A multi-task learning method for blind recognition of channel coding

Abstract

The application discloses a multi-task learning method for blind recognition of channel coding, and belongs to the field of wireless communication. The framework combines a classifier network with a decoder network and a shared feature extractor, can accurately recognize the coding mode and parameters of a candidate coding set receiving sequence in a complex communication environment, can accurately reject coding modes and parameters that do not exist in the candidate coding set, that is, realizes closed set / open set recognition of channel coding. In addition, the multi-task learning framework introduces CenterLoss in metric learning, learns the deep feature center of each coding class, and punishes the distance between the deep feature and the corresponding class feature center, so that the decision boundary of signal recognition is more compact and clear, and the channel coding recognition accuracy is further improved. In addition, the application automatically extracts relevant sequence features by using a neural network, thereby avoiding the tedious process of manually extracting features in the prior art.

Description

TECHNICAL FIELD

[0001] The application discloses a general multi-task learning framework for open / closed set identification of channel coding, and belongs to the field of wireless communication. BACKGROUND

[0002] In order to ensure the reliability in the information transmission process, improve the information transmission rate and the spectrum resource utilization rate, wireless communication has undergone continuous development and innovation for more than a hundred years, and many modern communication technologies such as source coding and decoding, encryption and decryption, channel error correction coding have appeared. In order to reduce the influence of interference in the channel on the reliability of communication, various channel coding theories and algorithms are proposed, and are widely used in many practical systems such as satellite communication, deep space communication and mobile communication.

[0003] The primary application scenario of channel coding blind identification analysis is in non-cooperative reconnaissance communication. In a cooperative communication system, both the transmitter and the receiver know the type and parameters of the channel coding used, so the receiver can use this information to correctly decode. However, in a non-cooperative communication scenario, the receiver needs to reconstruct the transmitted information from the intercepted signal without any prior knowledge of the channel coding. Therefore, for a non-cooperative receiver, it is crucial to correctly identify the type and coding parameters of the channel coding.

[0004] There are few studies on coding recognition in existing research, and the recognition of coding methods is only focused on some common linear block codes, convolutional codes, etc., and there is little mention of Turbo codes, LDPC codes and the latest designed Polar codes commonly used in modern communication systems. And the solution to the parameter identification problem (such as code length, code rate) is only for the same coding type, and there is no unified parameter identification method for different coding methods. The literature "B. Shen, C. Huang, W. Xu, T. Yang and S. Cui. Blind Channel Codes Recognition via Deep Learning. IEEE Journal on Selected Areas in Communications, 2021" proposes three neural network recognizers for detecting whether the received sequence is encoded by a Polar code. This recognition algorithm follows a closed set setting, in which the prior knowledge of all test coding types is assumed to be available during training. However, in practice, there are many types of channel coding, and when deploying the recognition algorithm, these recognition systems are likely to observe test samples belonging to unknown classes. Due to the closed set assumption, the recognition system will incorrectly identify test samples from unknown classes as samples belonging to one of the known closed set classes, which does not meet the requirements of non-cooperative communication.

[0005] The application proposes a general multi-task learning framework for open and closed set identification of channel coding type and coding parameters. SUMMARY

[0006] For the problem of channel coding blind identification, the purpose of the application is to propose a general multi-task learning framework, which is composed of a shared feature extractor network, a decoder network and a classifier network, for open and closed set identification of channel coding type and parameters. In addition, the Center Loss in metric learning is introduced in the network feature space to enhance the feature discrimination between known closed set data and unknown open set data, which can establish a close and clear boundary between known classes and unknown classes. For a trained multi-task learning network, the channel coding type (parameters) in the closed set candidate set can be identified and classified by the feature extractor plus the classifier. At the same time, the feature extractor plus the decoder can reject the test samples of unknown classes through the reconstruction error, realizing the open set identification of channel coding type (parameters). Therefore, the multi-task learning framework can accurately identify the encoding mode and parameters of the received sequence in a complex communication environment, and accurately reject the encoding mode and parameters that do not appear in the candidate encoding set.

[0007] The technical scheme of the application is a multi-task learning method for channel coding blind identification, which comprises the following steps:

[0008] Step 1: Construct a multi-task learning framework, which includes a shared feature extractor network a decoder network and a classifier network

[0009] The shared feature extraction network is a ResNet network with multiple residual blocks, and the function is represented as: wherein is the space of the intercepted sequence after coding, is the deep network feature space;

[0010] The classifier is an MLP (Multi-layer Perceptron) classification network with Softmax, and the function is represented as: wherein the deep network feature space, a label space for all K channel encoding types (parameter types), K being the number of identified categories of candidate encoding sets;

[0011] the decoder for a ResNet network with multiple residual blocks, denoted as: wherein the deep network feature space, a sequence reconstruction space corresponding to all input encoding sequences;

[0012] Step 2: generate training samples and test samples of the multi-task learning framework;

[0013] Step 2.1: training sample generation: select a candidate encoding set requiring closed set recognition, and generate channel encoding sequences to be recognized;

[0014] Step 2.2: test sample generation: for closed set recognition, test samples are generated according to the training sample process, and for open set recognition, test sample encoding sequences are generated from candidate encoding sets and non-candidate encoding sets;

[0015] Step 3: introduce a Center loss loss function and train the multi-task learning framework, i.e., a shared feature extractor network, a decoding network and a classification network;

[0016] Step 3.1: let and be a sample in any known class and its corresponding label; be the deep network output feature after the feature extractor ; be the reconstructed encoding sequence of the input encoding sequence after the feature extractor and the decoder ; in addition, be the predicted class probability vector obtained by the feature extractor and the classifier ; therefore, the loss function of the multi-task learning framework consists of three parts: the cross-entropy classification loss function of the classifier network the L 1 regular reconstruction loss of the decoder and the center loss introduced in the feature space Therefore, the final loss function of the multi-task learning framework is:

[0017]

[0018] wherein α, λ, β are three constants for balancing the three loss functions, Used for closed set recognition and classification tasks. Used for refactoring tasks. This ensures that the inter-class distance of the encoded signal in the deep feature space increases, while the intra-class distance decreases.

[0019]

[0020]

[0021]

[0022] In the formula, N represents the batch size. The cross-entropy classification loss function can be expressed as: Where K is the number of recognition categories in the candidate coding set. It is sample X i The corresponding label y i one-hot encoded vector, Indicates sample X i The predicted class probability vector, For L 1 The regular reconstruction loss can be expressed as in Input sample; z i For X i The depth feature vector of the sample, y i For X i Class tags, Centered on its class;

[0023] Step 3.2: To obtain the three networks in the multi-task learning framework The corresponding optimal network parameters Θ f ,Θ g and Θ c Randomly select a batch of training datasets for input. Calculate the loss function The parameters of the multi-task learning framework are optimized using the adaptive moment estimation algorithm (Adam). Through repeated iterations in small batches, the parameters of the multi-task learning framework gradually converge and approach the optimal value.

[0024] Step 4: Performance Testing of the Multi-Task Learning Framework

[0025] Step 4.1: For the closed set recognition test, the test steps are as follows: Input the test sample X into the feature extractor-classifier channel. Generate class probability estimation vectors Recorded as Where P(H) n|X) represents the probability that the test sample X is of the i-th class; the final recognition classification result y pred may be expressed as

[0026] Step 4.2: For open set recognition of channel coding type (parameters), the test steps are as follows:

[0027] Step 4.2.1: All training set samples are passed through the feature extractor and the decoder, and the training set samples are the samples in the closed set, to obtain the reconstructed signal, through the formula The reconstruction error of the training set sample is calculated, and the decision threshold is denoted as 80% of the maximum reconstruction error in the training set;

[0028] Step 4.2.2: The test sample X is input into the feature extractor, and the test sample includes the closed set and the non-closed set sample, to obtain its deep feature Then input the classifier , to obtain:

[0029]

[0030] At the same time, the deep feature z is input into the decoder , to obtain its reconstructed signal The reconstruction error is calculated If , it is determined as an unknown class sample; if , it is determined as a known class sample, and the label is y pred .

[0031] In summary, due to the adoption of the above technical scheme, the beneficial effects of the present application are:

[0032] 1. The present application is directed to the open set recognition problem of channel coding, and for the first time proposes a multi-task learning framework, which is effective for the recognition of channel coding type and parameters (code rate, code length).

[0033] 2. The present application avoids the tedious process of manually extracting signal features, and uses the designed network feature extractor, which reduces the complexity.

[0034] 3. The present application introduces Center Loss in the feature space, which can optimize the decision boundary between known closed set data and unknown open set data, and can establish a close and clear boundary between known classes and unknown classes.

[0035] 4. For closed-set identification, the performance of the proposed multi-task learning framework outperforms the state-of-the-art ResNet closed-set identifier and traditional CNN (Convolutional Neural Network) closed-set identifier by introducing the Center Loss. For open-set identification, the performance of the proposed multi-task learning framework outperforms the traditional SoftMax open-set identifier. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 Flowchart for blind identification of channel coding

[0037] Figure 2 Network diagram of the multi-task learning framework.

[0038] Figure 3 (a) Closed-set identification performance of the multi-task learning framework for channel coding types under different SNRs. (b) Closed-set identification performance of the multi-task learning framework for Polar code rates under different SNRs. (c) Closed-set identification performance of the multi-task learning framework for Polar code lengths under different SNRs.

[0039] Figure 4 (a) Two-dimensional visualization scatter plot of deep features for closed-set identification of channel coding rate parameters by CNN identifier under 10 dB. (b) Two-dimensional visualization scatter plot of deep features for closed-set identification of channel coding rate parameters by ResNet identifier under 10 dB. (c) Two-dimensional visualization scatter plot of deep features for closed-set identification of channel coding rate parameters by the multi-task learning framework under 10 dB. DETAILED DESCRIPTION

[0040] In order to make the purpose, technical method and advantages of the present application clearer, the present application will be described in further detail below in combination with embodiments and drawings.

[0041] The present application proposes a general multi-task learning framework for blind identification of channel coding types and parameters of received sequences in non-cooperative communication. The multi-task framework is composed of a shared ResNet feature extraction network a ResNet decoding network and an MLP classification network for open-set and closed-set identification of channel coding types and parameters. Meanwhile, the Center Loss in the network feature space is introduced to further enhance the feature boundary between known closed-set data and unknown open-set data, and to establish a close and clear boundary between known classes and unknown classes. For a trained multi-task learning framework, the feature extractor and the classifier It can perform closed-set identification and classification on the channel coding type (parameters) in the candidate closed-set set. Simultaneously, it uses a feature extractor and decoder... Test samples of unknown classes can be rejected by reconstructing errors. Therefore, this multi-task learning framework can accurately identify the encoding methods and parameters of received sequences in complex communication environments, while accurately rejecting encoding methods and parameters that do not appear in the candidate encoding set.

[0042] The general multi-task channel coding identification framework method of the present invention comprises the following specific steps:

[0043] Step 1: If it is encoding type identification, generate received sequence samples encoded by different channel codes; if it is encoding parameter identification (code rate and code length), generate received sequence samples encoded by different encoding parameters.

[0044] Step 1.1: For a non-cooperative communication system, the source message is encoded into a codeword sequence c = [c1, c2, ..., c...]. M ], where M is the number of transmitted codewords, and c i =[c i1 ,c i2 ,…,c iN ] represents the i-th codeword, N is the length of the codeword, i = 1, 2, ..., M, c ij This represents the j-th symbol of the i-th codeword;

[0045] Step 1.2: These M codewords are modulated sequentially and then transmitted over the channel. Considering a circularly symmetric complex Gaussian noise (CSCG) channel, the received sequence observed at the receiver can be represented as x = [x1, x2, ..., x...]. M ],in

[0046] x i =c i +n i

[0047] In the formula, x i For the i-th received signal, n i Let be CSCG noise, whose elements are independent and identically distributed complex Gaussian random variables with mean 0 and variance N0.

[0048] Step 1.3: In a non-cooperative communication system, the observed received sequence is a portion of the entire received sequence x, denoted as X. This can be represented as...

[0049] X = [x ij ,x ij+1 ,…,x sq ]

[0050] In the formula, x sqis the th symbol of the th received encoded sequence (1≤i≤s≤K, 1≤j,q≤N). x ij is the start point of the intercepted received sequence.

[0051] Step 1.4: Based on the corresponding start point, intercept the sequence backwards with length L to get the received sequence of the channel encoded sequence.

[0052] Step 1.5: According to different encoding methods, code rate, code length and signal-to-noise ratio of the channel, generate the intercepted sequence, and generate the candidate parameter set of the data set, which is given as

[0053]

[0054] In the formula, is the closed set of channel encoding type recognition, and x is the Cartesian product, and are the closed set of code rate and code length, respectively, is the set of channel signal-to-noise ratio values. In order to generate a sample of the data set, an element of is randomly selected, and a intercepted sequence (training sample) of the multi-task learning recognition framework is obtained according to steps 1.3 and 1.4.

[0055] Step 2: Randomly extract 90% of the generated sequence as the training set, and randomly extract the remaining part as the validation set, of which the validation set is used to adjust the network hyperparameters (including learning rate, batch size, iteration number and dimension of each weight matrix) in the training process. For the test set, if it is a closed set recognition test, the test sample is generated according to the training sample process; if it is an open set recognition, the test sample encoded sequence should be generated from the candidate encoding set and the non-candidate encoding set.

[0056] Step 3: Construct a multi-task learning recognition framework. A multi-task learning recognition framework is composed of three networks: a feature extractor a decoder and a classifier which is composed of multiple ResNet residual blocks and can be represented by the function where is the space of the intercepted encoded sequence, is the deep network feature space. which is also composed of multiple ResNet residual blocks, is an MLP (Multi-layer Perceptron) network with Softmax. Similarly, and can be represented as where is the sequence reconstruction space corresponding to all input encoded sequences, Encode the label space for all K channel coding types (parameter types), where K is the number of identified categories of candidate coding sets. The multi-task learning identification framework is shown in Figure 2 , and the parameters are shown in Table 1.

[0057] Step 4: Train the multi-task learning identification framework.

[0058] Step 4.1: Let Θ f , Θ g and Θ c be the parameters of . The classification loss is denoted by , which penalizes the network, i.e., the feature extractor-classifier network , for misclassifying known class samples. The reconstruction loss is denoted by , which penalizes the network, i.e., the feature extractor-decoder network Let and be a sample in any known class and its corresponding label. is the reconstructed input after passing through the feature extractor and the decoder . In addition, is the predicted class probability vector after passing through the feature extractor and the classifier . For an input sequence sample of a training data set with a Batch-size of N, the loss function can be represented as

[0059]

[0060]

[0061] In the formula, l c is the cross-entropy classification loss function, and l r is the L 1 regularization reconstruction loss. The definitions of l c and l r are as follows:

[0062]

[0063]

[0064] where is the one-hot encoding vector of the label y i . The predicted class probability vector for the training sample X i , and the reconstructed sequence of the input sample coding sequence. is the reconstructed sequence of the input sample coding sequence.

[0065] Step 4.2: Introduce Center Loss to enhance the feature boundary between known closed set data and unknown open set data, reduce intra-class distance, and increase inter-class distance. Center Loss can be expressed as

[0066]

[0067] where z i is the deep feature vector of the sample, y i is the class label of X i , and c i is the class center of y .

[0068] Step 4.3: The final loss function of the multi-task learning framework can be expressed as

[0069]

[0070] Here, α, λ, β are three constants used to balance the three loss functions.

[0071] Step 4.4: To obtain the optimal network parameters Θ f , Θ g , and Θ c of the multi-task learning framework, a batch of training data is randomly selected from the training samples, and the one-dimensional sequence of length L is converted into a two-dimensional sequence of N x N, which is input into the network The loss function is calculated Then, the multi-task learning network parameters are optimized using the Adam algorithm. Through repeated iteration of small batches, the parameters of the multi-task learning network gradually converge and approach the optimal value.

[0072] Step 5: To test the performance of the trained multi-task network model, the following method is used to evaluate its recognition performance.

[0073] Step 5.1: Closed set recognition for channel coding type (parameters): input the test sample X (all samples in the closed set candidate set) into the feature extractor-classifier channel to generate a class probability estimation vector denoted as where P(H n |X) represents the probability that the test sample X is of the i-th class; the final recognition classification result y pred can be expressed as

[0074]

[0075] Step 5.2: Open set recognition for channel coding type (parameters), the test steps are as follows:

[0076] Step 5.2.1: All training set samples (samples in the closed set) are passed through the feature extractor plus decoder to obtain the reconstructed signal, and the reconstruction error of the training set samples is calculated by the formula . Let the decision threshold be 80% of the maximum reconstruction error in the training set.

[0077] Step 5.2.2: The test sample X (closed set / non-closed set sample) is input into the feature extractor to obtain its deep feature Then input into the classifier , to obtain At the same time, the deep feature z is input into the decoder , to obtain its reconstructed signal Calculate the reconstruction error If , it is determined as an unknown class sample; if , it is determined as a known class sample, and the label is y pred .

[0078] The algorithm flowchart of steps 1-5 above is shown in Figure 1 .

[0079] The advantages of the present application are:

[0080] 1. The present application proposes a multi-task learning framework for channel coding open set identification for the first time, which is effective for both channel coding type and parameter (code rate, code length) identification.

[0081] 2. The present application avoids the tedious process of manually extracting signal features, and uses the designed network feature extractor to reduce complexity.

[0082] 3. The present application introduces Center Loss in the feature space, which can optimize the decision boundary between known closed set data and unknown open set data, and can establish a close and clear boundary between known classes and unknown classes.

[0083] 4. For closed set identification, the multi-task learning framework of the present application outperforms the current advanced ResNet closed set identifier by introducing Center Loss; for open set identification, the multi-task learning framework outperforms the traditional SoftMax open set identifier.

[0084] Embodiment

[0085] Taking LDPC, Polar, Turbo code, BCH code as an example, the specific identification process of the present application is described.

[0086] Dataset: Set the channel coding type closed set candidate set as For LDPC codes, set and Parameters are from IEEE 802.11 and IEEE 802.16 protocols. For Polar codes, set and For Turbo codes, set and Parameters are from LTE protocol. For BCH codes, set and

[0087] Network parameter setting: In the multi-task learning framework, the loss balance constants are set as α = 0.2, λ = 0.5, and β = 0.5. The learning rate is set as 0.001, and the training data set-size is set as 64. The generated sequence is truncated as L = 8192, converted into a two-dimensional 64x64 input feature extraction network, and trained in the multi-task recognition framework. The trained multi-task learning framework can complete the closed set recognition task and the open set recognition task. Among them, for type recognition, the training set size is 246000 samples, of which LDPC, Polar, Turbo, BCH are 66000 samples respectively. For parameter recognition, taking the code rate and code length recognition of Polar as an example, the training set size is 246000 samples, of which, for code rate recognition, 1 / 8, 2 / 8, 3 / 8, and 4 / 8 code rates are 66000 samples respectively; for code length recognition, 256, 512, 1024, and 2048 code lengths are 66000 samples respectively.

[0088] Embodiment recognition results:

[0089] Test closed set recognition performance: In order to test the closed set performance, we selected the currently popular CNN recognizer, ResNet recognizer as the benchmark. For the test set, we selected test data under different signal-to-noise ratios, and the signal-to-noise ratios were {-4, -2, 0, 2, 4, 6, 8, 10}, and for each test signal-to-noise ratio, 36000 test samples were generated. The test samples (only samples in the above closed set candidate set) are input into the feature extractor + classifier channel Figure 3 (a), (b), and (c) respectively show the recognition results of channel coding type, Polar coding code rate, and code length. We find that whether it is channel coding type recognition or coding parameter recognition, the ResNet closed set recognizer with Center Loss in the multi-task learning framework outperforms the currently popular CNN and ResNet recognizer. And when more than 4dB, the type recognition accuracy can reach 90%; when more than 2dB, the code rate and code length recognition can reach 90%. In order to illustrate the effectiveness of the present application,Figure 4 (a) (b) (c) respectively draw the visual t-sne representation of the deep features of the CNN identifier, the ResNet identifier and the multi-task learning framework identifier in the Polar code rate identification under 10 dB, it can be seen that the features extracted by the ResNet identifier and the multi-task learning framework identifier are obviously more different than the features extracted based on the CNN, and in addition, the features of the CenterLoss feature extraction network of the application have better intra-class similarity and greater inter-class difference than the ResNet identifier. This shows that for closed set identification, the multi-task identification framework is superior to the current most advanced CNN and ResNet identifier.

[0090] Test open set identification performance: In order to test the performance of open set identification, we introduce the encoding type and parameters that do not appear in the closed set candidate set for testing, and the number of test samples is the same as that of closed set identification. For open set identification of encoding type, introduce convolutional code and RS code as unknown test encoding type. At the same time, for open set identification of encoding parameters, introduce Polar code rate 5 / 8, 6 / 8 as unknown test code rate, and introduce Polar code length 128, 4096 as unknown test code length. In order to illustrate the open set identification performance of the application, the current popular SoftMax open set identifier is used as a benchmark comparison, which uses the hard threshold of the SoftMax of the logit layer to identify unknown test samples (in this experiment, it is set to 0.5). Table 2, Table 3, Table 4 respectively show the identification performance of channel encoding type open set identification, Polar encoding rate open set identification and Polar code length open set identification. We find that the open set identification performance of the application is superior to the SoftMax open set identifier in the scene of 0-10 dB, and this advantage is more obvious at low signal-to-noise ratio.

[0091] After closed set and open set identification test, the closed set identification performance of the multi-task learning framework exceeds the current advanced ResNet closed set identifier and the traditional CNN closed set identifier; for open set identification, the open set performance of the multi-task learning framework exceeds the traditional SoftMax open set identifier. Therefore, the multi-task learning framework can accurately identify the encoding mode and parameters of the received sequence in a complex communication environment while accurately rejecting the encoding mode and parameters that do not appear in the candidate encoding set.

[0092] Table 1 is the network parameters of the multi-task learning framework under different signal-to-noise ratios.

[0093]

[0094] Table 2 is the open set identification performance results of the multi-task learning framework for channel encoding type under different signal-to-noise ratios.

[0095] Test signal to noise ratio (dB) 0 2 4 6 8 10 SoftMax method 31.2％ 35.6％ 60.2％ 65.2％ 70.4％ 70.6％ Multi-task recognition framework 75.2％ 78.5％ 84.4％ 87.4％ 91.0％ 92.3％

[0096] Table 3 is the performance results of the multi-task learning framework for the Polar code rate open-set identification under different signal-to-noise ratios.

[0097] Test signal to noise ratio (dB) 0 2 4 6 8 10 SoftMax method 50.6％ 55.4％ 68.9％ 76.2％ 76.5％ 77.8％ Multi-task recognition framework 78.2％ 80.9％ 90.5％ 95.8％ 97.5％ 98.2％

[0098] Table 4 is the performance results of the multi-task learning framework for the Polar code length open-set identification under different signal-to-noise ratios.

[0099] Test signal to noise ratio (dB) 0 2 4 6 8 10 SoftMax method 54.3％ 54.4％ 67.8％ 77.4％ 79.6％ 80.5％ Multi-task recognition framework 75.2％ 82.4％ 91.6％ 95.9％ 98.5％ 98.5％

Claims

1. A multi-task learning method for blind recognition of channel coding, comprising: Step 1: constructing a multi-task learning framework, the multi-task learning framework comprising: a shared feature extractor network a decoder network and a classifier network The shared feature extraction network For a ResNet network with multiple residual blocks, the function is represented as: Wherein is the encoded truncated sequence space, is the deep network feature space; The classifier is an MLP classification network with Softmax, denoted as a function: where is the deep network feature space, is the label space of all K channel coding types, K is the number of identified categories of the candidate coding set; The decoder For a ResNet network with multiple residual blocks, it can be represented as: wherein is a deep network feature space, is a sequence reconstruction space corresponding to all input encoded sequences; Step 2: generating training samples and test samples of the multi-task learning framework; Step 2.1: training sample generation: selecting a candidate coding set requiring closed set recognition, and generating channel coding sequences to be recognized; Step 2.2: test sample generation: for closed set recognition, test samples are generated according to the training sample flow, and for open set recognition, test sample coding sequences are generated from the candidate coding set and non-candidate coding set; Step 3: introducing a Centerloss loss function and training the multi-task learning framework, i.e. a shared feature extraction network, a decoding network and a classification network; Step 3.1: Set and is a sample in an arbitrary known class and its corresponding label; is the deep network output feature of the sample after passing through the feature extractor ; is the input encoding sequence of the sample after passing through the feature extractor and the decoder ; in addition, is the predicted class probability vector of the sample after passing through the feature extractor and the classifier ; therefore, the loss function of the multi-task learning framework is composed of three parts: the cross-entropy classification loss function of the classifier network the L 1 regular reconstruction loss of the decoder , and the center loss introduced in the feature space ; therefore, the final loss function of the multi-task learning framework is: wherein a, l, b are three constants for balancing the three loss functions, for the closed set recognition classification task, for the reconstruction task, to ensure that the distance between classes in the deep feature space is large and the distance within the class is small. where N denotes the size of training batch, is the cross-entropy classification loss function, which can be represented as where K is the number of identification categories of the candidate encoding set, is the one-hot encoding vector of the sample X i corresponding label y i , is the predicted class probability vector of the sample X i , is the L 1 regular reconstruction loss, which can be represented as where is the input sample; z i is the deep feature vector of the sample X i , y i is the class label of the sample X i , is the class center thereof; Step 3.2: In order to obtain the optimal network parameters Θ corresponding to the three networks in the multi-task learning framework f , Θ g and Θ c , a batch of training data sets is randomly selected as input , and the loss function is calculated . The adaptive moment estimation algorithm (Adam) is used to optimize the parameters of the multi-task learning framework. Through repeated iteration in small batches, the parameters of the multi-task learning framework gradually converge and approach the optimal value. Step 4: performance test of the multi-task learning framework Step 4.1 : The test against the closed set recognition test is as follows: input the test sample X into the feature extractor-classifier channel producing a class probability estimate vector denoted by where P(H n |X) represents the probability that the test sample X is of the i-th class; the final recognition classification result y pred may be represented as Step 4.2: for open set recognition of channel coding types, the test steps are as follows: Step 4.2.1: Pass all training set samples, which are samples in the closed set, through the feature extractor plus decoder to get the reconstructed signal, by the formula Calculate the reconstruction error of the training set samples, and record the decision threshold is the 80% of the maximum reconstruction error in the training set; Step 4.2.2: input the test sample X into the feature extractor, the test sample including the closed set and non-closed set samples, to obtain the deep features thereof Then input the classifier In this way, the closed set and non-closed set samples are classified Meanwhile, the depth feature z is input into a decoder to obtain a reconstruction signal thereof The reconstruction error is calculated If it is determined as an unknown class sample; if it is determined as a known class sample, and the label is y pred .

Images