Portable intelligent signal processing system suitable for rapid detection and identification of communication signals

Abstract

The invention discloses a portable intelligent signal processing system suitable for rapid detection and identification of communication signals. The portable intelligent signal processing system comprises a portable computer hardware device and an upper computer software system. Hardware equipment of the portable computer adopts a CPU + NPU computing architecture; the portable computer hardware device further comprises an internal memory, an SSD hard disk, a battery, an LED displayer and an IO interface. The upper computer software system comprises a data import and export module, a preprocessing module, a neural network deployment module, an intelligent reasoning module, a storage and data management module and a man-machine interaction module. According to the method, an intelligent signal processing technology is introduced, and a data set is constructed through fine signal analysis, identification and labeling of offline signals; a deep network is trained on the basis, and the ability of an existing analyst is transferred to a deep neural network; rapid analysis and identification of communication reconnaissance signals can be completed, knowledge transfer iteration improvement of manual offline analysis is realized, and intelligence generation efficiency is effectively improved.

Description

Technical Field

[0001] The present invention relates to intelligent signal processing technology and application system, belonging to the technical field of communication countermeasure and electromagnetic spectrum management, and specifically to a portable intelligent signal processing system suitable for rapid detection and identification of communication signals. Background Art

[0002] Modern information warfare is extremely dependent on communication means. Communication signals, as carriers of air traffic control, voice commands, tactical command, situation transmission and other important parameter information transmission, play an extremely important strategic role and are the technical basis of network-centric warfare. In the current electronic warfare communication reconnaissance and confrontation process, each link needs to be manually processed and verified, which has poor timeliness and cannot meet the requirements of "actual combat" and "minute-level" intelligence support, and cannot guide electronic countermeasure operations in wartime. On the other hand, if the reconnaissance signal is interfered by large noise or the reconnaissance distance is far, the quality of the collected signal will seriously decline, and it is not easy to extract and effectively identify the signal features by manual means.

[0003] The use of artificial intelligence technology is an important way to overcome the above difficulties and pain points and achieve rapid signal identification in a complex spectrum environment. In August 2017, DARPA announced the launch of the "RF Machine Learning System RFMLS" project, which aims to use the capabilities of machine learning systems to study the composition of the radio spectrum, including understanding the types of signals occupying the spectrum, distinguishing important signals from background signals, identifying signals that do not conform to the rules, and using such systems to identify radio frequency signals from IoT devices, and distinguish these signals from signals that attempt to invade these devices. The U.S. Navy developed a machine learning algorithm for the EA-18G, hoping to gain a decisive military advantage in the electromagnetic spectrum through electromagnetic mobile warfare. The prerequisite is to be able to perceive the electromagnetic situation, which requires the use of advanced data processing methods to parse a large amount of unstructured data. The U.S. Army also uses an electronic warfare system with artificial intelligence algorithms. In 2018, the Army Rapid Capabilities Office also sponsored the "Blind Signal Classification Challenge", using signal processing and artificial intelligence technology to explore the correct detection and classification of the most radio signals.

[0004] There are many literatures on the detection and recognition of communication signals using neural network-based methods, but they do not involve specific engineering implementations. There are also few references to the development of corresponding portable intelligent signal processing platforms. Summary of the invention

[0005] In view of the problems of high manual participation in existing communication signal intelligence reconnaissance and insufficient signal detection and recognition performance, the present invention proposes a portable intelligent signal processing system suitable for rapid detection and recognition of communication signals, including a portable computer hardware device and a host computer software system; the portable computer hardware device adopts a CPU+NPU computing architecture; the portable computer hardware device also includes a memory, an SSD hard disk, a battery, an LED display, and an IO interface; the details are as follows:

[0006] The upper computer software system includes a data import and export module 1, a preprocessing module 2, a neural network deployment module 3, an intelligent reasoning module 4, a storage and data management module 5, and a human-computer interaction module 6; the details are as follows:

[0007] (1) Data import and export module 1

[0008] The data import and export module 1 receives the data import interaction instruction of the human-computer interaction module 6 as input, and imports the actual collected signal data from the ground storage array, the task display and control computer and other external devices into the system through the interfaces such as USB / Gigabit network port / 10G optical port; the data import and export module 1 passes the file path and the associated structure of the imported data to the storage and data management module 5;

[0009] The data import and export module 1 receives the data export interaction instruction of the human-computer interaction module 6 as input, obtains the path and association structure of the target file from the storage and data management module 5, and exports the file to the external device through the interfaces such as USB / Gigabit network port / 10G optical port; the data import and export module 1 feeds back the export progress, error control and other information to the human-computer interaction module 6;

[0010] (2) Preprocessing module 2

[0011] The preprocessing module 2 receives the data preprocessing interaction instruction and preprocessing parameters from the human-computer interaction module 6 as input, preprocesses the data to be processed input from the storage and data management module 5, and feeds back the file list obtained after preprocessing to the human-computer interaction module 6; the preprocessing result output by the preprocessing module 2 is output to the storage and data management module 5, and also output to the intelligent reasoning module 4;

[0012] (3) Neural Network Deployment Module 3

[0013] The neural network deployment module 3 receives the neural network deployment interaction instruction and deployment parameter settings of the human-computer interaction module 6, compiles the network model stored in the SSD hard disk of the portable computer, and outputs the compiled network model to the intelligent reasoning module 4; the neural network deployment module 3 feeds back the selectable network model list and path, network model type and parameter amount, and post-processing required for the network model to the human-computer interaction module 6;

[0014] (4) Intelligent Reasoning Module 4

[0015] The intelligent reasoning module 4 receives the reasoning interaction instruction from the human-computer interaction module 6, receives the preprocessing result from the preprocessing module 2 and the network model from the neural network deployment module 3, and completes the reasoning calculation on the dedicated NPU of the hardware device; the reasoning detection result obtained by the intelligent reasoning module 4 is output to the human-computer interaction module 6 and also output to the storage and data management module 5;

[0016] (5) Storage and data management module 5 and human-computer interaction module 6

[0017] The interactive control relationship between the storage and data management module 5 and the human-computer interaction module 6 and other modules is as described in the first four modules.

[0018] In one embodiment of the present invention, the data import and export module 1 is specifically as follows:

[0019] (a) Data import and export interfaces include Gigabit Ethernet interface and USB3.0; Ethernet is used to remotely access and import the original data collected by the ground data storage array and the mission system display and control computer; USB3.0 is used to build a shared network to import the original data collected by the ground data storage array and the mission system display and control computer;

[0020] (b) There are two types of imported data: one is the actual raw data, and the other is the data exchanged with the same system. Importing refers to importing from external devices to local devices;

[0021] There are three types of exported data: real-time raw data, pre-processed time-frequency matrix data, and inference result data. Export refers to exporting local files to external devices.

[0022] A workflow of a portable intelligent signal processing system suitable for rapid detection and identification of communication signals is proposed, which specifically includes the following steps:

[0023] Step 1: Data import

[0024] The data import and export module 1 receives the data import interaction instruction of the human-computer interaction module 6, and imports the actual collected signal data from the ground storage array, the mission display and control computer and other external devices through the interfaces such as USB / Gigabit Ethernet port / 10G optical port; the imported data is input into the storage and data management module 5, and the signal data is stored in the hard disk of the hardware device and the data is managed; the data import and export module 1 receives the data export interaction instruction of the human-computer interaction module 6, receives the data file path given by the storage and data management module 5, and exports the file to the external device through the interfaces such as USB / Gigabit Ethernet port / 10G optical port;

[0025] Step 2: Data Preprocessing

[0026] The preprocessing module 2 receives the data preprocessing interaction instructions and preprocessing parameters from the human-computer interaction module 6, and processes the data to be processed input from the storage and data management module 5; the preprocessing results output by the preprocessing module 2 are managed by the storage and data management module 5, and output to the intelligent reasoning module 4 for subsequent reasoning;

[0027] Step 3: Neural Network Deployment

[0028] The neural network deployment module 3 receives the neural network deployment interaction instruction and deployment parameter settings from the human-computer interaction module 6, compiles the network model stored in the SSD hard disk of the portable computer, and outputs the compiled network model to the intelligent reasoning module 4;

[0029] Step 4: Intelligent Reasoning

[0030] The preprocessed time-frequency matrix is ​​fed into the deployed neural network for inference. The NPU completes the inference calculation after the time-frequency matrix is ​​input into the neural network and outputs it.

[0031] The intelligent reasoning module 4 receives the reasoning interaction instructions from the human-computer interaction module 6, receives the preprocessing results from the preprocessing module 2 and the network model from the neural network deployment module 3, and completes the reasoning calculation on the dedicated NPU of the hardware device; the reasoning detection results obtained by the NPU are output to the human-computer interaction module 6 for visual display, and are also output to the storage and data management module 5 for result storage and management.

[0032] In another embodiment of the present invention, the preprocessing process involved in the second step "data preprocessing" is specifically as follows:

[0033] (1) Data reading and data shaping

[0034] The signal data is stored in the portable computer SSD hard disk in binary file IFD / dat format. Each file IFD / dat consists of a file header with a fixed byte length and a subsequent acquisition data segment. The file header stores parameters such as sampling bandwidth, sampling rate and center frequency. The details are as follows:

[0035] (a) Preprocessing module 2 reads the header file of the signal data from the SSD hard disk of the portable computer and records the acquisition time, sampling bandwidth, sampling rate and center frequency for subsequent display;

[0036] (b) The default format of the collected data segments in the signal data file in the SSD hard disk of the portable computer is a complex form, which is divided into two channels, I channel and Q channel, and is stored in an alternating form of IQIQIQ. The pre-processing module 2 reads the collected data segments and performs data shaping on them, converting the collected data segments into the form of I+jQ, where j is an imaginary unit, for subsequent time-frequency conversion processing;

[0037] (c) The data reading form adopts one of the following two methods: one is that the preprocessing module 2 automatically loads all signal data from the portable computer SSD hard disk into the memory at one time; the other is that a certain signal data is manually loaded from the portable computer SSD hard disk into the memory;

[0038] (2) Signal preprocessing

[0039] Perform signal preprocessing operations such as filtering, denoising, and resampling on the data in the form of I+jQ obtained in the previous step to obtain preprocessed data. The preprocessing operations can be expanded according to needs;

[0040] (3) Data segmentation

[0041] First, segment the preprocessed data; when segmenting the data, set the length and step size of each segment;

[0042] (4) Time-frequency transformation

[0043] After the data is segmented, each segment of data is transformed into a time-frequency matrix to obtain a time-frequency matrix of the same size as the training environment, as follows:

[0044] (a) Short-time Fourier transform (STFT) is used as the default time-frequency transform. The configurable parameters include window type, window overlap, and discrete Fourier transform (DFT) points. The time-frequency matrix obtained after the time-frequency transform is a complex-valued matrix.

[0045] (b) the complex value matrix obtained by the time-frequency transformation is sequentially subjected to squaring / modulo, logarithmic, and normalization processing, and the processed matrix becomes a real value matrix;

[0046] (5) Time-frequency matrix storage

[0047] The real-valued time-frequency matrix obtained in the previous step is stored in two forms: image and matrix. Image storage is to further convert the time-frequency matrix into uint8 data format, so that the matrix value is transformed into the image value range of [0,255], and then saved in image format; matrix storage is to directly save the original time-frequency matrix.

[0048] In yet another embodiment of the present invention, in step (4) "time-frequency transform" (a), the window type is selected as a Hamming window, and the window overlap and discrete Fourier transform are set to 256 points and 512 points respectively.

[0049] In another embodiment of the present invention, in the fourth step "intelligent reasoning", reasoning adopts a multiple NPU parallel reasoning architecture, that is, the input time-frequency matrix is ​​sent to multiple NPUs at the same time, and multiple NPUs calculate simultaneously; different network models are supported to be deployed on different NPUs during reasoning.

[0050] The communication signal intelligent processing system first converts the collected signal time series into a time-frequency matrix through preprocessing, and then inputs it into the neural network deployed at the back end for reasoning, completing the rapid detection, identification and time-frequency parameter estimation of wireless communication signals, and can also summarize and visualize the detection results. The platform is equipped with interactive and friendly host computer software and UI interface, which can realize neural network deployment, data import, preprocessing, reasoning, storage, human-computer interaction and other functions. The system introduces intelligent signal processing technology, and constructs a data set through fine signal analysis, recognition and annotation of offline signals; on this basis, it trains deep networks and transfers the capabilities of existing analysts to deep neural networks; it can complete the rapid analysis and identification of communication reconnaissance signals, realize the iterative improvement of knowledge transfer of manual offline analysis, and effectively improve the efficiency of intelligence generation. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 A connection diagram showing a portable intelligent signal processing system suitable for rapid detection and identification of communication signals according to the present invention;

[0052] Figure 2 The present invention shows the workflow of a portable intelligent signal processing system suitable for rapid detection and identification of communication signals;

[0053] Figure 3 The pre-processing process is shown. DETAILED DESCRIPTION

[0054] 1. A portable intelligent signal processing system suitable for rapid detection and identification of communication signals;

[0055] The invention provides a portable intelligent signal processing system suitable for rapid detection and identification of communication signals, which consists of two parts: a portable computer hardware device and a host computer software system.

[0056] The portable computer hardware device adopts the CPU+NPU computing architecture, and can support the deployment of models of various neural network frameworks such as Pytorch, Caffe, Darknet, etc., and supports multi-core multi-algorithm parallel acceleration, multi-core single algorithm acceleration, large image partition multi-core acceleration, multi-algorithm simultaneous loading, runtime algorithm switching (ms level), dual algorithm one-key configuration continuous reasoning, etc. The portable computer hardware device also includes memory, SSD hard disk, battery, LED display, IO interface, etc., as known to those skilled in the art.

[0057] The upper computer software system includes data import and export module 1, preprocessing module 2, neural network deployment module 3, intelligent reasoning module 4, storage and data management module 5, and human-computer interaction module 6. The connection relationship between the six modules and the portable computer hardware device is as follows: Figure 1 shown.

[0058] 1. Data import and export module 1

[0059] The data import and export module 1 receives the data import interaction instruction of the human-computer interaction module 6 as input, and imports the actual collected signal data from the ground storage array, the mission display and control computer and other external devices into the system through the USB / Gigabit Ethernet port / 10G optical port and other interfaces. In addition, while importing, the data import and export module 1 will pass the file path and associated structure of the imported data to the storage and data management module 5, which will manage it in a unified manner to realize subsequent convenient data file retrieval, screening and other operations.

[0060] For data export, the data import and export module 1 receives the data export interaction instruction of the human-computer interaction module 6 as input, obtains the path and association structure of the target file from the storage and data management module 5, and exports the file to an external device through an interface such as USB / Gigabit Ethernet port / 10G optical port. Similarly, during the data export process, the data import and export module 1 feeds back information such as export progress and error control to the human-computer interaction module 6 and then displays it on the interface.

[0061] (1) Data import and export interfaces include Gigabit Ethernet interface and USB3.0. Ethernet can be used to remotely access and import the real-time raw data of the ground data storage array and the mission system display and control computer, and USB3.0 can be used to build a shared network to import the real-time raw data of the ground data storage array and the mission system display and control computer;

[0062] (2) There are two types of imported data: one is the actual raw data, and the other is the data exchanged with the same system. Importing refers to importing from external devices to local devices. The two types of data import methods are shown in the following table.

[0063]

[0064]

[0065] There are three types of export data: real-time raw data, pre-processed time-frequency matrix data, and inference result data. Export refers to exporting local files to external devices. The export method is independent or joint export of the three types of data, as shown in the following table.

[0066]

[0067] (3) Supported import file formats: including but not limited to IFD, dat, bin, mat and other formats, which can be converted to each other.

[0068] 2. Preprocessing module 2

[0069] The preprocessing module 2 receives the data preprocessing interaction instruction and preprocessing parameters from the human-computer interaction module 6 as input, preprocesses the data to be processed input from the storage and data management module 5, and feeds the file list obtained after preprocessing back to the human-computer interaction module 6 for display. In addition, the preprocessing result output by the preprocessing module 2 is not only managed by the storage and data management module 5, but also output to the intelligent reasoning module 4 for subsequent reasoning.

[0070] 3. Neural Network Deployment Module 3

[0071] The neural network deployment module 3 receives the neural network deployment interaction instruction and deployment parameter settings of the human-computer interaction module 6 as input, compiles the network model stored in the portable computer SSD hard disk, and outputs the compiled network model to the intelligent reasoning module 4. In addition, the neural network deployment module 3 also feeds back information such as the list and path of selectable network models, the type and parameter amount of network models, and the post-processing required for the network model to the human-computer interaction module 6 for display on the interface.

[0072] 4. Intelligent reasoning module 4

[0073] The intelligent reasoning module 4 receives the reasoning interaction instruction of the human-computer interaction module 6 as input. On this basis, the intelligent reasoning module 4 also takes the preprocessing result from the preprocessing module 2 and the network model from the neural network deployment module 3 as input to complete the reasoning calculation on the dedicated NPU of the hardware device. The reasoning detection result obtained by the intelligent reasoning module 4 is not only output to the human-computer interaction module 6 for visual display, but also output to the storage and data management module 5 for result storage and management.

[0074] 5. Storage and data management module 5 and human-computer interaction module 6

[0075] The interactive control relationship between the storage and data management module 5 and the human-computer interaction module 6 and other modules is as described in the first four modules.

[0076] 2. Workflow of portable intelligent signal processing system suitable for rapid detection and identification of communication signals;

[0077] The workflow of the system from importing raw signal acquisition data to signal detection, recognition and reasoning is as follows: Figure 2 As shown, there are four steps in total: data import, data preprocessing, neural network deployment, and intelligent reasoning.

[0078] Step 1: Data import

[0079] Data import involves the data import and export module 1, the storage and data management module 5 and the human-computer interaction module 6. The data import and export module 1 receives the data import interaction instruction of the human-computer interaction module 6 as input, and imports the actual collected signal data from the ground storage array, the mission display and control computer and other external devices through interfaces such as USB / gigabit network port / 10G optical port. The imported data is input into the storage and data management module 5, and the signal data is stored in the hard disk of the hardware device while data management is performed. In addition, for data export, the data import and export module 1 receives the data export interaction instruction of the human-computer interaction module 6 as input, and also takes the data file path given by the storage and data management module 5 as input, and exports the file to the external device through interfaces such as USB / gigabit network port / 10G optical port.

[0080] Step 2: Data Preprocessing

[0081] Data preprocessing involves the data import and export module 2, the intelligent reasoning module 4, the storage and data management module 5 and the human-computer interaction module 6. The preprocessing module 2 receives the data preprocessing interaction instructions and preprocessing parameters of the human-computer interaction module 6 as input, and processes the data to be processed input from the storage and data management module 5. In addition, the preprocessing results output by the preprocessing module 2 are not only managed by the storage and data management module 5, but also output to the intelligent reasoning module 4 for subsequent reasoning.

[0082] The preprocessing process involved in the preprocessing module 2 is as follows Figure 3 As shown, it mainly includes five steps: data reading and data shaping, signal preprocessing, data segmentation, time-frequency transformation, and time-frequency matrix storage, as follows:

[0083] (1) Data reading and data shaping

[0084] The signal data is stored in the portable computer SSD hard disk in binary file IFD / dat format. Each file IFD / dat consists of a file header with a fixed byte length and a subsequent acquisition data segment. The file header stores parameters such as sampling bandwidth, sampling rate, and center frequency. The details are as follows:

[0085] (a) The preprocessing module 2 reads the header file of the signal data from the SSD hard disk of the portable computer and records the parameters such as the acquisition time, sampling bandwidth, sampling rate and center frequency for subsequent display.

[0086] (b) The default format of the collected data segments in the signal data file in the SSD hard disk of the portable computer is complex form, which is divided into two paths, I path and Q path (I path and Q path are well known to those skilled in the art), and is stored in an alternating form of IQIQIQ. The preprocessing module 2 reads the collected data segments and performs data shaping on them, converting the collected data segments into the form of I+jQ (j is an imaginary unit) for subsequent time-frequency transformation processing.

[0087] (c) There are two ways to read data. One is that the preprocessing module 2 automatically loads all signal data from the portable computer SSD hard disk into the memory at one time; the other is that a certain signal data can be manually loaded from the portable computer SSD hard disk into the memory.

[0088] (2) Signal preprocessing

[0089] The data in the form of I+jQ obtained in the previous step is subjected to signal preprocessing operations such as filtering, denoising, and resampling to obtain preprocessed data, and the preprocessing operations can be expanded according to needs. “Signal preprocessing operations such as filtering, denoising, and resampling” are well known to those skilled in the art and will not be described again.

[0090] (3) Data segmentation

[0091] A fixed-size time-frequency matrix corresponds to a fixed-length time series. Therefore, the preprocessed data is segmented first. When segmenting the data, the length and step size of each segment can be set, for example, the length of each segment is 60,000 points and the step size is 30,000 points.

[0092] (4) Time-frequency transformation

[0093] After the data is segmented, each segment of data is transformed into a time-frequency matrix to obtain a time-frequency matrix of the same size as the training environment, as follows:

[0094] (a) Short-time Fourier transform (STFT) is used as the default time-frequency transform. The configurable parameters include window type, window overlap, and discrete Fourier transform (DFT) point number. For example, the window type is Hamming window, and the window overlap and discrete Fourier transform are set to 256 points and 512 points, respectively. The time-frequency matrix obtained after the time-frequency transform is a complex-valued matrix.

[0095] (b) The complex value matrix obtained by the time-frequency transformation is sequentially subjected to squaring / modulo, logarithmic, and normalization processing, and the processed matrix becomes a real value matrix. The squaring / modulo, logarithmic, and normalization processing are well known to those skilled in the art and will not be described again.

[0096] Time-frequency transformation is the key to obtaining the inference time-frequency matrix.

[0097] (5) Time-frequency matrix storage

[0098] The real-valued time-frequency matrix obtained in the previous step is stored in two forms: image and matrix. Image storage is to further convert the time-frequency matrix into uint8 data format (this data format is well known to those skilled in the art), so that the matrix values ​​are transformed into the image value range of [0,255], and then saved in image format, such as .png or .jpg format. Matrix storage is to directly save the original time-frequency matrix.

[0099] Step 3: Neural Network Deployment

[0100] The neural network deployment involves a neural network deployment module 3, an intelligent reasoning module 4, a storage and data management module 5 and a human-computer interaction module 6. The neural network deployment module 3 receives the neural network deployment interaction instructions and deployment parameter settings of the human-computer interaction module 6 as input, compiles the network model stored in the portable computer SSD hard disk (the present invention uses the neural network compilation tool supporting the HKN201NPU chip developed by Xi'an Xiangteng Microelectronics Technology Co., Ltd. for compilation), and outputs the compiled network model to the intelligent reasoning module 4.

[0101] The types of neural networks that can be deployed include CNN (Pytorch), etc. The network deployment precision supports float16 and int8.

[0102] Step 4: Intelligent Reasoning

[0103] The time-frequency matrix obtained by preprocessing is sent to the deployed neural network for reasoning (the process of inputting the time-frequency matrix into the neural network to obtain the output is called reasoning, and this term is well known to those skilled in the art). The NPU completes the reasoning calculation after the time-frequency matrix is ​​input into the neural network, and outputs it. The classification of the signal type and the regression estimation of the signal parameters (the parameters involve the signal spectrum occurrence position, bandwidth, duration, etc.) can be output, involving the preprocessing module 2, the neural network deployment module 3, the intelligent reasoning module 4, the storage and data management module 5 and the human-computer interaction module 6.

[0104] The intelligent reasoning module 4 receives the reasoning interaction instruction from the human-computer interaction module 6 as input. On this basis, the intelligent reasoning module 4 also takes the preprocessing result from the preprocessing module 2 and the network model from the neural network deployment module 3 as input to complete the reasoning calculation on the dedicated NPU of the hardware device. The reasoning detection result obtained by the intelligent reasoning module 4 from the NPU is not only output to the human-computer interaction module 6 for visual display, but also output to the storage and data management module 5 for result storage and management.

[0105] Inference uses a multi-NPU parallel inference architecture, that is, the input time-frequency matrix is ​​sent to multiple NPUs at the same time, and multiple NPUs calculate simultaneously. Different network models can be deployed on different NPUs during inference. Inference with int8 and fp16 precision is supported.

[0106] The intelligent processing platform suitable for rapid detection and identification of communication signals first converts the collected signal time series into a time-frequency matrix through preprocessing, and then inputs it into the neural network deployed at the back end for reasoning, and finally completes the rapid detection, identification and time-frequency parameter estimation of wireless signals, and can also summarize and visualize the detection results. The data import and export module realizes the import of real-time raw data and interactive data with the same system in a variety of ways; at the same time, it realizes the export of real-time raw data, preprocessed time-frequency matrix data, and reasoning result data. The preprocessing module uses short-time Fourier transform to preprocess and calculate the input wireless acquisition data to form a time-frequency matrix. The neural network deployment module completes the deployment of the neural network after training. The intelligent reasoning module sends the time-frequency matrix to the deployed neural network for reasoning, and the NPU completes the calculation, which can output the classification of signal types and signal parameter regression estimation (parameters involve the position, bandwidth, duration, etc. of the signal spectrum). The storage and data management module stores and manages the input raw acquisition data, the preprocessed time-frequency matrix, and the results after reasoning, while ensuring fast retrieval; the human-computer interaction module provides human-computer interaction to achieve fast and friendly operation.

Claims

1. A portable intelligent signal processing system suitable for rapid detection and identification of communication signals, comprising a portable computer hardware device and a host computer software system; the portable computer hardware device adopts a CPU+NPU computing architecture; the portable computer hardware device also includes a memory, an SSD hard disk, a battery, an LED display, and an IO interface; the characteristics are: The upper computer software system includes a data import and export module (1), a preprocessing module (2), a neural network deployment module (3), an intelligent reasoning module (4), a storage and data management module (5), and a human-computer interaction module (6); the details are as follows: (1)Data import and export module (1) The data import and export module (1) receives the data import interaction instruction of the human-computer interaction module (6) as input, and imports the actual collected signal data from external devices such as the ground storage array and the task display and control computer into the system through the interfaces such as USB / Gigabit network port / 10G optical port; the data import and export module (1) transmits the file path and associated structure of the imported data to the storage and data management module (5); The data import and export module (1) receives the data export interaction instruction of the human-computer interaction module (6) as input, obtains the path and association structure of the target file from the storage and data management module 5, and exports the file to an external device through the interfaces such as USB / Gigabit Ethernet port / 10G optical port; The data import and export module (1) feeds back information such as export progress and error control to the human-computer interaction module (6); (2) Preprocessing module (2) The preprocessing module (2) receives the data preprocessing interaction instruction and preprocessing parameters of the human-computer interaction module (6) as input, preprocesses the data to be processed input from the storage and data management module (5), and feeds back the file list obtained after the preprocessing to the human-computer interaction module (6); The preprocessing result output by the preprocessing module (2) is output to the storage and data management module (5) and also output to the intelligent reasoning module (4); (3) Neural Network Deployment Module (3) The neural network deployment module (3) receives the neural network deployment interaction instruction and deployment parameter settings from the human-computer interaction module (6), compiles the network model stored in the SSD hard disk of the portable computer, and outputs the compiled network model to the intelligent reasoning module (4); the neural network deployment module (3) feeds back the selectable network model list and path, network model type and parameter quantity, and post-processing required for the network model to the human-computer interaction module (6); (4) Intelligent reasoning module (4) The intelligent reasoning module (4) receives the reasoning interaction instruction from the human-computer interaction module (6), receives the preprocessing result from the preprocessing module (2) and the network model from the neural network deployment module (3), and completes the reasoning calculation on the dedicated NPU of the hardware device; the reasoning detection result obtained by the intelligent reasoning module (4) is output to the human-computer interaction module (6) and also output to the storage and data management module (5); (5) Storage and data management module (5) and human-computer interaction module (6) The interactive control relationship between the storage and data management module (5) and the human-computer interaction module (6) and other modules is as described in the first four modules.

2. The portable intelligent signal processing system suitable for rapid detection and identification of communication signals as claimed in claim 1, characterized in that: The data import and export module (1) is as follows: (a) Data import and export interfaces include Gigabit Ethernet interface and USB3.0; Ethernet is used to remotely access and import the original data collected by the ground data storage array and the mission system display and control computer; USB3.0 is used to build a shared network to import the original data collected by the ground data storage array and the mission system display and control computer; (b) There are two types of imported data: one is the actual raw data, and the other is the data exchanged with the same system. Importing refers to importing from external devices to local devices; There are three types of exported data: real-time raw data, pre-processed time-frequency matrix data, and inference result data. Export refers to exporting local files to external devices.

3. The workflow of a portable intelligent signal processing system suitable for rapid detection and identification of communication signals is characterized in that: The specific steps include: Step 1: Data import The data import and export module (1) receives the data import interaction instruction of the human-computer interaction module (6), and imports the actual collected signal data from external devices such as the ground storage array and the mission display and control computer through interfaces such as USB / Gigabit Ethernet port / 10G optical port; The imported data is input into the storage and data management module (5), the signal data is stored in the hard disk of the hardware device and the data is managed; the data import and export module (1) receives the data export interaction instruction of the human-computer interaction module (6), receives the data file path given by the storage and data management module (5), and exports the file to the external device through the interfaces such as USB / Gigabit Ethernet port / 10G optical port; Step 2: Data Preprocessing The preprocessing module (2) receives the data preprocessing interaction instruction and preprocessing parameters from the human-computer interaction module (6), and processes the data to be processed input from the storage and data management module (5); The preprocessing result output by the preprocessing module (2) is managed by the storage and data management module (5) and output to the intelligent reasoning module (4) for subsequent reasoning; Step 3: Neural Network Deployment The neural network deployment module (3) receives the neural network deployment interaction instruction and deployment parameter settings from the human-computer interaction module (6), compiles the network model stored in the SSD hard disk of the portable computer, and outputs the compiled network model to the intelligent reasoning module (4); Step 4: Intelligent Reasoning The preprocessed time-frequency matrix is ​​fed into the deployed neural network for inference. The NPU completes the inference calculation after the time-frequency matrix is ​​input into the neural network and outputs it. The intelligent reasoning module (4) receives the reasoning interaction instruction from the human-computer interaction module (6), receives the preprocessing result from the preprocessing module (2) and the network model from the neural network deployment module (3), and completes the reasoning calculation on the dedicated NPU of the hardware device; the reasoning detection result obtained by the NPU is output to the human-computer interaction module (6) for visual display, and is also output to the storage and data management module (5) for result storage and management.

4. According to the workflow of the portable intelligent signal processing system suitable for rapid detection and identification of communication signals as claimed in claim 3, the preprocessing process involved in the second step "data preprocessing" is specifically as follows: (1) Data reading and data shaping The signal data is stored in the portable computer SSD hard disk in binary file IFD / dat format. Each file IFD / dat consists of a file header with a fixed byte length and a subsequent acquisition data segment. The file header stores parameters such as sampling bandwidth, sampling rate and center frequency. The details are as follows: (a) The preprocessing module (2) reads the header file of the signal data from the SSD hard disk of the portable computer and records the acquisition time, sampling bandwidth, sampling rate and center frequency for subsequent display; (b) The default format of the collected data segments in the signal data file in the SSD hard disk of the portable computer is a complex form, which is divided into two paths, I and Q. When stored, they are stored in an alternating form of IQIQIQ. The preprocessing module (2) reads the collected data segments and reshapes them, converting the collected data segments into the form of I+jQ, where j is an imaginary unit, for subsequent time-frequency conversion processing; (c) The data reading form adopts one of the following two methods: one is that the preprocessing module (2) automatically loads all signal data from the portable computer SSD hard disk into the memory at one time; the other is that a certain signal data is manually loaded from the portable computer SSD hard disk into the memory; (2) Signal preprocessing Perform signal preprocessing operations such as filtering, denoising, and resampling on the data in the form of I+jQ obtained in the previous step to obtain preprocessed data. The preprocessing operations can be expanded according to needs; (3) Data segmentation First, segment the preprocessed data; when segmenting the data, set the length and step size of each segment; (4) Time-frequency transformation After the data is segmented, each segment of data is transformed into a time-frequency matrix to obtain a time-frequency matrix of the same size as the training environment, as follows: (a) Short-time Fourier transform (STFT) is used as the default time-frequency transform. The configurable parameters include window type, window overlap, and discrete Fourier transform (DFT) points. The time-frequency matrix obtained after the time-frequency transform is a complex-valued matrix. (b) the complex value matrix obtained by the time-frequency transformation is sequentially subjected to squaring / modulo, logarithmic, and normalization processing, and the processed matrix becomes a real value matrix; (5) Time-frequency matrix storage The real-valued time-frequency matrix obtained in the previous step is stored in two forms: image and matrix. Image storage is to further convert the time-frequency matrix into uint8 data format, so that the matrix value is transformed into the image value range of [0,255], and then saved in image format; matrix storage is to directly save the original time-frequency matrix.

5. According to the workflow of the portable intelligent signal processing system suitable for rapid detection and identification of communication signals as described in claim 4, in step (4) "time-frequency transformation" (a), the window type selects the Hamming window, and the window overlap and discrete Fourier transform are set to 256 points and 512 points respectively.

6. According to the workflow of the portable intelligent signal processing system suitable for rapid detection and identification of communication signals as described in claim 3, in the fourth step "intelligent reasoning", reasoning adopts a multiple NPU parallel reasoning architecture, that is, the input time-frequency matrix is ​​sent to multiple NPUs at the same time, and multiple NPUs calculate simultaneously; different network models can be deployed on different NPUs during reasoning.

Images