it is the cumulative distribution function of the standard normal distribution with a mean of 60dB.
[0072] S3: Assume that there are already samples Then we have:
[0073]
[0074] in, Indicates that under a certain set of parameters θ, the sample is obtained by sampling The probability of , N is the number of sampling points; i represents the i-th iteration.
[0075] S4: Based on the Bayesian conditional probability formula, we have:
[0076]
[0077] in, Indicates that in the existing observation samples p(θ) represents the prior distribution of the parameter θ.
[0078] In the test, it is assumed that the prior distribution p(θ) of the subject's hearing parameter θ is uniformly distributed, so it can be obtained that The distribution of theta That is, the estimation of model parameters is obtained through test samples.
[0079] Based on the proposed subject response probability model and parameter estimation through the model, this method introduces active learning into the sampling process to maximize the amount of information obtained in each sampling, further accelerating the PTC test. The goal of active learning is to find the next sampling point x next , so that its information about θ is maximized. next The calculation is based on the mutual information formula:
[0080]
[0081] in, is Shannon entropy, which is used to measure the uncertainty of a random variable;
[0082] is the mutual information of the sampling point x to the parameter θ under the current sampling result; the specific calculation is approximate calculation using the following formula:
[0083]
[0084]
[0085] For detailed information of formula (4), please refer to the document Houlsby, N., Huszár, F., Ghahramani, Z., & Lengyel, M. (2011). Bayesian Active Learning for Classification and Preference Learning. ArXiv, abs / 1112.5745.
[0086] In the subject response probability model of this application, e ,Le ,k 1 ,k 2 The four parameters have the same weight, which often results in the mutual information being large at the highest points at both ends of the PTC "V" curve. However, in the actual application of dead zone detection, we are more concerned about where the tip of the PTC curve is, which corresponds to the parameter f. e Relative f sig Therefore, in the technical solution of this application, an attention mechanism is added on the basis of mutual information. Calculate the current f e The marginal distribution Afterwards according to The value of is superimposed with a Gaussian distribution as the attention weight. The attention mechanism makes the model pay more attention to f e Testing points at nearby frequencies are more likely to find the lowest point of the "V"-shaped curve.
[0087] S5: Introducing the attention mechanism into the subject response probability model, we obtain a PTC detection method based on active-learning, namely the active-PTC model.
[0088] Then, the i-th frequency f i The calculation method of the attention weight A at is:
[0089]
[0090] in, Indicates that the mean is lnf e , a normal distribution with a variance of 0.01 is lnf i The probability of
[0091] Represents the current observation sample; Indicates the current f e The marginal distribution of σ 2 Represents variance.
[0092] After calculating the attention at all frequencies, we normalize it to the interval [0,1] to obtain the attention matrix A. In order to retain the model's ability to explore the entire space, a weight of 0.5 is reserved for all spaces to prevent the model from continuously detecting in a small area.
[0093] S6: Stimulus parameter x for the next test of the model next The calculation method is:
[0094]
[0095] Where I is the identity matrix; is the mutual information of the sampling point x to the parameter θ under the current sampling result.
[0096] S7: Before each test, set the pure tone signal parameter f sig and L sig and narrowband noise test parameter f e ,L e ,k 1 ,k 2 ,f mask ,l mask Implement initialization testing after testing the scope;
[0097] Initialization tests include:
[0098] Take 10dB as L mask The step size is, when testing the pure tone signal frequency f sig , pure tone signal frequency and f e The mean of the minimum values in the test range Pure tone signal frequency and f e The mean of the maximum values in the test range The test responses at three frequencies are tested. Each time a frequency is tested, the measurement of this frequency point is stopped when the test response changes (i.e., from "hearable" to "inaudible" or from "inaudible" to "hearable"), so that the initial series of sampling results can be obtained. The sampling results obtained by initialization Send to middle.
[0099] In the present application, the initialization method can ensure that subsequent iterations have good initial values, thereby ensuring that subsequent calculations can reach stability faster.
[0100] In this embodiment, the test range set includes:
[0101] f e and f mask The test range is set at the pure tone signal frequency f sig A location one octave nearby;
[0102] L e and L mask The test range is set to pure tone signal strength L sig Below 15dB to 90dB;
[0103] k 1 and k 2 The test range is set to 10 -3 Up to 10 -0.5 .
[0104] S8: Pass Calculate the y corresponding to x; calculate the current posterior parameter distribution After that, we calculate
[0105] S9: Calculate the next stimulation parameter x using the next stimulation parameter formula next ;
[0106]
[0107] S10: Set the stimulus parameter x next As input x, steps S8 to S9 are executed repeatedly until the following stop condition is met, and the test stops;
[0108] The stopping conditions are:
[0109] Three consecutive tests is greater than the preset stopping confidence threshold, and the three f obtained in these three tests e The difference between the values is less than f e 1 / 20 of the octave.
[0110] Among them, the stopping confidence threshold is adaptively set according to the test needs, and the accuracy of the test and the test implementation time are controlled by the stopping confidence threshold. Generally speaking, the larger the stopping confidence threshold, the longer the implementation time and the higher the accuracy; in this embodiment, the stopping confidence threshold is set to 0.5.
[0111] In the technical solution of this application, by setting the stop condition, it is ensured that the test time can be saved while achieving sufficiently accurate results. Among them, the condition "three consecutive tests "greater than the preset stopping confidence threshold" ensures that the test values are sufficiently concentrated, and the condition "the three f obtained in three tests are e The difference between the values is less than f e The 1 / 20" of the octave ensures the stability of the test value, and the combination of the two ensures the accuracy of the test value.
[0112] In specific implementation, the stop condition can also be set to be determined by the experimenter's subjective judgment. If the experimenter believes that a relatively satisfactory result has been obtained during the experiment, the test can be stopped manually.
[0113] S11: construct the basic diagram of PTC curve;
[0114] The PTC curve basic graph has frequency on the horizontal axis and intensity on the vertical axis;
[0115] During the test, when y=0, the corresponding x=[f mask ,L mask] are marked on the PTC curve basic diagram, and all points are fitted to obtain the subject response probability model curve as the test result.
[0116] Step S8 can also be implemented by the following steps:
[0117] a1: Get the f in this test respectively e 、f mask , L e , L mask , k 1 and k 2 The scope of the test;
[0118] a2: The intensity parameter L e and L mask According to the equal interval, the frequency parameter f e and f mask According to the exponential equal intervals, 25 data points were taken respectively;
[0119] Use f e 、f mask , L e , L mask All data points are used to construct a 6-dimensional parameter matrix P;
[0120] In the parameter matrix P, each element [f e ,L e ,k 1 ,k 2 ,f mask ,L mask ] all conform to the calculation relationship of P(y=1|x,θ);
[0121] a3: When step S8 needs to be executed, the stimulation parameter x=[f mask ,L mask ]Query in the parameter matrix P and get the corresponding
[0122] That is, after testing the subject with parameter x and getting response y, (x, y) is stored in the current result After looking up the table from the matrix P, we can get Then update the current posterior parameter distribution according to equation (3) according to Calculated And the mutual information is calculated according to equation (4): Then determine the next test parameter x according to equation (8): next , x next Repeat the above process as x until the test stop condition is met.
[0123] In the technical solution of the present application, by setting a 6-dimensional parameter matrix P, it is ensured that in the process of calculating mutual information, only a table lookup is required, and the values in the parameter space do not need to be calculated sequentially. The calculation of each test can be completed within 1 second, so that the test subject does not need to wait between each two test points during the test. The calculation speed is greatly improved.
[0124] In the specific implementation process, the current mutual information as well as The software interface shows that the experimental operator can manually set the test parameters based on this information. If no setting is made, the software will automatically calculate the parameters x. next The next test is performed until the stop condition is met or the experiment operator manually stops the test.
[0125] For specific results based on the active-PTC model test in this application, refer to Figure 3 , the dead zone edge frequency can be obtained by measuring the tip frequency.
[0126] Figure 2 In the figure, the horizontal axis is frequency and the vertical axis is intensity. All the stimulus parameters involved in the test are x = [f mask ,L mask ] are marked in the figure, where the origin ● represents y=1 (marked as heard in the figure), and x represents y=0 (marked as did not hear in the figure). All points where y=0 are fitted can obtain the V-shaped broken line of the result obtained by the active-PTC model test in this application.
[0127] After using the technical solution of the present invention, the same test, if using the existing technology, each test of the PTC curve takes 9 minutes, using the technical solution of the present application, a single PTC test only takes about 6 minutes, saving more than 30% of the time. Considering that two or more PTC tests are usually required to determine the dead zone edge frequency in dead zone detection, the technical solution of the present application can save a lot of testing time.
[0128] Figure 4 Two box plots are made based on the test results after 16 subjects were tested using the technical solution in this application and the prior art solution. The horizontal axis in the figure is the test method used, and the vertical axis is the time taken for a single test. There is a statistically significant difference between the two, that is, the statistical p value in the figure is less than 0.0001.
[0129] At the same time, the test results of the present invention are well consistent with the test results of the prior art, and the peak frequencies of the PTC curves obtained by the test (corresponding to the dead zone edge frequencies) are basically consistent. Figure 5As shown in the figure, a box plot is made after using the same equipment, using the technical solution in this application and the prior art solution to test 16 subjects to obtain the tip frequency of the PTC curve. The horizontal axis in the figure is the test method used, and the vertical axis is the tip frequency of the PTC curve. It can be seen that there is no statistically significant difference in the test results of the two test methods (the statistical p value in the figure is equal to 0.15).
[0130] The present application proposes an active-PTC detection method based on active learning, constructs a probability model of subject response for describing PTC, obtains model parameters by parameter estimation, and obtains the subject's PTC curve by sampling some sample points, thereby accelerating the test. The present invention introduces the active learning method into the sampling process of parameter estimation. Active-PTC uses the active learning method to obtain as much information as possible in each sampling, greatly speeds up the test, solves the problem of the PTC method being too time-consuming, and achieves high frequency accuracy characteristics. Compared with traditional PTC detection methods, active-PTC can detect the dead zone edge frequency more quickly, and because of the introduction of the probability model, the experimenter can control the confidence level required to stop the experiment, thereby achieving a balance between test time and accuracy.
Claims
1. A cochlear dead zone edge detection method based on PTC, It is characterized in that It includes the following steps: S1: Construct a subject response probability model based on the PTC principle; The subject response probability model is a V-shaped curve based on the parameter θ = [f e ,L e ,k 1 ,k 2 ] represents the model parameters of the subject response probability model; Among them, f e is the frequency at the tip of the V-shaped curve, L e is the sound intensity at the tip of the V-shaped curve; (f e ,L e ) is the vertex coordinate of the subject response probability model curve; k 1 represents the forward slope of the subject response probability model curve, k 2 represents the backward slope of the subject response probability model curve; S2: Based on the subject response probability model, the probability P(y=1|x,θ) that the subject hears a pure tone is expressed as: Where Φ(x) is the cumulative distribution function of the standard normal distribution; f mask is the masking sound frequency; x=[f mask ,L mask ] represents the stimulus parameter, f mask is the frequency of the masking noise, L mask is the intensity of the masking noise; f e is the frequency at the tip of the V-shaped curve, L e is the sound intensity at the tip of the V-shaped curve; y = 1 means that the subject can hear the pure tone under the condition of stimulation parameter x; y = 0 means that the subject cannot hear the pure tone under the condition of stimulation parameter x; σ is a hyperparameter; S3: Assume that there are already samples Then we have: in, Indicates that under a certain set of parameters θ, the sample is obtained by sampling The probability of , N is the number of sampling points, i represents the i-th iteration; S4: Based on the Bayesian conditional probability formula, we have: in, Indicates that in the existing observation samples p(θ) represents the prior distribution of the parameter θ. S5: Introduce the attention mechanism into the subject response probability model, the i-th frequency f i The calculation method of the attention weight A at is: in, Indicates that the mean is lnf e , a normal distribution with a variance of 0.01 is lnf i The probability of Represents the current observation sample; Indicates the current f e The marginal distribution of σ 2 represents variance; S6: Stimulus parameter x for the next test of the model next The calculation method is: Where I is the identity matrix; is the mutual information of the sampling point x to the parameter θ under the current sampling result; S7: Before each test, set the pure tone signal parameter f sig and L sig and narrowband noise test parameter f e ,L e ,k 1 ,k 2 ,f mask ,l mask Implement initialization testing after testing the scope; The initialization test includes: Take 10dB as L mask The step size is, when testing the pure tone signal frequency f sig , pure tone signal frequency and f e The mean of the minimum values in the test range Pure tone signal frequency and f e The mean of the maximum values in the test range The test responses at three frequencies. Each time a frequency is tested, the measurement of this frequency point is stopped when the test response changes once. This way, the initial series of sampling results can be obtained. The sampling results obtained by initialization Send to middle; S8: Pass Calculate the y corresponding to x; calculate the current posterior parameter distribution After that, we calculate S9: Calculate the next stimulation parameter x using the next stimulation parameter formula next ; S10: Set the stimulus parameter x next As input x, steps S8 to S9 are executed repeatedly until the following stop condition is met, and the test stops; The stopping conditions are: Three consecutive tests is greater than the preset stopping confidence threshold, and the three f obtained in these three tests e The difference between the values is less than f e 1 / 20 of the octave; S11: construct the basic diagram of PTC curve; The abscissa of the PTC curve basic diagram is frequency, and the ordinate is intensity; During the test, when y=0, the corresponding x=[f mask ,L mask ] are marked on the PTC curve base diagram, and all points are fitted to obtain the subject response probability model curve as the test result.
2. According to the PTC-based cochlear dead zone edge detection method of claim 1, Features: In step S6, the mutual information The calculation method is: in, is Shannon entropy, which is used to measure the uncertainty of a random variable; is the mutual information of the sampling point x to the parameter θ under the current sampling result; it is specifically calculated using the following formula:
3. According to the PTC-based cochlear dead zone edge detection method of claim 1, Features: The stopping confidence threshold is set to 0.
5.
4. According to the PTC-based cochlear dead zone edge detection method of claim 1, Features: Step S8 can also be implemented by the following steps: a1: Get the f in this test respectively e 、f mask , L e , L mask , k 1 and k 2 The test scope of a2: In the test range, the intensity parameter L e and L mask According to the equal interval, the frequency parameter f e and f mask Take 25 data points at equal intervals according to the index; Use f e 、f mask , L e , L mask All data points are used to construct a 6-dimensional parameter matrix P; In the parameter matrix P, each group of elements [f e ,L e ,k 1 ,k 2 ,f mask ,L mask ] all conform to the calculation relationship of P(y=1|x,θ); a3: When step S8 needs to be executed, the stimulation parameter x=[f mask ,L mask ] Query in the parameter matrix P and get the corresponding 5. According to the PTC-based cochlear dead zone edge detection method of claim 1, Features: The test scope includes: f e and f mask The test range is set at the pure tone signal frequency f sig A location one octave nearby; L e and L mask The test range is set to pure tone signal strength L sig Below 15dB to 90dB; k 1 and k 2 The test range is set to 10 -3 Up to 10 -0.5 .
