Waveform detection method for anomalies caused by single event upsets of radio frequency devices

Abstract

This invention relates to an anomaly detection method for radio frequency (RF) devices, specifically a waveform detection method for anomalies caused by single-event upsets (SETs) in RF devices. It solves the technical problem of accurately, quickly, and adaptively determining the type of error caused by SETs when RF devices are subjected to single-event radiation in space. The method includes the following steps: 1) obtaining a training set; 2) extracting feature vectors from the waveform data in the training set using a wavelet scattering convolutional neural network; 3) inputting the feature vectors into a gated recurrent unit (GRU) neural network for feature classification training; 4) connecting the trained GRU neural network to a linear layer and setting the linear layer according to the label format; 5) inputting the waveform data to be detected into the wavelet scattering convolutional neural network, then sequentially passing it through the trained GRU neural network and the linear layer, and outputting the waveform type classification result.

Description

Technical Field

[0001] This invention relates to an anomaly detection method for radio frequency devices, specifically to a waveform detection method for anomalies caused by single-event upsets in radio frequency devices. Background Technology

[0002] High-energy particle radiation in space can cause single-event effects (SEE) on the logic units of radio frequency (RF) devices. Most errors caused by SEE are soft SEE errors (i.e., errors that do not damage the hardware circuitry itself, but only affect stored data or circuit states, and can be recovered to a normal state through system reset, power-on, or rewriting). SEE flips in the configuration bits of a device's configuration register are a cause of soft SEE errors. SEE flips can cause device logic malfunctions, calculation errors, functional abnormalities, and even system crashes. Due to advancements in manufacturing processes and increasingly higher device integration, the effects of SEE flips in space are aggravated, posing a severe challenge to the reliability of aerospace applications and placing higher demands on the detection of SEE flips in RF devices.

[0003] Existing single-event upset detection methods for radio frequency devices generally employ the following three approaches:

[0004] 1. Similar to the readback refresh technology for FPGA single-event upsets (SEORs), SEORs in RF devices are usually caused by high-energy particle radiation that flips the configuration bits in the device's configuration register. Therefore, the configuration bit data can be read back through the configuration interface of the configuration register and compared with the original data to determine whether an SEOR has occurred and to identify the error type (Li Mingkun. Research on Single-Event Detection and Repair Technology of Zynq Configuration Memory [D]. Harbin Institute of Technology, 2017). However, this method has shortcomings. The responsive access mechanism of RF devices makes the readback method of the configuration register too time-consuming and unable to detect errors in a timely manner. Moreover, this method is an indirect error detection method and cannot guarantee that an error in the configuration register will necessarily be reflected in the waveform error.

[0005] 2. Radio frequency (RF) devices generally have built-in anomaly monitoring interfaces. These interfaces are at default values ​​during normal operation, and these values ​​will change once an anomaly occurs. Therefore, the monitoring interface can be used to determine whether a single-event upset (SHO) has occurred in the RF device. However, the monitoring content of the built-in monitoring interface of the RF device is relatively fixed, resulting in low SHO detection efficiency. In addition, this function must be used in a specific operating mode of the RF device, limiting its application scope. Polling all monitoring positions is time-consuming and cannot control the fault tolerance range.

[0006] 3. When the received waveform information of the RF device is known in advance, the waveform shape can be plotted using the waveform data, and a fixed frequency range, amplitude range, and phase difference can be given. By detecting whether the frequency, amplitude, and phase of the received waveform are within the given range, it can be determined whether a single-event upset has occurred. The drawback of this method is that it requires prior knowledge of the received waveform information when detecting whether the frequency, amplitude, and phase of the received waveform are within the given range, limiting its application scope. It cannot be used for complex waveforms where the frequency, amplitude, and phase values ​​are not fixed. Summary of the Invention

[0007] This invention provides a waveform detection method for anomalies caused by single-event upsets in radio frequency devices, in order to solve the technical problem that when radio frequency devices are subjected to single-event radiation in space and undergo single-event upsets, it is impossible to accurately, quickly, and adaptively determine the type of anomaly caused by the upset through the received carrier waveform.

[0008] The technical solution of this invention is:

[0009] A waveform detection method for anomalies caused by single-event upsets in radio frequency devices, characterized by the following steps:

[0010] 1) Obtain the labeled training set

[0011] A training set of waveform data for single-event upset detection of radio frequency devices is established, and the data in the training set are labeled according to the waveform type. The training set contains waveform data of existing waveform types for anomalies caused by single-event upset of radio frequency devices. The label is a 1*T vector, where T represents the number of waveform features.

[0012] 2) A wavelet scattering convolutional neural network is used to extract features from the waveform data in the training set to obtain feature vectors; all feature vectors of the same waveform data are concatenated into an input matrix, and the input matrices of all waveform types have the same size; the wavelet scattering convolutional neural network has at least two layers; the feature vector refers to the feature vector extracted from the waveform data output by each layer of the neural network;

[0013] 3) Input the input matrices of all waveform types obtained from the training set into the gated recurrent unit neural network for feature classification training to obtain the trained gated recurrent unit neural network;

[0014] 4) Connect the output of the gated recurrent unit neural network to the linear layer, and set the linear layer according to the format of the label so that the format of the output waveform type classification result is consistent with the label format;

[0015] 5) Input the waveform data to be detected into the wavelet scattering convolutional neural network, and then pass it through the trained gated recurrent unit neural network and the linear layer in sequence to output the waveform type classification result of the waveform data to be detected.

[0016] Further, in step 2), the feature vector S(k) of the wavelet scattering convolutional neural network is formulated as follows:

[0017]

[0018] f(k) represents the waveform data at time k.

[0019] * indicates a convolution operation;

[0020] |·| represents the modulo operation;

[0021] Let be a wavelet function, where λ n Represents the variation of the scattering path of the nth layer at different scales and directions, n≥1, λ n =2 j γ, j represents the scale, and γ represents the direction;

[0022] ψ γ (2 -j k) is the result of binary scaling and rotation of the Morlet wavelet at scale j at time k;

[0023] For scaling function, The result is obtained by binary scaling of the Gabor wavelet at scale j at time k.

[0024] Further, in step 2), the Gabor wavelet The formula is:

[0025]

[0026] Where δ0∈(0,1];

[0027] The formula for the Morlet wavelet ψ(k) is:

[0028]

[0029] Where β is an adjustable parameter used to adjust ∑ψ(k)=0, and ω0 is a dimensionless frequency, taking positive integer values.

[0030] Furthermore, in step 2), the wavelet scattering convolutional neural network has three layers.

[0031] Further, in step 3), the gated recurrent unit neural network includes N GRUs connected sequentially along the signal output direction, and the output of the last GRU is connected to the input of the linear layer;

[0032] The input matrix obtained in step 2) is fed into the first GRU, and after being processed by N GRUs in sequence, the processing result is output from the last GRU for feature classification training.

[0033] Furthermore, in step 3), the gated recurrent unit neural network includes 256 GRUs.

[0034] Further, in step 1), the label is a 1*4 vector [θ1,θ2,θ3,θ4], where θ1 indicates whether the waveform is complete (1 for complete, 0 for incomplete); θ2 indicates whether the waveform has spikes (1 for spikes, 0 for no spikes); θ3 indicates whether the waveform is orthogonal (1 for orthogonal, 0 for non-orthogonal); and θ4 indicates whether the waveform is aliased (1 for aliased, 0 for no aliased).

[0035] Further, in step 1), a training set of single-event upset detection waveform data for radio frequency devices is established using a simulated fault injection method; the waveform data is single-cycle waveform data.

[0036] The beneficial effects of this invention are:

[0037] 1. This invention relates to a waveform detection method for anomalies caused by single-event upsets in radio frequency (RF) devices. The method extracts features from the received waveform data of the RF device using a wavelet scattering convolutional neural network and then classifies the features using a gated recurrent unit (GRU) neural network. This method quickly obtains the waveform type results for each data point. It can be used to quickly identify whether a device has experienced a single-event upset and the type of error, achieving high accuracy. Furthermore, it is not limited by the device's operating mode and does not require prior knowledge of the received waveform information. The method makes judgments adaptively based on the received data and has universal applicability.

[0038] 2. The present invention provides a waveform detection method for anomalies caused by single-event flips in radio frequency devices. The wavelet scattering convolutional neural network used does not require training, has a fast processing time, and can produce good feature extraction results, which facilitates the improvement of subsequent classification accuracy. It also performs well with low data volume.

[0039] 3. The present invention provides a waveform detection method for anomalies caused by single-event upsets in radio frequency devices. Since it is necessary to identify the waveform to obtain the waveform type, including abnormal waveform types, this method uses a gated recurrent unit neural network to classify the waveform type after feature extraction. The gated recurrent unit neural network requires less training time and has fewer parameters than other types of recurrent neural networks, making it faster to train and less prone to overfitting.

[0040] 4. This invention addresses waveform detection methods for anomalies caused by single-event upsets in radio frequency devices. It employs a wavelet scattering network to extract features and then uses a gated recurrent neural network for feature classification, thus improving accuracy. The most important criterion for normal communication is whether the waveform changes. Traditional methods suffer from uncertainty, where changes in register values ​​do not necessarily alter the waveform. For example, changes to certain non-readable bits can lead to waveform errors that cannot be detected during readback, while changes to some readable bits may occur without a waveform change, leading to incorrect judgments as if no error readback has occurred. This invention avoids this problem.

[0041] 5. The present invention provides a waveform detection method for anomalies caused by single-event upsets in radio frequency devices. Compared with the traditional method, which requires sending a readback command and waiting for a response when reading back the configuration register, resulting in a long process and inability to detect errors in a timely manner, the present invention directly judges the error by receiving data without waiting for a response, thereby shortening the error detection time and improving the error detection efficiency. Attached Figure Description

[0042] Figure 1 This is a schematic diagram of the wavelet scattering convolutional neural network structure in an embodiment of the waveform detection method for anomalies caused by single-event upsets in radio frequency devices according to the present invention.

[0043] Figure 2 This is a flowchart of the gated recurrent unit neural network and linear layer in an embodiment of the waveform detection method for anomalies caused by single-event upsets in radio frequency devices according to the present invention.

[0044] Figure 3 This is a partial schematic diagram of a gated recurrent neural network in an embodiment of the waveform detection method for anomalies caused by single-event upsets in radio frequency devices according to the present invention.

[0045] Figure 4 This is a schematic diagram of the principle of the gated recurrent unit neural network in an embodiment of the waveform detection method for anomalies caused by single-event upsets in radio frequency devices according to the present invention. Detailed Implementation

[0046] The main idea of ​​this invention is to use a wavelet scattering convolutional neural network to extract features from labeled input waveform data, obtaining feature vectors. These feature vectors are then fed into a gated recurrent unit neural network for feature classification, and finally, after passing through a linear layer, the classified result is obtained. The invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0047] This invention provides a waveform detection method for anomalies caused by single-event upsets in radio frequency devices, comprising the following steps:

[0048] 1) Obtain the labeled training set

[0049] A training set of waveform data for single-event upset detection of radio frequency devices was established using a simulated fault injection method. The training set consists of 1000 sets of single-cycle waveform data. The data in the training set were labeled according to waveform type. The training set includes waveform data for all waveform types related to anomalies caused by single-event upsets of radio frequency devices. The label is a 1*4 vector [θ1, θ2, θ3, θ4], where θ1 indicates whether the waveform is complete (1 for complete, 0 for incomplete); θ2 indicates whether the waveform has glitches (1 for glitches, 0 for no glitches); θ3 indicates whether the waveform is orthogonal (1 for orthogonal, 0 for non-orthogonal); and θ4 indicates whether the waveform is aliased (1 for aliased, 0 for no aliased).

[0050] 2) A wavelet scattering convolutional neural network is used to extract features from the waveform data in the training set to obtain feature vectors; all feature vectors of the same waveform data are concatenated into an input matrix of the same size; the wavelet scattering convolutional neural network has three layers; the feature vector refers to the feature vector extracted from the waveform data output by each layer of the neural network.

[0051] The eigenvector formula for a wavelet scattering convolutional neural network is:

[0052]

[0053] f(k) represents the waveform data at time k.

[0054] * indicates a convolution operation;

[0055] |·| represents the modulo operation;

[0056] Let be a wavelet function, where λ n Represents the variation of the scattering path of the nth layer at different scales and directions, n≥1, λ n =2 j γ, j represents the scale, and γ represents the direction;

[0057] ψ γ (2 -j k) is the result of binary scaling and rotation of the Morlet wavelet at scale j at time k;

[0058] For scaling function, The result is obtained by binary scaling of the Gabor wavelet at scale j to k.

[0059] The construction of wavelet scattering convolutional neural networks mainly involves the following steps:

[0060] (1) Construct the scaling function. The general expression of the scaling function is:

[0061]

[0062] If the maximum scale is J, then the maximum scale function is:

[0063]

[0064] In this embodiment, the scaling function is a Gabor wavelet. The formula is:

[0065]

[0066] Where δ0∈(0,1).

[0067] (2) Constructing wavelet functions

[0068] The general expression for the wavelet function is:

[0069] ψ j,γ (k)=2 -2j ψ γ (2 -j k)

[0070] Where j represents the scale (j≤J) and γ represents the direction, and in this embodiment, γ=3.

[0071] The wavelet function used is the Morlet wavelet, and the formula for the Morlet wavelet ψ(k) is:

[0072]

[0073] Where β is an adjustable parameter, used to adjust ∑ψ(k)=0, and ω0 is a dimensionless frequency that can take positive integer values.

[0074] In other embodiments, other fundamental waves, such as Haar wavelets or Marr wavelets, can also be used for wavelet scattering transform.

[0075] (3) Constructing wavelet transform

[0076] The wavelet transform of the received waveform data is as follows:

[0077] W j,γ f(k)=f(k)*ψ j,γ (k)

[0078] Where f(k) is the waveform data at time k, and * denotes convolution operation.

[0079] (4) Constructing the modular operator

[0080] To obtain low-frequency information, modulo operations are performed on the wavelet transform:

[0081] U λ f(k)=|f(k*ψ λ (k)|

[0082] Where λ = 2 j γ.

[0083] (5) Constructing wavelet scattering transform

[0084] The wavelet scattering transform formula is as follows:

[0085]

[0086] Among them, S j (λ) is called the scattering operator and is also the final set of features.

[0087] (6) Constructing a wavelet scattering convolutional neural network

[0088] After a signal undergoes wavelet scattering transform, the wavelet characteristic coefficients become relatively stable due to averaging operations, but this also suppresses high frequencies, resulting in some loss and reducing the discriminative power of the wavelet characteristic coefficients. To recover high-frequency information, the wavelet modulus can be further decomposed at a relatively larger scale (which must be smaller than the maximum scale). However, such features still lack translation invariance, so modulus operations and averaging (low-pass filtering) are needed to achieve stability of the characteristic coefficients.

[0089] Each feature averaging process involves the loss of high-frequency information; therefore, to ensure energy conservation, iterative processing of the upper-layer wavelet scattering coefficients is necessary. Thus, the feature vector formula for the wavelet scattering convolutional neural network is:

[0090]

[0091] Let the scattering path P = {λ1, λ2, λ3, ... λ m}, where m is the maximum path depth and n is the path depth, both n and m are positive integers, and 1 ≤ n ≤ m. According to the above formula, the following can be achieved: Figure 1 The wavelet scattering convolutional neural network shown is divided into three layers. The output of each layer in the diagram is the extracted feature vector S(k). All feature vectors S(k) are concatenated into an input matrix x of the same size. k Postscript:

[0092] x k ∈R l×d

[0093] R represents the set of matrices, and l and d represent the number of rows and columns of the matrices, respectively.

[0094] Wavelet scattering convolutional neural networks utilize the concept of wavelet transform. They use a scaling function to represent the original signal. As the scale decreases, the representation of the original signal becomes increasingly coarse and ambiguous, with a growing difference from the original signal. To address this, a wavelet function is introduced to represent the difference between the scaling function representation and the original signal. Finally, the combination of the scaling function and the wavelet function accurately represents the original signal. The coefficients applied to the scaling function and the wavelet function are the result of the wavelet transform. These coefficients can then be used to reconstruct the original signal. Its advantages include being unaffected by translation, rotation, scaling, or deformation, possessing rich feature information representation, and overcoming the time-varying nature of wavelet transform.

[0095] 3) Input the input matrices of all waveform types obtained from the training set into the gated recurrent unit neural network for feature classification training to obtain the trained gated recurrent unit neural network.

[0096] A gated recurrent unit (GRU) neural network consists of N GRUs connected sequentially along the signal output direction. The output of the last GRU is connected to the input of the linear layer. After the input matrix is ​​processed by each GRU, it outputs the corresponding hidden state h. The first hidden state h0 is the initial value, which is 0 or a random number. Figure 2 As shown, the input matrix obtained in step 2) is fed into the first GRU. After passing through N GRUs, the processing result is output from the last GRU. The processing result is processed by a linear layer to obtain a waveform type classification result that is consistent with the label format.

[0097] The structure of a gated recurrent unit neural network is as follows: Figure 3 As shown, a GRU is used as a unit, and the units are cascaded. The output of the previous unit is used as the input of the next unit for continuous iteration. The current input matrix x k ∈R l×d After input from the first GRU layer, it passes through each subsequent GRU layer in sequence, h t-1 ∈R p×d for h t The hidden state passed down from the previous node contains relevant information about the previous node. The GRU then receives the hidden state h that is passed to the next node. t The input matrix is ​​processed through a gated recurrent unit neural network to obtain the final hidden state h. N The final output is processed through a linear layer to obtain the classification result. In this embodiment, 256 GRUs are used, i.e., N=256.

[0098] The internal operation principle of each GRU is as follows: Figure 4 As shown, the main operations include updating the gate and resetting the gate.

[0099] Calculate the update gate z using the following formula.t :

[0100] z t =σ(W (z) x k +U (z) h t-1 )

[0101] Where x k Given the input matrix W (z) ∈R p×l U (z) ∈R p×p Let h be the weight matrix. t-1 The information stored is from the previous time step t-1. The update gate adds these two pieces of information and feeds them into σ (the sigmoid activation function), compressing the activation result to between 0 and 1. The update gate helps the model decide how much past information to pass to the future, or how much information from the previous and current time steps needs to be passed on.

[0102] Calculate the reset gate r using the following expression. t :

[0103] r t =σ(W (r) x k +U (r) h t-1 )

[0104] Reset door r t This primarily determines how much past information needs to be forgotten. The expression is the same as the update gate expression, except that the weight matrix of the linear transformation is replaced by W. (r) ∈R p×l U (r) ∈R p×p h t-1 and x k First, a linear transformation is performed, and then the results are added together and fed into σ (the sigmoid activation function) to output the activation value.

[0105] When using the reset gate, new memory content will use the reset gate to store past related information. Its calculation expression is:

[0106] h' t =tanh(Wx k +U(r t ⊙h t-1 ))

[0107] Enter x k Information h from the previous time step t-1 First, a linear transformation is performed, that is, multiplying by the weight matrix W∈R. p×l and U∈Rp×p ⊙ represents the Hadamard product, i.e., r t with h t-1 The Hadamard product is the element-wise product of the elements. Since the reset gate calculated earlier is a vector of 0s and 1s, it measures the size of the gating. For example, if the gating value for an element is 0, it means that the information for that element is completely forgotten. This Hadamard product determines which previous information to retain and which to forget. The results of these two calculations are then added together and fed into the tanh (hyperbolic tangent) activation function.

[0108] In the final step, the network needs to calculate h. t This vector will retain the information of the current cell and pass it to the next cell. In this process, an update gate is needed, which determines the current memory content h. t 'and the previous time step h t-1 What information needs to be collected? This process can be represented as:

[0109] h t =z t ⊙h t-1 +(1-z t )⊙h' t

[0110] z t To update the activation result of the gate, it also controls the inflow of information in the form of gating. t with h t-1 The Hadamard product represents the information retained from the previous time step to the final memory. This information, plus the information retained from the current memory to the final memory, equals the output of the final gated loop unit.

[0111] Each GRU does not erase previous information over time; instead, it retains relevant information and passes it on to the next unit, thus utilizing all information and avoiding the vanishing gradient problem.

[0112] 4) Connect the output of the gated recurrent unit neural network to the linear layer, and set the linear layer according to the label format so that the format of the output waveform type classification result is consistent with the label format.

[0113] 5) Input the waveform data to be detected into the wavelet scattering convolutional neural network, and then pass it through the trained gated recurrent unit neural network and the linear layer in sequence. The linear function maps it into a 1*4 vector and outputs the waveform type classification result of the waveform data to be detected.

Claims

1. A waveform detection method for anomalies caused by single-event upsets in radio frequency devices, characterized in that, Includes the following steps: 1) Obtain the labeled training set A training set of waveform data for single-event upset detection of radio frequency devices is established, and the waveform data in the training set are labeled according to the waveform type. The training set contains waveform data of existing waveform types that are oriented towards anomalies caused by single-event upset of radio frequency devices. The label is a 1*T vector, where T represents the number of waveform features. The label is a 1*4 vector. ,in, This indicates whether the waveform is complete; 1 indicates complete, and 0 indicates incomplete. This indicates whether the waveform has glitches; 1 indicates glitches and 0 indicates no glitches. This indicates whether the waveforms are orthogonal; 1 indicates orthogonal waveforms, and 0 indicates non-orthogonal waveforms. This indicates whether the waveform is aliased; 1 indicates waveform aliasing, and 0 indicates waveform without aliasing. A training set of single-event upset detection waveform data for radio frequency devices is established using a simulated fault injection method; the waveform data is single-cycle waveform data. 2) A wavelet scattering convolutional neural network is used to extract features from the waveform data in the training set to obtain feature vectors; all feature vectors of the same waveform data are concatenated into an input matrix, and the input matrices of all waveform types have the same size; the wavelet scattering convolutional neural network has at least two layers; the feature vector refers to the feature vector extracted from the waveform data output by each layer of the wavelet scattering convolutional neural network. 3) Input the input matrices of all waveform types obtained from the training set into the gated recurrent unit neural network for feature classification training to obtain the trained gated recurrent unit neural network; 4) Connect the output of the gated recurrent unit neural network to the linear layer, and set the linear layer according to the format of the label so that the format of the output waveform type classification result is consistent with the label format; 5) Input the waveform data to be detected into the wavelet scattering convolutional neural network, and then pass it through the trained gated recurrent unit neural network and the linear layer in sequence to output the waveform type classification result of the waveform data to be detected.

2. The waveform detection method for anomalies caused by single-event upsets in radio frequency devices according to claim 1, characterized in that, In step 2), the formula for the feature vector S(k) of the wavelet scattering convolutional neural network is: ； The input is the waveform data at time k; This represents the convolution operation; For modulo operation; Let be a wavelet function, where This represents the variation of the scattering path at different scales and directions in the nth layer, where n ≥ 1. j represents the scale. Indicates direction; , The result is obtained by binary scaling and rotation of the Morlet wavelet at scale j at time k; φ j For scaling function, ,;φ( k) is the result of binary scaling of the Gabor wavelet at scale j at time k.

3. The waveform detection method for anomalies caused by single-event upsets in radio frequency devices according to claim 2, characterized in that, In step 2), the formula for the Gabor wavelet φ(k) is: ； in, ; The formula for the Morlet wavelet ψ(k) is: ； in, This is an adjustable parameter used for adjustment. , For dimensionless frequency, Take a positive integer.

4. The waveform detection method for anomalies caused by single-event upsets in radio frequency devices according to claim 3, characterized in that: In step 2), the wavelet scattering convolutional neural network has three layers.

5. The waveform detection method for single-event upset anomalies in radio frequency devices according to any one of claims 1-4, characterized in that, In step 3), the gated recurrent unit neural network includes N GRUs connected sequentially along the signal output direction, and the output of the last GRU is connected to the input of the linear layer. The input matrix obtained in step 2) is fed into the first GRU, and after being processed by N GRUs in sequence, the processing result is output from the last GRU for feature classification training.

6. The waveform detection method for anomalies caused by single-event upsets in radio frequency devices according to claim 5, characterized in that: In step 3), the gated recurrent unit neural network includes 256 GRUs.

Images