An acceleration method for fusion of incomplete information in evidence theory based on FPGA

Abstract

The application discloses a non-complete information fusion accelerated calculation method based on FPGA in evidence theory, which allocates the non-complete information fusion accelerated calculation task to two parts of Progarmmable Logic (PL) and Processing System (PS) in a ZYNQ embedded system, wherein the PS part is based on an ARM SOC to realize modeling of non-complete information, and the PL part is based on FPGA to realize acceleration of the evidence theory, and the PL and the PS are connected through an Advanced eXtensible interface (AXI) interface. The application proposes a hardware architecture for non-complete information fusion accelerated calculation, which can effectively solve the problem of "focal element explosion" possibly occurring in evidence combination, and improve the efficiency and speed of the evidence theory fusion processing.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, specifically relating to an FPGA-based method for accelerating computation by fusing incomplete information in evidence theory. Background Technology

[0002] Dempster's evidence theory, as an effective method for expressing and processing uncertain information, has been widely applied in information fusion fields such as fault diagnosis, risk analysis, and multi-attribute decision-making. In the fusion process, evidence theory uses Dempster's combination rule within the same identification framework to fuse various evidence bodies into a new evidence body, providing a reference for decision-making. However, the number of focal elements in the identification framework grows exponentially; that is, when there are n elements in the identification framework, the maximum number of focal elements can be 2^n. n This leads to an exponential increase in the computational complexity of evidence theory fusion rules with the number of focal elements, causing a "focal element explosion" problem in evidence combination and limiting the application of evidence theory. Therefore, finding an efficient algorithm to accelerate evidence processing has become an urgent need. To avoid excessive computational complexity due to an excessive number of focal elements, many scholars have proposed near-computation methods to reduce computational speed by compressing the number of focal elements. However, while this algorithm can improve computational speed, it sacrifices the accuracy of the results, which limits the application of evidence theory in practical work. Summary of the Invention

[0003] To overcome the shortcomings of existing technologies, this invention provides an FPGA-based method for accelerating the fusion of incomplete information in evidence theory. This method distributes the task of accelerating the fusion of incomplete information to two parts of the ZYNQ embedded system: the Programmable Logic (PL) and the Processing System (PS). The PS part uses an ARM-based SOC to model incomplete information, while the PL part uses an FPGA to accelerate evidence theory. The PL and PS are connected via an Advanced deXtensible interface (AXI). This invention proposes a hardware architecture for accelerating the fusion of incomplete information, which can effectively solve the "focal explosion" problem that may occur in evidence combination, and improve the efficiency and speed of evidence theory fusion processing.

[0004] The technical solution adopted by this invention to solve its technical problem is as follows:

[0005] Step 1: In the Vitis development environment, create a new application project and divide the dataset containing m samples and n categories into a training set and a test set according to a set ratio; the training set is used to build a Gaussian model of each category of data on k attributes, and the test set is used to verify the effectiveness of the model;

[0006] Based on the number of sample attributes, a DS evidence theory identification framework is determined; the set 2 consists of each element of the identification framework. Ω , is defined as:

[0007]

[0008] Among them, {H1}, {H2}, {H k} represent sets of single elements, and Ω represents a set containing all elements;

[0009] Step 2: Based on the training set sample data, calculate the mean μ and standard deviation σ of each category of samples on each attribute, and construct n×k Gaussian models accordingly; the membership function of the Gaussian model is expressed as follows:

[0010]

[0011] Step 3: Calculate the matching value between the test sample and the Gaussian model of each attribute;

[0012] Suppose that the cross values ​​of a test sample with a certain attribute and n Gaussian models, from low to high, are denoted as w1, w2, ... w m The corresponding subset of the identification framework is denoted as {{C1,C2,...C}. m},{C2,...C m}, ..., {C m-1 C m}{C m The GBPA calculation process is as follows:

[0013]

[0014] The lowest point w1 represents the set of multiple subsets {C1,C2,...C}. m BPA support for}, w m -w m-1 Represents the highest single subset {C} m The level of BPA support for} This is the BPA support for the unknown category; calculate the GBPA of the sample on each of the k attributes in turn, and denote it as GBPA. k ;

[0015] Step 4: Write RTL code to calculate the conflict coefficient between the two pieces of evidence;

[0016] If the intersection of subset A1 in Evidence 1 and subset A2 in Evidence 2 is an empty set, then the corresponding m1(A1) and m2(A2) are multiplied. After calculating the confidence product of all non-overlapping focal elements between the two pieces of evidence, the results are added together to obtain the conflict coefficient between the two pieces of evidence. The conflict coefficient is calculated as follows:

[0017]

[0018] Step 5: Write RTL code to calculate the fusion value of the two pieces of evidence;

[0019] During the calculation, if the non-empty subset A1 of Evidence 1 and the non-empty subset A2 of Evidence 2 have an intersection H1, m1(A1) and m2(A2) are multiplied. After calculating all operations involving the intersection H1 in both pieces of evidence, the results are added together to obtain the support strength of the fused evidence regarding subset H1. The formula for calculating the fusion of k pieces of evidence is defined as follows:

[0020]

[0021] Step 6: Perform Modelsim functional simulation on the evidence theory fusion RTL code in Steps 4 and 5 to verify whether the fusion results of the two pieces of evidence conform to the theoretical calculation. If not, return to Step 4 for modification; if they conform, encapsulate the evidence theory IP core of the code in Steps 4 and 5 according to the IP core creation wizard. When encapsulating the IP core, first complete the configuration of AXI interface type, mode, data width and number of registers, then configure the register read and write functions to support the sending and receiving of data at the PS and PL ends; then instantiate the evidence theory code and set the devices supported by the IP core, interface clock, and reset signal; finally, compile and package to generate the evidence theory IP core.

[0022] Step 7: Build the minimum hardware system and connect the PL and PS components;

[0023] Step 8: Calculate the evidence theory fusion results for the test samples on each attribute;

[0024] Step 9: [Regarding...] The decision is made by ranking the elements and selecting the element with the highest confidence score. If... If the value is the largest, the sample is considered to be of an unknown category.

[0025] Step 10: Based on step 9, complete the category determination for all samples and send the final result to the host computer via serial port for display; based on the display result, obtain the data classification accuracy.

[0026] Preferably, the conflict coefficient K represents the degree of conflict between pieces of evidence; the larger K is, the greater the degree of conflict.

[0027] Preferably, step 7 is as follows:

[0028] Step 7-1: Build the ZYNQ embedded minimum system on the ZYNQ development board, denoted as the PS system; the system includes an "ARM Cortex-A9" functional unit, a DDR3 controller, a UART controller, and an AXI interface for interaction between the PL and PS;

[0029] Step 7-2: Create a Vivado project, add the evidence theory IP core generated in Step 6, then add the PS system generated in Step 7-1, and connect the ZYNQ embedded minimum system on the PS side and the IP core on the PL side through the AXI interface to form a minimum hardware system;

[0030] Step 7-3: Generate the top-level HDL module and Bitstream file, export hardware platform resources, and obtain the xsa file.

[0031] Preferably, step 8 is as follows:

[0032] Step 8-1: In the Vitis environment, add the hardware platform xsa obtained in step 7, including the IP core, serial port and ZYNQ embedded minimum system;

[0033] Step 8-2: Write a software program to implement communication between PS and PL, including setting register addresses and configuring register transmission and reception; the receive register is used to receive the collision coefficient and fusion result sent from the PL to the PS, and the send register is used for the PS to send the data to be fused to the PL.

[0034] Step 8-3: Run the software program. The GBPA data generated in Step 4 is sent to the IP core through the AXI interface. The IP core performs the fusion operation through evidence theory fusion. After the application reads the data from the register, it divides the received data by the conflict coefficient to calculate the fused result m(A) for each piece of evidence. The operation rules are as follows:

[0035]

[0036] Complete the fusion of k pieces of evidence in sequence.

[0037] The beneficial effects of this invention are as follows:

[0038] 1. This invention designs a high-efficiency evidence theory IP core based on the FPGA in the ZYNQ system. Compared with traditional serial computing, this IP core can process multiple data simultaneously through parallel computing, significantly improving the computation speed of evidence theory and effectively alleviating the "focal explosion" problem. Furthermore, based on the ZYNQ PS system, this invention completes the modeling of incomplete information and communicates with the PL through the AXI interface, realizing integrated processing of incomplete information modeling, evidence fusion, and decision-making. To improve portability, this invention encapsulates the evidence theory fusion rules into an IP core, making it applicable to other FPGA series. This means that regardless of whether using the ZYNQ series or other FPGA series, this IP core can be easily applied, thereby fully leveraging the advantages of evidence theory.

[0039] 2. By leveraging the parallel computing capabilities of FPGAs, this method can process incomplete information more quickly, providing more reliable results. As an application platform, the ZYNQ platform boasts abundant peripheral resources, providing excellent support for the application of evidence theory. Whether in automotive electronics, video surveillance, automation control, or other fields, applying evidence theory on the ZYNQ platform can fully utilize its advantages to achieve efficient and accurate data processing and decision-making. Attached Figure Description

[0040] Figure 1 This is a system block diagram of the present invention. Detailed Implementation

[0041] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0042] To avoid the focal explosion problem and excessive computational complexity in evidence theory, this invention proposes an accelerated computation method for incomplete information fusion in evidence theory based on FPGA. By accelerating evidence theory implementation in a ZYNQ embedded system, the focal explosion problem can be effectively addressed, providing efficient and accurate computational results. Furthermore, this invention utilizes a ZYNQ PS system, i.e., an ARM-based SOC system, to complete the modeling of incomplete information and communicates with the PL via an AXI interface, achieving integrated processing of incomplete information modeling, evidence fusion, and decision-making. This invention will promote the application of evidence theory in practical work, especially its widespread application in the field of IoT devices.

[0043] like Figure 1 As shown, a method for accelerating computation by fusing incomplete information in evidence theory based on FPGA is characterized by the following steps:

[0044] Step 1: In the Vitis development environment, create a new application project. Divide the dataset containing m samples and n categories into a training set and a test set according to a certain ratio. The training set is used to construct Gaussian models for each category of data on k attributes, and the test set is used to verify the effectiveness of the models. Based on the number of sample attributes, determine the DS evidence theory identification framework. The set of elements of the identification framework constitutes 2. Ω , is defined as:

[0045]

[0046] In this embodiment, an iris flower dataset is used, containing 150 samples divided into three categories: mountain iris, variegated iris, and Virginia iris. Each category contains 50 samples, and each sample includes four attributes: sepal length, sepal width, petal length, and petal width. The dataset is divided into training and testing sets in a 6:4 ratio. The recognition frame size is consistent with the number of sample categories, both being 3. The set of elements in the recognition frame is as follows:

[0047]

[0048] Step 2: Based on the training set sample data, calculate the mean u and standard deviation σ of each category of samples on each attribute, and construct n×k Gaussian models accordingly. The membership function of the Gaussian model is expressed as follows:

[0049]

[0050] In this embodiment, the mean and standard deviation of the three categories of data on the four attributes of sepal length, sepal width, petal length, and petal width are shown in Tables 1, 2, 3, and 4.

[0051] Table 1 Mean and standard deviation of calyx length attributes

[0052]

[0053] Table 2 Mean and standard deviation of calyx width attributes

[0054]

[0055] Table 3 Mean and Standard Deviation of Petal Length Attribute

[0056]

[0057] Table 4 Mean and Standard Deviation of Petal Width Attribute

[0058]

[0059] Step 3: Calculate the matching value between the test sample and the Gaussian model for each attribute. Assume that the cross-cross values ​​of a test sample with n Gaussian models for a certain attribute, from low to high, are denoted as w1, w2, ... w m The corresponding subset of the identification framework is denoted as {{C1,C2,...C}. m},{C2,...C m}, ..., {C m-1 C m}{C m The calculation process for Generalized Basic ProbabilityAssignment (GBPA) is as follows.

[0060]

[0061] The lowest point w1 represents the set of multiple subsets {C1,C2,...C}. m BPA support for}, w m -w m-1 Represents the highest single subset {C} m The BPA support of the unknown class is denoted as m({φ}). The GBPA of the sample on each of the k attributes is calculated sequentially and denoted as GBPA. k .

[0062] In this embodiment, *Iris truncata* and *Iris variegated* in the iris flower dataset are considered as known categories and used to construct a Gaussian model. *Iris virginiana* is considered as an unknown category and used as test data. During testing, a test sample 1 (calyx length = 4.9, calyx width = 3.1, petal length = 1.5, petal width = 0.2) is selected from the *Iris truncata* test dataset, and a test sample 2 (calyx length = 7.7, calyx width = 3, petal length = 6.1, petal width = 2.3) is selected from the *Iris virginiana* test dataset to calculate the GPBA values ​​of each sample on the four attributes. The results are shown in Tables 5 and 6.

[0063] Table 5. Test samples from the Iris test dataset.

[0064]

[0065] Table 6. Test Samples from the Virginia Iris Test Dataset

[0066]

[0067] Step 4: Write RTL code to calculate the conflict coefficient between the two pieces of evidence. If the intersection of subset A1 in evidence 1 and subset A2 in evidence 2 is an empty set, multiply the corresponding m1(A1) and m2(A2). After calculating all intersections between the two pieces of evidence... After the calculations, the results are added together to obtain the conflict coefficient between the two pieces of evidence. The conflict coefficient is calculated as follows:

[0068]

[0069] K represents the degree of conflict between pieces of evidence; the larger K is, the greater the degree of conflict.

[0070] In this embodiment, during the calculation of Sample 1, the conflict coefficients of evidence for attribute 1 and attribute 2 are first calculated. Then, based on the fusion calculation result, the conflict coefficient is calculated with attribute 3, and so on. The conflict coefficients between each attribute are shown in Table 7.

[0071] Table 7 Conflict Coefficient

[0072]

[0073] Step 5: Write RTL code to calculate the fusion value of the two pieces of evidence. During the calculation process, if the non-empty subset A1 of Evidence 1 and the non-empty subset A2 of Evidence 2 have an intersection H1, m1(A1) and m2(A2) are multiplied. After calculating all operations involving the intersection H1 in both pieces of evidence, the results are added together to obtain the support strength of the fused evidence regarding subset H1. Utilizing the parallel computing advantage of the FPGA on the PL side, the calculation results of other subsets are obtained simultaneously according to the evidence fusion rules. The k evidence fusion formulas are defined as follows:

[0074]

[0075] In this embodiment, the fusion results of each attribute during the calculation of sample 1 are shown in Table 8:

[0076] Table 8 Fusion Results

[0077]

[0078] Step Six: Perform Modelsim functional simulation on the evidence theory fusion RTL code from Steps Four and Five to verify whether the fusion results of the two pieces of evidence conform to the theoretical calculations. If incorrect, return to Steps Four and Five for modification. If correct, encapsulate the evidence theory IP core of the code from Steps Four and Five according to the IP core creation wizard. During IP core encapsulation, first configure the AXI interface type, mode, data width, and number of registers. Then configure register read / write functions to support data transmission and reception at the PS and PL ends. Next, instantiate the evidence theory code and set the devices supported by the IP core, interface clock, reset signal, etc. Finally, compile and package to generate the evidence theory IP core.

[0079] Step 7: Build the minimum hardware system and connect the PL and PS components. This includes:

[0080] S1: Build a ZYNQ embedded minimum system on the ZYNQ development board, denoted as the PS system. This system includes an "ARMCortes-A9" functional unit, a DDR3 controller, a UART controller, and an AXI interface for interaction between the PL and PS.

[0081] S2: Create a Vivado project, add the evidence theory IP core generated in step six, then add the PS system generated in the previous step, and connect the ZYNQ embedded minimum system on the PS side and the IP core on the PL side through the AXI interface to form a minimum hardware system.

[0082] S3: Generate the top-level HDL module and Bitstream file, export hardware platform resources, and obtain the xsa file.

[0083] Step 8: Calculate the evidence theory fusion results for the test samples across all attributes. This includes:

[0084] S1: In the Vitis environment, add the hardware platform xsa obtained in step seven, including the IP core, serial port and ZYNQ embedded minimum system.

[0085] S2: Write software programs to implement communication between the PS and PL, including setting register addresses and configuring register transmission and reception. The receive register is used to receive the collision coefficient and fusion result sent from the PL to the PS, while the send register is used by the PS to send the data to be fused to the PL.

[0086] S3: Run the software program. The GBPA data generated in step four is sent to the IP core via the AXI interface. The IP core performs the fusion operation through evidence theory fusion. After the application reads the data from the register, it divides the received data by the conflict coefficient to calculate each fused result m(A). The operation rules are as follows.

[0087]

[0088] Complete the fusion of k pieces of evidence in sequence.

[0089] In this embodiment, five registers are used to receive data sent from the PL end to the PS end, and eight registers are used by the PS end to send data to be fused to the PL end. The evidence fusion results of each attribute during the calculation of Sample 1 and Sample 2 are shown in Tables 9 and 10.

[0090] Table 9. Evidence fusion results for each attribute in Sample 1.

[0091]

[0092]

[0093] Table 10. Evidence fusion results for each attribute in Sample 2

[0094]

[0095] Step 9: Sort the fusion results of each attribute of the sample and assign the category corresponding to the maximum value to the sample.

[0096] In this embodiment, in sample 1, m (Iris tectorum) is the largest and is considered to be of the Iris tectorum category; in sample 2, The sample with the highest value is considered to be of an unknown category, and the system's results are consistent with the theoretical results.

[0097] Step 10: Based on Step 9, complete the category determination for all samples and send the final results to the host computer via serial port for display. Based on the displayed results, the data classification accuracy is calculated, as shown in Table 11.

[0098] Table 11 Classification Accuracy

[0099]

[0100] After implementing evidence theory fusion on FPGA, a comparative analysis was conducted on the evidence theory computation time of FPGA and PC. The PC parameters were: 12th Gen Intel(R) Core(TM) i5-12400 2.50GHz processor, 16GB memory, and MATLAB 2021 software; the FPGA model was XC7z020clg400-2, and the development software was Vivado 2020 and Vitis 2020. By reading the program execution time, the PC execution time was 0.0064018s, while the FPGA execution time was 0.004s, a reduction of 0.0024 seconds compared to the high-performance PC.

Claims

1. An FPGA-based acceleration method for non-perfect information fusion in Dempster-Shafer theory, characterized in that, Comprising the following steps: Step 1: In the Vitis development environment, a new application project is established, a data set containing m samples and n categories is divided into a training set and a test set according to a set proportion; Wherein the training set is used to build a Gaussian model of each category data on k attributes, and the test set is used to verify the effectiveness of the model; According to the number of sample attributes, a D-S evidence theory identification framework is determined; a set 2 composed of elements of the identification framework Ω is defined as: wherein {H1}, {H2}, {H k} denote sets of single elements, respectively, and Ω denotes a set containing all elements. Step 2: Based on the training set sample data, the mean μ and standard deviation σ of each category sample on each attribute are calculated to build n x k Gaussian models; the Gaussian model membership function is represented as follows: Step 3: Calculate the matching value of the test sample and each attribute Gaussian model; Assume that the intersection value of a test sample on a certain attribute with n Gaussian models is recorded from low to high as w1, w2,... w m The corresponding recognition framework subsets are recorded as {{C1, C2,... C m}, {C2,... C m},..., {C m-1 , C m}, {C m}}, then the GBPA calculation process is as follows: where w1 represents the BPA support for the most subset {C m} and w m represents the BPA support for the least subset {C m-1}. m is the BPA support for the unknown class; and the GBPA of the sample on the k attributes is calculated in turn, denoted as GBPA k .​ Step 4: Write RTL code to calculate the conflict coefficient of the two pieces of evidence; If the intersection of the subset A1 in evidence 1 and the subset A2 in evidence 2 is empty, multiply m1(A1) and m2(A2), and after calculating the belief product of all non-intersecting focal elements between the two pieces of evidence, add the results to obtain the conflict coefficient of the two pieces of evidence; the conflict coefficient is calculated as follows: Step 5: Write RTL code to calculate the fusion value of the two pieces of evidence; In the calculation process, if the non-empty subset A1 in evidence 1 and the non-empty subset A2 in evidence 2 have an intersection H1, multiply m1(A1) and m2(A2), and after calculating all operations containing the intersection H1 in the two pieces of evidence, add the results to obtain the support degree of the two pieces of evidence after fusion with respect to the subset H1; the fusion calculation formula of the k pieces of evidence is defined as follows: Step 6: Perform modelsim functional simulation on the evidence theory fusion RTL code in steps 4 and 5 to verify whether the fusion result of the two pieces of evidence conforms to the theoretical operation, if not, return to step 4 for modification; if so, according to the IP core creation wizard, encapsulate the evidence theory IP core for the codes in steps 4 and 5; when encapsulating the IP core, first complete the configuration of the AXI interface type, mode, data width and register quantity, and then configure the register read-write function to support the transmission and reception of PS and PL end data; then instantiate the evidence theory code and set the devices, interface clock and reset signal supported by the IP core, and finally compile and package to generate the evidence theory IP core; Step 7: Build a minimum hardware system and connect the PL and PS parts; Step 8: Calculate the evidence theory fusion result of the test sample on each attribute; Step 9: [Regarding...] The decision is made by ranking the elements and selecting the element with the highest confidence score. If... If the value is the largest, the sample is considered to be of an unknown category. Step 10: Based on step 9, complete the category judgment of all samples, and send the final result to the host computer through the serial port for display; according to the display result, the data classification accuracy is obtained.

2. The method of claim 1, wherein the method is characterized in that, The conflict coefficient K represents the conflict degree between the evidence, the larger K is, the greater the conflict degree is.

3. The method of claim 1, wherein the method is characterized in that, The step 7 is specifically as follows: Step 7-1: Build a ZYNQ embedded minimum system on the ZYNQ development board, denoted as the PS system; the system includes "ARMCortes-A9" functional unit, DDR3 controller, UART controller, and AXI interface for PL and PS interaction; Step 7-2: Establish a vivado project, add the evidence theory IP core generated in step 6, then add the PS system generated in step 7-1, and complete the connection between the PS end of the ZYNQ embedded minimum system and the IP core on the PL end through the AXI interface to form a minimum hardware system; Step 7-3: Generate a top-level HDL module and a Bitstream file, export the hardware platform resources, and obtain an xsa file.

4. The method of claim 3, wherein the method is characterized in that, The step 8 is specifically as follows: Step 8-1: In the Vitis environment, add the hardware platform xsa obtained in step 7, including the IP core, the serial port and the ZYNQ embedded minimum system; Step 8-2: Write a software program to realize the communication between PS and PL, including setting the register address and configuring the register transceiver; wherein the receiving register is used to receive the conflict coefficient and the fusion result sent by the PL end to the PS end, and the sending register is used to send the data to be fused by the PS end to the PL end; Step 8-3: Run the software program, the GBPA data generated in step 4 is sent to the IP core through the AXI interface, and the fusion operation is completed through the evidence theory fusion IP core; after the application program reads the data of the register, the received data is divided by the conflict coefficient to calculate the fusion result m(A) of each evidence; the operation rule is as follows: The fusion of k pieces of evidence is completed in turn.

Images