Performance debugging method for microwave filter fault diagnosis based on graph embedding

Abstract

The application discloses a kind of based on graph embedding microwave filter fault diagnosis's performance debugging method, first, the admittance parameter feature extraction network of complex convolution coding subnetwork and complex deconvolution decoding subnetwork cascade is built, then combined with the microwave filter structure to be diagnosed and debugged graph convolution inversion network is built, generates admittance parameter training set and scattering parameter training set, finally alternately trains admittance parameter feature extraction network and graph convolution inversion network, carries out fault diagnosis and debugging to microwave filter.The present application is an unsupervised model, which reduces the difficulty of collecting a large number of data sets;And in the process of fault diagnosis and debugging, without any iterative optimization process, a large amount of diagnosis and debugging time can be saved;The model constructed in combination with the physical structure of the filter can be used for fault diagnosis and debugging of microwave filters of any structure, and has higher applicability.

Description

Technical Field

[0001] This invention belongs to the field of physical technology, and more specifically relates to a performance tuning method for microwave filter fault diagnosis based on graph embedding in the field of electrical digital data processing technology. This invention can diagnose faults in microwave filters using graph embedding, and based on the fault diagnosis, tune the electrical performance of the microwave filter. Background Technology

[0002] Microwave filters, mainly including cavity filters and LC filters, are key frequency selection devices in wireless communication base stations and satellite communication systems. They are widely used in the RF front-end of satellite and wireless communication systems, and their filtering performance determines the frequency selection quality. In the actual design and production of microwave filters in engineering practice, due to physical dimensional errors in design software and manufacturing processes, as well as assembly errors of debugging components during manufacturing, the output response of microwave filters often fails to perfectly match the design requirements. This results in electrical performance not meeting design and service requirements, i.e., filter failure. Such failures severely affect the filtering performance of filters. Therefore, rapid filter fault diagnosis and debugging based on the diagnosed faults are indispensable links in the production, manufacturing, and service of microwave filters. The coupling matrix is ​​a key intermediate parameter for filter fault diagnosis. Computer-aided diagnostic methods based on the Cauchy method are prone to ill-conditioned matrices when fitting high-order filter coupling matrices, making it difficult to diagnose and quickly debug filters of order 10 and above. Furthermore, computer-aided diagnostic methods based on vector fitting require multiple iterative fittings to construct the coupling matrix, and the time for each iteration increases with the filter order. In addition, the filter coupling matrix constructed using the above methods requires corresponding folding and rotation to obtain the coupling matrix corresponding to the designed topology. Machine learning-based methods require constructing corresponding adjustment values ​​as labels, and many methods convert scattering parameters into real numbers before feature extraction, thus losing potential features.

[0003] China University of Geosciences (Wuhan) disclosed a method for debugging microwave filters using a multi-feature fusion modeling approach in its patent application "A Multi-Feature Fusion Modeling and Debugging Method for Microwave Filters" (Application No.: 202110128530.4, Publication No.: CN 113158541 A). The method's implementation steps are as follows: First, the adjustable components of the microwave filter are debugged, and the corresponding debugging amounts are measured. Simultaneously, the scattering parameters of the filter under this debugging state are measured using a Vector Network Analyzer (VNA), generating a dataset of scattering parameters and debugging amounts. Then, a convolutional neural network model is built, and the generated dataset is used to train the convolutional neural network model, realizing the mapping between the filter's scattering parameters and debugging amounts. Finally, the microwave filter is debugged based on the trained convolutional neural network model. The shortcomings of this method are as follows: since the convolutional neural network debugging model is a supervised training model, the debugging amount of each filter needs to be measured as a label when training the debugging model. However, measuring the debugging amount in each debugging state requires a lot of manpower, making it difficult to create a large dataset to train the debugging model, which makes it difficult to implement in practical applications. In addition, the feature extraction units used in the debugging model built by this method are all real number convolutional layers, while the scattering parameters are complex numbers, which leads to the loss of some information during the feature extraction process. These two shortcomings reduce the mapping accuracy of the entire model and decrease the debugging efficiency.

[0004] Suzhou Xinmai Intelligent Technology Co., Ltd. disclosed a method for debugging the performance of a cavity filter using a fused knowledge graph in its patent application "An Intelligent Auxiliary Debugging Method and Device for a Cavity Filter" (Application No.: 202110609307.1, Publication No.: CN 113255222 A). The method's implementation steps are as follows: First, the cavity debugging amount and scattering parameters are associated through a knowledge graph, along with the corresponding filter characteristics (model, batch, etc.); then, the residual between each adjustment amount is calculated, and the coupling matrix of the scattering parameters is solved using circuit synthesis methods. A residual network is constructed to map the residual and the coupling matrix; finally, the adjustment amount corresponding to the scattering parameters is solved through this mapping relationship to achieve filter debugging. The shortcomings of this method are that, since the knowledge graph used is only used to associate all scattering parameters, tuning parameters, and intrinsic characteristics, without considering the filter's own cavity design structure, it cannot solve for the coupling matrix of filters with arbitrary topologies. In addition, the traditional circuit synthesis method is used to solve the coupling matrix, which requires matrix decomposition and a large number of optimization iterations. These two shortcomings make it easy to produce ill-conditioned matrices when calculating high-order filter tuning parameters, and the large number of iterations increases the tuning time, which cannot meet the requirements of real-time testing and adjustment. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of the existing technologies mentioned above by proposing a performance debugging method for fault diagnosis of graph-embedded microwave filters. This method solves the problems of difficulty in generating large datasets, loss of some complex features during scattering parameter feature extraction, inability to solve the coupling matrix of arbitrary structure filters, ill-conditioned coupling matrix of high-order filters, and long debugging time caused by numerous optimization iterations.

[0006] The approach to achieving the objective of this invention is to construct an admittance parameter feature extraction network and a scattering parameter generation network based on graph embedding principles, thereby enabling filter fault diagnosis and performance tuning. First, the filter to be diagnosed and tuned is abstracted as an undirected graph, where each cavity of the filter corresponds to a node in the undirected graph, and each coupling path of the filter corresponds to an edge in the undirected graph. Then, an admittance parameter feature extraction network, comprising a complex convolutional coding sub-network and a complex deconvolutional decoding sub-network, is constructed to embed features for the nodes corresponding to each cavity of the filter to be diagnosed and tuned. Finally, a scattering parameter generation network, comprising a graph convolutional inversion sub-network and a physical evolution sub-network, is constructed. Based on the feature embedding of each node, the graph convolutional inversion sub-network calculates the coupling matrix corresponding to the current tuning state for fault diagnosis. The physical evolution sub-network further calculates the scattering parameters and tuning parameters corresponding to the current tuning state using the coupling matrix for filter tuning. Since both the admittance parameter feature extraction network and the scattering parameter generation network in this invention are unsupervised networks, they can be directly trained using the constructed dataset. No additional data labels are required during training, solving the problem of difficulty in creating large datasets in existing technologies. Because the complex convolutional encoding subnetwork and complex deconvolutional decoding subnetwork in the admittance parameter feature extraction network constructed by the method of this invention both employ complex convolutional and complex deconvolutional modules, complex features can be extracted during the feature extraction process without converting the scattering parameters to real numbers, thus solving the problem of partial loss of complex features during scattering parameter feature extraction. Since the graph convolutional inversion subnetwork in the scattering parameter generation network constructed by the method of this invention integrates the physical structure of the filter, generating an adjacency matrix with the filter's debugging cavity as nodes, it solves the problem of not being able to solve the coupling matrix of arbitrary structure filters and addresses the problem of ill-conditioned matrices when calculating high-order filter coupling matrices. Since both the admittance parameter feature extraction network and the scattering parameter generation network in the method of this invention are neural network models, after training, filter fault diagnosis involves matrix operations, eliminating the need for additional iterative optimization processes and solving the problem of long debugging times caused by extensive optimization iterations required in existing methods.

[0007] The specific steps for implementing the method of the present invention include the following:

[0008] Step 1, Construct an admittance parameter feature extraction network:

[0009] Step 1.1: Construct a complex convolutional coding subnetwork consisting of three cascaded complex convolutional modules with identical structures. Each complex convolutional module consists of two cascaded complex convolutional layers.

[0010] The kernel size of the two complex convolutional layers in the first complex convolutional module is set to 3×1, the number of channels is set to 64 and 128 respectively, the stride is set to 1, and the activation function is implemented by the CRelu function.

[0011] In the second complex convolution module, the kernel size of the two complex convolutional layers is set to 3×1, the number of channels is set to 256 and 128 respectively, the stride is set to 1, and the activation function is implemented by the CRelu function.

[0012] In the third complex convolution module, the kernel size of the two complex convolutional layers is set to 3×1, the number of channels is set to 64 and n+2 respectively, the value of n is equal to the order of the filter, the stride is set to 1, and the activation function is implemented by the CRelu function.

[0013] Step 1.2: Construct a complex deconvolution decoding subnetwork consisting of three cascaded complex deconvolution modules. Each of the three complex deconvolution modules consists of two cascaded complex deconvolution layers.

[0014] In the first complex deconvolution module, the kernel size of the two complex deconvolution layers is set to 3×1, the number of channels is set to 64 and 128 respectively, the stride is set to 1, and the activation function is implemented by the CRelu function.

[0015] In the second complex deconvolution module, the kernel size of the two complex deconvolution layers is set to 3×1, the number of channels is set to 256 and 128 respectively, the stride is set to 1, and the activation function is implemented by the CRelu function.

[0016] In the third complex deconvolution module, the kernel size of the two complex deconvolution layers is set to 3×1, the number of channels is set to 64 and 3 respectively, the stride is set to 1, and the activation function is implemented by the CRelu function.

[0017] Step 1.3: Concatenate the complex convolutional coding subnetwork and the complex deconvolutional decoding subnetwork to form an admittance parameter feature extraction network;

[0018] Step 2, construct the graph convolutional inversion network:

[0019] Step 2.1: Abstract the microwave filter into an undirected graph. The source end, load end, and each cavity of the filter correspond to nodes in the undirected graph, and each mutually coupled path and cross-coupled path corresponds to an edge in the undirected graph. This results in an undirected graph corresponding to the filter structure, and the adjacency matrix of the undirected graph is generated.

[0020] Step 2.2: Construct a graph convolution inversion network consisting of three cascaded graph convolutional layers. Use undirected graph adjacency matrices in the adjacency matrices of the first to third graph convolutional layers. Set the number of nodes to n1+2 and the number of channels to n1+2, 32, and n1+2 respectively. The value of n1 is equal to that of n.

[0021] Step 3, Generate the training set:

[0022] Step 3.1: Randomly adjust each cavity of a microwave filter at least 30 times, and simultaneously collect the scattering parameters for each adjustment using a VNA, obtaining 30×n² sets of scattering parameters, where n² is equal to n; the forward reflection coefficient S in each set of scattering parameters... 11 and forward transmission coefficient S 21 The scattering parameter set that makes up this group, and the 30×n2 groups of scattering parameter sets are combined to form the scattering parameter training set D. S ;

[0023] Step 3.2: Convert each set of scattering parameters into admittance parameters, obtaining 30×n² sets of converted admittance parameters. Then, extract the positive input admittance Y from each set of converted admittance parameters. 11 Positive transfer admittance Y 21 The set of admittance parameters, along with the corresponding sampling frequency f, constitutes this set of admittance parameters. These 30×n² sets of admittance parameters form the admittance parameter training set D. Y ;

[0024] Step 4: Alternately train the admittance parameter feature extraction network and the graph convolutional inversion network:

[0025] Step 4.1, train the admittance parameter training set D Y The input is fed into the admittance parameter feature extraction network, the complex convolutional coding sub-network outputs the admittance features, and the complex deconvolutional coding sub-network outputs the predicted admittance parameters; the training set D is used to extract the admittance parameters. Y As the admittance parameter label for the admittance parameter feature extraction network, the mean squared error loss function is selected, the loss Loss1 between all predicted admittance parameters and the corresponding input admittance parameter label is calculated, and the parameters in each layer of the admittance parameter feature extraction network are updated using the backpropagation gradient descent method.

[0026] Step 4.2: Input the admittance features output by the complex convolutional coding subnetwork into the graph convolutional inversion network, and output the training set D of all admittance parameters. YThe corresponding prediction coupling matrix is ​​used to calculate the training set D of all admittance parameters. Y The corresponding predicted scattering parameters; the training set D S As the scattering parameter labels for the graph convolutional inversion network, the mean square error loss function is selected, and the loss Loss2 between all predicted scattering parameters and the corresponding input scattering parameter labels is calculated. Then, the parameters in each layer of the graph convolutional inversion network are updated using the backpropagation gradient descent method.

[0027] Step 4.3: Repeat steps 4.1 and 4.2 until both Loss1 and Loss2 converge, to obtain the trained admittance parameter feature extraction network and graph convolution inversion network.

[0028] Step 5: Perform fault diagnosis and debugging on the microwave filter to be diagnosed and debugged:

[0029] Step 5.1: Measure the scattering parameters of the microwave filter to be fault-diagnosed and debugged using a VNA. Following the same method as in Step 3.2, convert the measured scattering parameters into admittance parameters and then input them into the trained admittance parameter feature extraction network to output admittance features. Input the admittance features into the trained graph convolution inversion network to output the coupling matrix M corresponding to the measured scattering parameters. Calculate the coupling matrix between M and the microwave filter design. The difference is used to obtain the difference matrix;

[0030] Step 5.2: Determine whether the absolute values ​​of all elements in the difference matrix are all less than ε, where ε is the error value set according to the actual filtering requirements in the range of [0.05, 0.3]. If yes, proceed to step 6; otherwise, determine that the microwave filter is faulty and proceed to step 5.3.

[0031] Step 5.3: Using the coupling matrix M and the spatial mapping method, calculate the adjustment amount of the current state of the microwave filter. Adjust the microwave filter according to the adjustment amount and then execute step 5.1.

[0032] Step 6: Determine that the microwave filter is fault-free, end the debugging process, and obtain a fault-free microwave filter.

[0033] Compared with the prior art, the present invention has the following advantages:

[0034] First, this invention constructs an admittance parameter feature extraction network consisting of a cascaded complex convolutional coding sub-network and a complex deconvolutional decoding sub-network. All feature extraction modules in this network are complex convolutional and complex deconvolutional modules, overcoming the problem of partial feature loss in existing methods when extracting complex features. Since it is no longer necessary to convert the scattering parameters to real numbers, the admittance parameter feature extraction network can directly extract complex features, preserving them and overcoming the ill-conditioned nature of high-order filter coupling matrices in existing technologies. This enables fault diagnosis and performance tuning of high-order filters.

[0035] Secondly, this invention constructs a graph convolutional inversion network based on the structure of the microwave filter to be diagnosed and debugged. The filter to be diagnosed and debugged is abstracted as an undirected graph, with each cavity of the filter corresponding to a node in the undirected graph, and each coupling path of the filter corresponding to an edge in the undirected graph. This overcomes the problem in existing technologies of being unable to extract the coupling matrix of filters with arbitrary structures, enabling this invention to achieve fault diagnosis and performance debugging of filters with arbitrary structures, thus broadening its applicability.

[0036] Third, the admittance parameter feature extraction network and graph convolution inversion network constructed in this invention are unsupervised networks that do not require additional data labels during training and can be trained directly using the input data. This overcomes the difficulty of generating large datasets in existing technologies, making the dataset generation process simple and feasible, and reducing the cost of training the admittance parameter feature extraction network and scattering parameter generation network. Furthermore, after the admittance parameter feature extraction network and graph convolution inversion network are trained, all computational diagnostic processes are matrix calculations, eliminating the need for additional iterative optimization processes. This overcomes the problem of existing methods requiring extensive optimization and iterative computation during debugging, resulting in shorter computation time and higher real-time performance during debugging. Attached Figure Description

[0037] Figure 1 This is a flowchart of the method of the present invention;

[0038] Figure 2 This is a schematic diagram of the adjacency matrix constructed in conjunction with the physical structure of the filter in an embodiment of the present invention, wherein, Figure 2 (a) is a schematic diagram of the sixth-order cavity filter to be debugged. Figure 2 (b) is an abstract undirected graph of the sixth-order cavity filter to be debugged;

[0039] Figure 3 This is a schematic diagram illustrating the connection method between the admittance parameter feature extraction network and the graph convolution inversion network in an embodiment of the present invention. Detailed Implementation

[0040] The technical solution and effects of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0041] Reference Figure 1 A sixth-order cavity filter is selected as an example for fault diagnosis and performance debugging. The implementation steps of the present invention will be further described.

[0042] Step 1: Construct an admittance parameter feature extraction network.

[0043] Step 1.1: Construct a complex convolutional encoding subnetwork consisting of three complex convolutional modules, each of which is composed of two cascaded complex convolutional layers.

[0044] In the first complex convolution module, the kernel size of the two complex convolutional layers is 3×1, the number of channels is 64 and 128 respectively, the stride is 1, and the activation function is implemented by the CRelu function.

[0045] In the second complex convolution module, the kernel size of the two complex convolutional layers is 3×1, the number of channels is 256 and 128 respectively, the stride is 1, and the activation function is implemented by the CRelu function.

[0046] In the third complex convolution module, the kernel size of the two complex convolutional layers is 3×1, and the number of channels is 64 and n+2 respectively. The value of n depends on the order of the filter. In this embodiment of the invention, the value of n is 6, the stride is 1, and the activation function is implemented by the CRelu function.

[0047] Step 1.2: Construct a complex deconvolution decoding subnetwork consisting of three cascaded complex deconvolution modules. Each of the three complex deconvolution modules is composed of two cascaded complex deconvolution layers.

[0048] In the first complex deconvolution module, the kernel size of the two complex deconvolution layers is 3×1, the number of channels is 64 and 128 respectively, the stride is 1, and the activation function is implemented by the CRelu function.

[0049] In the second complex deconvolution module, the kernel size of the two complex deconvolution layers is 3×1, the number of channels is 256 and 128 respectively, the stride is 1, and the activation function is implemented by the CRelu function.

[0050] In the third complex deconvolution module, the kernel size of the two complex deconvolution layers is 3×1, the number of channels is 64 and 3 respectively, the stride is 1, and the activation function is implemented by the CRelu function.

[0051] Step 1.3: Concatenate the complex convolutional coding subnetwork and the complex deconvolutional decoding subnetwork to form an admittance parameter feature extraction network.

[0052] Step 2: Construct a graph convolutional inversion network.

[0053] Reference Figure 2 The process of constructing a graph convolutional inversion subnetwork is described in detail. Figure 2 (a) is a schematic diagram of a sixth-order cavity microwave filter in an embodiment of the present invention. Figure 2 In (a), S and L represent the source and load terminals of the filter, respectively, and A, B, C, D, E, and F represent the six tuning cavities of the filter. Each adjacent cavity of the filter has a mutual coupling path, and there is a cross-coupling path between cavity C and cavity F. This filter can be abstracted as... Figure 2 (b) represents an undirected graph, where the source, load, and each cavity of the filter correspond to nodes in the undirected graph, and each coupled path and cross-coupled path corresponds to an edge in the undirected graph.

[0054] Construct a graph convolutional inversion network consisting of three cascaded graph convolutional layers. The adjacency matrices in the first to third graph convolutional layers are all 8×8 undirected graph adjacency matrices, with 8 nodes in each layer and 8, 32, and 8 channels respectively. Figure 3 The diagram shows the connection between the admittance parameter feature extraction network and the graph convolution inversion network. The input of the graph convolution inversion network is the output of the complex convolutional coding subnetwork in the admittance parameter feature extraction network.

[0055] The rows of the undirected graph adjacency matrix represent the node numbers in the undirected graph corresponding to the sixth-order cavity microwave filter. Each column represents the connection relationship between nodes in the undirected graph corresponding to the filter. The matrix is ​​generated as follows: if there is an edge between nodes in the undirected graph corresponding to the filter, the corresponding element in the adjacency matrix is ​​1; otherwise, the corresponding element is 0. The resulting undirected graph adjacency matrix for the sixth-order cavity microwave filter is as follows:

[0056]

[0057] Among them, A a Let S represent the adjacency matrix of the undirected graph corresponding to the sixth-order cavity microwave filter, where S, L, A, B, C, D, E, and F represent the node numbers of the undirected graph corresponding to the sixth-order cavity microwave filter.

[0058] Step 3: Generate the training dataset.

[0059] Step 3.1: To ensure the convergence of the constructed admittance parameter feature extraction network and graph convolution inversion network, achieving higher fault diagnosis accuracy and better debugging results, the sixth-order cavity microwave filter debugging bolts in this embodiment are randomly adjusted 300 times. Simultaneously, scattering parameters are collected each time using a VNA, resulting in 300 sets of scattering parameters. The forward reflection coefficient S in each set of scattering parameters is... 11 and forward transmission coefficient S 21 This set of scattering parameters is composed of 300 sets of scattering parameters, forming a scattering parameter training set D. S .

[0060] Step 3.2: Convert each set of scattering parameters to admittance parameters according to the following formula to obtain 300 sets of converted admittance parameters:

[0061]

[0062]

[0063] Among them, Y 11 i and Y 21 i Let S represent the positive input admittance and positive transfer admittance in the admittance parameter group after the scattering parameter transformation of the i-th group, i = 1, 2, 3, ..., 300. 11 i S 22 i S 21 i S 12 i Let represent the forward reflection coefficient, reverse reflection coefficient, forward transmission coefficient, and reverse transmission coefficient in the i-th group of scattering parameters, respectively.

[0064] The positive input admittance Y in each set of transformed admittance parameters 11 Positive transfer admittance Y 21 The set of admittance parameters, along with the corresponding sampling frequency f, constitutes this set of admittance parameters. These 300 sets of admittance parameters form the admittance parameter training set D. Y .

[0065] Step 4: Alternately train the admittance parameter feature extraction network and the graph convolution inversion network.

[0066] Step 4.1: Set the training hyperparameters. Set the batch size to 20 and the learning rate Lr to 0.00001.

[0067] Step 4.2, train the admittance parameter training set D YThe admittance parameter feature extraction network is fed into the complex convolutional coding sub-network, and the admittance features output by the complex deconvolutional coding sub-network are then fed into the complex deconvolutional coding sub-network to obtain the predicted admittance parameters. The admittance parameter feature extraction network is an unsupervised learning network that uses the training set D... Y As the admittance parameter labels for the admittance parameter feature extraction network, the mean squared error loss function is selected as the loss function, and the loss Loss1 between all predicted admittance parameters and the corresponding input admittance parameter labels is calculated.

[0068] The mean squared error loss function is as follows:

[0069]

[0070] Where Loss1 represents the loss between all predicted admittance parameters and their corresponding input admittance parameter labels, ∑ represents the summation operation, |·| represents the absolute value operation, and N represents the admittance parameter training set D input to the admittance parameter feature extraction network. Y The total number of samples, in this embodiment of the invention, N is 300, and k represents the admittance parameter training set D. Y The sample numbers, k = 1, 2, 3, ..., 300, y k The training set D represents the admittance parameters. Y The admittance parameter predicted for the k-th sample, x k The training set D represents the admittance parameters. Y The admittance parameter label corresponding to the kth sample in the dataset.

[0071] Step 4.3: Input the admittance features into the graph convolutional inversion network to obtain the training set D of all admittance parameters. Y The corresponding prediction coupling matrix is ​​calculated using the following formula for the training set D of all admittance parameters. Y Corresponding predicted scattering parameters:

[0072]

[0073]

[0074] Among them, M k S represents the prediction coupling matrix corresponding to the k-th sample. 21 k (f) and S 11 k (f) respectively represent the results of M kIn the calculated predicted scattering parameters, the forward reflection coefficient and forward transmission coefficient are used, j represents the imaginary unit, R1 and R2 represent the source impedance and load impedance of the microwave filter, respectively, and BW and f0 represent the desired bandwidth and center frequency, respectively; all are filter design parameters. In this embodiment of the invention, R1 = 1, R2 = 1, BW = 0.2 GHz, f0 = 9.97 GHz, I represents an (n+2) order identity matrix, and R represents the normalized impedance matrix with a size of (n+2) × (n+2), where the element R 11 =1,R n+2,n+2 =1, and all other elements are 0.

[0075] Step 4.4: Calculate all predicted scattering parameters and the corresponding scattering parameter D. S The mean squared error loss Loss2 is calculated between the two networks. Using backpropagation gradient descent, in each iteration, the loss Loss1 of the admittance parameter feature extraction network is first calculated, and the parameters in each layer of the admittance parameter feature extraction network are updated. Then, the admittance parameter feature extraction network is input into the graph convolutional inversion network to calculate the loss Loss2, and the parameters in each layer of the graph convolutional inversion network are updated. This iteration continues until both Loss1 and Loss2 converge, resulting in the trained admittance parameter feature extraction network and graph convolutional inversion network.

[0076] Step 5: Perform fault diagnosis and debugging on the sixth-order cavity microwave filter.

[0077] Step 5.1: Measure the scattering parameters of the sixth-order cavity microwave filter to be fault-diagnosed and debugged using a VNA. Following the same method as in Step 3.2, convert the measured scattering parameters into admittance parameters and then input them into the trained admittance parameter feature extraction network and graph convolution inversion network. Calculate the coupling matrix M corresponding to the measured scattering parameters, and then calculate the coupling matrix between M and the sixth-order cavity microwave filter design. The difference is used to obtain the difference matrix.

[0078] Step 5.2: Determine whether the absolute values ​​of all elements in the difference matrix are all less than ε. If so, proceed to step 6; otherwise, determine that the microwave filter is faulty and proceed to step 5.3. In this embodiment of the invention, ε = 0.1.

[0079] Step 5.3: Using the coupling matrix M and the spatial mapping method, calculate the adjustment amount of the current state of the microwave filter. Adjust the microwave filter according to the adjustment amount and then execute step 5.1.

[0080] Step 6: Determine that the microwave filter is fault-free, end the debugging process, and obtain a fault-free microwave filter.

Claims

1. A performance debugging method for microwave filter fault diagnosis based on graph embedding, characterized in that, The microwave filter is abstracted as an undirected graph, a conductance parameter feature extraction network is constructed by cascading a complex convolution coding subnetwork and a complex deconvolution decoding subnetwork, and a graph convolution inversion network is constructed in combination with the structure of the microwave filter to be diagnosed and debugged; the steps of the performance debugging method include the following: Step 1, constructing a conductance parameter feature extraction network: Step 1.1, building a complex convolution coding subnetwork composed of three complex convolution modules with the same structure in cascade, each complex convolution module being composed of two complex convolution layers in cascade; The convolution kernel size of the two complex convolution layers in the first complex convolution module is set to 3x1, the channel number is set to 64 and 128 respectively, the step is set to 1, and the activation function is realized by a CRelu function; The convolution kernel size of the two complex convolution layers in the second complex convolution module is set to 3x1, the channel number is set to 256 and 128 respectively, the step is set to 1, and the activation function is realized by a CRelu function; The convolution kernel size of the two complex convolution layers in the third complex convolution module is set to 3x1, the channel number is set to 64 and n+2 respectively, n being equal to the order of the filter, the step is set to 1, and the activation function is realized by a CRelu function; Step 1.2, building a complex deconvolution decoding subnetwork composed of three complex deconvolution modules in cascade, each complex deconvolution module being composed of two complex deconvolution layers in cascade; The convolution kernel size of the two complex deconvolution layers in the first complex deconvolution module is set to 3x1, the channel number is set to 64 and 128 respectively, the step is set to 1, and the activation function is realized by a CRelu function; The convolution kernel size of the two complex deconvolution layers in the second complex deconvolution module is set to 3x1, the channel number is set to 256 and 128 respectively, the step is set to 1, and the activation function is realized by a CRelu function; The convolution kernel size of the two complex deconvolution layers in the third complex deconvolution module is set to 3x1, the channel number is set to 64 and 3 respectively, the step is set to 1, and the activation function is realized by a CRelu function; Step 1.3, cascading the complex convolution coding subnetwork and the complex deconvolution decoding subnetwork to form a conductance parameter feature extraction network; Step 2, constructing a graph convolution inversion network: Step 2.1, abstracting the microwave filter as an undirected graph, the source end and the load end of the filter and each cavity of the filter corresponding to the nodes in the undirected graph, each mutual coupling path and cross coupling path corresponding to the edges in the undirected graph, obtaining an undirected graph corresponding to the structure of the filter, and generating an adjacency matrix of the undirected graph; the rows of the adjacency matrix represent the nodes in the undirected graph of the microwave filter, each column represents the connection relationship between the nodes in the undirected graph of the microwave filter, if there is an edge between the nodes in the undirected graph corresponding to the microwave filter, the element at the corresponding position in the adjacency matrix is 1, and if there is no edge between the nodes in the undirected graph, the element at the corresponding position in the adjacency matrix is 0; Step 2.2, a graph convolution inversion network is built by cascading three graph convolution layers, the adjacency matrices in the first to third graph convolution layers are set as undirected graph adjacency matrices, the number of nodes is set as n1+2, and the number of channels is set as n1+2, 32, and n1+2, respectively, and n1 is equal to n; Step 3, generating a training set: Step 3.1, randomly debug each cavity of a microwave filter for at least 30 times, and collect the scattering parameters of each debugging through the VNA to obtain 30×n2 groups of scattering parameters, the value of n2 is equal to n; the forward reflection coefficient S 11 and the forward transmission coefficient S 21 of each group of scattering parameters constitute the scattering parameter set of the group, and 30×n2 groups of scattering parameter sets constitute a scattering parameter training set D S ; Step 3.2, converting each group of scattering parameters into admittance parameters to obtain a converted 30xn2 groups of admittance parameters, and calculating a forward input admittance Y 11 , a forward transfer admittance Y 21 and a corresponding sampling frequency f to form the group of admittance parameter sets, and grouping the 30xn2 groups of admittance parameter sets to form an admittance parameter training set D Y ; Step 4, alternately training the admittance parameter feature extraction network and the graph convolution inversion network: Step 4.1, training set D of admittance parameters Y The input is input into the admittance parameter feature extraction network, the complex convolutional coding subnetwork outputs the admittance feature, and the complex deconvolutional coding subnetwork outputs the predicted admittance parameter; the training set D Y As the admittance parameter label of the admittance parameter feature extraction network, the mean square error loss function is selected, the loss Loss1 between all predicted admittance parameters and the corresponding input admittance parameter label is calculated, and the parameters in each layer of the admittance parameter feature extraction network are updated by using the back propagation gradient descent method. Step 4.2: Input the admittance features output by the complex convolutional coding subnetwork into the graph convolutional inversion network, and output the training set D of all admittance parameters. Y The corresponding prediction coupling matrix is ​​used to calculate the training set D of all admittance parameters. Y The corresponding predicted scattering parameters; the training set D S As the scattering parameter labels for the graph convolutional inversion network, the mean square error loss function is selected, and the loss Loss2 between all predicted scattering parameters and the corresponding input scattering parameter labels is calculated. Then, the parameters in each layer of the graph convolutional inversion network are updated using the backpropagation gradient descent method. Step 4.3, repeating steps 4.1 and 4.2 until Loss1 and Loss2 converge, obtaining the trained admittance parameter feature extraction network and the graph convolution inversion network; Step 5, fault diagnosis and debugging of the microwave filter to be diagnosed and debugged: Step 5.1, measure the scattering parameters of the microwave filter to be diagnosed and debugged with the VNA, in the same way as step 3.2, convert the measured scattering parameters into admittance parameters, and then input them into the trained admittance parameter feature extraction network to output the admittance features; input the admittance features into the trained graph convolution inversion network to output the coupling matrix M corresponding to the measured scattering parameters, calculate the difference between M and the design coupling matrix M of the microwave filter, and obtain the difference matrix . Step 5.2, judging whether the absolute values of all elements in the difference matrix are less than epsilon, epsilon is an error value set in the range of [0.05, 0.3] according to actual filtering requirements, if yes, executing step 6, otherwise, judging the microwave filter as faulty and executing step 5.3; Step 5.3, using the coupling matrix M to apply the space mapping method to calculate the debugging amount of the current state of the microwave filter, and debugging the microwave filter according to the debugging amount, and then executing step 5.1; Step 6, judging the microwave filter as fault-free, ending the debugging, and obtaining a fault-free microwave filter.

2. The performance debugging method for microwave filter fault diagnosis based on graph embedding according to claim 1, characterized in that, The conversion of each group of scattering parameters into admittance parameters in step 3.2 is obtained by the following formula: wherein Y 11 i and Y 21 i respectively represent the forward input admittance and the forward transfer admittance in the converted admittance parameter group of the i-th scattering parameter group, i represents the serial number of the sample in the admittance parameter training set D S 11 i , S 22 i , S 21 i , S 12 i respectively represent the forward reflection coefficient, the backward reflection coefficient, the forward transmission coefficient, and the backward transmission coefficient in the i-th scattering parameter group.​ 3. The method of claim 1, wherein, The loss Loss1 between all predicted admittance parameters and corresponding input admittance parameter labels in step 4.1 is obtained by the following formula: wherein, Loss1 represents the loss between all predicted admittance parameters and the admittance parameter labels corresponding to the inputs, ∑ represents a summation operation, |·| represents an absolute value operation, N represents the total number of samples in the admittance parameter training set D Y input into the admittance parameter feature extraction network, n2 represents the number of samples in the admittance parameter training set D Y , k = 1, 2, 3,..., 30 x n2, y k represents the predicted admittance parameter of the kth sample in the admittance parameter training set D Y , x k represents the admittance parameter label corresponding to the kth sample in the admittance parameter training set D Y .

4. The method of claim 1, wherein, The calculation of all admittance parameters training set D is described in step 4.2 Y The corresponding predicted scattering parameters are implemented by the following equations: wherein M k represents the predicted coupling matrix corresponding to the kth sample, S 21 k (f) and S 11 k (f) respectively represent the forward reflection coefficient and the forward transmission coefficient in the predicted scattering parameters calculated by M k , j represents the imaginary unit, R1 and R2 represent the source impedance and the load impedance of the microwave filter respectively, BW and f0 represent the desired bandwidth and the center frequency respectively, both are the design parameters of the filter, I represents the n+2 order unit matrix, R represents the normalized impedance matrix with the size of (n+2) x (n+2), wherein the elements R 11 = 1, R n+2,n+2 = 1, and other elements are all 0.

5. The method for performance debugging of microwave filter fault diagnosis based on graph embedding according to claim 1, characterized in that, The loss Loss2 between all predicted scattering parameters and corresponding input scattering parameter labels in step 4.2 is obtained by the following formula: wherein Loss2 represents a loss value between all predicted scattering parameters and corresponding scattering parameter labels of the input, s k represents the kth sample predicted scattering parameter in the admittance parameter training set D Y represents the kth sample predicted scattering parameter in the admittance parameter training set D k represents the kth sample corresponding scattering parameter label in the scattering parameter training set D S represents the kth sample corresponding scattering parameter label in the scattering parameter training set D

Images