Large-bandwidth signal classification method based on multi-dimensional variable matrix operation

Abstract

The invention aims to provide a large-bandwidth signal classification method based on multi-dimensional variable matrix operation, which can solve the problem of limited identification and classification efficiency of accurately and quickly detecting and identifying various target signals in a large bandwidth under the traditional large-bandwidth spectrum monitoring. The method has the following characteristics that the narrow-band channelization function efficiency of a large-bandwidth signal is improved through channelization segmentation efficient matrix structure operation and a hierarchical filtering mode; by constructing a variable-length feature matrix template and adopting a halving and fine search strategy, the efficiency of parallel realization of feature sequence template matching of thousands of narrow-band signals is improved; according to different practical application scenes, the dimension of the feature matrix can be dynamically adjusted to realize the dynamic expansion of the matching length and the type of the template, and the requirements of different processing durations and bandwidths of data are met. The method is high in flexibility, low in complexity, high in target signal recognition and classification performance and particularly suitable for a high-speed parallel signal recognition and classification system.

Description

Technical Field

[0001] The present invention belongs to the technical field of digital signal processing, and particularly relates to a large-bandwidth signal classification method based on multi-dimensional variable matrix operations. Background Art

[0002] The air wireless electromagnetic signal environment is dense, including signals of various frequency bands such as short waves, ultra-short waves, and microwaves. The communication frequency band is wide, the electromagnetic environment is complex, and various signals such as broadcast signals, civil signals, and industrial noises are mixed together. The target signal is hidden in the crowded complex electromagnetic environment. In the actual electromagnetic environment, being able to directly and quickly detect and identify the target radio signal in a large broadband environment is of great significance for subsequent signal processing.

[0003] The mainstream methods for traditional large-bandwidth signal detection and identification are to search for the large-bandwidth spectrum energy, and then analyze the signals one by one; or use improved means such as dynamic thresholds and signal background iteration as the thresholds for detecting signals. Especially for weak signals, the detection probability of signals is improved. However, these methods all have the same defect that they first detect the presence and position of signals under broadband, and then filter and analyze the signals at that position to complete the signal identification and classification. The above serial processing mechanisms all have the same defect that the processing bandwidth is relatively narrow, and it is difficult to directly identify and classify signals under broadband. In terms of processing mechanisms and speeds, it is difficult to meet the identification requirements of target signals in a large-bandwidth complex electromagnetic environment, and false alarms and missed alarms are likely to occur.

[0004] In summary, the signal processing system under communication large bandwidth needs to be able to process a large amount of data in an extremely short time. The existing technologies cannot meet the requirements of accurately and quickly detecting and identifying target signals under communication large bandwidth. It is necessary to design a high-speed parallel operation algorithm, which puts forward higher requirements for the accuracy and real-time performance of high-speed signal processing design, and meets the requirements of being able to detect and identify target signals in real time and quickly under communication large bandwidth. Summary of the Invention

[0005] The purpose of the present invention is to provide a large-bandwidth signal classification method based on multi-dimensional variable matrix operations, which can solve the problem of limited recognition speed of accurately and quickly detecting and identifying multiple target signals in traditional large-bandwidth signal classification and recognition methods.

[0006] In the first aspect of the embodiment of the present invention, a large-bandwidth signal classification method based on multi-dimensional variable matrix operations is disclosed, including:

[0007] S1, obtaining a discrete signal to be classified, a sampling rate Fs, the total number D of filtering channels, and the number of two-stage filtering channels; the number of two-stage filtering channels includes the number NUM1 of primary filtering channels and the number NUM2 of secondary filtering channels; the length of the discrete signal to be classified is Len;

[0008] S2, obtain the feature template matrix; the dimension of the feature template matrix is M×N, where N represents the category of the feature signal and M represents the length of the feature signal; the column vector of the feature template matrix is a feature signal;

[0009] S3, based on the feature template matrix, classify the discrete signal to be classified to obtain the category value of the discrete signal to be classified;

[0010] The classifying the discrete signal to be classified based on the feature template matrix to obtain the category value of the discrete signal to be classified includes:

[0011] S31, design a first-level filter matrix and a second-level filter matrix based on the sampling rate Fs, the total number of filter channels D, and the number of two-level filter channels;

[0012] S32, process the discrete signal to be classified using the first-level filter matrix to obtain a first filtered output matrix;

[0013] S33, process the first filtered output matrix using the second-level filter matrix to obtain a two-dimensional data matrix;

[0014] S34, perform matching calculation processing on the two-dimensional data matrix based on the feature template matrix to obtain a set of matching quantities;

[0015] S35, perform category extraction processing on the set of matching quantities to obtain the category value of the discrete signal to be classified.

[0016] The designing a first-level filter matrix and a second-level filter matrix based on the sampling rate Fs, the total number of filter channels D, and the number of two-level filter channels includes:

[0017] S311, construct a first-level filter matrix according to the sampling rate Fs and the number of first-level filter channels NUM1;

[0018] S312, construct a second-level filter matrix according to the sampling rate Fs and the number of second-level filter channels NUM2.

[0019] The constructing a first-level filter matrix according to the sampling rate Fs and the number of first-level filter channels NUM1 includes:

[0020] Evenly divide the effective bandwidth corresponding to the sampling rate Fs into NUM1 adjacent sub-bands;

[0021] Set the first-level filter order constant len_filter1 to len_filter1 = 2D;

[0022] For the frequency band range of each sub-band, using the digital filter design method, a filter vector with an order of len_filter1 / NUM1 for each sub-band is designed;

[0023] Using the filter vectors of all sub-bands, a first-level filter matrix is constructed.

[0024] The second-level filter matrix is constructed according to the sampling rate Fs and the number of second-level filtering channels NUM2, including:

[0025] The effective bandwidth corresponding to the sampling rate Fs is evenly divided into NUM2 adjacent sub-bands;

[0026] Set the second-level filtering order constant len_filter2 to len_filter2 = 4D;

[0027] For the frequency band range of each sub-band, using the digital filter design method, a filter vector with an order of len_filter2 / NUM2 for each sub-band is designed;

[0028] Using the filter vectors of all sub-bands, a second-level filter matrix is constructed.

[0029] Processing the discrete signal to be classified using the first-level filter matrix to obtain a first filtered output matrix, including:

[0030] Represent the discrete signal to be classified as a digital matrix matching the dimension of the first filtered output matrix;

[0031] Using the first-level filter matrix, perform parallel filtering on the column vectors of the digital matrix to obtain a first output matrix;

[0032] Perform FFT processing on each column vector of the first output matrix to obtain a first filtered output matrix.

[0033] Processing the first filtered output matrix using the second-level filter matrix to obtain a two-dimensional data matrix, including:

[0034] Using the second-level filter matrix, perform parallel filtering on the column vectors of the first filtered output matrix to obtain a second output matrix;

[0035] Perform FFT processing on each column vector of the second output matrix to obtain a two-dimensional data matrix.

[0036] Performing matching calculation processing on the two-dimensional data matrix based on the feature template matrix to obtain a set of matching quantities, including:

[0037] S341. Perform dimensionality expansion processing on the two-dimensional data matrix and the feature template matrix respectively to obtain corresponding four-dimensional data matrices and four-dimensional feature template matrices;

[0038] S342. Perform first matching calculation processing on the four-dimensional data matrix and the four-dimensional feature template matrix to obtain a first matching four-dimensional matrix;

[0039] S343. Perform second matching calculation processing on the four-dimensional data matrix and the four-dimensional feature template matrix to obtain a second matching four-dimensional matrix;

[0040] S344. Use the first matching four-dimensional matrix and the second matching four-dimensional matrix to construct a matching quantity set.

[0041] The expression for the first matching calculation processing is:

[0042] c = |(FFT(a * conj(b))) 2 ,

[0044] where c is the first matching four-dimensional matrix, conj is conjugate calculation, a is the four-dimensional data matrix, and b is the four-dimensional feature template matrix;

[0045] The expression for the second matching calculation processing is:

[0046]

[0047] where S a is the normalized four-dimensional data matrix, S b is the normalized four-dimensional feature template matrix, S ab is the cross-correlation four-dimensional matrix, and coff is the first matching four-dimensional matrix.

[0048] Performing class extraction processing on the matching quantity set to obtain the class value of the discrete signal to be classified includes:

[0049] For each value of the third dimension P1 of the first matching four-dimensional matrix, respectively obtain the corresponding three-dimensional matrix, and count the total number of elements in the three-dimensional matrix that are greater than a preset first discrimination value, and use the total number as the first matching value corresponding to the value of the third dimension P1;

[0050] Use the first matching values corresponding to all values of the third dimension P1 of the first matching four-dimensional matrix to construct a first matching vector;

[0051] For each value of the third dimension P2 of the second matching four-dimensional matrix, respectively obtain the corresponding three-dimensional matrix, and count the total number of elements in the three-dimensional matrix that are greater than a preset second discrimination value, and use the total number as the second matching value corresponding to the value of the third dimension P2;

[0052] Construct a corresponding second matching vector by using all the second matching values corresponding to the values of the third dimension P2 of the second matching four-dimensional matrix.

[0053] Perform joint calculation processing on the first matching vector and the second matching vector to obtain the class value of the discrete signal to be classified.

[0054] The beneficial effects of the present invention are as follows:

[0055] The present invention uses an efficient matrix parallel operation method of hierarchical channelization to implement the splitting function of large-bandwidth signals. Among them, hierarchical channelization can effectively reduce the demand for high-order filters and improve the subsequent calculation efficiency; construct multiple groups of feature matrix templates, an efficient matrix matching method, and parallelly implement the feature sequence template matching of thousands of narrow-band signals. The feature matrix templates can be dynamically adjusted according to needs to improve flexibility and achieve an efficient matching function. Adopt a halving and fine search strategy to improve the operation efficiency; the matrix dimension is variable. For different actual application scenarios, adjust the signal data matrix dimension to meet the needs of different signal sampling durations and bandwidths. At the same time, the characteristic template matrix can also be dynamically added and changed to realize the dynamic expansion of the matching length and type of the template.

[0056] To solve the problem of classifying and identifying multiple signals under the traditional serial processing system broadband, the present invention directly omits the signal search process, performs channelization splitting on large-bandwidth signals, splits the large-bandwidth signals into thousands of narrow-band signals, and parallelly performs the target signal recognition and classification of thousands of narrow-band signals. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 It is a schematic diagram of the complex signal channelization even permutation method of the present invention;

[0058] Figure 2 It is a schematic diagram of the complex signal channelization receiving structure of the present invention;

[0059] Figure 3 It is a schematic diagram of the matrix parallel batch processing model;

[0060] Figure 4 It is the feature sequence template, efficient matrix data organization and operation structure;

[0061] Figure 5 It is a flowchart of a large-bandwidth signal classification method based on multi-dimensional variable matrix operation proposed by the present invention;

[0062] Figure 6 It is the true frequency domain diagram of the large-bandwidth signal;

[0063] Figure 7 It is the DDC spectrum diagram after the large-bandwidth signal is split;

[0064] Figure 8 It is the actual classification effect diagram of the large-bandwidth signal. Specific implementation manners

[0065] To better understand the content of the present invention, an embodiment is given here.

[0066] Figure 1 It is the schematic diagram of the complex signal channelized even arrangement mode of the present invention; Figure 2 It is the schematic diagram of the complex signal channelized receiving structure of the present invention; Figure 3 It is the schematic diagram of the matrix parallel batch processing model; Figure 4 It is the feature sequence template, efficient matrix data organization and operation structure; Figure 5 It is the flow chart of a large-bandwidth signal classification method based on multi-dimensional variable matrix operation proposed by the present invention; Figure 6 It is the true frequency domain diagram of the large-bandwidth signal; Figure 7 It is the DDC frequency spectrum diagram after the large-bandwidth signal is segmented; Figure 8 It is the actual classification effect diagram of the large-bandwidth signal.

[0067] In the first aspect of the embodiment of the present invention, a large-bandwidth signal classification method based on multi-dimensional variable matrix operation is disclosed, which omits the signal search process, directly uses an efficient matrix operation structure to perform multiple narrow-band DDC segmentation on a large-bandwidth channel, has a real-time parallel processing mechanism for the full-spectrum signal, segments the signal under a large bandwidth into thousands of narrow-band channels, and performs real-time continuous acquisition, segmentation, parallel analysis and processing to ensure that the signal is continuous without omission.

[0068] In the first aspect of the embodiment of the present application, a large-bandwidth signal classification method based on multi-dimensional variable matrix operation is disclosed, including:

[0069] S1. Obtain the discrete signal to be classified, the sampling rate Fs, the total number D of filtering channels, and the number of two-stage filtering channels; the number of two-stage filtering channels includes the number NUM1 of the first-stage filtering channels and the number NUM2 of the second-stage filtering channels; the length of the discrete signal to be classified is Len;

[0070] S2. Obtain the feature template matrix; the dimension of the feature template matrix is M×N, N represents the category of the feature signal, and M represents the length of the feature signal; the column vector of the feature template matrix is a feature signal;

[0071] S3. Based on the feature template matrix, perform classification processing on the discrete signal to be classified to obtain the category value of the discrete signal to be classified;

[0072] The performing classification processing on the discrete signal to be classified based on the feature template matrix to obtain the category value of the discrete signal to be classified includes:

[0073] S31. Design a first - stage filter matrix and a second - stage filter matrix based on the sampling rate Fs, the total number of filtering channels D, and the number of two - stage filtering channels.

[0074] S32. Process the discrete signal to be classified using the first - stage filter matrix to obtain a first filtered output matrix.

[0075] S33. Process the first filtered output matrix using the second - stage filter matrix to obtain a two - dimensional data matrix.

[0076] S34. Perform matching calculation processing on the two - dimensional data matrix based on the feature template matrix to obtain a set of matching quantities.

[0077] S35. Perform category extraction processing on the set of matching quantities to obtain the category value of the discrete signal to be classified.

[0078] The design of the first - stage filter matrix and the second - stage filter matrix based on the sampling rate Fs, the total number of filtering channels D, and the number of two - stage filtering channels includes:

[0079] S311. Construct a first - stage filter matrix according to the sampling rate Fs and the number of first - stage filtering channels NUM1.

[0080] S312. Construct a second - stage filter matrix according to the sampling rate Fs and the number of second - stage filtering channels NUM2.

[0081] The construction of the first - stage filter matrix according to the sampling rate Fs and the number of first - stage filtering channels NUM1 includes:

[0082] Evenly divide the effective bandwidth corresponding to the sampling rate Fs into NUM1 adjacent sub - frequency bands.

[0083] Set the first - stage filtering order constant len_filter1 as len_filter1 = 2D.

[0084] For the frequency band range of each sub - frequency band, use the digital filter design method to design a filtering vector with an order of len_filter1 / NUM1 for each sub - frequency band.

[0085] Construct a first - stage filter matrix using the filtering vectors of all sub - frequency bands.

[0086] The construction of the second - stage filter matrix according to the sampling rate Fs and the number of second - stage filtering channels NUM2 includes:

[0087] Evenly divide the effective bandwidth corresponding to the sampling rate Fs into NUM2 adjacent sub - frequency bands.

[0088] Set the second - stage filtering order constant len_filter2 as len_filter2 = 4D;

[0089] For the frequency band range of each sub - band, using the digital filter design method, design a filtering vector with the order of len_filter2 / NUM2 for each sub - band;

[0090] Use the filtering vectors of all sub - bands to construct a second - stage filter matrix;

[0091] The processing of the discrete signal to be classified using the first - stage filter matrix to obtain the first filtered output matrix includes:

[0092] Represent the discrete signal to be classified as a digital matrix matching the dimension of the first filtered output matrix;

[0093] Use the first - stage filter matrix to perform parallel filtering on the column vectors of the digital matrix to obtain the first output matrix;

[0094] Perform FFT processing on each column vector of the first output matrix to obtain the first filtered output matrix;

[0095] The processing of the first filtered output matrix using the second - stage filter matrix to obtain a two - dimensional data matrix includes:

[0096] Use the second - stage filter matrix to perform parallel filtering on the column vectors of the first filtered output matrix to obtain the second output matrix;

[0097] Perform FFT processing on each column vector of the second output matrix to obtain a two - dimensional data matrix;

[0098] The matching calculation processing of the two - dimensional data matrix based on the feature template matrix to obtain a set of matching quantities includes:

[0099] S341, perform dimension expansion processing on the two - dimensional data matrix and the feature template matrix respectively to obtain the corresponding four - dimensional data matrix and four - dimensional feature template matrix;

[0100] S342, perform the first matching calculation processing on the four - dimensional data matrix and the four - dimensional feature template matrix to obtain the first matching four - dimensional matrix;

[0101] S343, perform the second matching calculation processing on the four - dimensional data matrix and the four - dimensional feature template matrix to obtain the second matching four - dimensional matrix;

[0102] S344, use the first matching four - dimensional matrix and the second matching four - dimensional matrix to construct a set of matching quantities.

[0103] The expression of the first matching calculation process is:

[0104] c = |(FFT(a * conj(b))) 2 ,

[0105] where c is the first matching four-dimensional matrix, conj is the conjugate calculation, a is the four-dimensional data matrix, and b is the four-dimensional feature template matrix;

[0106] The expression of the second matching calculation process is:

[0107]

[0108] where S a is the normalized four-dimensional data matrix, S b is the normalized four-dimensional feature template matrix, S ab is the cross-correlation four-dimensional matrix, and coff is the first matching four-dimensional matrix.

[0109] The dimension expansion processing of the two-dimensional data matrix and the feature template matrix respectively to obtain the corresponding four-dimensional data matrix and four-dimensional feature template matrix includes:

[0110] The two-dimensional data matrix and the feature template matrix are respectively replicated in the extended dimension, and the dimensions of the obtained four-dimensional matrices are both D_valid×(Len / D)×N×M; the dimension of the two-dimensional data matrix is D_valid×(Len / D); the elements of the obtained four-dimensional data matrix are a(:,:,m,n) = A, and the elements of the four-dimensional feature template matrix are b(k,l,:,:) = B, where k, l, m, and n respectively represent the first dimension value, the second dimension value, the third dimension value, and the fourth dimension value of the four-dimensional matrix, A represents the two-dimensional data matrix, and B represents the feature template matrix; a(:,:,m,n) = A means that for any combination of m and n within their value ranges, the value of a(:,:,m,n) is A, and the meaning of b(k,l,:,:) = B is the same as above. In a*conj(b), conj(b) means taking the conjugate of each element in b, and a*conj(b) means multiplying the elements with the same subscripts in matrices a and b. After taking the conjugate of the elements in b and multiplying them with the elements in a, the elements with the corresponding same subscripts in the new four-dimensional matrix are obtained. Let c1(k,l,m,n) = a(k,l,m,n)*conj(b(k,l,m,n)), and c1 is the four-dimensional matrix obtained by the operation of a*conj(b). For FFT(a*conj(b)), the vectors c1(k,l,m,:) in the fourth dimension of matrix c1 are respectively subjected to FFT calculations to obtain new vectors in the fourth dimension. After performing FFT calculations on all the vectors in the fourth dimension of matrix c1 and updating, the new four-dimensional matrix FFT(a*conj(b)) is obtained. After taking the modulus value of each element of the four-dimensional matrix FFT(a*conj(b)) and then calculating the square value, |FFT(a*conj(b))| is obtained. 2 , that is, the four-dimensional matrix c.

[0111] The is to respectively calculate the means of the vectors in the fourth dimension of matrices a and b to obtain and Then, cumulative processing is performed on the fourth dimension value n to obtain the four-dimensional matrix S a and S b , specifically,

[0112] The Its specific expression is:

[0113]

[0114] Among them, S ab is a four-dimensional matrix, and S ab (k,l,m,1) is an element in S ab .

[0115] The It is to perform the above operations on each element in S ab , S a and S b to obtain the corresponding elements in the four-dimensional matrix coff. Specifically coff(k, l, m, 1) represents the corresponding element in the four-dimensional matrix coff.

[0116] The class extraction process for the set of matching quantities to obtain the class value of the discrete signal to be classified includes:

[0117] For each value of the third dimension P1 of the first matching four-dimensional matrix, the corresponding three-dimensional matrix is respectively obtained, and the total number of elements in the three-dimensional matrix that are greater than a preset first discrimination value is statistically obtained, and the total number is used as the first matching value corresponding to the value of the third dimension P1;

[0118] Using the first matching values corresponding to all values of the third dimension P1 of the first matching four-dimensional matrix, a first matching vector is constructed;

[0119] For each value of the third dimension P2 of the second matching four-dimensional matrix, the corresponding three-dimensional matrix is respectively obtained, and the total number of elements in the three-dimensional matrix that are greater than a preset second discrimination value is statistically obtained, and the total number is used as the second matching value corresponding to the value of the third dimension P2;

[0120] Using the second matching values corresponding to all values of the third dimension P2 of the second matching four-dimensional matrix, a corresponding second matching vector is constructed;

[0121] Performing joint calculation processing on the first matching vector and the second matching vector to obtain the class value of the discrete signal to be classified;

[0122] The expression of the joint calculation processing is:

[0123]

[0124] where N is the length of the second matching vector, μ i is the i-th item of the first matching vector, v i is the i-th item of the second matching vector, and I is the class value of the discrete signal to be classified.

[0125] For each value of the third dimension P1 of the first matching four-dimensional matrix, the corresponding three-dimensional matrix is respectively obtained. When the value of P1 is y1, the obtained corresponding three-dimensional matrix is c(:, :, y1, :); the situation for the second matching four-dimensional matrix is the same as above.

[0126] The parallel filtering of the column vectors of the digital matrix by using the first-level filter matrix to obtain the first output matrix includes:

[0127] Using each row vector of the first - stage filter matrix to perform a convolution operation with the column vector of the same serial number in the digital matrix, to obtain the column vector of the corresponding serial number in the first output matrix;

[0128] Performing FFT processing on each column vector of the first output matrix respectively to obtain the first filtered output matrix, which means performing FFT processing on each column vector of the first output matrix respectively to obtain the corresponding column vectors, and using all the column vectors to construct the first filtered output matrix;

[0129] The parallel filtering and FFT processing process of the second - stage filter matrix is the same as the above process.

[0130] For each value of the third dimension P1 of the first matching four - dimensional matrix, obtaining the corresponding three - dimensional matrix respectively means that when the value of the third dimension P1 ranges from 1 to N, obtaining the corresponding three - dimensional matrix AP2(:,:,P1,:), where AP2 represents the first matching four - dimensional matrix, and the processing process of the second matching four - dimensional matrix is the same as the above process.

[0131] Using the filtering vectors of all sub - bands to construct the first - stage or second - stage filter matrix means constructing the first - stage or second - stage filter matrix with the filtering vectors of the sub - bands as column vectors;

[0132] The first discrimination value can be 18;

[0133] The first discrimination value can be 22;

[0134] The parallel filtering is to use the column vectors of the digital matrix as signals, and each column vector is filtered by the column vectors of the first - stage or second - stage filter matrix with the corresponding column serial numbers respectively to obtain the output column vectors, and using all the output column vectors to construct the first output matrix or the second output matrix;

[0135] The digital filter design method is the FIR digital filter design method.

[0136] The frequency band range of the i - th sub - band is [(i - 1)×Fs / (2×NUM1), i×Fs / (2×NUM1)], or [(i - 1)×Fs / (2×NUM2), i×Fs / (2×NUM2)];

[0137] The effective bandwidth is 1 / 2 of the sampling rate Fs;

[0138] The discrete signal to be classified is obtained by digitally sampling the signal received by the communication receiver;

[0139] In the second aspect of the embodiments of the present invention, an embodiment of a large-bandwidth signal classification method based on multi-dimensional variable matrix operation is disclosed as follows Figure 5 As shown, to implement this method, the following steps are adopted:

[0140] Step 51: Determine the parameters. The large-bandwidth IQ sampling duration Len = 32768×1024, the sampling rate Fs = 40.96 MHz, the effective bandwidth BW = 28.5 MHz, the number of DDC channels D = 1024 for channel splitting, and the number of channels for two-stage filtering are NUM1 = 16 and NUM1 = 64 respectively;

[0141] Step 52: Design the two-stage filters for NUM1 = 16 and NUM1 = 64 respectively according to the channel splitting parameters determined in Step 51. Design the prototype low-pass filter according to the preset passband-stopband relationship to obtain the filter coefficients and orders len_filter1 = 2048 and len_filter2 = 4096;

[0142] Step 53: Organize the data into a matrix of 1×Len, the channelized matrix structure. After two-stage filtering, output the split DDC data matrix 2D×(Len / D), and the effective bandwidth data matrix D_valid×(Len / D).

[0143] In this embodiment, the input IQ length Len = 32768×1024, D = 1024, and the large-bandwidth IQ data sampling rate is Fs = 40.96 MHz. Since the way of two-stage filtering with road delay is adopted, the data dimension can be expressed as Len = 32768×(16*64). From Figure 2 In the processing structure of channelized high-efficiency matrix data organization and operation, the steps of the first-stage filtering are as follows: First, when NUM1 = 16 for the signal, use the Reshape function to convert the signal into a two-dimensional matrix with the matrix dimension (Len / NUM1)×NUM1; Second, use the polyphase decomposition of the Reshape function for the prototype low-pass filter coefficients, and the decomposed matrix is (Len_filter1 / NUM1)×NUM1; Then use the Filter function to implement the parallel filtering function between the signal dimension (Len / NUM1)×NUM1 and the filter dimension (Len_filter1 / NUM1)×NUM1, and the dimension after parallel filtering is (Len / NUM1)×NUM1; Finally, use the FFT function to perform the parallel FFT on the result of (Len / NUM1)×NUM1, and output the final result (Len / NUM1)×NUM1.

[0144] The secondary filtering is carried out in the same way. According to the effective bandwidth, the filtering results are organized into a valid data matrix of D_valid×(Len / D), where D_valid = 1425 and Len / D = 32768. The sampling rate of each channel after segmentation is Fs / D = 40.96 MHz / 1024 = 40 KHz;

[0145] Step 54: Construct a feature template matrix, M×N. Group templates of different lengths and perform a normalization operation on the feature templates;

[0146] In this embodiment, the constructed template lengths are 128, 64, 32, etc., and the number of templates is also different. The constructed feature template matrices are three types, namely 128×22, 64×4, and 32×8. At the same time, a normalization operation is performed on the matrix;

[0147] Step 55: Matrix calculation of the feature template matching algorithm. According to the data dimension and the feature template dimension, organize the matching correlation map results into a four-dimensional data structure D_valid×O×P×Q. Adopt a search strategy combining half-detection and fine-detection to improve the operation efficiency;

[0148] In this embodiment, according to the principle of feature template matching, as Figure 4 shown, the data matrix implementation structure of the template matching function after channelized segmentation of the signal. It mainly includes the following parts. Match the input two-dimensional data matrix D_valid×(Len / D) = 1024×32768 with three feature sequence templates (dimensions are 128×22, 64×4, 32×8) respectively. The matrix matching formula is xconvfft = |(FFT(a*conj(b))) 2 , where a and b are the input data matrix and the feature template matrix respectively, where, Organize the matching correlation map results into a four-dimensional data structure D_valid×O×P×Q, where D_valid = 1024: the number of segmented channels, O: the number of data groups for a single match, which can be set arbitrarily according to the actual situation, P: the number of target templates to be matched, which are 22, 4, 8, and Q: the length of the template to be matched, which are 128, 64, 32. Adopt a search strategy combining half-detection and fine-detection to improve the operation efficiency. The half-length setting in this embodiment is 64.

[0149] Step 56: Do you need to change the bandwidth and sampling duration?

[0150] In this embodiment, for different actual application scenarios, if there is a need to adjust the sampling duration and bandwidth requirements of the input signal, then execute Step 51 to re-adjust and organize the data matrix as required.

[0151] Step 57: Is the feature template matrix newly added or changed?

[0152] In this embodiment, for different actual application scenarios, if it is necessary to dynamically add and change the feature template matrix to realize the dynamic expansion of the matching length and types of the template, then execute Step 54, modify the corresponding parameters, and reorganize the feature template matrix.

[0153] For the heterogeneous parallel processing platform using CPU + GPU in this embodiment, within 800 ms, for 3000 channels sliced under a 28.5 MHz bandwidth, the matching test time for the characteristic sequences of 60 types of signals is as Figure 6 shown, and the actual classification effect of the signals under large bandwidth is as Figure 7 shown, and the running time situation is shown in Table 1.

[0154] Table 1 Test results of the algorithm running speed in this embodiment

[0155]

[0156] The efficient classification and recognition method developed by the present invention for a large bandwidth high-speed parallel computing platform, the efficient matrix operation structure is adapted to data matrixization, channelization operation matrixization, parameter extraction, matching matrixization, etc. This method is applicable to the software developed for the parallel platform and has the advantages of strong flexibility and excellent performance.

[0157] In the third aspect of the embodiment of the present invention, a large bandwidth signal classification method based on multi-dimensional variable matrix operation is disclosed. The signal search process is omitted, and an efficient matrix operation structure is directly used to perform multiple narrowband DDC slicing on a large bandwidth channel, and a real-time parallel processing mechanism for the full-spectrum signal. The signal under large bandwidth is sliced into thousands of narrowband channels, and real-time continuous acquisition, slicing, and parallel analysis and processing are performed to ensure that the signal is continuous without omission, including the following steps:

[0158] S1) Construct an efficient matrix-based parallel operation structure to realize the hierarchical channelization function, perform channelization operation on the large bandwidth signal, and slice the large bandwidth signal into thousands of narrowband signals. Among them, the hierarchical channelization strategy can effectively reduce the demand for high-order filters and improve the subsequent calculation efficiency;

[0159] S2) Construct a feature matrix template with variable lengths for different target signals, and parallelly implement the feature sequence template matching for thousands of narrowband signals. The feature matrix template can be dynamically adjusted according to requirements to realize the efficient matching function;

[0160] S3) For different actual application scenarios, adjust the dimension of the signal data matrix to meet the requirements of different sampling time lengths and bandwidths of the signal. At the same time, it is also possible to dynamically add and change the feature template matrix to realize the dynamic expansion of the matching length and types of the template.

[0161] Further, the step S1) includes:

[0162] S11) Input and collect wideband IQ data with a sampling rate of Fs and 1×Len sampling points. According to the complex signal channelized even-channel arrangement method used as shown below, the principle is as follows: Assume that the normalized frequency range of the complex signal is 0 to 2π, and the center frequencies ω Figure 3 of the Kth channel (k = 0, 1,... K-1) are respectively: ω k (k) = 2πk / K. When the critical decimation K = D, according to the derivation process, the mathematical model of the complex signal channelized receiver based on the channelized structure is obtained as shown in k and the derivation formula (1). Assume that the filter length is 1×len_filter, and let len_filter = MD, where M is an integer. Divide the time-domain impulse response h(n) of the prototype low-pass filter into D groups by uniform decimation every M points. The input original sequence is S(n), where n = mD, S Figure 1 (m) = S(mD - p), h P (m) = h(mD + p). Then the output of the kth channel is: P (m) = DFT[S

[0163] y k (m) * h P (m)] P (m)] k

[0164] The decimation filter for each channel is no longer the original low-pass filter, but the polyphase branch of this filter, and its computational complexity is reduced to 1 / D, greatly improving the real-time processing ability of this channelization technology. In addition, the FFT algorithm can further accelerate the operation speed.

[0165] S12) According to the algorithm implementation process of step S11), after analysis, converting it into an efficient matrix implementation structure of the channelization algorithm mainly includes the following parts: D-channel delay of the signal, downsampling of filter coefficients, filtering, and DFT operation. As shown in the figure is the processing process of channelized matrix data and operations. According to formula (3), from data arrangement to processing, matrix acceleration operation libraries can be used to achieve it. The multi-dimensional matrix operation function libraries Reshape, Filter, and FFT can be directly called to complete the channelization operation. For the signal in the transition band after filtering, the present invention uses the processing method of double mixing to cover the transition band.

[0166] S13) According to the efficient matrix implementation structure of the channelization algorithm in step S12), as shown in Figure 2The processing diagram of the efficient matrix-based data organization and operation of the shown channelization algorithm. The input large-bandwidth IQ data has a length of 1×Len, and the prototype low-pass filter has a length of 1×Len_filter. First, for the D-path delay operation of the signal, the Reshape function is used to convert the signal into a two-dimensional matrix with dimensions (Len / D)×D. Second, the prototype low-pass filter coefficients are decomposed using the polyphase decomposition of the Reshape function, and the decomposed matrix has dimensions (Len_filter / D)×D. Then, the Filter function is used to implement the parallel filtering function between the signal dimension (Len / D)×D and the filter dimension (Len_filter / D)×D, and the dimension after parallel filtering is (Len / D)×D. Finally, the FFT function is used to perform the parallel FFT on the result of (Len / D)×D, and the final result of (Len / D)×D is output.

[0167] Specifically, in step S12), the secondary mixing processing method can use any method that can obtain a signal covering the transition band of the filter.

[0168] S14) Input a signal with a sampling length of 1×Len. The number of stages of the hierarchical polyphase filtering used is set to two levels, and the two-level filters need to be designed separately. If the number of channels divided at one time is assumed to be D, then the number of channels divided at the two levels needs to satisfy: D = NUM1×NUM2. The number of channels divided at the first level is NUM1, and the sampling rate of different channels after division is: Fs / NUM1, and the length is (1×Len) / NUM1. On the basis of the division at the first level, the number of channels divided at the second level is NUM2, and the sampling rate of different channels after division is: Fs / (NUM1×NUM2), and the length is (1×Len) / (NUM1*NUM2).

[0169] S15) According to the division result of step S13), that is, the two-dimensional matrix has dimensions 2D×(Len / D), and considering the large-bandwidth effective bandwidth value BW, the matrix formed by reselecting and combining the useful signal channels has dimensions D_valid×(Len / D).

[0170] Furthermore, step S2) includes:

[0171] S21) Construct a feature template matrix. Group the feature templates with different lengths and perform power normalization operations. The template has dimensions: M×N, where M represents the length of the feature template and N represents the number of feature templates.

[0172] S22) According to the feature template matching principle, it is converted into an efficient matrix implementation structure. It mainly includes the following parts. As shown in the figure, the input two-dimensional data matrix D_valid×(Len / D) is matched with the matching feature sequence template, and the matrix matching formula is xconvfft = (FFT(a*conj(b))) 2 , where a and b are the input data matrix and the feature template matrix respectively, where, The relevant graph results of the matching are organized into a four-dimensional data structure D_valid×O×P×Q, where D_valid: the number of segmented channels, O: the number of data groups for a single match, P: the number of target templates to be matched, that is, N described in S21), Q: the length of the template to be matched, that is, M described in S21);

[0173] S23) Construct multiple groups of feature matrix templates. Therefore, an efficient matching function can be realized;

[0174] Particularly for the data organization matrix and operation matrix implementation methods in S22), the Figure 3 matrix parallel batch processing model shown is used. The model mainly includes multi-dimensional data types: thread, block, and grid, which respectively represent the number of threads or the number of thread blocks in the x, y, and z directions. Multiple threads form a thread block, and multiple thread blocks form a grid. For example, the matrix batch processing kernel function is defined as: Kernel_name<<<gridDim, blockDim, SharedMemorySize, stream>>>(args…), where blockDim and gridDim define the dimensions of the thread grid and block, determining the total number of threads started. According to the resource allocation of the kernel function, the corresponding computing units are generated, and reasonable allocation of computing units can effectively improve the parallel processing speed. In this embodiment, according to the data and algorithm structure, its tasks are divided into parallel operations, the data is organized into a multi-dimensional structure, mapped to the corresponding grid, thread block, and number of threads for parallel operations, and the corresponding signal processing algorithm is adjusted into the operation structure of multi-dimensional data.

[0175] S24) For S22), combined with Figure 4The data matrix of the template matching function after the channelized splitting signal shown. Since the matching result is organized into a four-dimensional data structure D_valid×O×P×Q, where O is the number of data groups for a single match, the optimization method used is: using the correlation between symbol transmissions, adopting a search method combining halving and fine detection. The four-dimensional data structure D_valid×(O / 2)×P×Q formed by the halving detection of the matching result has its computational amount reduced by half. According to the result after the halving detection, not only will D_valid decrease significantly, but the value of O will also decrease, effectively improving the search efficiency while ensuring the search accuracy.

[0176] Further, the step S3) includes:

[0177] S31) For different actual application scenarios, if it is necessary to adjust the requirements for the sampling duration and bandwidth of the input signal, then execute step S1), re-adjust and organize the data matrix as required, otherwise execute step S32):

[0178] S32) At the same time, if it is necessary to dynamically add and change the characteristic template matrix to achieve dynamic expansion of the matching length and types of templates, then execute step S2), modify the corresponding parameters, otherwise the step ends.

[0179] In summary, a large-bandwidth signal classification method based on multi-dimensional variable matrix operations of the present invention is adapted to the current high-speed parallel computing system, improving the recognition and classification efficiency of detecting and identifying specific target signals under a large bandwidth.

[0180] The above are only the embodiments of the present invention and are not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the scope of the claims of the present invention.

Claims

1. A large bandwidth signal classification method based on multi-dimensional variable matrix operation, characterized in that: include: S1, obtain the discrete signal to be classified, the sampling rate Fs, the total number of filter channels D and the number of two-stage filter channels; The number of two-stage filtering channels includes the number of primary filtering channels NUM1 and the number of secondary filtering channels NUM2; the length of the discrete signal to be classified is Len; S2, obtain the feature template matrix; The dimension of the feature template matrix is ​​M×N, where N represents the category of the feature signal and M represents the length of the feature signal; The column vector of the feature template matrix is ​​a feature signal; S3, based on the feature template matrix, classify the discrete signal to be classified to obtain the category value of the discrete signal to be classified.

2. The large bandwidth signal classification method based on multi-dimensional variable matrix operation as claimed in claim 1, characterized in that: The step of performing classification processing on the discrete signal to be classified based on the feature template matrix to obtain the category value of the discrete signal to be classified includes: S31, based on the sampling rate Fs, the total number of filter channels D and the number of two-stage filter channels, a primary filter matrix and a secondary filter matrix are designed; S32, using a first-level filter matrix to process the discrete signal to be classified to obtain a first filter output matrix; S33, using a secondary filter matrix, processing the first filter output matrix to obtain a two-dimensional data matrix; S34, performing matching calculation processing on the two-dimensional data matrix based on the feature template matrix to obtain a matching quantity set; S35, performing category extraction processing on the matching quantity set to obtain the category value of the discrete signal to be classified.

3. The large bandwidth signal classification method based on multi-dimensional variable matrix operation as claimed in claim 2, characterized in that: The first-stage filter matrix and the second-stage filter matrix are designed based on the sampling rate Fs, the total number of filter channels D and the number of two-stage filter channels, including: S311, constructing a first-level filter matrix according to the sampling rate Fs and the number of first-level filter channels NUM1; S312, constructing a secondary filter matrix according to the sampling rate Fs and the number of secondary filter channels NUM2.

4. The large bandwidth signal classification method based on multi-dimensional variable matrix operation as claimed in claim 3, characterized in that: The first-level filter matrix is ​​constructed according to the sampling rate Fs and the number of first-level filter channels NUM1, including: The effective bandwidth corresponding to the sampling rate Fs is evenly divided into NUM1 adjacent sub-bands; Set the first-level filter order constant len_filter1 to len_filter1=2D; For the frequency range of each sub-band, a filter vector with an order of len_filter1 / NUM1 is designed by using a digital filter design method; The filter vectors of all sub-bands are used to construct a first-level filter matrix.

5. The method for classifying large bandwidth signals based on multi-dimensional variable matrix operations as claimed in claim 4, characterized in that: The secondary filter matrix is ​​constructed according to the sampling rate Fs and the number of secondary filter channels NUM2, including: The effective bandwidth corresponding to the sampling rate Fs is evenly divided into NUM2 adjacent sub-bands; Set the secondary filter order constant len_filter2 to len_filter2=4D; For the frequency band range of each sub-band, a filter vector with an order of len_filter2 / NUM2 is designed by using a digital filter design method; The filter vectors of all sub-bands are used to construct a secondary filter matrix.

6. The method for classifying large bandwidth signals based on multi-dimensional variable matrix operations as claimed in claim 5, characterized in that: The method of processing the discrete signal to be classified by using a primary filter matrix to obtain a first filter output matrix includes: Representing the discrete signal to be classified as a digital matrix matching the dimension of the first filter output matrix; Using a primary filter matrix, parallel filtering is performed on the column vectors of the digital matrix to obtain a first output matrix; Perform FFT processing on each column vector of the first output matrix to obtain a first filtered output matrix.

7. The method for classifying large bandwidth signals based on multi-dimensional variable matrix operations as claimed in claim 5, characterized in that: The method of processing the first filter output matrix by using a secondary filter matrix to obtain a two-dimensional data matrix includes: Using a two-stage filter matrix, parallel filtering is performed on the column vectors of the first filter output matrix to obtain a second output matrix; Perform FFT processing on each column vector of the second output matrix to obtain a two-dimensional data matrix.

8. The method for classifying large bandwidth signals based on multi-dimensional variable matrix operations as claimed in claim 5, characterized in that: The matching calculation process is performed on the two-dimensional data matrix based on the feature template matrix to obtain a matching quantity set, including: S341, performing dimension expansion processing on the two-dimensional data matrix and the feature template matrix respectively to obtain corresponding four-dimensional data matrix and four-dimensional feature template matrix; S342, performing a first matching calculation process on the four-dimensional data matrix and the four-dimensional feature template matrix to obtain a first matching four-dimensional matrix; S343, performing a second matching calculation process on the four-dimensional data matrix and the four-dimensional feature template matrix to obtain a second matching four-dimensional matrix; S344: construct a matching quantity set using the first matching four-dimensional matrix and the second matching four-dimensional matrix.

9. The large bandwidth signal classification method based on multi-dimensional variable matrix operation as claimed in claim 8, characterized in that: The expression of the first matching calculation process is: c=|(FFT(a*conj(b))) 2 , Wherein, c is the first matching four-dimensional matrix, conj is the conjugate calculation, a is the four-dimensional data matrix, and b is the four-dimensional feature template matrix; The expression of the second matching calculation process is: Among them, S a is the normalized four-dimensional data matrix, S b is the normalized four-dimensional feature template matrix, S ab is the cross-correlation four-dimensional matrix, and coff is the first matching four-dimensional matrix.

10. The large bandwidth signal classification method based on multi-dimensional variable matrix operation as claimed in claim 8, characterized in that: The performing category extraction processing on the set of matching quantities to obtain the category value of the discrete signal to be classified includes: For each value of the third dimension P1 of the first matching four-dimensional matrix, respectively obtain a corresponding three-dimensional matrix, count the total number of elements in the three-dimensional matrix that are greater than a preset first discrimination value, and use the total number as the first matching value corresponding to the value of the third dimension P1; A first matching vector is constructed by using first matching values ​​corresponding to all values ​​of the third dimension P1 of the first matching four-dimensional matrix; For each value of the third dimension P2 of the second matching four-dimensional matrix, respectively obtain a corresponding three-dimensional matrix, count the total number of elements in the three-dimensional matrix that are greater than a preset second discrimination value, and use the total number as the second matching value corresponding to the value of the third dimension P2; Using the second matching values ​​corresponding to all values ​​of the third dimension P2 of the second matching four-dimensional matrix, a corresponding second matching vector is constructed; The first matching vector and the second matching vector are jointly calculated to obtain a category value of the discrete signal to be classified.

Images