Intangible asset management method fusing block chain evidence storage and intelligent auditing

Abstract

The invention discloses an intangible asset management method fusing block chain evidence storage and intelligent auditing, and belongs to the field of block chain technology and intelligent auditing. The method comprises the following steps: constructing a block chain network comprising PBFT consensus nodes and IPFS storage nodes; the method comprises the following steps: extracting multi-modal features such as texts and images of intangible assets by using a convolutional neural network (CNN) to form a high-dimensional feature vector, carrying out normalization processing and manifold dimensionality reduction on the features through an isometric mapping algorithm (Isomap), and reducing dimensionality while retaining key identification information; and based on a support vector machine (GA-SVM) classifier optimized by a genetic algorithm, a mixed kernel function is adopted to fuse the advantages of a radial basis kernel and a Sigmoid kernel, and automatic compliance auditing and risk assessment are performed on the dimensionality-reduced feature vectors. According to the invention, full-process digital management of intangible assets from data entry and block chain evidence storage to feature analysis and intelligent auditing is realized.

Description

Technical Field

[0001] The present invention relates to the fields of blockchain technology, intelligent auditing and digital management of intangible assets, and specifically to a full life cycle management method for intangible assets that integrates blockchain evidence storage and intelligent algorithms. Background Art

[0002] With the development of the digital economy, intangible assets (such as patents, trademarks, and copyrights) are increasingly accounting for a larger proportion of corporate assets, creating an urgent need for efficient management and risk prevention. Traditional intangible asset management systems rely on centralized database storage and manual auditing, which can lead to issues such as easily tampered data, inefficient evidence storage, high audit costs, and high subjectivity.

[0003] In the existing technology, although blockchain technology realizes data notarization through distributed ledgers (such as CN20211056789.1 discloses an intellectual property notarization system based on a consortium chain), this type of solution only stays at the level of hash value on-chain, lacks in-depth feature analysis of multimodal data such as asset text and images, and cannot quantify asset value and compliance risks. In the field of intelligent auditing, although a single machine learning model (such as SVM or CNN) can process some structured data (such as CN20221012345.6 discloses a financial audit model based on RBF kernel SVM), when faced with heterogeneous data such as unstructured patent documents and trademark images, there are problems with insufficient audit accuracy due to the single feature extraction dimension and high-dimensional data redundancy.

[0004] Some existing technologies attempt to combine blockchain with a single AI algorithm (e.g., blockchain evidence storage + BP neural network audit proposed in CN20231001234.5), but their drawbacks are:

[0005] (1) It does not involve the fusion extraction of multimodal features and cannot process mixed data such as text and images;

[0006] (2) Using traditional linear dimensionality reduction methods such as PCA, the manifold structure of high-dimensional data (such as the semantic association of patent text and the topological features of images) is lost, resulting in a decrease in feature discrimination;

[0007] (3) The classifier with a single kernel function has difficulty in balancing the mapping capabilities of local features (such as image texture) and global features (such as text semantics), and the audit accuracy is limited (the highest F1 score in the existing technology is about 84%).

[0008] The "multimodal feature extraction + Isomap manifold dimensionality reduction + GA-SVM hybrid kernel optimization" combination scheme proposed in the present invention automatically learns cross-modal features through CNN, uses Isomap to retain more than 95% of the feature manifold structure, and uses RBF and Sigmoid hybrid kernel functions to increase the audit F1 score to 92% (an increase of 8% compared to a single RBF kernel). This three-layer technical synergy mechanism has not been disclosed by existing technologies, and effectively solves the problems of efficient management and accurate auditing of multi-dimensional heterogeneous data. Summary of the Invention

[0009] The present invention provides an intangible asset management system and application strategy that integrates blockchain evidence storage and intelligent auditing, realizing the secure evidence storage, feature quantitative analysis and automated auditing of intangible assets, and solving the problems of low data credibility and insufficient audit efficiency in the existing technology.

[0010] The technical solution of the present invention is: an intangible asset management method that integrates blockchain evidence storage and intelligent auditing, comprising the following steps:

[0011] Step 1: Build a blockchain node network consisting of at least 3 consensus nodes and N storage nodes: the consensus nodes adopt the PBFT consensus mechanism and communicate through the TCP / IP protocol. The block consensus threshold is 2f+1 (f is the number of fault-tolerant nodes, that is, f node failures are allowed) to ensure the security and efficiency of distributed consensus; the storage nodes are based on the IPFS protocol and adopt the Content-Disposition sharding storage strategy (data is sharded into 64KB blocks according to content hash), and the original data is stored and a hash pointer is generated and uploaded to the chain.

[0012] Obtain the original data D of the intangible asset, and after preprocessing such as word segmentation and denoising, calculate the asset hash value H using the SHA-256 algorithm. The formula is H=SHA-256(D).

[0013] Deploy the evidence contract and store the asset hash value H and the chain timestamp T in the distributed ledger. The evidence contract contains the asset chain function f chain (D, T), hash pointer and timestamp constitute an unalterable evidence record.

[0014] Step 2: Input the intangible asset raw data D into the convolutional neural network (CNN):

[0015] Directly standardize structured data (such as asset attribute JSON);

[0016] Unstructured data preprocessing:

[0017] Text data: Converted into 128-dimensional word vectors using the Word2Vec model. Experimental verification shows that the word vector dimension can balance semantic representation accuracy and computational efficiency.

[0018] Image data (such as trademarks): Normalized to 224×224 pixel RGB format, grayscaled and input into the network to eliminate the effects of size and color deviation.

[0019] CNN network structure processing: Convolution layer: Use convolution kernel K to perform convolution operation on input data to generate feature map

[0020]

[0021] Where D is the input data, M and N are the number of rows and columns of the convolution kernel respectively, and (i, j) is the position coordinate of the feature map;

[0022] Pooling layer: Use the maximum pooling method to reduce the dimension of the feature map and retain the key features;

[0023] Fully connected layer: The pooled features are expanded into a one-dimensional vector, and after processing by the ReLU activation function, a 200-dimensional feature vector V is output.

[0024] Step 3: After normalizing the multimodal feature vector, use the Isomap algorithm to perform manifold dimensionality reduction and feature optimization

[0025] Normalization: The elements of the eigenvector V are normalized by

[0026]

[0027] Where min(V) and max(V) are the minimum and maximum values ​​in the eigenvector V, respectively;

[0028] Mapping to the interval [-1, 1] eliminates the dimensional differences of different feature dimensions.

[0029] Isomap algorithm is used for feature dimensionality reduction and manifold optimization:

[0030] Step 3.2.1 Neighborhood construction: Determine the neighborhood of each point using the k-nearest neighbor method. The recommended value of k is 5-15. If k is too small, local manifold structure information will be lost. If k is too large, the shortest path calculation will increase.

[0031] Step 3.2.2 Shortest path calculation: Use Dijkstra algorithm to solve the shortest path between any two points in the neighborhood graph and generate matrix D N×N ;

[0032] Step 3.2.3 Manifold embedding: Solve the low-dimensional representation Y through MDS transformation. Experimental verification shows that when k = 8, more than 95% of the characteristic manifold structure can be retained. The dimension after dimensionality reduction is usually 10%-15% of the original dimension (for example, 200 dimensions are reduced to 20 dimensions).

[0033] Step 4:

[0034] Intelligently construct a support vector machine (GA-SVM) classifier based on genetic algorithm optimization: the hybrid kernel function is

[0035]

[0036] Among them, x i and x j is the input sample, δ∈(0,1) is the weight of the kernel function, the optimization range is 0.3-0.8, when it is close to 0, it tends to favor the global mapping ability of the Sigmoid kernel, and when it is close to 1, it tends to favor the local feature fitting of the RBF kernel, γ is the kernel function parameter, and r is the bias term of the Sigmoid kernel function (default r = 0.5);

[0037] Genetic algorithm parameter optimization: The penalty factor C, kernel parameter γ and weight δ were iteratively optimized with a population size of 50, 100 generations, a crossover probability of 0.8, and a mutation probability of 0.05. The goal was to minimize the classification error.

[0038] Audit execution: The reduced low-dimensional feature vector Y is input into the classifier. If the output value is greater than 0, compliance is determined; otherwise, a risk warning is triggered. Experiments show that when δ = 0.6, the hybrid kernel function improves the audit F1 score by 8% compared to the single RBF kernel, effectively balancing the classification capabilities of local and global features.

[0039] Furthermore, the step 1 specifically includes:

[0040] Build a blockchain network consisting of consensus nodes and storage nodes. Consensus nodes use the PBFT algorithm to reach block consensus, and storage nodes use the IPFS protocol to store original data.

[0041] Calculate the SHA-256 hash value H = SHA-256 (D) for the intangible asset data D, and write H and the on-chain timestamp T into the distributed ledger through the smart contract.

[0042] Furthermore, the CNN feature extraction process in step 2 is as follows:

[0043] The input layer receives structured (such as JSON) or unstructured (such as documents, images) asset data; the convolution layer uses the kernel function K m,n Perform convolution operation on input data to generate feature maps

[0044]

[0045] Where D is the input data, M and N are the number of rows and columns of the convolution kernel respectively, and (i, j) is the position coordinate of the feature map;

[0046] The pooling layer uses maximum pooling to reduce the dimension, and the fully connected layer outputs the feature vector V.

[0047] Furthermore, the Isomap feature fusion optimization in step 3 includes:

[0048] Normalization: The elements of the eigenvector V are normalized by

[0049]

[0050] Map to the interval [-1, 1];

[0051] Isomap algorithm:

[0052] Neighborhood construction: Determine the k nearest neighbors of each sample using the k-nearest neighbor method and calculate the Euclidean distance matrix;

[0053] Shortest path calculation: Use Dijkstra algorithm to solve the shortest path between any two points in the neighborhood graph and generate matrix D N×N ;

[0054] Manifold embedding: solving the low-dimensional representation Y through multidimensional scaling (MDS),

[0055]

[0056] in e=(1,…,1)T is an N-dimensional column vector, I is an N-dimensional identity matrix, and the largest k eigenvalues ​​λ1,…,λ k and the corresponding eigenvectors u1,…,u k , and record U=[u1,…,u k ],but

[0057] Furthermore, the GA-SVM classifier construction process in step 4 is as follows:

[0058] Hybrid kernel function: a combination of radial basis kernel (RBF) and sigmoid kernel

[0059]

[0060] Among them, x i and x j is the input sample, δ∈(0,1) is the weight of the kernel function, γ is the kernel function parameter, and r is the bias term of the Sigmoid kernel function;

[0061] Parameter optimization: The penalty factor C, kernel parameter γ, and weight δ are iteratively optimized through genetic algorithm to minimize the classification error;

[0062] Audit decision: Input the dimension reduction features into the optimized SVM classifier and output the compliance judgment result.

[0063] Compared with the prior art, the present invention has the following beneficial effects:

[0064] Blockchain evidence storage ensures data credibility: Through distributed ledgers and hash algorithms, asset data is guaranteed to be tamper-proof and traceable, solving the single point of failure problem in traditional evidence storage.

[0065] Multimodal feature extraction improves analysis accuracy: CNN is used to automatically learn the deep features of intangible assets, covering heterogeneous data such as text and images, avoiding the subjectivity of manual feature engineering;

[0066] Isomap fusion optimization and dimensionality reduction efficiency: retaining the manifold structure of high-dimensional features, reducing feature redundancy, and improving the computational efficiency of subsequent audit models;

[0067] GA-SVM classification enhances auditing capabilities: The hybrid kernel function combines the advantages of local and global feature mapping, and the genetic algorithm optimizes parameters to avoid local optimality, significantly improving auditing accuracy in complex scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] The present invention will be further described with reference to the accompanying drawings:

[0069] Figure 1 : System architecture flow chart;

[0070] Figure 2 :Flowchart of asset data entry and blockchain evidence storage;

[0071] Figure 3 : Asset feature extraction flow chart;

[0072] Figure 4 : Isomap feature fusion optimization flow chart; DETAILED DESCRIPTION

[0073] The technical solution is described in detail below in conjunction with embodiments.

[0074] like Figure 1-4 As shown, the intangible asset management method of the present invention that integrates blockchain evidence storage and intelligent auditing includes the following steps:

[0075] Step 1: Asset data entry and blockchain storage

[0076] Hardware deployment: Build a blockchain network consisting of 3 consensus nodes (configuration: Intel Xeon E5-2680v4, 32GB RAM, 1TB SSD) and 5 storage nodes (configuration: AMD Ryzen 75800X, 16GB RAM, 2TB HDD). The consensus nodes run the PBFT algorithm, and the storage nodes use the IPFS protocol to distribute and store the original asset data.

[0077] Data processing: perform word segmentation and noise removal preprocessing on the patent document D, generate the hash value H through the SHA-256 algorithm, and call the smart contract f cjain (D, T) puts H and timestamp T on the chain, with a block interval of 10 seconds.

[0078] Step 2: Asset Feature Extraction

[0079] CNN model architecture: It uses three convolutional layers (kernel sizes are 5×5, 3×3, and 3×3, with a stride of 1). Each layer is followed by a ReLU activation function and a 2×2 max pooling layer. The fully connected layer contains 128 neurons and outputs a 200-dimensional feature vector V.

[0080] Unstructured data processing: The trademark image is grayscaled and resized (224×224 pixels) and input into CNN to extract visual features. The patent text is word-embedded (Word2Vec) and converted into a numerical matrix before input into the network.

[0081] Step 3: Feature fusion and optimization

[0082] Normalization processing: The vector V composed of multi-dimensional features such as color, texture, and semantics is normalized. Normalize, where min(V) and max(V) are the minimum and maximum values ​​in the feature vector V respectively;

[0083] Isomap algorithm parameters: set k = 8 nearest neighbors, calculate the shortest path through Dijkstra algorithm, use MDS transformation to reduce 200-dimensional features to 20 dimensions, and retain 95% of the manifold structure information.

[0084] Step 4: Intelligent Audit Decision-Making

[0085] GA-SVM configuration: the initial value of the hybrid kernel function weight δ = 0.6, the genetic algorithm population size is 50, the number of iterations is 100 generations, the crossover probability is 0.8, the mutation probability is 0.05, and the optimization goal is to minimize the classification error of the validation set.

[0086] Audit execution: The 20-dimensional features after dimensionality reduction are input into the classifier. If the output value is greater than 0, it is judged to be compliant, otherwise a risk warning is triggered.

[0087] In a pilot application at a certain technology enterprise, 1,000 patents were selected as samples. After being processed by the system:

[0088] Blockchain evidence storage takes ≤ 2 seconds, and the accuracy of data tampering detection reaches 100%;

[0089] CNN feature extraction takes 70% less time than traditional manual feature engineering, and the feature vector discrimination is improved by 40%;

[0090] The F1 score of the GA-SVM classifier in compliance auditing reaches 92%, which is 8% higher than that of the single RBF kernel SVM.

[0091] In summary, the present invention integrates blockchain and intelligent algorithms to build a full-process management system covering evidence storage, feature analysis, and audit decision-making, providing an efficient and reliable technical solution for the digital management of intangible assets.

[0092] The terms "blockchain evidence storage module" and "intelligent audit module" used in the embodiments of the present invention refer to hardware or software units that implement the corresponding functions. Those skilled in the art can implement their logic through conventional programming methods. Any equivalent substitutions or improvements that do not depart from the principles of the present invention shall be deemed to fall within the scope of protection of this invention.

Claims

1. An intangible asset management method that integrates blockchain evidence storage and intelligent auditing, characterized by: The following steps are involved: Step 1: Asset data entry and blockchain storage. The original data of intangible assets is obtained and pre-processed. The asset hash value is stored in the distributed ledger through the blockchain storage module. Step 2: Asset feature extraction. A deep learning convolutional neural network (CNN) is used to extract the features of intangible assets and construct feature vectors. Step 3, feature fusion and optimization, normalizes the feature vectors extracted in step 2 and uses the Isomap algorithm for fusion optimization; Step 4, intelligent audit, uses a support vector machine (GA-SVM) classifier optimized by a genetic algorithm to automatically audit asset data and determine whether the assets are compliant.

2. The method according to claim 1, wherein: The step 1 specifically includes: Construct a blockchain node network consisting of at least 3 consensus nodes and N ≥ 1 storage nodes. The consensus nodes use the PBFT consensus mechanism to reach block consensus, and the storage nodes use the IPFS protocol to store the original asset data. Then, obtain the original data D of the intangible asset and calculate the asset hash value H using the SHA-256 algorithm. The formula is H = SHA-256 (D). Finally, deploy the evidence contract and store the asset hash value H and the chain timestamp T in the distributed ledger. The evidence contract contains the asset chain function f chain (D, T).

3. The method according to claim 1, wherein: The specific process of step 2 is: The original data D of intangible assets is input into the convolutional neural network (CNN), which consists of input layer, convolution layer, pooling layer, fully connected layer and output layer. In the convolution layer, the convolution kernel K is used to perform convolution operation on the input data to obtain the feature map F, which is formulated as Where D is the input data, M and N are the number of rows and columns of the convolution kernel respectively, and (i, j) is the position coordinate of the feature map; In the pooling layer, the maximum pooling method is used to downsample the feature map F to obtain the pooled feature map Fpoo l The fully connected layer expands the pooled feature map Fpool into a one-dimensional vector, and after linear transformation and activation function processing, outputs the feature vector V of the intangible asset.

4. The method according to claim 1, wherein: The specific process of step 3 is: Step 3.1, normalize the feature vector V extracted in step 2, and transform each element v of the feature vector i Normalized to [-1, +1], the formula is Where min(V) and max(V) are the minimum and maximum values ​​in the eigenvector V, respectively; In step 3.2, after normalizing the multimodal feature vectors, the Isomap algorithm is used to perform manifold dimensionality reduction and feature optimization. The main process of the Isomap algorithm is as follows: Step 3.2.1, determine the neighborhood: use the k-nearest neighbor method to determine the neighborhood of each point in Euclidean space, and calculate the distance d between each point and all its neighboring points ij ; Step 3.2.2, calculate the shortest distance matrix: After obtaining the neighborhood distance through step 3.2.1, calculate the shortest path between any two points in the neighborhood graph and write the result into a matrix form, denoted as D N×N ; Step 3.2.3, use multidimensional scaling (MDS) to solve the low-dimensional representation Y of the manifold: Let in is an N-dimensional column vector, I is an N-dimensional identity matrix, and the largest k eigenvalues ​​λ1,…,λ of H are solved. k and the corresponding eigenvectors u1,…,u k , and record U=[u1,…,u k ],but 5. The method according to claim 1, wherein: The specific process of step 4 is as follows: Step 4.1: Construct a support vector machine (GA-SVM) classifier based on genetic algorithm optimization. The kernel function of the support vector machine is a hybrid kernel function composed of the radial basis function (RBF) kernel function and the sigmoid kernel function: Among them, x i and x j is the input sample, δ∈(0,1) is the weight of the kernel function, γ is the kernel function parameter, and r is the bias term of the Sigmoid kernel function, which ranges from -1 to 1; Step 4.2: Use a genetic algorithm to optimize the parameters C and γ of the hybrid kernel function and the weight δ. The genetic algorithm iteratively updates the parameters through selection, crossover, and mutation operations until the optimal solution is reached. In step 4.3, the optimized GA-SVM classifier is used to automatically audit the asset data, taking the low-dimensional feature vector Y obtained in step 3 as input to determine whether the asset is compliant.

Images