Structural health monitoring signal compression sampling method and device

Abstract

The present invention provides a method and device for compressive sampling of structural health monitoring signals, and relates to the technical fields of signal processing and structural health monitoring. The method includes the design of a compressed sampling observation matrix of a deep convolutional generative adversarial network for transfer learning compression, the design of a network architecture of a deep convolutional generative adversarial network, the setting of network model parameters, the application of transfer learning, the determination of an optimal low-dimensional representation vector, etc. The present invention utilizes the powerful feature learning ability of a deep convolutional generative adversarial network to learn its internal structural features from the big data of monitoring signals, and applies it to compressed sampling, which greatly alleviates the instability of the decompression and reconstruction process, helps to break the high dependence of traditional methods on sparse bases, and avoids the tedious process of finding a suitable sparse base.

Description

Technical Field

[0001] The present invention relates to the technical field of signal processing and structural health monitoring, and in particular to a structural health monitoring signal compression sampling method and device. Background Art

[0002] During the long-term use of buildings and structures, their structural performance will inevitably be weakened by factors such as environmental erosion and human damage, which will increase the probability of safety accidents. Therefore, it is of great significance to carry out health monitoring of engineering structures, identify safety hazards of structures by monitoring various indicators of civil engineering structures during use, issue early warnings for possible safety problems, and take corresponding measures in a timely manner.

[0003] Since structural health monitoring has the characteristics of integrating modern sensing, data transmission, signal processing, structural state assessment and other complex cutting-edge technologies, its application also faces many challenging problems. Among them, high-frequency sampling of the dynamic response data of the structure is an important part of evaluating the structural state. A large amount of monitoring data has brought huge work pressure to the data acquisition and signal processing of the structural health monitoring system. Therefore, an effective data compression and reconstruction method is urgently needed.

[0004] However, the compression sampling technology must be based on the sparse characteristics of the signal, that is, it is necessary to find a suitable sparse basis to make the signal sparse enough under this basis, and embed the sparse regular term explicit expression to perform sparse optimization. However, in reality, the sparse basis of many structural health monitoring signals is difficult to find, which makes the traditional compression sampling method difficult to apply. Therefore, it is necessary to develop a compression sampling method that can ensure the accuracy of signal reconstruction without spending a lot of time to find a suitable sparse basis. Summary of the invention

[0005] In view of the above problems, the present invention provides a method and device for compressing and sampling structural health monitoring signals.

[0006] On the one hand, the present invention provides a method for compressing and sampling structural health monitoring signals, including: obtaining a health monitoring signal of a structure to be tested and establishing a signal data set; initializing a first generator and a first discriminator and constructing a generative adversarial network; randomly generating a first characterization vector with an initial dimension, inputting the first characterization vector into the first generator, and preliminarily training the generative adversarial network using the signal data set to obtain a trained second generator and a second discriminator; randomly generating a set of second characterization vectors with an initial dimension, inputting the second characterization vector into a second generator, generating corresponding real-like signals, and calculating the similarity between the real-like signals and the signal data set as the generation capability of the second generator; uniformly reducing the initial dimension of the second characterization vector successively, and calculating the second characterization vector after each dimension reduction. The vector is input into the current generator, and the feature extraction model is used to perform transfer learning training on the generative adversarial network. After each transfer learning training, it is judged whether the generation ability of the current generator and the dimension of the current second characterization vector meet the corresponding conditions respectively. If so, the transfer learning training is stopped to obtain the trained third generator and the third discriminator; a linear observation matrix is ​​used to compress the health monitoring signal and add a random noise signal to obtain compressed sampling data; a set of third characterization vectors with intrinsic dimension are randomly generated as the input of the third generator, and the gradient descent algorithm is used to iteratively optimize the third characterization vector based on the compressed sampling data, and the optimized third characterization vector is re-input into the third generator to generate a decompressed reconstructed signal corresponding to the health monitoring signal.

[0007] On the other hand, the present invention provides a structural health monitoring signal compression sampling device, including: a monitoring signal acquisition module, used to acquire the health monitoring signal of the structure to be tested and establish a signal data set; a network construction module, used to initialize the first generator and the first discriminator, and construct a generative adversarial network; a preliminary training module, used to randomly generate a first characterization vector with an initial dimension, input the first characterization vector into the first generator, and use the signal data set to preliminarily train the generative adversarial network to obtain a trained second generator and a second discriminator; a generation capacity calculation module, used to randomly generate a set of second characterization vectors with an initial dimension, input the second characterization vector into the second generator, generate corresponding real-like signals, and calculate the similarity between the real-like signals and the signal data set as the generation capacity of the second generator; a transfer learning training module, used to uniformly increase the initial dimension of the second characterization vector successively. Decrease, input the second characterization vector after each dimension decrease into the current generator, use the feature extraction model to perform transfer learning training on the generative adversarial network, and judge whether the generation ability of the current generator and the dimension of the current second characterization vector meet the corresponding conditions after each transfer learning training. If so, stop the transfer learning training to obtain the trained third generator and the third discriminator; the compression sampling module is used to use a linear observation matrix to compress the health monitoring signal and add a random noise signal to obtain compressed sampling data; the decompression and reconstruction module is used to randomly generate a set of third characterization vectors with intrinsic dimension as the input of the third generator, use the gradient descent algorithm and iteratively optimize the third characterization vector based on the compressed sampling data, and re-input the optimized third characterization vector into the third generator to generate a decompressed and reconstructed signal corresponding to the health monitoring signal.

[0008] Compared with the prior art, the structural health monitoring signal compression sampling method and device provided by the present invention have at least the following beneficial effects:

[0009] (1) Compared with general data compression methods, the compressed sampling method provided by the present invention does not require complete original data during the acquisition process, and only retains a small amount of valid data to express the signal, which occupies less resources and has a lower probability of packet loss during transmission, while being able to achieve a higher-precision reconstruction effect.

[0010] (2) Compared with the traditional compressed sampling method, the present invention does not need to monitor the sparse characteristics of the signal, but uses the powerful feature learning ability of the deep convolutional generative adversarial network to learn its internal structural characteristics from the monitoring signal big data and apply it to compressed sampling, which greatly alleviates the instability of the decompression and reconstruction process, helps to break the traditional method's high dependence on the sparse basis and avoids the tedious process of finding a suitable sparse basis.

[0011] (3) The present invention introduces a transfer learning method, which significantly reduces the dimension of the solution space while ensuring the efficiency of signal reconstruction. This alleviates the reconstruction process's requirements for the completeness of the observed data, improves the signal compression ratio, and further reduces resource consumption in the actual sampling process, greatly improving the sampling and transmission efficiency of structural health monitoring signal data. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] The above contents and other objects, features and advantages of the present invention will become more apparent through the following description of the embodiments of the present invention with reference to the accompanying drawings, in which:

[0013] Figure 1 A schematic diagram showing the working principle of a method and device for compressive sampling of structural health monitoring signals according to an embodiment of the present invention is shown;

[0014] Figure 2 Schematically shows a flow chart of a method for compressive sampling of structural health monitoring signals according to an embodiment of the present invention;

[0015] Figure 3 A specific operation flow chart of the structural health monitoring signal compression sampling method according to an embodiment of the present invention is schematically shown;

[0016] Figure 4 A flowchart schematically shows a preliminary training process according to an embodiment of the present invention;

[0017] Figure 5 A flowchart of a similarity calculation process according to an embodiment of the present invention is schematically shown;

[0018] Figure 6 A flowchart of an iterative optimization process according to an embodiment of the present invention is schematically shown;

[0019] FIG. 7( a1 ) and FIG. 7( a2 ) schematically show reconstruction effect diagrams of simulated sinusoidal signals according to an embodiment of the present invention;

[0020] FIG. 7( b1 ) and FIG. 7( b2 ) schematically show the reconstruction effect diagram of the health monitoring signal of the actual structure according to the embodiment of the present invention;

[0021] Figure 8 The block diagram of the structural health monitoring signal compression sampling device according to an embodiment of the present invention is schematically shown. DETAILED DESCRIPTION

[0022] Below, embodiments of the present invention will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the present invention. In the following detailed description, for ease of explanation, many specific details are set forth to provide a comprehensive understanding of embodiments of the present invention. However, it is apparent that one or more embodiments may also be implemented without these specific details. In addition, in the following description, descriptions of known structures and technologies are omitted to avoid unnecessary confusion of concepts of the present invention.

[0023] The terms used herein are only for describing specific embodiments and are not intended to limit the present invention. The terms "comprise", "include", etc. used herein indicate the existence of the features, steps, operations and / or components, but do not exclude the existence or addition of one or more other features, steps, operations or components.

[0024] All terms (including technical and scientific terms) used herein have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted as having a meaning consistent with the context of this specification, and should not be interpreted in an idealized or overly rigid manner.

[0025] When using expressions such as "at least one of A, B, and C, etc.", they should generally be interpreted according to the meaning of the expression commonly understood by those skilled in the art (for example, "a system having at least one of A, B, and C" should include but is not limited to a system having A alone, B alone, C alone, A and B, A and C, B and C, and / or A, B, C, etc.).

[0026] Some block diagrams and / or flow charts are shown in the accompanying drawings. It should be understood that some blocks or combinations thereof in the block diagrams and / or flow charts can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, so that these instructions can create a device for implementing the functions / operations described in these block diagrams and / or flow charts when executed by the processor. The technology of the present invention can be implemented in the form of hardware and / or software (including firmware, microcode, etc.). In addition, the technology of the present invention can take the form of a computer program product on a computer-readable storage medium storing instructions, which can be used by an instruction execution system or used in conjunction with an instruction execution system.

[0027] The compressed sampling method based on deep convolutional generative adversarial network not only constrains the reconstructed signal to have correct features through the generative model, but also significantly reduces the dimension of the solution space to the space of low-dimensional representation vectors, greatly reducing the ill-posedness of the reconstruction process. The lower the dimension of the low-dimensional representation vector, the lower the requirement of the compressed sampling method for the completeness of the observed data. Considering that directly using lower-dimensional representation vectors will increase the difficulty of deep convolutional generative adversarial network to learn data features, this method introduces transfer learning in the training process of deep convolutional generative adversarial network, and gradually compresses the number of dimensions of low-dimensional representation vectors while ensuring that the network model can correctly capture the data features, thereby enhancing the reconstruction performance of the above-mentioned compressed sampling method for highly compressed structural health monitoring signals.

[0028] At present, in the field of structural health monitoring, the signal compression sampling method of deep convolutional generative adversarial network based on transfer learning compression breaks through the sparsity constraint of the original signal. It not only uses the deep generative network model to get rid of the sparsity requirement of the original signal, but also compresses the low-dimensional representation dimension of the network model through transfer learning, which greatly improves the reconstruction accuracy and compression ratio of the signal. At the same time, it has strong universality and can be better applied to actual engineering in the field of structural health monitoring.

[0029] Figure 1 The working principle diagram of the structural health monitoring signal compression sampling method and device according to an embodiment of the present invention is schematically shown.

[0030] like Figure 1 As shown in the figure, ε in the figure represents the random noise generated during sampling, observation, and transmission. Since the monitoring signals in practice are not highly sparse, it is difficult for traditional compressed sampling methods to ensure accurate reconstruction of the original signal at a high compression ratio. Therefore, it is necessary to develop a compressed sampling method that can accurately reconstruct the original signal from a small amount of compressed sampling data. The compressed sampling method of structural health monitoring signals using deep convolutional generative adversarial networks based on transfer learning compression can be parallelized on high-performance hardware and has extremely high computational efficiency. Since there are a large number of structural health monitoring signals that need to be analyzed in actual engineering, the decompression and reconstruction algorithm is required to have the ability to reconstruct signals in real time, which is suitable for this method.

[0031] The following will be based on Figure 1 The working principle of the description is Figure 2 7 and 8 describe in detail the structural health monitoring signal compression sampling method according to an embodiment of the present invention.

[0032] Figure 2 The flowchart of the structural health monitoring signal compression sampling method according to an embodiment of the present invention is schematically shown. Figure 3The specific operation flow chart of the structural health monitoring signal compression sampling method according to an embodiment of the present invention is schematically shown.

[0033] Combination Figure 2 and Figure 3 As shown, the structural health monitoring signal compression sampling method of this embodiment may include operations S220 to S270.

[0034] In operation S220 , a health monitoring signal of the structure to be tested is acquired and a signal data set is established.

[0035] The health monitoring signal may be, for example, an acceleration response signal d monitored by the structure to be tested, which is used to train a subsequently established generative adversarial network.

[0036] In operation S220, a first generator and a first discriminator are initialized to construct a generative adversarial network.

[0037] In an embodiment of the present invention, the network parameters of the first generator and the first discriminator are randomly initialized based on a zero-mean Gaussian distribution. The network structures of the first generator and the first discriminator may include five one-dimensional convolutional layers, a batch normalization layer, and an activation function layer in sequence, and the generative adversarial network uses an Adam optimizer to reversely update the network parameters of the first generator and the first discriminator.

[0038] In operation S230, a first characterization vector having an initial dimension is randomly generated, the first characterization vector is input into a first generator, and a generative adversarial network is preliminarily trained using a signal data set to obtain a trained second generator and a second discriminator.

[0039] For example, a larger value k0 is selected as the initial dimension of the first representation vector z, and the first generator G and the first discriminator D of the deep convolutional generative adversarial network are constructed based on the one-dimensional convolutional layer to reduce the parameter scale of the network model and prevent overfitting during the training process.

[0040] The parameters are tuned by adversarial training so that the first generator G can generate more realistic real-like signals. After the training is completed, the network parameters of the trained second generator G and the second discriminator D are recorded as .

[0041] In operation S240, a set of second characterization vectors with an initial dimension is randomly generated, the second characterization vectors are input into a second generator, a corresponding real-like signal is generated, and the similarity between the real-like signal and the signal data set is calculated as the generation capability of the second generator.

[0042] Train a generator with better generation effect After that, a set of second representation vectors with the same dimension as the first representation vector is randomly generated to generate the corresponding real-like signals. The similarity between the real-like signals and the signal dataset can be used as a quantitative index SI0 to evaluate the generation ability of the second generator.

[0043] In operation S250, the initial dimension of the second characterization vector is uniformly reduced each time, and the second characterization vector after each dimension reduction is input into the current generator. The generative adversarial network is trained for transfer learning using a feature extraction model. After each transfer learning training, it is determined whether the generation capability of the current generator and the dimension of the current second characterization vector meet the corresponding conditions. If so, the transfer learning training is stopped to obtain a trained third generator and a third discriminator.

[0044] After the aforementioned preliminary training, the generative adversarial network needs to be trained through transfer learning to obtain a third generator and a third discriminator with better generation effect.

[0045] In operation S260, a linear measurement matrix is ​​used to compress the health monitoring signal and add a random noise signal to obtain compressed sampling data.

[0046] In operation S270, a set of third characterization vectors with intrinsic dimensions is randomly generated as input of the third generator, a gradient descent algorithm is used and the third characterization vectors are iteratively optimized based on the compressed sampling data, and the optimized third characterization vectors are re-input into the third generator to generate a decompressed reconstructed signal corresponding to the health monitoring signal.

[0047] Through the embodiments of the present invention, in order to solve the problem that the existing structural health monitoring signal compression sampling method requires the original signal to have a high sparsity, when the sparsity of the signal cannot meet the requirements, problems such as low reconstruction accuracy and slow calculation speed will occur. The present invention adopts a deep convolutional generative adversarial network to learn the internal structural characteristics of the monitoring signal from the big data, breaking the high dependence of the signal decompression and reconstruction on sparsity, and gradually compresses the low-dimensional representation dimension of the network model based on the transfer learning method, so as to achieve high-precision reconstruction with only a small amount of compressed sampling data, which can be well applied to practical engineering.

[0048] Figure 4 The flowchart of the preliminary training process according to the embodiment of the present invention is schematically shown.

[0049] like Figure 4 As shown, the above operation S230 of inputting the first characterization vector into the first generator and using the signal data set to preliminarily train the generative adversarial network may specifically include operations S2301 to S2303.

[0050] In operation S2301, a first characterization vector is input into a first generator, and a generated signal is output.

[0051] The first characterization vector z is randomly generated based on a standard Gaussian distribution, which is input into a first generator, and a generated signal having the characteristics of a real signal is output.

[0052] In operation S2302, the generated signal and the health monitoring signal in the signal data set are input into a first discriminator, and a discrimination result is output, where the discrimination result is the probability that the judgment signal comes from the signal data set.

[0053] The generated signal and the real signal d in the training data set are input into the first discriminator D for identification to determine the probability that the signal comes from the training data set.

[0054] In operation S2303, the generative adversarial network is preliminarily trained by minimizing the first objective function of the first generator and the second objective function of the first discriminator, and alternately updating the network parameters of the first generator and the first discriminator, respectively.

[0055] Furthermore, the first objective function of the first generator and the second objective function of the first discriminator are calculated according to the following formulas:

[0056]

[0057]

[0058] Among them, L G represents the first objective function of the first generator; z represents the first characterization vector; Indicates that the first characterization vector obeys its data distribution law Pz; G(z) represents the generated signal output by the first generator; D(G(z)) represents the discrimination result of the generated signal output by the first discriminator; L D represents the second objective function of the first discriminator; Indicates that the health monitoring signal d obeys its data distribution law p data (d); D(d) represents the discrimination result of the health monitoring signal output by the first discriminator.

[0059] Figure 5 The flowchart of the similarity calculation process according to the embodiment of the present invention is schematically shown.

[0060] like Figure 5 As shown, the calculation of the similarity between the real signal and the signal data set in the above operation S240 may specifically include operations S2401 to S2403.

[0061] In operation S2401, a pre-trained diagnostic classification model constructed based on a deep residual shrinkage network is introduced, and data truncation is performed before classifying the signal data set to obtain a feature extraction model.

[0062] The pre-trained diagnostic classification model is an open source network model that extracts the feature of the acceleration response signal obtained by data truncation. .

[0063] In operation S2402, a feature extraction model is used to extract high-dimensional features of the real-like signal and the signal data set, and the two high-dimensional features are fitted using a multivariate normal distribution to obtain a real-like signal fitting feature and a signal data set fitting feature.

[0064] Using feature extraction model , respectively extracting high-dimensional features of real-like signals and signal datasets: Then, the multivariate normal distribution is used to fit these two high-dimensional features respectively to obtain the real signal fitting features and the signal data set fitting features: .

[0065] In operation S2403 , the Fréchet distance between the fitting feature of the real-like signal and the fitting feature of the signal data set is calculated, and the Fréchet distance is determined as the similarity between the real-like signal and the signal data set.

[0066] Fréchet distance is a description of path space similarity proposed by a French mathematician in 1906. This description also takes into account the factor of path space distance and is more applicable to the similarity of spatial paths.

[0067] Specifically, the Fréchet distance between the fitting features of the real-like signal and the fitting features of the signal dataset is calculated according to the following formula:

[0068]

[0069] In the formula, μ r and∑ r Respectively represent the expectation and covariance matrix of the signal data set fitting characteristics; μ g and∑ g They respectively represent the expectation and covariance matrix of the fitting features of the real-like signal; Tr() represents the trace of the square matrix, that is, the sum of the main diagonal elements of the square matrix; SI0 represents the Fréchet distance, that is, the generation ability of the second generator.

[0070] Please continue reading Figure 3 In an embodiment of the present invention, the generation capability of the current generator and the dimension of the current second characterization vector in the above operation S250 each satisfy corresponding conditions, which may specifically include: the error between the generation capability of the current generator and the generation capability of the second generator is less than a preset error threshold; and the dimension of the current second characterization vector reaches a specified dimension.

[0071] Among them, the error between the generation capacity of the current generator and the generation capacity of the second generator is less than the preset error threshold, that is, , where ∈ is the preset error threshold.

[0072] Furthermore, the specified dimension is the intrinsic dimension of the signal data set. Under the intrinsic dimension, the dimension of the second characterization vector cannot be further reduced.

[0073] See also Figure 3 Specifically, the above operation S250 is to reduce the dimension k0 of the second representation vector in the current generator network architecture to k i , i=1,...,p, the rest of the network architecture remains unchanged, and p is the specified number of decrements. After each transfer learning training, the trained network parameters and , use the feature extractor model in transfer learning to perform transfer learning training on the generative adversarial network, specifically adjusting the network parameters of the part connected to the current second representation vector.

[0074] After each transfer learning training, the generation capacity of the current generator is calculated , based on this, adjust the hyperparameters of transfer learning training until , we get a generator that can generate a real-like signal with a lower-dimensional second representation vector Repeat the above process until the dimension of the second representation vector is k p′ Unable to further reduce, keep the network parameters of the generator G and the discriminator D at this time ( , abbreviated as θ G ,θ D ) no longer changes, thereby obtaining the trained third generator and the third discriminator.

[0075] In the embodiment of the present invention, the number of rows of the measurement matrix is ​​less than the number of columns, and the initial value of the measurement matrix is ​​randomly generated from a zero-mean Gaussian distribution.

[0076] Please continue reading Figure 3 Specifically, the above operation S260 is based on the linear measurement matrix , the health monitoring signal monitored by the structure to be tested, which is a high-dimensional original signal , compress to obtain the preliminary compressed sampling data y = Ax, . The number of rows m of the measurement matrix is ​​much smaller than the number of columns n, so as to achieve data compression. First, from a zero-mean Gaussian distribution The random noise signal ε is added in the compressed sampling process to increase the robustness of the model. Therefore, the compressed sampling data y = Ax + ε is obtained. .

[0077] Furthermore, in order to ensure that the original signal retains certain spatial position characteristics after compression sampling, the observation matrix Specifically, without loss of generality, assuming that the number of rows and columns of the observation matrix have the greatest common divisor (m, n), it can be converted into retaining only a few sub-blocks , and the rest of the elements are set to zero:

[0078]

[0079] The number of rows in the measurement matrix A is much smaller than the number of columns, so data compression is achieved. The measurement matrix is ​​equivalent to segmented compression of the original signal, thus retaining some spatial features.

[0080] Figure 6 The flowchart of the iterative optimization process according to the embodiment of the present invention is schematically shown.

[0081] like Figure 6 As shown, the above operation S270 adopts the gradient descent algorithm and iteratively optimizes the third characterization vector based on the compressed sampling data, which may specifically include operations S2701 to S2703.

[0082] In operation S2701, the third characterization vector is input into the third generator to obtain a corresponding generated signal.

[0083] The third characterization vector z (0) It is also randomly generated based on the standard Gaussian distribution, and its dimension is k determined above. p′ .

[0084] In operation S2702, a preset objective function is established according to the generated signal and the compressed sampling data.

[0085] Specifically, the preset objective function is calculated according to the following formula:

[0086]

[0087] Where y represents the compressed sampling data; A represents the observation matrix; q represents the total number of iterations of the gradient descent algorithm; z (j) represents the third representation vector after the j-th iteration optimization; represents the quasi-real signal output by the third representation vector in the third generator after the j-th iteration optimization; ||·||2 represents the norm operation; L j Represents the preset objective function after the j-th iterative optimization.

[0088] In operation S2703, the third characterization vector is iteratively optimized by minimizing a preset objective function.

[0089] Therefore, the third characterization vector is iteratively optimized through the gradient descent algorithm to achieve the approximation of the generated signal to the real health monitoring signal. The optimal third characterization vector is selected based on minimizing the preset objective function. , and obtain the decompressed and reconstructed signal corresponding to the health monitoring signal: .

[0090] Based on the above public content, please continue to combine Figure 3 A specific embodiment of the present invention performs compression sampling on signals in two situations: sinusoidal simulation signals and actual monitoring signals. The signal sequence lengths of the sinusoidal simulation signals and actual monitoring signals used are both 512 points.

[0091] The structural health monitoring signal compression sampling method provided by the present invention is used below to perform compression sampling, decompression and reconstruction on the structural health monitoring signal.

[0092] Specifically, the analog signal x is constructed as shown in the following formula:

[0093]

[0094] Among them, A i ~U(0,1),w i ~U(1,20), , U represents uniform distribution, A i represents the amplitude, w i Indicates frequency; Indicates phase.

[0095] The actual monitoring signal is the acceleration monitoring signal of a bridge, which is obtained by dividing the signal collected by the acceleration sensor. Both signals are divided into short signals with a sequence length of 512 points to form a signal data set.

[0096] First, establish the network architecture of the first generator G and the first discriminator D. The first generator G and the first discriminator D are composed of 5 one-dimensional convolutional layers, batch normalization layers and activation function layers. Then, the first generator G inputs a 100-dimensional first representation vector z randomly sampled from a zero-mean normal distribution, and outputs a short signal with a sequence length of 512 points; the first discriminator D inputs a 512-dimensional signal vector and outputs a value between 0 and 1. The optimizer uses the Adam optimizer to initialize the training parameters, train the deep convolutional generative adversarial network, and perform parameter tuning so that the trained second generator G can generate more realistic simulation signals, and maintain the network parameters of the second generator G and the second discriminator D. No longer changes.

[0097] Next, we train a second generator with better generation effect. After that, a set of second representation vectors is randomly generated to generate the corresponding real-like signals, and the similarity between them and the signal dataset is calculated as the quantitative index SI0 of the generator's generation ability;

[0098] Then, the dimension of the second representation vector is reduced from 100 to k i , i=1,...,p, the rest of the network architecture remains unchanged. Based on the trained network parameters and , use the feature extractor mode in transfer learning to retrain the network parameters of the part connected to the low-dimensional representation vector. During the training process, the generation capacity of the current generator is calculated after each iteration. , based on this, adjust the hyperparameters of transfer learning training until , we can obtain a generator that can generate real-like signals with lower-dimensional representation vectors Repeat the above process until the dimension of the low-dimensional representation vector is k p =10, it cannot be further reduced, and the network parameters θ of the third generator G and the third discriminator D are maintained at this time. G ,θ D No longer changing;

[0099] Again, based on the zero-mean Gaussian distribution Construct and transform the linear observation matrix , where the compressed observation matrix A contains several sub-blocks , a=1,...,128, and the rest of the elements are set to zero. In order to increase the robustness of the model, noise is added during the compression sampling process. , and obtain the compressed sampling data y=Ax+ε. Among them, the number of rows of the compressed observation matrix A is 128, the number of columns is 512, and the measurement matrix A satisfies that the number of rows is less than the number of columns to achieve the data compression effect, that is, the data is compressed by 4 times.

[0100] Finally, initialize the number of random vectors to 100 and randomly sample low-dimensional vectors with a dimension of 10 from a zero-mean normal distribution. Set various optimization parameters (optimization times, learning rate, etc.). For each set of low-dimensional vectors, input the compressed data y and the signal reconstruction generator , using the objective function, the low-dimensional representation vector is iteratively optimized through the gradient descent algorithm to achieve the approximation of the generated signal to the observed data. The optimal low-dimensional representation vector is selected based on the minimum objective function value , and get the decompressed reconstructed signal .

[0101] Figures 7(a1) and 7(a2) schematically show the reconstruction effect diagram of the simulated sinusoidal signal according to an embodiment of the present invention. Figures 7(b1) and 7(b2) schematically show the reconstruction effect diagram of the health monitoring signal of the actual structure according to an embodiment of the present invention.

[0102] Among them, Figure 7(a1) and Figure 7(b1) are both original high-dimensional signals, which are simulated sinusoidal signals and health monitoring signals, respectively, and Figure 7(a2) and Figure 7(b2) are both reconstructed signals, which are reconstructed signals corresponding to the simulated sinusoidal signals and health monitoring signals, respectively.

[0103] By comparing the reconstructed signal with the original high-dimensional signal, it can be seen that the method provided by the present invention can accurately reconstruct a quasi-real signal that retains the characteristics of the original high-dimensional signal at a high compression rate, and the reconstruction result has good robustness.

[0104] In summary, the method proposed in the present invention specifically includes the design of the compressed sampling observation matrix of the deep convolutional generative adversarial network compressed by transfer learning, the network architecture design of the deep convolutional generative adversarial network, the setting of network model parameters, the application of transfer learning, the determination of the optimal low-dimensional representation vector, etc. The signal reconstruction method based on the generative network model of the present invention not only gets rid of the sparsity requirement for the original signal, but also compresses the low-dimensional representation dimension of the network model through transfer learning, greatly improving the reconstruction accuracy and compression ratio of the signal. At the same time, the method of deep convolutional generative adversarial network adopts more convolution and transposed convolution layers, has fewer parameters, and has a faster reconstruction speed. The model transforms the observation matrix during the compressed sampling process, can retain some spatial features, is easier to decompress and reconstruct, and has a higher reconstruction accuracy.

[0105] Based on the above structural health monitoring signal compression sampling method, the present invention also provides a structural health monitoring signal compression sampling device. Figure 8 The device is described in detail.

[0106] Figure 8 The block diagram of the structural health monitoring signal compression sampling device according to an embodiment of the present invention is schematically shown.

[0107] like Figure 8 As shown, the structural health monitoring signal compression sampling device 800 of this embodiment includes a monitoring signal acquisition module 810, a network structure construction module 820, a preliminary training module 830, a generation capability calculation module 840, a transfer learning training module 850, a compression sampling module 860 and a decompression and reconstruction module 870.

[0108] The monitoring signal acquisition module 810 is used to acquire the health monitoring signal of the structure to be tested and establish a signal data set.

[0109] The network structure construction module 820 is used to initialize the first generator and the first discriminator to construct a generative adversarial network.

[0110] The preliminary training module 830 is used to randomly generate a first characterization vector with an initial dimension, input the first characterization vector into the first generator, and use the signal data set to preliminarily train the generative adversarial network to obtain a trained second generator and a second discriminator.

[0111] The generation capability calculation module 840 is used to randomly generate a set of second characterization vectors with an initial dimension, input the second characterization vectors into the second generator, generate corresponding real-like signals, and calculate the similarity between the real-like signals and the signal data set as the generation capability of the second generator.

[0112] The transfer learning training module 850 is used to uniformly reduce the initial dimension of the second characterization vector gradually, input the second characterization vector after each dimension reduction into the current generator, and use the feature extraction model to perform transfer learning training on the generative adversarial network. After each transfer learning training, it is determined whether the generation ability of the current generator and the dimension of the current second characterization vector meet the corresponding conditions. If so, the transfer learning training is stopped to obtain a trained third generator and a third discriminator.

[0113] The compression sampling module 860 is used to compress the health monitoring signal using a linear observation matrix and add a random noise signal to obtain compressed sampling data.

[0114] The decompression and reconstruction module 870 is used to randomly generate a set of third characterization vectors with intrinsic dimensions as input to the third generator, adopt a gradient descent algorithm and iteratively optimize the third characterization vectors based on compressed sampling data, and re-input the optimized third characterization vectors into the third generator to generate a decompressed and reconstructed signal corresponding to the health monitoring signal.

[0115] It should be noted that the implementation method of the device part is similar to the implementation method of the method part, and the technical effects achieved are also similar. For specific details, please refer to the above-mentioned method implementation method part, which will not be repeated here.

[0116] According to an embodiment of the present invention, any multiple modules among the monitoring signal acquisition module 810, the network structure construction module 820, the preliminary training module 830, the generation capability calculation module 840, the transfer learning training module 850, the compression sampling module 860 and the decompression reconstruction module 870 can be combined into one module for implementation, or any one of the modules can be split into multiple modules. Alternatively, at least part of the functions of one or more of these modules can be combined with at least part of the functions of other modules and implemented in one module. According to an embodiment of the present invention, at least one of the monitoring signal acquisition module 810, the network structure construction module 820, the preliminary training module 830, the generation capacity calculation module 840, the transfer learning training module 850, the compression sampling module 860 and the decompression reconstruction module 870 can be at least partially implemented as a hardware circuit, such as a field programmable gate array (FPGA), a programmable logic array (PLA), a system on a chip, a system on a substrate, a system on a package, an application-specific integrated circuit (ASIC), or can be implemented by hardware or firmware such as any other reasonable way of integrating or packaging the circuit, or in any one of the three implementation methods of software, hardware and firmware or in any appropriate combination of any of them. Alternatively, at least one of the monitoring signal acquisition module 810, the network structure construction module 820, the preliminary training module 830, the generation capacity calculation module 840, the transfer learning training module 850, the compression sampling module 860 and the decompression reconstruction module 870 can be at least partially implemented as a computer program module, and when the computer program module is run, the corresponding function can be executed.

[0117] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present invention. In this regard, each box in the flow chart or block diagram can represent a module, a program segment, or a part of a code, and the above-mentioned module, program segment, or a part of a code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram or flow chart, and the combination of the boxes in the block diagram or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0118] It will be appreciated by those skilled in the art that the features described in the various embodiments and / or claims of the present invention may be combined and / or combined in various ways, even if such combinations and / or combinations are not explicitly described in the present invention. In particular, the features described in the various embodiments and / or claims of the present invention may be combined and / or combined in various ways without departing from the spirit and teachings of the present invention. All of these combinations and / or combinations fall within the scope of the present invention.

[0119] The embodiments of the present invention are described above. However, these embodiments are only for the purpose of illustration, and are not intended to limit the scope of the present invention. Although each embodiment is described above, it does not mean that the measures in each embodiment cannot be used in combination. The scope of the present invention is defined by the attached claims and their equivalents. Without departing from the scope of the present invention, those skilled in the art may make various substitutions and modifications, which should all fall within the scope of the present invention.

Claims

1. A structural health monitoring signal compression sampling method, characterized in that: include: Obtain health monitoring signals of the structure to be tested and establish a signal data set; Initialize the first generator and the first discriminator, and construct a generative adversarial network, wherein the network structures of the first generator and the first discriminator each include five one-dimensional convolutional layers, a batch normalization layer, and an activation function layer in sequence, and the generative adversarial network uses an Adam optimizer to reversely update the network parameters of the generator and the discriminator; Randomly generate a first characterization vector with an initial dimension, input the first characterization vector into the first generator, and preliminarily train the generative adversarial network using the signal data set to obtain a trained second generator and a second discriminator; Randomly generate a set of second characterization vectors with the initial dimension, input the second characterization vectors into the second generator, generate corresponding real-like signals, and calculate the similarity between the real-like signals and the signal data set as the generation capability of the second generator; The initial dimension of the second characterization vector is uniformly reduced gradually, and the second characterization vector after each dimension reduction is input into the current generator, and the generative adversarial network is trained by transfer learning using a feature extraction model. After each transfer learning training, it is determined whether the generation capability of the current generator and the dimension of the current second characterization vector meet corresponding conditions respectively. If so, the transfer learning training is stopped to obtain a trained third generator and a third discriminator; Using a linear measurement matrix, compressing the health monitoring signal and adding a random noise signal to obtain compressed sampling data; Randomly generate a set of third characterization vectors with intrinsic dimensions as inputs of the third generator, adopt a gradient descent algorithm and iteratively optimize the third characterization vectors based on the compressed sampling data, and re-input the optimized third characterization vectors into the third generator to generate a decompressed reconstructed signal corresponding to the health monitoring signal; The calculating of the similarity between the real-like signal and the signal data set specifically includes: A pre-trained diagnostic classification model based on a deep residual shrinkage network is introduced, and data truncation is performed before classifying the signal data set to obtain a feature extraction model; Using the feature extraction model to extract high-dimensional features of the real-like signal and the signal data set respectively, and using multivariate normal distribution to fit the two high-dimensional features respectively, to obtain real-like signal fitting features and signal data set fitting features; Calculating the Fréchet distance between the fitting feature of the real-like signal and the fitting feature of the signal data set, and determining the Fréchet distance as the similarity between the real-like signal and the signal data set; The Fréchet distance between the real signal fitting feature and the signal data set fitting feature is calculated according to the following formula: in, and Respectively represent the expectation and covariance matrices of the fitted features of the signal dataset; and They represent the expectation and covariance matrices of the fitting features of the real-like signal respectively; Tr() represents the trace of the square matrix, that is, the sum of the main diagonal elements of the square matrix; represents the Fréchet distance, that is, the generation ability of the second generator.

2. The structural health monitoring signal compression sampling method according to claim 1, characterized in that: The health monitoring signal is an acceleration response signal monitored by the structure to be tested.

3. The structural health monitoring signal compression sampling method according to claim 1, characterized in that: The network parameters of the generator and the discriminator are randomly initialized based on a zero-mean Gaussian distribution, and the first characterization vector and the third characterization vector are randomly generated based on a standard Gaussian distribution.

4. The structural health monitoring signal compression sampling method according to claim 1, characterized in that: The step of inputting the first representation vector into the first generator and preliminarily training the generative adversarial network using the signal data set specifically includes: Inputting the first characterization vector into a first generator and outputting a generated signal; Inputting the generated signal and the health monitoring signal in the signal data set into a first discriminator, and outputting a discrimination result, wherein the discrimination result is a probability that the judgment signal comes from the signal data set; The generative adversarial network is preliminarily trained by minimizing a first objective function of the first generator and a second objective function of the first discriminator and alternately updating network parameters of the first generator and the first discriminator.

5. The method for compressing and sampling structural health monitoring signals according to claim 4, characterized in that: The first objective function of the first generator and the second objective function of the first discriminator are calculated according to the following formulas: in, represents the first objective function of the first generator; z represents the first characterization vector; Indicates that the first characterization vector obeys its data distribution law Pz; a generated signal representing an output of the first generator; represents a discrimination result of the generated signal output by the first discriminator; represents the second objective function of the first discriminator; Indicates that the health monitoring signal d obeys its data distribution law ; Represents the discrimination result of the health monitoring signal output by the first discriminator.

6. The structural health monitoring signal compression sampling method according to claim 1, characterized in that: The generation capability of the current generator and the dimension of the current second characterization vector each satisfy corresponding conditions, specifically including: The error between the generation capability of the current generator and the generation capability of the second generator is less than a preset error threshold; The dimension of the current second characterization vector reaches the specified dimension.

7. The method for compressive sampling of structural health monitoring signals according to claim 6, characterized in that: The specified dimension is the intrinsic dimension of the signal data set. Under the intrinsic dimension, the dimension of the second characterization vector cannot be further reduced.

8. The method for compressive sampling of structural health monitoring signals according to claim 1, characterized in that: The number of rows of the measurement matrix is ​​smaller than the number of columns, and the initial value of the measurement matrix is ​​randomly generated from a zero-mean Gaussian distribution.

9. The method for compressive sampling of structural health monitoring signals according to claim 1, characterized in that: The adopting a gradient descent algorithm and iteratively optimizing the third characterization vector based on the compressed sampling data specifically includes: Inputting the third characterization vector into the third generator to obtain a corresponding generated signal; Establishing a preset objective function according to the generated signal and the compressed sampling data; The third characterization vector is iteratively optimized by minimizing the preset objective function.

10. The structural health monitoring signal compression sampling method according to claim 9, characterized in that: The preset objective function is calculated according to the following formula: ,j=0,1,……,q Where y represents the compressed sampling data; A represents the observation matrix; q represents the total number of iterations of the gradient descent algorithm; represents the third representation vector after the j-th iteration optimization; represents the real-like signal output by the third representation vector in the third generator after the j-th iteration optimization; represents norm operation; Represents the preset objective function after the j-th iterative optimization.

11. A structural health monitoring signal compression sampling device, characterized in that: include: A monitoring signal acquisition module is used to obtain the health monitoring signal of the structure to be tested and establish a signal data set; A network structure construction module, used to initialize the first generator and the first discriminator, and construct a generative adversarial network, wherein the network structures of the first generator and the first discriminator each include five one-dimensional convolutional layers, a batch normalization layer, and an activation function layer in sequence, and the generative adversarial network uses an Adam optimizer to reversely update the network parameters of the generator and the discriminator; A preliminary training module, used for randomly generating a first characterization vector with an initial dimension, inputting the first characterization vector into the first generator, and using the signal data set to preliminarily train the generative adversarial network to obtain a trained second generator and a second discriminator; a generation capability calculation module, configured to randomly generate a set of second characterization vectors having the initial dimension, input the second characterization vectors into the second generator, generate corresponding real-like signals, and calculate the similarity between the real-like signals and the signal data set as the generation capability of the second generator; A transfer learning training module, used to uniformly reduce the initial dimension of the second characterization vector gradually, input the second characterization vector after each dimension reduction into the current generator, use the feature extraction model to perform transfer learning training on the generative adversarial network, and after each transfer learning training, determine whether the generation capability of the current generator and the dimension of the current second characterization vector meet the corresponding conditions respectively. If so, stop the transfer learning training to obtain a trained third generator and a third discriminator; A compression sampling module, used to compress the health monitoring signal using a linear observation matrix and add a random noise signal to obtain compressed sampling data; a decompression and reconstruction module, configured to randomly generate a set of third characterization vectors having an intrinsic dimension as inputs of the third generator, iteratively optimize the third characterization vectors based on the compressed sampling data using a gradient descent algorithm, and re-input the optimized third characterization vectors into the third generator to generate a decompressed and reconstructed signal corresponding to the health monitoring signal; The calculating of the similarity between the real-like signal and the signal data set specifically includes: A pre-trained diagnostic classification model based on a deep residual shrinkage network is introduced, and data truncation is performed before classifying the signal data set to obtain a feature extraction model; Using the feature extraction model to extract high-dimensional features of the real-like signal and the signal data set respectively, and using multivariate normal distribution to fit the two high-dimensional features respectively, to obtain real-like signal fitting features and signal data set fitting features; Calculating the Fréchet distance between the fitting feature of the real-like signal and the fitting feature of the signal data set, and determining the Fréchet distance as the similarity between the real-like signal and the signal data set; The Fréchet distance between the real signal fitting feature and the signal data set fitting feature is calculated according to the following formula: in, and Respectively represent the expectation and covariance matrices of the fitted features of the signal dataset; and They represent the expectation and covariance matrices of the fitting features of the real-like signal respectively; Tr() represents the trace of the square matrix, that is, the sum of the main diagonal elements of the square matrix; represents the Fréchet distance, that is, the generation ability of the second generator.

Images