A General Dynamic Compensation Method for Multiple Types of Sensors

Abstract

The present invention provides a general dynamic compensation method for multi-type sensors, belonging to the technical field of sensor compensation, including: obtaining dynamic calibration data of multiple types of sensors and their corresponding standard data; training a first generative adversarial network and migrating its model parameters to the corresponding modules of a second generative adversarial network; inputting the dynamic calibration data and standard data in the dataset after time-frequency conversion into the second generative adversarial network to obtain a single-sensor dynamic compensation model; inputting the loss function value, model network parameters, output value of the single-sensor dynamic compensation model, and multi-type sensor data set to be compensated into a meta-learning network to obtain a general sensor dynamic compensation model; inputting other sensor dynamic calibration data into the general sensor dynamic compensation model for training to obtain a compensation model applicable to this sensor. This method can perform signal compensation on multiple types of sensors.

Description

Technical Field

[0001] The present invention belongs to the technical field of sensor compensation, and particularly relates to a general dynamic compensation method for multiple types of sensors. Background Art

[0002] With the development of technology, people's exploration of nature has become more and more accurate, and the accuracy of accidental or sudden transient signals has received increasing attention. Common transient tests include processes such as rocket launches, weapon tests, building demolitions, and high-speed trains entering and exiting tunnels. At present, with the improvement of the performance of acquisition devices, the main bottleneck of the test bandwidth in transient test systems is the working frequency bands of various types of sensors. Their insufficient dynamic performance leads to signal distortion in the output, thereby causing dynamic errors in the test and affecting the test accuracy of the system. Due to the different types and ranges of physical quantities being tested, the types and ranges of sensors used are also different. Each type and range of sensor needs to consider its dynamic characteristics during transient testing, and targeted compensation is required for sensors that cannot meet the characteristics of transient signals. In order to reduce dynamic errors, currently, it is necessary to select an appropriate compensation algorithm for each sensor's dynamic characteristics to obtain a compensation model.

[0003] The commonly used compensation methods are divided into two categories: The first category is to first obtain the transfer model of the sensor system, and then construct a dynamic compensation system model based on the transfer model. This method can be traced back to 1984 at the earliest. David C. Hyland et al. proposed approximating the sensor as a second-order linear system and calculating its corresponding dynamic parameters to obtain a compensation equation. In recent years, foreign teams such as Maria Grazia and the National Key Laboratory of Electronic Test Technology in China have successively proposed higher-order sensor approximation systems and used the method of zero-pole cancellation to obtain a dynamic compensation model. Although this method can obtain a compensation model, during the obtaining process, it is necessary to manually adjust the compensation parameters, which limits the automation degree of this method and cannot achieve a general compensation method for multiple types of sensors.

[0004] The second category is not to rely on the transfer model of the sensor system, but to directly obtain the dynamic compensation model through a non-linear system. This method has been popularized with the research of swarm intelligence algorithms and neural network algorithms. Since Professor Xu Kejun used the PSO algorithm to achieve dynamic compensation of acceleration sensors in 2009, more and more domestic and foreign scholars have adopted this method. The system compensation order, speed, and accuracy have all been significantly improved.

[0005] Although the above methods have achieved automatic algorithm solving, the solving process requires preparing data for each sensor, which is difficult to apply and popularize in the scenario of batch compensation for a large number of sensors. Summary of the Invention

[0006] To overcome the deficiencies of the above-mentioned existing technologies, the present invention provides a general dynamic compensation method for multi-type sensors.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] A general dynamic compensation method for multi-type sensors, comprising:

[0009] Performing time-frequency conversion on the dynamic calibration data of multiple types of sensors and their corresponding standard data to form a data set after time-frequency conversion;

[0010] Training a first generative adversarial network, and migrating the hidden layer parameters of its generator G0, the hidden layer parameters of discriminator D0, and the output layer parameters of discriminator D0 to the corresponding modules of the second generative adversarial network as the initial parameters of the second generative adversarial network;

[0011] Inputting the standard data into discriminator D1 of the second generative adversarial network, and inputting the dynamic calibration data into generator G1 of the second generative adversarial network, and training generator G1 to obtain a single-sensor dynamic compensation model;

[0012] Taking the multi-type sensor data set to be compensated, the loss function value, the model network parameters, and the output value of the single-sensor dynamic compensation model in the single-sensor dynamic compensation model as the input end of the meta-learning network, and having the meta-learning network judge the quality of the single-sensor dynamic compensation model and reversely adjust the parameters in the single-sensor dynamic compensation model; after the meta-learning network is trained and the single-sensor dynamic compensation model is adjusted, a general sensor dynamic compensation model is obtained;

[0013] Training the general sensor dynamic compensation model with the dynamic calibration data of other sensors and their corresponding standard data to obtain a specific dynamic compensation model applicable to the other sensors, and using the specific dynamic compensation model to perform signal compensation on the other sensors.

[0014] Further, the formula for performing time-frequency conversion on the dynamic calibration data of multi-type sensors and their corresponding standard data is:

[0015]

[0016] where w is the frequency, t is the time, e -iwt is a complex function, f(t) is an arbitrary time-domain signal, and F(w) is the frequency-domain signal after Fourier transform.

[0017] Further, the first generative adversarial network is a voice enhancement network. The input of its generator G0 is noisy speech, and the input of its discriminator D0 is the clean speech corresponding to the noisy speech.

[0018] Further, the mathematical model of the discriminator D0 is:

[0019]

[0020] where p is the real sample, P data (p) is the spatial data distribution of the real sample, z is the generated sample, z ∼ P z (z) represents a certain random distribution that the random noise follows, and E is to calculate the expected value. Further, in the second generative adversarial network, the parameters of the input and output layers of its generator G1, the input layer of its discriminator D1, and the three hidden layers between the input and output layers of its generator G1 and the input layer of its discriminator D1 are iteratively updatable parameters, and the network parameters of all the remaining hidden layers directly migrate the training results of the first generative adversarial network without iteration.

[0021] Further, the meta-learning network includes:

[0022] A test data and compensation result quality and feature identification network layer; used to judge the compensation effect of the single-sensor dynamic compensation model on the compensation data set of multiple types of sensors;

[0023] A data feature, compensation result and generator parameter comprehensive analysis network layer; used to evaluate the model parameters in the single-sensor dynamic compensation model by using the output results of the test data and compensation result quality and feature identification network layer.

[0024] Further, the meta-learning network further includes:

[0025] A generator parameter feature analysis network layer, used to identify the model parameters of the single-sensor dynamic compensation model, and avoid the influence of the diversity of the dynamic calibration data of multiple types of sensors on the training results of the meta-learning network.

[0026] Further, the loss function of the meta-learning network is:

[0027]

[0028] where T i is the task of each training, P(T) is the task set composed of multiple tasks, and L Ti is the task T iThe loss function, where θ is the model parameter of the single-sensor dynamic compensation model obtained from the previous training, f(θ) is the parametric function of the model parameter of the single-sensor dynamic compensation model obtained from one training, and θ i ' is the optimal parameter obtained from this training, and α is the trainable parameter.

[0029] A general dynamic compensation method for multi-type sensors provided by the present invention has the following beneficial effects:

[0030] The present invention uses transfer learning to transfer a high-quality network that has been fully trained with big data samples to the small-sample field, avoiding the process of training a network from scratch with small samples, improving the utilization rate of small-sample data, and then optimizing some parameters or network structures to achieve the transfer of network functions; also using the meta-learning network mechanism, in the process of obtaining multiple sensor models, using a small number of base-class samples to train the meta-learner, recording the model change process and model features, obtaining the general sensor dynamic compensation model and outputting it; using the general sensor dynamic compensation model to provide a good initial parameter for the new model, and then using a small number of new samples to fine-tune the new model to quickly achieve parameter adaptation, improving the learning efficiency of the model for new samples, thus saving the training process that each sensor needs to start from scratch. Therefore, in addition to inheriting the high degree of automation of deep learning, this method only requires a small amount of dynamic calibration data and a small number of iterations for any sensor to be calibrated to obtain the model. It solves the problem that the sensor compensation model in the prior art cannot be applied and popularized in the scenario of batch compensation of a large number of sensors. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the embodiments of the present invention and its design scheme, the accompanying drawings required for this embodiment will be briefly introduced below. The accompanying drawings in the following description are only partial embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0032] Figure 1 It is the overall principle block diagram of a general dynamic compensation method for multi-type sensors according to an embodiment of the present invention;

[0033] Figure 2 It is the main process module diagram of the present invention;

[0034] Figure 3 It is the principle block diagram for obtaining a single-sensor dynamic compensation model based on a transfer generative adversarial network according to an embodiment of the present invention;

[0035] Figure 4 It is the structure diagram of the meta-learning network layer and loss function according to an embodiment of the present invention;

[0036] Figure 5Schematic diagram of the traditional sensor dynamic compensation method according to the embodiments of the present invention. Detailed implementation manners

[0037] In order to enable those skilled in the art to better understand the technical solutions of the present invention and be able to implement them, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention, and cannot be used to limit the protection scope of the present invention.

[0038] In addition, terms such as "first", "second", etc. are only used for descriptive purposes and cannot be understood as indicating or implying relative importance. In the description of the present invention, it should be noted that unless otherwise clearly specified or limited, the terms "connected" and "connected" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances. In the description of the present invention, unless otherwise stated, the meaning of "a plurality" is two or more, which will not be elaborated here.

[0039] Embodiment:

[0040] The present invention provides a general dynamic compensation method for multi-type sensors, specifically as Figure 1 shown, including:

[0041] Performing time-frequency conversion on the dynamic calibration data of multiple types of sensors and their corresponding standard data to form a data set after time-frequency conversion; training a first generative adversarial network, and migrating the hidden layer parameters of its generator, the hidden layer parameters of the discriminator, and the output layer parameters of the discriminator to the corresponding modules of the second generative adversarial network as the initial parameters of the second generative adversarial network; inputting the standard data into the discriminator D1 of the second generative adversarial network, inputting the dynamic calibration data into the generator G1 of the second generative adversarial network, and training the generator G1 to obtain a single-sensor dynamic compensation model; using the multi-type sensor data set to be compensated, the loss function value, the model network parameters, and the output value of the single-sensor dynamic compensation model in the single-sensor dynamic compensation model as the input end of the meta-learning network, and having the meta-learning network judge the quality of the single-sensor dynamic compensation model and reversely adjust the parameters in the single-sensor dynamic compensation model; after the meta-learning network is trained and the single-sensor dynamic compensation model is adjusted, a general sensor dynamic compensation model is obtained; using the dynamic calibration data of other sensors and their corresponding standard data to train the general sensor dynamic compensation model to obtain a specific dynamic compensation model applicable to other sensors, and using the specific dynamic compensation model to perform signal compensation on other sensors.

[0042] Compared with this method, the prior art has the following disadvantages:

[0043] 1. Difficult to automate and low solution accuracy

[0044] The existing methods require manual adjustment of compensation parameters, which brings inevitable errors, low compensation efficiency, and limits the automation degree of such methods.

[0045] 2. Low solution accuracy

[0046] The dynamic compensation models obtained by the existing methods have limited compensation accuracy due to the existence of multiple non-minimum phase systems.

[0047] 3. Lack of generality

[0048] Currently, the research on sensor compensation methods focuses on a certain type or a specific type of sensor. Due to the great differences in the dynamic characteristics of various types of sensors, corresponding compensation algorithms need to be proposed according to their specific characteristics.

[0049] 4. Lack of data volume

[0050] Limited by the high test cost, the existing sensor calibration data sets are limited, and a large number of repeated tests to obtain sufficient data affect the actual service life of the sensors.

[0051] The effects of the present invention:

[0052] Using a small sample size of the target sensor to obtain a general compensation model for multiple types and ranges of sensors;

[0053] For new types of sensors not covered, continue to train with a small amount of data to quickly solve a high-precision compensation model.

[0054] The following is the research basis for proposing the present invention:

[0055] Taking multiple types of sensors as an example, specifically three typical transient test sensors: acceleration sensors, pressure sensors, and thermocouples. And the dynamic characteristics of each type of sensor are different. The specific dynamic characteristics are shown in Table 1. As Figure 1 Shown are the bode diagrams of the three sensors and their compensation models obtained by traditional methods.

[0056] Table 1 Sensor type and dynamic characteristic table

[0057]

[0058] As can be seen from Table 1, the reasons for the dynamic errors generated by the three different types of sensors during the test are not the same. When studying a single type of sensor, it is necessary to broaden the working bandwidth of the acceleration sensor, suppress the resonance frequency of the pressure sensor, and reduce the rise time of the thermocouple. However, their compensation essential principles are the same, which is to construct a specific filter to realize the expansion and correction of the original system passband.

[0059] Based on this, the present invention realizes the research of a general compensation method by combining time-frequency transformation and transfer meta-learning algorithm.

[0060] The following are the principles and effects of the meta-learning used in the present invention:

[0061] In deep learning, the training unit is "data". The model is optimized through data, the loss function is calculated, and the mapping function between data is found. In meta-learning, the training unit is divided into two layers. The first layer is "task", and the second layer is "data" corresponding to each task. The process of meta-learning is that multiple tasks and corresponding training data are regarded as "training samples", and a "meta-learner" learns some "meta-knowledge" from these "training samples", so that good generalization effects can be obtained in new machine learning tasks, that is, "test samples". Therefore, the network based on meta-learning can not only train data with different attributes at the same time, but also reduce the training samples and obtain the optimal model parameters based on multiple tasks.

[0062] The following are the specific embodiments of the present invention:

[0063] The implementation scheme of the present invention is as Figure 2 shown. It is mainly divided into four steps: First, construct a dataset after the time-frequency conversion of multi-type sensors; second, obtain a dynamic compensation model for a single sensor for the data after the time-frequency conversion through the first-layer transfer generative adversarial network (GAN); then, during the training process of the first-layer network, use the training data, model parameters, and loss function results as the input of the meta-learning network to realize the training of the meta-learning network. Finally, use the obtained general compensation network to realize the dynamic compensation of the target sensor in the case of small samples.

[0064] 1) Unifying the sensor compensation target based on time-frequency conversion

[0065] The system uses a large amount of dynamic calibration data of the above three types of sensors with multiple ranges as the original training dataset. Due to the different dynamic characteristics of each sensor, the time domain of the data is quite different. Perform time-frequency conversion on all dynamic calibration data and their corresponding standard data to unify them into frequency domain features. The time-frequency conversion formula is as shown in formula (1).

[0066] (w represents frequency, t represents time, e -iwtFor any time-domain signal f(t) (a complex variable function), after Fourier transform, it becomes the frequency-domain signal F(ω).

[0067] Since the commonly used signals in dynamic calibration are mainly step signals and pulse signals. Their frequency-domain characteristics are close and definite. Due to the presence of dynamic errors in the test data, the frequency-domain characteristics of the data are different in a specific frequency band, but the main data trends and profiles are consistent. Thus, when all the data are converted to the frequency domain, the data characteristics are close and the dynamic defects are clear.

[0068] 2) Obtaining the single-sensor dynamic compensation model based on the transfer adversarial network

[0069] The learning network constructed under the meta-learning method is mainly divided into two layers. The bottom layer is the basic network, and the upper layer is the meta-learning network. The basic network layer of the present invention is composed of a transfer adversarial network, adopting the structure of a generative adversarial network, combining the means of transfer learning, and using a large amount of dynamic calibration data of a single sensor as the training data set. After multiple trainings, the corresponding dynamic compensation model is obtained. Repeat this process to complete the obtaining of the compensation models for multiple sensors of three types and multiple ranges described in the present invention.

[0070] The present invention uses speech enhancement as the transfer source network. The two networks before and after transfer use isomorphic networks, that is, generative adversarial networks with the same network structure and the same number of parameters.

[0071] The basic idea of the generative adversarial network is as Figure 3 shown. The generator G0 is mainly responsible for processing the noisy speech signal z(n), so that the denoised speech signal after passing through the generator G0 and the pure speech signal p(n) are used together as the input samples of the discriminator D0. The essence of the discriminator D0 is a binary classifier, and the specific mathematical model is shown in formula (2).

[0072]

[0073] In the above formula, p represents the real sample, P data (p) represents the spatial data distribution of the real sample, z represents the generated sample, z ∼ P z (z) represents that the random noise follows a certain random distribution (for example: normal distribution, Gaussian distribution, etc.), and E represents calculating the expected value.

[0074] It is mainly used to distinguish between the pure speech signal and the speech signal G0(z) enhanced by the generator. The goal is to make D0(p) approach 1 and make D0(G(z)) approach 0. The discrimination result of the discriminator D0 is returned to the generator G0 as the supervision result, and at the same time, the right or wrong of the discrimination result is returned to the discriminator D0.

[0075]

[0076] Based on Equation (2), by adjusting the generator G0 to minimize the overall result (where the loss function takes the opposite value), the optimal generator G0 can be obtained. This is the desired speech enhancement network model.

[0077] After obtaining the generator parameters with better speech enhancement effect, transfer learning operations are carried out. To ensure that the network maintains a high level of filter parameter identification ability and adaptability to new data distributions after transfer. As Figure 2 shown, in the present invention, only the input and output layers of the generator G1, the input layer of the discriminator D1, and the parameters of the two hidden layers connected to the three are set as iteratively updatable parameters after transfer.

[0078] The objective of retraining these three parts is to realize the network mapping relationship, map the dynamic compensation data distribution to the speech enhancement data space, and enable the network to still utilize the identification ability of the source data space.

[0079] The retraining part realizes the feature mapping from the target domain to the source domain, as shown in Equation (4):

[0080]

[0081] Among them, f(z) is the final identification ability of the network for the target domain, and R(z) is the mapping relationship that maps the target domain to the source domain, that is, the new network structure composed of the parameters to be obtained in the retraining part.

[0082] The mapping distance is defined as A, that is:

[0083]

[0084] Among them, the script A is the Borel set, and A is a subset of it. Equation (5) is to take all subsets and find the upper bound of the mapping distance. By training to reduce the upper bound to an acceptable error range, the purpose of transfer learning is achieved.

[0085] For the specific derivation process, assume a possible value of A:

[0086] A→I(h)={z∈Z:h(z)=1,h∈H}(6)

[0087] That is, record the A distance as the H distance:

[0088]

[0089] Since the upper bound needs to be taken, it is necessary to find that the set I(h) is (-∞, 0). Then it can be obtained:

[0090]

[0091] It can be seen that by training to obtain η that conforms to Equation (8), the upper bound of the mapping distribution distance can be minimized, that is, transfer learning can be achieved.

[0092] So far, the setting of the transfer network is completed. By training it with multiple dynamic calibration data from the same sensor, the dynamic compensation network of this sensor can be obtained, which is the trained generator G1.

[0093] 3) Based on meta-learning, the identification and feedback of the characteristics of the compensation networks of different types of sensors are realized

[0094] In the meta-learning upper-layer network of the present invention, during the training process of the basic network, the changes of the training model and the model parameters are recorded and analyzed, and the corresponding relationship between the change rules and the fitness function values is summarized, so as to learn a network that accelerates and promotes the optimization of the model parameters. The present invention further adds a data analysis network during the training process of the meta-learning network to improve the adaptability of the training network to different data types. The structure diagram of the final meta-learning network layer and the loss function is as Figure 4 shown.

[0095] During the process of obtaining the single-sensor compensation model, the network model iteration parameters, the corresponding dynamic calibration data, the network compensation output results (obtained from the compensation network and the calibration data), and the loss function results are uniformly used as the input data of the meta-learning network. The meta-learning network is divided into two main parts: one part is the data feature identification network, which takes the dynamic data and the compensation output results as inputs. It is composed of a multi-layer convolutional network, a linear fully connected network, and an activation function, and is capable of analyzing the features of the test data itself, obtaining the degree of dynamic feature loss, and at the same time analyzing the quality of the compensation results to determine the pros and cons of the compensation structure of the generator dynamic compensation network for the test data. On the one hand, the output results of this part of the network are sorted by the output network to organize the output data length and dimension, and then handed over to the loss function for evaluation and feedback. On the other hand, it is used as the input of the generator parameter comprehensive analysis network layer to provide a reference basis for the evaluation of the generator parameters.

[0096] The other part of the meta-learning network is the generator parameter feature analysis network layer. In order to analyze the characteristics of the generator parameters, the analysis network is divided into 6 independent analysis network layers and a comprehensive analysis network layer that finally summarizes and organizes the analysis results.

[0097] Taking the process of obtaining the compensation model of one type of sensor as a task T i , its model is the parametric function f with parameters θ θ , when the meta-learning network needs to adapt to the next type of sensor compensation model task T i ', the optimal parameter θ becomes θ i ', and the gradient descent for updating the parameters is required, that is:

[0098]

[0099] Among them, α is the step size parameter, and L(·) is the loss function. The parameters of the model are trained and optimized through the data set P(T) composed of multiple tasks. The specific process is:

[0100]

[0101] That is, the obtained network parameters can minimize the loss under all tasks.

[0102] Each independent analysis network layer uses an attention mechanism network structure, and the weight is adjusted by the comprehensive calculation results of the input generator loss function and the loss function of the network, so as to improve the influence of excellent network parameters on the network. The results of each iteration of the network are passed to their respective loss functions by the two output network layers to calculate the loss results and perform back propagation, adjust the parameters of the entire network, and optimize the parameters of the meta-learning network.

[0103] Finally, through the training of dynamic compensation processes of sensors of various types and ranges, the meta-learning network is able to evaluate the quality of the sensor compensation network and predict its evolutionary trend. This ability can guide the compensation training process of new sensors and improve iteration efficiency and compensation accuracy. After the above two main steps, the meta-learning network can be trained to obtain a network that is sensitive to the compensation model.

[0104] The dynamic calibration data of sensors outside the training set is input into the basic dynamic compensation model. After a small number of iterations supervised by the meta-learning network, a dynamic compensation model similar to that of a large number of trainings can be obtained, thereby greatly reducing the amount of training data and iterative calculations required in the dynamic compensation process. The automation level and model identification efficiency of sensor dynamic compensation based on deep learning are improved.

[0105] The following are the differences between the present invention and the prior art:

[0106] At present, the discussion on the dynamic compensation model of sensors is all about a certain sensor, and the dynamic compensation model is obtained based on its own dynamic characteristics and dynamic calibration data. The relevant research is centered on the method of obtaining the specific model. Figure 5 shown.

[0107] The purpose of the present invention is not to obtain a compensation model for a single sensor or a class of sensors, but to obtain a universal basic model for dynamic compensation of multiple types of sensors, to improve the comprehensive analysis and utilization capabilities of dynamic calibration data of multiple types of sensors, and to further study the common characteristics and essential causes of dynamic compensation processes of different types of sensors. The principle block diagram of the sensor dynamic compensation method proposed by the present invention is as follows: Figure 1 shown.

[0108] In order to implement a general compensation method for multi-type sensors, the present invention innovatively proposes to introduce the transfer meta-learning method in the field of deep learning into the field of sensor dynamic compensation, and proposes several targeted improvements according to the characteristics of sensor dynamic calibration.

[0109] First, in order to unify the compensation process into a homogeneous task model suitable for transfer meta-learning processing, a time-frequency transformation scheme is proposed to perform time-frequency transformation on the test data, and the determination of the sensor compensation model is unified into the determination of the specific filter model parameters, realizing the unification of the compensation purposes for multi-type sensors.

[0110] Second, in order to avoid the influence of the diversity of sensor dynamic calibration data types on the training results of the meta-learning network, a data analysis module is added to the meta-learning network. Through the data analysis module, the dynamic calibration data types can be identified. At the same time, the influence weight of the data difference caused by the passband characteristics of the sensor on the network is enhanced, and the influence weight of the characteristics of the data itself on the network is reduced, improving the adaptability of the meta-learning network to various calibration signals.

[0111] It should be emphasized that the general dynamic compensation method described in the present invention does not obtain a general solution for dynamic compensation directly applicable to all sensors. Instead, through the meta-learning network structure, the general characteristics of the dynamic compensation models of various types of sensors are refined. For sensors outside the training set, it is necessary to continue to perform a limited number of iterations (the data volume is much smaller than that required for traditional method training, with a difference of about two orders of magnitude) through a small amount of dynamic calibration data under the general compensation network framework described in the present invention to obtain the corresponding specific dynamic compensation model.

[0112] The above-described embodiments are only the preferred specific embodiments of the present invention, and the protection scope of the present invention is not limited thereto. Any simple changes or equivalent replacements of the technical solutions that can be obviously obtained by those skilled in the art within the technical scope disclosed by the present invention all belong to the protection scope of the present invention.

Claims

1. A general dynamic compensation method for multi-type sensors, characterized in that Including: Performing time-frequency conversion on the dynamic calibration data of multiple types of sensors and their corresponding standard data to form a data set after time-frequency conversion; Training a first generative adversarial network, and migrating the hidden layer parameters of its generator G0, the hidden layer parameters of discriminator D0, and the output layer parameters of discriminator D0 to the corresponding modules of the second generative adversarial network as the initial parameters of the second generative adversarial network; Inputting the standard data into discriminator D1 of the second generative adversarial network, and inputting the dynamic calibration data into generator G1 of the second generative adversarial network, and training generator G1 to obtain a single-sensor dynamic compensation model; Taking the multi-type sensor data set to be compensated, the loss function value, model network parameters, and output value of the single-sensor dynamic compensation model in the single-sensor dynamic compensation model as the input end of the meta-learning network, and the meta-learning network discriminates the quality of the single-sensor dynamic compensation model and reversely adjusts the parameters in the single-sensor dynamic compensation model; After the meta-learning network is trained and the single-sensor dynamic compensation model is adjusted, a general sensor dynamic compensation model is obtained; Among them, the meta-learning network includes: A test data and compensation result quality and feature identification network layer for discriminating the compensation effect of the single-sensor dynamic compensation model on the multi-type sensor data set to be compensated; A data feature, compensation result and generator parameter comprehensive analysis network layer for using the output result of the test data and compensation result quality and feature identification network layer to evaluate the model parameters in the single-sensor dynamic compensation model; A generator parameter feature analysis network layer for identifying the model parameters of the single-sensor dynamic compensation model to avoid the influence of the diversity of the dynamic calibration data of multiple types of sensors on the training result of the meta-learning network; Training the general sensor dynamic compensation model with the dynamic calibration data of other sensors and their corresponding standard data to obtain a specific dynamic compensation model applicable to the other sensors, and using the specific dynamic compensation model to perform signal compensation on the other sensors.

2. The general dynamic compensation method for multiple types of sensors according to claim 1, wherein The formula for performing time-frequency conversion on the dynamic calibration data of multiple types of sensors and their corresponding standard data is: where, w is the frequency, t is the time, and e -iwt is a complex function, f(t) is an arbitrary time-domain signal, and F(w) is the frequency-domain signal after Fourier transform.

3. A general dynamic compensation method for multi-type sensors according to claim 1, characterized in that The first generative adversarial network is a voice enhancement network, the input of its generator G0 is noisy speech, and the input of its discriminator D0 is the clean speech corresponding to the noisy speech.

4. A general dynamic compensation method for multi-type sensors according to claim 3, characterized in that The mathematical model of the discriminator D0 is: Among them, is the real sample, is the spatial data distribution of the real sample, is the generated sample, represents a certain random distribution that the random noise follows, is to calculate the expected value.

5. A general dynamic compensation method for multi-type sensors according to claim 1, characterized in that, In the second generative adversarial network, the parameters of the input and output layers of its generator G1, the input layer of discriminator D1, and the three hidden layers between the input and output layers of generator G1 and the input layer of discriminator D1 are iteratively updatable parameters, and the network parameters of all other hidden layers directly migrate the training results of the first generative adversarial network without iteration.

6. A general dynamic compensation method for multi-type sensors according to claim 1, characterized in that The loss function of the meta-learning network is: Among them, T i is the task for each training, and P(T) is the task set composed of multiple tasks. is the loss function of task T i , θ is the model parameter of the single-sensor dynamic compensation model obtained from the previous training, f(θ) is the parameterization function of the model parameter of the single-sensor dynamic compensation model obtained from one training. is the optimal parameter obtained from this training, and α is the trainable parameter.

Images