A signal detection and recognition method based on dynamic label distillation

Abstract

The present invention discloses a signal detection and recognition method based on dynamic label distillation, which relates to the field of signal detection and recognition. The method comprises: constructing a training data set and a teacher network, and pre-training the teacher network on the training data set to determine the pre-trained teacher network; extracting dynamic labels using the training data set and the pre-trained teacher network; training a student network using the dynamic labels, and updating the network parameters of the student network until the training is completed, and using the trained student network for signal detection and recognition based on time-frequency diagrams. Compared with the prior art, the present invention has good universality and low cost, and is particularly suitable for signal detection and recognition scenarios based on time-frequency diagrams.

Description

Technical Field

[0001] The present invention relates to the technical field of signal detection and identification, and more specifically, to a signal detection and identification method based on dynamic label distillation. Background Art

[0002] Signal detection and identification technology plays a vital role in the field of signal and information processing. Its core lies in accurately and quickly identifying and analyzing target signals to provide support for subsequent decision-making and control. For example, in the field of wireless communications, signal detection and identification results can help wireless communication systems cope with signal interference in multi-user environments and ensure the reliability and stability of communications; in the field of spectrum resource management, signal detection and identification is used to monitor spectrum usage, identify illegal or unauthorized signals, protect spectrum resources for legitimate users, and effectively maintain normal communication order.

[0003] In recent years, with the development of artificial intelligence technology, signal detection and recognition methods based on deep learning are becoming a research hotspot. Deep learning models trained with massive electromagnetic signal data can quickly and accurately complete signal detection and recognition. Signal detection and recognition technology based on time-frequency diagrams has a broad application basis. Time-frequency diagrams are easy to obtain, and the use of deep learning image target detection methods can utilize the detection and recognition of multiple signals in a wider detection frequency range.

[0004] In order to further improve the accuracy of signal detection and recognition, the number of parameters and computational complexity of deep target detection network models are gradually increasing. The implementation and application of complex structure deep network models require high-memory, high-throughput servers. Summary of the invention

[0005] In order to overcome the defects of high cost and low versatility in the above-mentioned prior art, the present invention provides a signal detection and recognition method based on dynamic label distillation.

[0006] In order to solve the above technical problems, the technical solution of the present invention is as follows:

[0007] In a first aspect, a signal detection and recognition method based on dynamic tag distillation comprises:

[0008] Constructing a training data set and a teacher network, and pre-training the teacher network on the training data set to determine a pre-trained teacher network; wherein the training data set includes a time-frequency diagram and signal labeling information in the corresponding time-frequency diagram, and the signal labeling information includes a signal category labeling and a signal position labeling;

[0009] Extracting dynamic labels using the training data set and the pre-trained teacher network;

[0010] The dynamic labels are used to train the student network, and the network parameters of the student network are updated until the training is completed. The trained student network is used for signal detection and recognition based on time-frequency diagrams.

[0011] In a second aspect, a computer program product comprises a computer program or computer executable instructions, wherein when the computer program or computer executable instructions are executed by a processor, the method described in the first aspect is implemented.

[0012] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0013] The present invention discloses a signal detection and recognition method based on dynamic label distillation. By constructing a "teacher network-student network" architecture, a pre-trained teacher network pre-trained on a training data set is used to extract dynamic labels, and the dynamic labels are used to guide the training of the student network, so that a lightweight student network is obtained after the training is completed and used for signal detection and recognition. Compared with the prior art, the present invention has good universality and low cost, and can be widely used in signal detection and recognition scenarios based on time-frequency diagrams. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 This is a flow chart of the signal detection and identification method in Example 1 of the present application;

[0015] Figure 2 This is a schematic diagram of training data annotation in Example 1 of the present application;

[0016] Figure 3 This is a schematic diagram of the dynamic tag generation process in Example 1 of the present application. DETAILED DESCRIPTION

[0017] The terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the terms used in this way can be interchanged in appropriate cases, which is only to describe the distinction mode adopted by the objects of the same attributes in the embodiments of the present application when describing. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, so that the process, method, system, product or equipment containing a series of units need not be limited to those units, but may include other units that are not clearly listed or inherent to these processes, methods, products or equipment. The term "determine" widely covers various actions, and may include acquisition, calculation, calculation, processing, derivation, investigation, search (for example, search in a table, database or other data structure), ascertainment, and similar actions, and may also include reception (for example, receiving information), access (for example, accessing data in a memory) and similar actions, and may also include generation, creation, establishment and similar actions, as well as parsing, selection, selection and similar actions, etc. The relevant definitions of other terms will be given in the following description.

[0018] It should be noted that when an element is considered to be "connected" to another element, it can be directly connected to the other element, or connected to the other element through an intermediate element. In addition, the "connection" in the following embodiments should be understood as "electrical connection", "communication connection", etc. if there is transmission of electrical signals or data between the connected objects.

[0019] The drawings are for illustrative purposes only and should not be construed as limiting the present patent;

[0020] In order to better illustrate the present embodiment, some parts in the drawings may be omitted, enlarged or reduced, and do not represent the size of the actual product;

[0021] It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0022] To facilitate implementation by non-technical personnel, some of the technologies in this application are described:

[0023] Knowledge distillation: Knowledge distillation aims to use a high-precision and high-complexity teacher model (T-Model) to guide the optimization process of a lightweight student model (S-Model). The output of the teacher model is used to generate soft labels, and the student model optimizes its own parameters to make the output results as close as possible to the soft labels of the teacher model, thereby realizing the transfer of the "knowledge" information of the teacher model to the student model.

[0024] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0025] Example 1

[0026] As described in the background technology, in order to further improve the accuracy of signal detection and recognition, the number of parameters and the amount of computation of the deep target detection network model gradually increase. On resource-constrained computing devices, the deep target detection network model cannot be directly deployed, and the complex network structure needs to be lightweight.

[0027] In the prior art, it has been seen that complex network models are pruned to obtain lightweight models. However, the performance of the model will be significantly reduced after pruning, so some researchers have further trained the pruned network to make up for the performance loss. By repeating the process of "structural pruning-fine-tuning training" many times, the network structure scale can meet the predetermined requirements; the performance of the lightweight network obtained in this way is limited by the performance of the original complex network model, and the lightweight effect of the model is easily affected by the network structure and data characteristics. In addition, the time spent on repeated fine-tuning training is considerable, which is not conducive to the rapid deployment and application of lightweight models.

[0028] In addition, the existing methods of using knowledge distillation to achieve model lightweighting generally force the student network to directly imitate the final output or intermediate layer output of the teacher network. If the architecture or performance of the student and teacher networks are significantly different, then this hard constraint on the output value will be too strict. Although it is possible to bridge the gap in structure or performance between the student and teacher networks by introducing an intermediate network, this approach will introduce more hyperparameters, larger memory, longer training time, and complicate the process.

[0029] This embodiment provides a signal detection and recognition method based on dynamic tag distillation. Figure 1 ,include:

[0030] Constructing a training data set and a teacher network, and pre-training the teacher network on the training data set to determine a pre-trained teacher network; wherein the training data set includes a time-frequency diagram and signal labeling information in the corresponding time-frequency diagram, and the signal labeling information includes a signal category labeling and a signal position labeling;

[0031] Extracting dynamic labels using the training data set and the pre-trained teacher network;

[0032] The dynamic labels are used to train the student network, and the network parameters of the student network are updated until the training is completed. The trained student network is used for signal detection and recognition based on time-frequency diagrams.

[0033] It should be emphasized that this embodiment proposes a method for implementing a lightweight network model for signal detection and recognition. On the basis of the "teacher network-student network" architecture, dynamic label generation is used to generate new annotation information for the training of the student network, and the link weights of the student network are updated through training; by applying the trained student network, signal analysis and detection based on time-frequency graphs can be completed. The method has better universality and can break through the limitations of existing technologies on network structure and output constraints. The time cost of training is low, and the trained student network is particularly suitable for deployment on resource-constrained computing devices.

[0034] It should be noted that the teacher network is composed of an image target detection network with high accuracy and high computational complexity, including but not limited to the target detection network of the YOLO series. The student network can be obtained by the teacher network based on the model pruning method, or a lightweight network model different from the teacher network can be used, including but not limited to ShuffleNet and MobileNet.

[0035] Those skilled in the art will appreciate that Figure 2 The signal labeling information can be implemented by marking the position of each signal on the corresponding time-frequency diagram with a rectangular box, and associating the signal category label with each rectangular box. As a non-limiting example, it can be expressed in the form of diagonal vertices [x1, y1, x2, y2] or in the form of center point and width and height [x0, y0, w, h].

[0036] It should also be understood that the training termination condition for the teacher network and / or student network in the present application may be that the number of training times reaches a predetermined value, or that the corresponding loss function tends to converge, which is set by technical personnel in this field according to actual conditions and is not limited here.

[0037] In some preferred embodiments, determining a pre-trained teacher network comprises:

[0038] The data of the training data set are sent to the teacher network in batches, the loss function is calculated and the network parameters are updated until the pre-training is completed, and the teacher network with the network parameters finally updated is used as the pre-trained teacher network; wherein,

[0039] For each batch of N training data, the loss function is expressed as follows:

[0040] L teacher =L cls +L loc

[0041] L cls Represents the signal detection, recognition and classification loss, and its expression is:

[0042]

[0043] In the formula, is the classification label of the nth sample, represents the probability that the nth sample predicted by the teacher network is the kth signal category;

[0044] L loc Represents the signal detection, recognition and localization loss, for each prediction window of the teacher network and the matching marked window B gt , whose expression is:

[0045]

[0046] In the formula, express and B gt IoU value of express The center point and B gt The distance from the center point; express and B gt The diagonal length of the smallest enclosed rectangle.

[0047] In some preferred embodiments, the dynamic tag is extracted, see Figure 3 ,include:

[0048] Extract the detection results of the pre-trained teacher network on the training data set as the detection result set D = {D i |i=1,2,…,N},D i Indicates the corresponding detection result of the i-th time-frequency graph among the N time-frequency graphs of the current training batch; wherein the detection result includes the signal detection category and signal detection position corresponding to the time-frequency graph;

[0049] Specifically, the set D can be further expressed as D i ={(c i,k ,l i,k )|k=1,2,…}, where c i,k , l i,l Respectively represent the category and position of the kth signal in the i-th time-frequency diagram;

[0050] Extract the signal annotation information corresponding to the N time-frequency diagrams in the training data set as the annotation data set G={G i |i=1,2,…,N},G i Indicates the corresponding signal labeling information of the i-th time-frequency diagram;

[0051] Specifically, the set G can be further expressed as in Respectively represent the category and position of the kth labeled signal in the i-th time-frequency diagram;

[0052] Determine a matching cost according to the detection result set D and the labeled data set G;

[0053] The first k smallest window sets of the matching costs are extracted as the dynamic labels.

[0054] In some optional embodiments, the expression of the matching cost is:

[0055]

[0056] In the formula, c i,pq represents the matching cost determined according to the p-th detection result and the q-th signal labeling information; Represents the matching cost measured by the angle of the target area saliency; Represents the matching cost measured from the perspective of the classification result; Represents the matching cost measured by the angle of the positioning result.

[0057] In some specific implementations, all matching cost values ​​can form a label matching cost matrix C i :

[0058]

[0059] The label matching cost matrix contains the matching costs of all P detection results and Q labeled data.

[0060] Further, using the dynamic labels to train the student network includes:

[0061] For the N time-frequency graphs of the current training batch:

[0062] The window set with the first k smallest matching costs in the i-th time-frequency graph, i.e., the dynamic label, is B i ={(p,q)|c i,pq ∈Top-k minimum values}; then for the N time-frequency graphs in a single training batch, a window set B with the minimum matching cost can be constructed = {B1, B2,…, B N};

[0063] Using B i The window contained in calculates the loss function of the student network:

[0064] L student =L cls-dyn +L loc-dyn

[0065] Where, L cls-dyn represents the signal detection and recognition classification loss calculated using the optimal dynamic label, L loc-dyn Indicates signal detection, identification and positioning loss;

[0066] According to the loss function L student The network parameters of the student network are updated until the training is completed.

[0067] Furthermore, the expression of the signal detection recognition classification loss is as follows:

[0068]

[0069] In the formula, is the classification label of the nth training sample calculated using dynamic label generation, The probability that the nth sample predicted by the student network is the kth signal category; and,

[0070] The expression of the signal detection identification positioning loss is as follows:

[0071]

[0072] In the formula, for and B gt The IoU value of for The center point and B gt The distance from the center point, for and B gt The diagonal length of the smallest enclosed rectangle.

[0073] Furthermore, the expression for measuring the matching cost from the perspective of the target region saliency (referred to as “target region saliency cost”) is:

[0074]

[0075] Where, for the i-th time-frequency diagram, l i,p and Represent the p-th detection box and the q-th annotation box respectively; Indicates l i,p and The distance between the center points of the two windows, Indicates l i,p and The area occupied by the union of two windows.

[0076] It should be noted that l i,p and The closer the spatial distance is, the greater the saliency cost of the target area. The closer the value is to 0, the more it will be reflected in c i,pq The smaller the value of is, the smaller the matching cost between the p-th detection result and the q-th labeled data is.

[0077] Furthermore, the expression for measuring the matching cost from the perspective of the classification result (referred to as “classification cost”) is:

[0078]

[0079] Where, for the i-th time-frequency diagram, represents the p-th detection box l i,p The confidence of the predicted k-th signal class, represents the p-th detection box l i,p Assigned to the qth annotation box The true confidence value of the k-th signal category; Indicates l i,p and The window overlap value IoU is Indicates l i,p Whether the classification label of is the k-th signal category.

[0080] Furthermore, the expression for measuring the matching cost (referred to as “positioning cost”) from the perspective of the positioning result is:

[0081]

[0082] Where, for the i-th time-frequency diagram, Indicates l i,p and The overlap ratio of the two windows; where

[0083]

[0084] In the formula, |·| represents the area of ​​the graphic region, for l i,p and The minimum bounding rectangle of the two windows.

[0085] It should be noted that l i,p and The higher the overlap, The closer the value is to 0, the smaller the positioning cost is.

[0086] Example 2

[0087] This embodiment provides a computer-readable storage medium, on which is stored at least one instruction, at least one program, code set or instruction set, and the at least one instruction, at least one program, code set or instruction set is loaded and executed by a processor, so that the processor executes part or all of the steps of the method provided in Example 1 of the present application.

[0088] It is understood that the storage medium may be transient or non-transient. Exemplarily, the storage medium includes, but is not limited to, a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and other media that can store program codes.

[0089] Exemplarily, the processor may be a central processing unit (CPU), a microprocessor (MPU), a digital signal processor (DSP), or a field programmable gate array (FPGA).

[0090] Exemplarily, the read-only memory includes but is not limited to MASK ROM, PROM, EPROM, EEPROM, Flash, etc.

[0091] Exemplarily, the random access memory includes but is not limited to DRAM, SRAM, SDRAM, DDR SDRAM, etc.

[0092] In some examples, a computer program product is provided, which can be implemented in hardware, software or a combination thereof. As a non-limiting example, the computer program product can be embodied as the storage medium, or as a software product, such as an SDK (Software Development Kit).

[0093] As a non-limiting example, a computer program product is provided, the computer program product includes a computer program or a computer executable instruction, the computer program or the computer executable instruction is stored in a computer-readable storage medium. A processor of an electronic device reads the computer program or the computer executable instruction from the computer-readable storage medium, and the processor executes the computer executable instruction, so that the electronic device performs some or all steps of the method described in the embodiment of the present application.

[0094] In some examples, a computer program is provided, comprising a computer-readable code, and when the computer-readable code is run in a computer device, a processor in the computer device executes a part or all of the steps in the method.

[0095] This embodiment also proposes an electronic device, including a memory and a processor, wherein the memory stores at least one instruction, at least one program, a code set or an instruction set, and when the processor executes the at least one instruction, at least one program, a code set or an instruction set, it implements part or all of the steps of the method described in Example 1.

[0096] In some examples, a hardware entity of the electronic device is provided, including: a processor, a memory and a communication interface; wherein the processor generally controls the overall operation of the electronic device; the communication interface is used to enable the electronic device to communicate with other terminals or servers through a network; the memory is configured to store instructions and applications executable by the processor, and can also cache data to be processed or processed by the processor and various modules in the electronic device (including but not limited to image data, audio data, voice communication data and video communication data), which can be implemented by flash memory (FLASH), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM) or random access memory (RAM).

[0097] The processor may include one or more processing elements. Thus, the processor may include one or more integrated circuits (ICs) configured to perform the functions of the processor. In addition, each integrated circuit may include a circuit (e.g., a first circuit, a second circuit, and other circuits, etc.) configured to perform the functions of the processor.

[0098] Furthermore, data may be transmitted between the processor, the communication interface and the memory via a bus, which may include any number of interconnected buses and bridges, and the bus connects various circuits of one or more processors and memories together.

[0099] It can be understood that the options in the above-mentioned embodiment 1 are also applicable to this embodiment, so they will not be described again here.

[0100] The same or similar reference numerals correspond to the same or similar components;

[0101] The terms used to describe the positional relationship in the drawings are only used for illustrative purposes and should not be construed as limiting the present application;

[0102] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application may be combined with each other.

[0103] In different specific implementations, the method or system described in the present application can be implemented in software, hardware or a combination thereof. In addition, the order of the steps of the method can be changed, and various elements can be added, reordered, combined, omitted, modified, etc.

[0104] Obviously, the above-mentioned embodiments of the present application are merely examples for clearly illustrating the present application, and are not intended to limit the implementation methods of the present application, and are not intended to limit the present application. For those of ordinary skill in the art, other different forms of changes or modifications can be made based on the above description. Each discrete structure / function module or unit can be integrated together to form an independent part, or each module can exist separately, or two or more modules can be integrated to form an independent part, and the structure and function of the discrete components can be implemented as a combined structure or component. It is not necessary and impossible to enumerate all the implementation methods here. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application should be included in the scope of protection of the claims of the present application.

Claims

1. A signal detection and recognition method based on dynamic tag distillation, characterized in that: include: Constructing a training data set and a teacher network, and pre-training the teacher network on the training data set to determine a pre-trained teacher network; wherein the training data set includes a time-frequency diagram and signal labeling information in the corresponding time-frequency diagram, and the signal labeling information includes a signal category labeling and a signal position labeling; Extracting dynamic labels using the training data set and the pre-trained teacher network; The dynamic labels are used to train the student network, and the network parameters of the student network are updated until the training is completed. The trained student network is used for signal detection and recognition based on time-frequency diagrams.

2. According to claim 1, a signal detection and identification method based on dynamic tag distillation is characterized in that: The extracting of dynamic tags comprises: Extract the detection results of the pre-trained teacher network on the training data set as the detection result set D = {D i |i=1,2,…,N},D i Indicates the corresponding detection result of the i-th time-frequency graph among the N time-frequency graphs of the current training batch; wherein the detection result includes the signal detection category and signal detection position corresponding to the time-frequency graph; Extract the signal annotation information corresponding to the N time-frequency diagrams in the training data set as the annotation data set G={G i |i=1,2,…,N},G i Indicates the corresponding signal labeling information of the i-th time-frequency diagram; Determine a matching cost according to the detection result set D and the labeled data set G; The first k smallest window sets of the matching costs are extracted as the dynamic labels.

3. The signal detection and identification method based on dynamic tag distillation according to claim 2 is characterized in that: The expression of the matching cost is: In the formula, c i,pq represents the matching cost determined according to the p-th detection result and the q-th signal labeling information; Represents the matching cost measured by the angle of the target area saliency; Represents the matching cost measured from the perspective of the classification result; Represents the matching cost measured by the angle of the positioning result.

4. The signal detection and identification method based on dynamic tag distillation according to claim 3 is characterized in that: The expression for measuring the matching cost from the angle of the target region saliency is: Where, for the i-th time-frequency diagram, l i,p and Represent the p-th detection box and the q-th annotation box respectively; Indicates l i,p and The distance between the center points of the two windows, Indicates l i,p and The area occupied by the union of two windows.

5. The signal detection and identification method based on dynamic tag distillation according to claim 3 is characterized in that: The expression for measuring the matching cost from the perspective of the classification result is: Where, for the i-th time-frequency diagram, represents the p-th detection box l i,p The confidence of the predicted k-th signal class, represents the p-th detection box l i,p Assigned to the qth annotation box The true confidence value of the k-th signal category; Indicates l i,p and The window overlap value IoU is Indicates l i,p Whether the classification label of is the k-th signal category.

6. The signal detection and identification method based on dynamic tag distillation according to claim 3 is characterized in that: The expression for measuring the matching cost with respect to the positioning result is: Where, for the i-th time-frequency diagram, Indicates l i,p and The overlap ratio of the two windows; where In the formula, |·| represents the area of ​​the graphic region, for l i,p and The minimum bounding rectangle of the two windows.

7. The signal detection and identification method based on dynamic tag distillation according to claim 3 is characterized in that: The method of training the student network using the dynamic labels includes: For the N time-frequency graphs of the current training batch: The window set with the first k smallest matching costs in the i-th time-frequency graph, i.e., the dynamic label, is B i ={(p,q)|c i,pq ∈Top-k smallest values}; Using B i The window contained in calculates the loss function of the student network: L student =L cls-dyn +L loc-dyn Where, L cls-dyn represents the signal detection and recognition classification loss calculated using the optimal dynamic label, L loc-dyn Indicates signal detection, identification and positioning loss; According to the loss function L student The network parameters of the student network are updated until the training is completed.

8. The signal detection and identification method based on dynamic tag distillation according to claim 7 is characterized in that: The expression of the signal detection recognition classification loss is as follows: In the formula, is the classification label of the nth training sample calculated using dynamic label generation, The probability that the nth sample predicted by the student network is the kth signal category; and, The expression of the signal detection identification positioning loss is as follows: In the formula, for and B gt The IoU value of for The center point and B gt The distance from the center point, for and B gt The diagonal length of the smallest enclosed rectangle.

9. A signal detection and identification method based on dynamic tag distillation according to any one of claims 1 to 8, characterized in that: The step of determining a pre-trained teacher network comprises: The data of the training data set are sent to the teacher network in batches, the loss function is calculated and the network parameters are updated until the pre-training is completed, and the teacher network with the network parameters finally updated is used as the pre-trained teacher network; wherein, For each batch of N training data, the loss function is expressed as follows: L teacher =L vls +L loc L cls Represents the signal detection, recognition and classification loss, and its expression is: In the formula, is the classification label of the nth sample, represents the probability that the nth sample predicted by the teacher network is the kth signal category; L loc Represents the signal detection, recognition and localization loss, for each prediction window of the teacher network and the matching marked window B gt , whose expression is: In the formula, express and B gt IoU value of express The center point and B gt The distance from the center point; express and B gt The diagonal length of the smallest enclosed rectangle.

10. A computer program product comprising a computer program or computer executable instructions, characterized in that: When the computer program or computer executable instructions are executed by a processor, the method according to any one of claims 1 to 9 is implemented.

Images