Industrial intelligence data processing method and equipment based on AI intelligent agent

Abstract

The invention provides an industrial information data processing method and equipment based on an AI agent, and the method comprises the steps: obtaining industrial information data; generating a text embedding vector and a position coding vector according to the industrial information data; splicing the text embedding vector and the position coding vector to obtain a semantic enhancement embedding vector; processing the semantic enhancement embedded vector by using a tense convolutional network, and capturing a local fluctuation feature vector and a global trend feature vector; splicing the local fluctuation feature vector and the global trend feature vector to obtain a trend prediction vector; splicing the semantic enhancement embedding vector with the trend prediction vector, and outputting a trend analysis feature vector; inputting a semantic relation graph constructed based on the semantic enhancement embedding vector into the graph attention network, and outputting a semantic association feature vector; and carrying out feature intersection on the trend analysis feature vector and the semantic association feature vector, and outputting a collaborative feature vector, thereby improving the accuracy of industrial intelligence prediction.

Description

Technical Field

[0001] This invention relates to the field of intelligent industrial intelligence, and more specifically, to an industrial intelligence data processing and analysis method, device, system, computer-readable storage medium, and its applications based on AI intelligent agents. Background Technology

[0002] Industry intelligence serves as a crucial foundation for strategic decision-making, risk warning, and market insight, and its analysis process has always been a key focus for both research and industry. Current technologies largely rely on web crawlers and natural language processing methods to acquire information from channels such as news reports, industry reports, corporate websites, and social media. This information is then used for semantic representation through word vectors or pre-trained language models, combined with time series models for trend prediction.

[0003] Chinese invention patent CN120562406A discloses a multi-agent collaborative decision-making intelligence defense method, system, device, and storage medium. This method uses a hash function to generate intelligence fingerprints, then uses these fingerprints to calculate intelligence risk values, and performs defense operations based on whether the risk value exceeds a preset threshold, thereby achieving real-time closed-loop intelligence defense. However, the existing technology still has the following problems:

[0004] 1. Existing methods struggle to distinguish between the main content of a webpage and noisy text, resulting in a lack of distinctiveness in embedding.

[0005] 2. Methods such as LSTM are difficult to effectively model long-term dependencies and non-stationarity. Although TCN can alleviate the lag problem, it has overshoot or delayed response to abrupt inflection points.

[0006] 3. Single models are difficult to cover multi-dimensional intelligence features. Ensemble learning and multi-task learning methods suffer from inter-agent conflicts, computational redundancy, or insufficient transfer, making it difficult to achieve efficient collaboration.

[0007] 4. Conventional flat classification ignores the hierarchical relationship between categories, which can easily confuse similar categories.

[0008] Therefore, exploring how to use deep learning methods for industry intelligence mining and analysis is of great significance. Summary of the Invention

[0009] The method of this invention processes industry intelligence data based on AI intelligent agents, and then predicts the processing results according to primary and secondary classifications to obtain analysis results, which can fully explore the value of industry intelligence data.

[0010] This application discloses a method for processing industrial intelligence data based on AI intelligent agents, including:

[0011] S1: Obtain industry intelligence data, which includes a DOM tree structure;

[0012] S2: Based on the text content of the DOM tree and the position information of the nodes in the DOM tree, generate a text embedding vector and a position encoding vector;

[0013] S3: Concatenate the text embedding vector and the position encoding vector to obtain the semantically enhanced embedding vector;

[0014] S4: Use a temporal convolutional network to process the semantically enhanced embedding vector to capture local fluctuation feature vectors and global trend feature vectors;

[0015] S5: Concatenate the local fluctuation feature vector with the global trend feature vector to obtain the trend prediction vector;

[0016] S6: Concatenate the semantic enhancement embedding vector with the trend prediction vector to output the trend analysis feature vector; input the semantic relationship graph constructed based on the semantic enhancement embedding vector into the graph attention network to output the semantic association feature vector;

[0017] S7: Perform feature cross-referencing on the trend analysis feature vector and the semantic association feature vector to output a collaborative feature vector.

[0018] Furthermore, the specific steps S2 and S3 include:

[0019] S21: The DOM tree is a hierarchical structure representation. The text content of the DOM tree with different hierarchical structures is segmented to obtain the text segmentation result. The text segmentation result is used to obtain the text embedding vector of different levels through the embedding matrix. According to the hierarchy of the DOM tree nodes, the position encoder captures the structural position information of the nodes in the DOM tree to generate position encoding vectors of different levels.

[0020] S31: Concatenate the text embedding vectors of different levels and the position encoding vectors of different levels to obtain the semantically enhanced embedding vector.

[0021] Furthermore, the method for obtaining the embedding matrix is ​​selected from one or more of the following: Word2vec, BERT, Sentence-BERT, SentenceTransformers; and / or the position encoder is selected from one or more of the following: trained position encoding, trigonometric function position encoding, relative position encoding.

[0022] Furthermore, the frequency of each word in the text segmentation results is counted and the probability distribution of each word is calculated. The text information entropy of each word is calculated based on the probability distribution of each word. The text information entropy is scaled and adjusted to obtain semantic weight coefficients at different levels.

[0023] The semantic weight coefficients are used to perform a weighted summation on the splicing result to obtain a semantically enhanced embedding vector that integrates textual semantics and structural information.

[0024] Furthermore, the specific steps of S4 include:

[0025] S41: Combine the semantic enhancement embedding vector with all semantic enhancement embedding vectors within a time τ days of the semantic enhancement embedding vector to form a recent historical embedding vector sequence, and input the recent historical embedding vector sequence into the first temporal convolutional network to extract local fluctuation features and output a local fluctuation feature vector.

[0026] S42: Combine the semantic enhancement embedding vector with all semantic enhancement embedding vectors within 2τ days of the semantic enhancement embedding vector to form a long-term historical embedding vector sequence, and input the long-term historical embedding vector sequence into the second temporal convolutional network to extract global trend features and output a global trend feature vector;

[0027] The τ is a natural integer greater than 7;

[0028] Furthermore, the gradient of the global trend feature vector with respect to time t is calculated, and the gradient is then calculated. The norm is mapped to a trend change sensitivity factor using the Sigmoid function and a sensitivity coefficient λ, where the sensitivity coefficient λ ranges from [0.5, 5]. The local fluctuation feature vector is concatenated and fused with the global trend feature vector, and the concatenation and fusion result is weighted element by element using the trend change sensitivity factor to output a trend prediction vector.

[0029] Furthermore, the AI ​​agent includes a trend analysis agent, a semantic association agent, and a domain expert agent;

[0030] The trend analysis agent is used to concatenate the semantically enhanced embedding vector with the trend prediction vector, capture temporal dependencies and trend features through a Transformer encoder, and output a trend analysis feature vector.

[0031] The semantic association agent is used to input semantic enhancement embedding vectors into the semantic association agent and output semantic association feature vectors. The specific steps are as follows: a semantic relationship graph is constructed based on the semantic enhancement embedding vectors. The points of the semantic relationship graph are the semantic enhancement embedding vectors, and the edges represent the semantic or syntactic relationships between the semantic enhancement embedding vectors. The semantic relationship graph is input into a graph attention network to output semantic association feature vectors.

[0032] The domain expert agent is used to obtain a task description, and the task description is used to obtain a task embedding vector through the embedding matrix. The task embedding vector and the semantically enhanced embedding vector are input into a hybrid expert layer to obtain expert features. The hybrid expert layer includes a gating network and multiple experts. The expert features are routed to different domain expert networks through the gating network to output a domain expert feature vector.

[0033] The trend analysis feature vector, the semantic association feature vector, and the domain expert feature vector are multiplied element by element to achieve feature cross-multiplication, and a collaborative feature vector containing domain expert features is output through a linear projection layer.

[0034] Furthermore, the collaborative feature vector is input into the dynamic task allocation mechanism. The dynamic task allocation mechanism refers to extracting sample task demand vectors from the collaborative feature vectors through a mapping weight matrix, calculating the KL divergence between the sample task demand vectors and the capability prototype vectors of each AI agent, and obtaining the allocation weight matrix through normalization using a function and a temperature coefficient. The AI ​​agent capability prototype vectors refer to trend analysis feature vectors, semantic association feature vectors, and domain expert feature vectors.

[0035] The trend analysis feature vector, the semantic association feature vector, and the domain expert feature vector are weighted according to the weighting matrix to obtain a weighted enhanced feature vector.

[0036] The weighted enhanced feature vectors of all samples are averaged and concatenated, and then a gated vector is generated by an activation function.

[0037] The trend analysis feature vector is compressed using a multilayer perceptron to obtain compressed trend features. The compressed semantic expert features are obtained by concatenating the compressed semantic association feature vector and the domain expert feature vector. The compressed trend features and the compressed semantic expert features are weighted and summed using the gate vector and the complement vector of the gate vector. After layer normalization, a collaborative feature vector with gate fusion is output.

[0038] This application also discloses a method for constructing an industry intelligence analysis model based on AI intelligent agents, including:

[0039] Acquire industry intelligence data, including tags;

[0040] The industry intelligence data is processed using an AI-based intelligent agent-based industry intelligence data processing method to obtain a collaborative feature vector;

[0041] The collaborative feature vector is input into a fully connected layer to obtain a predicted classification result. The predicted classification result is compared with the label to generate a total loss function. The total loss function is optimized to obtain an industry intelligence analysis model.

[0042] Furthermore, the labels include primary labels and secondary labels.

[0043] Furthermore, the specific steps for generating the total loss function are as follows: inputting the collaborative feature vector into the first fully connected layer, calculating the first-level category prediction probability distribution through the activation function, concatenating the collaborative feature vector with the first-level category prediction probability and inputting it into the second fully connected layer, calculating the second-level category prediction probability distribution through the activation function, jointly calculating the first-level category prediction loss and the second-level category prediction loss through the weighted cross-entropy loss function, and balancing the importance of the first-level category prediction loss and the second-level category prediction loss using the hierarchical weight coefficients, thereby obtaining the total loss function;

[0044] The range of the hierarchical weight coefficient is [0,1].

[0045] Furthermore, the method also includes feedback-driven online optimization, which calculates the gradient of the total loss function, applies consistency constraints to the predicted classification results to update the model parameters, and updates the agent capability prototype vector to achieve online model optimization.

[0046] This application also discloses an industry intelligence data analysis method based on AI intelligent agents, including:

[0047] Obtain industry intelligence data;

[0048] The industry intelligence data is processed using an AI-based intelligent agent-based industry intelligence data processing method to obtain a collaborative feature vector;

[0049] The collaborative feature vector is input into the industry intelligence analysis model to obtain the predicted classification result.

[0050] This application also discloses a computing device for processing industrial intelligence data based on AI intelligent agents, the device comprising: a memory and a processor;

[0051] The memory is used to store program instructions;

[0052] The processor is used to call program instructions, which, when executed, implement an AI-based intelligent agent-based industrial intelligence data processing method, or an AI-based intelligent agent-based industrial intelligence analysis model construction method, or implement the steps of the industrial intelligence data analysis method.

[0053] This application also discloses a computer-readable storage medium including a computer program that, when executed by a processor, implements the AI-based intelligent agent-based industrial intelligence data processing method, or the AI-based intelligent agent-based industrial intelligence analysis model construction method, or the steps of the industrial intelligence data analysis method.

[0054] Advantages of this application:

[0055] 1. This application innovatively calculates the semantic weight coefficient of each level of the DOM tree through text information entropy, effectively controlling the output of information and improving the semantic representation capability of the model;

[0056] 2. This application innovatively constructs local fluctuation channels and global trend channels through temporal convolutional networks, and introduces trend change sensitive factors to accurately capture local fluctuations and global change trends in industry data, thereby improving feature extraction capabilities;

[0057] 3. This application innovatively enables multiple intelligent agents to cooperate heterogeneously, and improves feature complementarity and resource utilization through dynamic task allocation;

[0058] 4. This application creatively adopts a hierarchical classification head combined with first- and second-level label dependencies to enhance category differentiation; at the same time, it combines an online feedback optimization mechanism to achieve continuous adaptive updates of the model, which powerfully promotes the development of the industry intelligence sector. Attached Figure Description

[0059] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0060] Figure 1 This is a schematic flowchart of the industry intelligence data processing method based on AI intelligent agents provided in the embodiments of the present invention;

[0061] Figure 2 This is a schematic flowchart of the method for constructing an industry intelligence analysis model based on AI intelligent agents provided in this embodiment of the invention;

[0062] Figure 3 This is a schematic flowchart of the industry intelligence data analysis method based on AI intelligent agents provided in the embodiments of the present invention;

[0063] Figure 4 This is a comparison chart of the noise filtering ratios of different methods provided in the embodiments of the present invention;

[0064] Figure 5 This is a performance comparison chart of different modeling methods provided in the embodiments of the present invention;

[0065] Figure 6 This is a comparative diagram of the industry intelligence trend prediction effect provided in the embodiments of the present invention;

[0066] Figure 7This is a schematic diagram of a computing device for processing industrial intelligence data based on AI intelligent agents, provided in an embodiment of the present invention;

[0067] Figure 8 This is a schematic diagram of the industry intelligence data analysis interface based on AI intelligent agents provided in an embodiment of the present invention;

[0068] Figure 9 This is a schematic diagram of the human-computer interaction interface for industry intelligence data analysis based on AI intelligent agents provided in an embodiment of the present invention. Detailed Implementation

[0069] To enable those skilled in the art to better understand the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0070] In some of the processes described in the specification, claims, and accompanying drawings of this invention, multiple operations are included in a specific order. However, it should be clearly understood that these operations may not be performed in the order they appear herein, or may be performed in parallel. The operation numbers, such as S1, S2, etc., are merely used to distinguish different operations and do not themselves represent any execution order. Furthermore, these processes may include more or fewer operations, and these operations may be performed sequentially or in parallel.

[0071] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0072] Figure 1 This is a schematic flowchart of an industry intelligence data processing method based on AI intelligent agents provided in an embodiment of the present invention. Specifically, it includes the following steps:

[0073] S1: Obtain industry intelligence data, which includes a DOM tree structure;

[0074] In one embodiment, industry intelligence data is primarily collected from industry-related websites via targeted web crawlers, including news portals, industry report platforms, corporate websites, forums, and social media, covering content across multiple dimensions such as technology, market, and policy; the collected raw data includes HTML webpage content, text information, and DOM tree structure.

[0075] In one embodiment, the collected data is preprocessed by removing irrelevant HTML tags, advertisements, duplicate content, and low-quality text; extracting the node text from the DOM tree; and then using the Jieba word segmentation tool to segment the text content and generate word sequences.

[0076] In a specific embodiment, when using the Jieba word segmentation tool to segment a web page node, if the text content of the web page node is “based on the analysis of AI intelligent agents”, the word segmentation result obtained by the Jieba word segmentation tool is [“based on”, “AI”, “intelligent agent”, “of”, “analysis”].

[0077] In one embodiment, the collected data is labeled, a process manually performed by a team of domain experts based on a multi-dimensional classification system of industry intelligence.

[0078] The primary categories include major categories such as technology, market, and policy. For example, "technology" corresponds to the field of technological innovation, and "market" corresponds to economic trends.

[0079] Secondary categories are further subdivisions of primary categories, such as "AI chips" and "artificial intelligence algorithms" under "technology," and "supply and demand analysis" and "competitive landscape" under "market."

[0080] In one specific embodiment, the processed data is divided into a training set, a validation set, and a test set in a 14:3:3 ratio.

[0081] S2: Based on the text content of the DOM tree and the position information of the nodes in the DOM tree, generate a text embedding vector and a position encoding vector;

[0082] In one embodiment, the DOM tree is a hierarchical structure representation. The text content of the DOM tree with different hierarchical structures is segmented to obtain text segmentation results. The text segmentation results are then used to obtain text embedding vectors of different levels through an embedding matrix. Based on the hierarchy of the DOM tree nodes, a position encoder is used to capture the structural position information of the nodes in the DOM tree to generate position encoding vectors of different levels.

[0083] In one specific embodiment, the method for obtaining the embedding matrix is ​​selected from one or more of the following: Word2vec, BERT, Sentence-BERT, SentenceTransformers; and / or the position encoder is selected from one or more of the following: trained position encoding, trigonometric function position encoding, relative position encoding.

[0084] In one embodiment, the text content in the DOM tree node is segmented into words to obtain the text segmentation result. The probability of each word is calculated based on its frequency of occurrence in the text segmentation result. The text information entropy is calculated based on the probability distribution of all words to quantify the information density of the text. The specific formula is as follows:

[0085] ,in, This represents the index of each word in the word segmentation result; The text segmentation result represents a sequence of words obtained by segmenting the text content in the DOM tree nodes. Represents the results of text segmentation. Information entropy is used to quantify the information density of text; This represents a logarithmic function, with the default base being a natural number constant. Indicates the first Each word in the text segmentation results The probability of it appearing in the following is calculated as follows:

[0086] ,in, Represents the results of text segmentation. The total number of all words in the text; It is the first Each word in the text segmentation results The number of times it appears in the text.

[0087] In a specific embodiment, assume the text segmentation result [“AI”, “intelligent agent”, “analysis”, “AI”], then for the first word “AI”, ,but .

[0088] In one embodiment, the scaling of text information entropy at all depths is adjusted using an entropy sensitivity coefficient, and the semantic weight coefficient for each depth is calculated using the Softmax function to strengthen the weights of nodes with high information entropy, as shown below:

[0089] ,in, Indicates the first Semantic weight coefficients for depth are used to weight embedding vectors at different depths; This is the entropy sensitivity coefficient, which controls the scaling degree of information entropy. The larger the value, the more prominent the weight of high-entropy nodes. It is preferred to set it to 1.0. For the first Deep textual information entropy, For the first Text segmentation results at depth; Indicates the first Deep textual information entropy, calculation method and information entropy same.

[0090] In one specific embodiment, depth The node text is "Homepage|Products", and the word segmentation results in... [Homepage, Products], Computing ,but .

[0091] In a specific embodiment, semantic weight coefficients The calculation method is based on the design of the Softmax function. By considering the differences in information density at different depths in the webpage structure, it improves the discriminativeness of the embedded representation through weighted aggregation. It is used to convert information entropy values ​​into probability distributions. Since Softmax is usually used for multi-class output, it amplifies the weight of high information entropy nodes by combining the entropy sensitivity coefficient, thereby strengthening information-rich content, such as information-rich main text, and suppressing low information density nodes, such as low information density advertising text.

[0092] In one specific embodiment, a position encoding vector is generated using a sine function based on the node's hierarchy in the DOM tree to capture the node's structural position information within the DOM tree, represented as:

[0093] ,in, Indicates the first The depth position encoding vector of the first Each dimension value represents the structural position information of a node in the DOM tree; This represents the level of a node in the DOM tree, starting from the root node, where level 1 is the root node. The level increments by 1 for each subsequent level, and so on. If a node is a grandchild of the root node, its level is... Represents the sine function; This represents the dimension index of the position encoding vector. For the embedding dimension; it should be noted that the input of the sine function... The use of 10000 as the base in the denominator is an empirical value, designed to generate sine waves of different frequencies at different locations, thereby encoding positional information. The empirical value of 10000 allows the model to capture positional relationships in long sequences, ensuring the uniqueness and smoothness of the encoding, and enabling the model to effectively handle DOM trees of various depth ranges.

[0094] S3: Concatenate the text embedding vector and the position encoding vector to obtain the semantically enhanced embedding vector;

[0095] In one embodiment, the text embedding vectors of different levels and the position encoding vectors of different levels are concatenated to obtain the semantically enhanced embedding vector, represented as:

[0096]

[0097] in, This represents a semantically enhanced embedding vector that can fuse textual and structural information to represent the comprehensive embedding of nodes; Indicates the first Deep text segmentation results The embedding vector is obtained through the embedding matrix, which is a trainable parameter that is optimized during training. Its function is to map word indices to vectors, using random initialization as the initial value and updating it through training. Indicates a splicing operation; Indicates the first The depth-based location encoding vector is the entire location encoding vector, composed of... Composed of several dimensions, Indicates the first one Each dimension value.

[0098] In one embodiment, a comparative experiment was conducted on the Word2Vec method based on word vectors, the GloVe method based on global vectors, the BERT method based on bidirectional encoder representation, and the technology proposed in this invention. The results are shown in Table 1 and... Figure 4 As shown, the experiment tested the noise filtering effect on three common categories in industry intelligence (technology, market, and policy). The evaluation index was the percentage of noise filtering effect; a higher value indicates a stronger ability to retain effective information and remove noise. The experimental results show that the technology proposed in this invention significantly outperforms other comparative methods in all three intelligence categories. The results show that the conventional Word2Vec and GloVe methods have relatively low noise filtering effects, indicating that they struggle to effectively distinguish core semantics from noise when processing multi-source heterogeneous data like industry intelligence. The BERT method shows a significant improvement over the previous two methods, demonstrating that its powerful semantic understanding ability improves the noise filtering effect to some extent, but it still has limitations. This invention's technology, by dynamically adjusting the embedding weights by combining the depth of the document object model tree nodes and the text information entropy, can more accurately strengthen the weight of the main content and suppress low-information-density nodes. Therefore, it achieves the highest noise filtering effect value in all categories, demonstrating its significant advantage in processing web crawler data.

[0099] Table 1 Comparison of Noise Filtering Effects

[0100] Word2Vec Glove BERT This application technology 66.2% 71.7% 78.6% 90.7% market 62.5% 64.8% 74.1% 85.4% policy 61.8% 62.6% 69.9% 83.8%

[0101] S4: Use a temporal convolutional network to process the semantically enhanced embedding vector to capture local fluctuation feature vectors and global trend feature vectors;

[0102] In one embodiment, step S4 specifically includes:

[0103] S41: Combine the semantic enhancement embedding vector with all semantic enhancement embedding vectors within a time τ days of the semantic enhancement embedding vector to form a recent historical embedding vector sequence, and input the recent historical embedding vector sequence into the first temporal convolutional network to extract local fluctuation features and output a local fluctuation feature vector.

[0104] In a specific embodiment, a temporal convolutional network is used to process recent historical embedded vector sequences, capturing local fluctuations and rapid change patterns in the short term, and outputting a local fluctuation feature vector, represented as follows:

[0105] ,in, This represents a local volatility feature vector, characterizing local volatility features in the short term, such as rapid change patterns in prices, indicators, or sentiment. The parameter is A temporal convolutional network for extracting local fluctuations; This is the set of parameters for a temporal convolutional network used to extract local fluctuations; Indicates from Time's up The semantically enhanced embedding vector sequence at time step is a semantically enhanced embedding vector. In the time window sequence within; The length of the local time window is set according to the data characteristics and task requirements, such as for daily frequency data. Indicates 30 days; For time indexing.

[0106] In one specific embodiment, the parameter is: The temporal convolutional network consists of two causal convolutional layers, two dilated convolutional layers, and one normalization layer. Causal convolutions ensure minimal loss of future information, while dilated convolutions expand the receptive field. Parameters include convolutional kernel weights and biases, and the ReLU activation function is used.

[0107] S42: Combine the semantic enhancement embedding vector with all semantic enhancement embedding vectors within 2τ days of the semantic enhancement embedding vector to form a long-term historical embedding vector sequence, and input the long-term historical embedding vector sequence into the second temporal convolutional network to extract global trend features and output a global trend feature vector;

[0108] The τ is a natural integer greater than 7;

[0109] In one specific embodiment, a temporal convolutional network is used to process long-term historical embedding vector sequences, capturing macroscopic, slowly changing global trends, and outputting a global trend feature vector, represented as:

[0110] ,in, It represents the global trend feature vector, characterizing the long-term global trend vector, such as an overall upward or downward trend, periodic patterns, etc. The parameter is Temporal convolutional networks are used to extract global trends. This is the set of parameters for a temporal convolutional network used to extract global trends. Indicates from Time's up Semantic-enhanced embedding vector sequence at each time step.

[0111] In a specific embodiment, the parameter is... The temporal convolutional network structure is similar, but the parameters are... Temporal convolutional networks use a large dilation coefficient to capture long-term dependencies, and the dilation coefficient can be set in an exponential growth manner.

[0112] In one specific embodiment, the gradient of the global trend feature vector with respect to time is calculated, through its... The norm measures the rate of change of a trend. Using the sigmoid function and a sensitivity coefficient, the gradient norm is mapped to a trend change sensitivity factor, expressed as:

[0113] ,

[0114] in, This represents a trend change sensitivity factor, used to adjust prediction weights, and is a scalar between 0 and 1. This represents the sensitivity coefficient, which controls the strength of the gradient's influence on the factor value. Its value range is [0, 1.0], with the preferred value being... express Norm; Represents the global trend feature vector The gradient is the global trend feature vector. The gradient over time, representing the rate of change of a trend, is formulated as follows:

[0115] , The sign for partial derivatives; The global trend feature vector at any given time; The global trend feature vector at any given time; For time step.

[0116] In one specific embodiment, considering the sensitivity to trend changes, the model is made to adaptively focus on abrupt inflection points. The trend change sensitivity factor is calculated using the Sigmoid activation function, and based on this, the following is employed: The norm measures the intensity of change and controls sensitivity through the sensitivity coefficient. The gradient norm is mapped to the (0,1) interval. When the gradient norm is large, it indicates that the trend change is drastic and the trend change sensitivity factor is close to 1, which strengthens the weight of the current prediction. When the gradient norm is small, the trend change sensitivity factor is close to 0, which suppresses the current prediction.

[0117] In a specific embodiment, to ensure that the dimensions of the local fluctuation feature vector and the global trend feature vector match, the output dimensions of the temporal convolutional network used to extract local fluctuations and the temporal convolutional network used to extract global trends need to be set to be the same. Specifically, the same number of convolutional filters are set to ensure that the output dimensions are the same.

[0118] S5: Concatenate the local fluctuation feature vector with the global trend feature vector to obtain the trend prediction vector;

[0119] In one specific embodiment, the local fluctuation feature vector and the global trend feature vector are concatenated, fused using a gated linear layer and a sigmoid activation function, and the fusion result is weighted element-by-element using a trend change sensitivity factor to output a trend prediction vector, the formula of which is:

[0120] ,in, This represents the trend prediction vector, which characterizes the embedding of future trends. This represents the Sigmoid activation function; Let be the weight matrix of the gated linear layer, which are trainable parameters; is the bias vector of the gated linear layer, which are trainable parameters; This indicates element-wise multiplication.

[0121] S6: Concatenate the semantic enhancement embedding vector with the trend prediction vector to output the trend analysis feature vector; input the semantic relationship graph constructed based on the semantic enhancement embedding vector into the graph attention network to output the semantic association feature vector;

[0122] S7: Perform feature cross-referencing on the trend analysis feature vector and the semantic association feature vector to output a collaborative feature vector.

[0123] In one embodiment, the AI ​​agent includes a trend analysis agent, a semantic association agent, and a domain expert agent.

[0124] In one embodiment, the trend analysis agent is used to concatenate the semantically enhanced embedding vector with the trend prediction vector, capture temporal dependencies and trend features through a Transformer encoder, and output a trend analysis feature vector.

[0125] In one specific embodiment, the trend analysis feature vector is obtained by the following formula:

[0126] ,in, This represents the feature vector for trend analysis, used to capture time-series trend information; The parameter is The Transformer encoder is the same as the standard Transformer encoder, including a multi-head attention layer, a feedforward neural network, and a residual connection structure; This is the set of trainable parameters for the Transformer.

[0127] The semantic association agent is used to input semantic enhancement embedding vectors into the semantic association agent and output semantic association feature vectors. The specific steps are as follows: a semantic relationship graph is constructed based on the semantic enhancement embedding vectors. The points of the semantic relationship graph are word segments, and the edges represent the semantic or syntactic relationships between the word segments. The semantic relationship graph is input into a graph attention network to output semantic association feature vectors.

[0128] In one specific embodiment, the semantic association feature vector is obtained by the following formula:

[0129] ,in, It represents a semantic association feature vector, which captures the semantic association information of words in their neighborhood; The parameter is The graph attention network, like the standard graph attention network, uses an attention mechanism to aggregate neighbor node information; The set of trainable parameters for a graph attention network; Represents a semantically enhanced embedding vector; This represents the function for constructing a semantic relation graph.

[0130] In one specific embodiment By using a co-occurrence matrix to count the frequency of word pairs appearing together within a specific window, and by using dependency parsing to analyze the grammatical dependencies between words, for example, for the sentence "AI intelligent agent analysis industry", the co-occurrence matrix may show that "AI" and "intelligent agent" co-occur frequently, and dependency parsing may show that "AI" modifies "intelligent agent".

[0131] In a specific implementation, a semantic relation graph construction function is defined. The implementation method is as follows:

[0132] First, word co-occurrence frequency is calculated based on a sliding window, and syntactic relations are extracted from the dependency parsing results;

[0133] Then, the co-occurrence frequency is normalized into a weight, and a fixed weight is assigned to the existing dependency relationship. The two are added together to obtain the comprehensive edge weight.

[0134] Finally, an empirical threshold is set, such as a weight ≥ 0.5, and only edges with weights higher than this threshold are retained, thereby constructing a semantic relationship graph.

[0135] The nodes of the semantic relation graph represent words, and the edges represent semantic or grammatical relationships between words. The input is an embedding vector, and the output is a semantic relation graph, representing the semantic relation graph construction function.

[0136] In one specific embodiment, the co-occurrence matrix counts the frequency of word pairs co-occurring within a specific window, and dependency parsing analyzes the grammatical dependencies between words. For example, for the sentence "AI intelligent agent analysis industry", the co-occurrence matrix may show that "AI" and "intelligent agent" co-occur frequently, and dependency parsing may show that "AI" modifies "intelligent agent".

[0137] The domain expert agent is used to obtain a task description, which includes task ID information; the task ID information is passed through the embedding layer to obtain a task embedding vector; the task embedding vector and the semantically enhanced embedding vector are input into a hybrid expert layer to obtain expert features, which includes a gating network and multiple experts, and the expert features are routed to different domain expert networks through the gating network to output a domain expert feature vector;

[0138] In one specific embodiment, the domain expert feature vector is obtained using the following formula:

[0139] , This represents the feature vector of a domain expert, which contains domain-specific information. The parameter is The hybrid expert layer includes multiple expert networks and a gating network. The expert networks specifically adopt a fully connected layer structure, and the gating network calculates expert weights based on the task embedding vector and the semantic enhancement embedding vector. A set of trainable parameters for a hybrid expert layer; This represents the task embedding vector, used for routing to different domain experts.

[0140] In one specific embodiment, the task embedding vector The calculation is shown in the following formula:

[0141] ,

[0142] in, Represents the task embedding vector; Indicates the embedding layer; This indicates the task ID information.

[0143] The trend analysis feature vector, the semantic association feature vector, and the domain expert feature vector are multiplied element by element to achieve feature cross-multiplication, and a collaborative feature vector containing domain expert features is output through a linear projection layer.

[0144] In one specific embodiment, the collaborative feature vector is obtained by the following formula:

[0145] ,in, This represents a collaborative feature vector that can integrate trend, semantic, and domain features; is the weight matrix of the linear projection layer, and are trainable parameters; The bias vector of the linear projection layer is a trainable parameter; This indicates element-wise multiplication.

[0146] In one embodiment, the collaborative feature vector is input into a dynamic task allocation mechanism. The dynamic task allocation mechanism refers to extracting sample task demand vectors from the collaborative feature vectors through a mapping weight matrix, calculating the KL divergence between the sample task demand vectors and the capability prototype vectors of each AI agent, and obtaining the allocation weight matrix through normalization using a function and a temperature coefficient. The AI ​​agent capability prototype vectors refer to trend analysis feature vectors, semantic association feature vectors, and domain expert feature vectors.

[0147] In one specific embodiment, the weight allocation matrix is ​​obtained by the following formula:

[0148] ,in, The first element of the weight distribution matrix represents the weight distribution matrix. Line number The element of the column, i.e., the first The sample is assigned to the first The weights of each agent; For sample index; For indexing intelligent agents, Each corresponds to one of three intelligent agents; This represents the KL divergence, used to measure the difference between two probability distributions; Indicates the first Each sample's task requirement vector represents its task characteristics. The collaborative feature vectors are mapped using a weight matrix. The mapping weight matrix is ​​obtained by nonlinear mapping extraction and is a trainable parameter that is learned during training. The collaborative feature vector can be mapped to the task requirement space to represent the task characteristics of the sample so as to match it with the agent's capability prototype. The KL divergence is then calculated to achieve dynamic task allocation. Specifically, random initialization is used as the initial value and updated through training. Indicates the first A prototype vector of an agent's capabilities represents the core capability characteristics of the agent. Indicates the first A prototype vector of the capabilities of an agent; To distinguish from The intelligent agent index; This is a temperature coefficient that controls the smoothness of the distribution. Its value ranges from [0, 10]. A larger value results in a more uniform distribution. (Preferred value would be...) .

[0149] The trend analysis feature vector, the semantic association feature vector, and the domain expert feature vector are weighted according to the weighting matrix to obtain a weighted enhanced feature vector.

[0150] In one specific embodiment, the weighted enhanced feature vector is obtained by the following formula:

[0151] , , ,in, The first element of the weight distribution matrix represents the weight distribution matrix. The element in the first column of the row, i.e. Each sample is assigned a weight to the first agent; The first element of the weight distribution matrix represents the weight distribution matrix. The element in the second column of the row, i.e. The weights assigned to the second agent for each sample; The first element of the weight distribution matrix represents the weight distribution matrix. The element in the 3rd column of the row, i.e. The weights assigned to the third agent for each sample; For the first The trend analysis feature vector of the nth sample is for the nth sample. The trend feature vector of each sample, i.e., the trend feature vector In the Specific instances on a sample; For the first The semantic association feature vector of the nth sample is for the nth sample. The semantic association feature vector of each sample, i.e., the semantic association feature vector In the Specific instances on a sample; For the first The domain expert feature vector for the nth sample is for the nth sample. The domain expert feature vector of each sample, i.e., the domain expert feature vector In the Specific instances on a sample; For the first Weighted enhanced trend analysis feature vectors for each sample; For the first The weighted enhanced semantic association feature vector of each sample; For the first Weighted enhanced domain expert feature vectors for each sample.

[0152] The weighted enhanced feature vectors of all samples are averaged and concatenated, and then a gated vector is generated by an activation function.

[0153] In one specific embodiment, the activation function is the Sigmoid activation function.

[0154] In one embodiment, the gating vector is obtained by the following formula:

[0155] ,in, This represents the gate vector, used to control the weights of feature fusion; is the weight matrix of the gated layer, which are trainable parameters; The bias vector of the gated layer is a trainable parameter; To obtain the weighted enhanced trend analysis feature vector, it is calculated by averaging the weighted features of all samples in the current batch:

[0156] , To obtain the weighted and enhanced semantic association feature vector, it is calculated by averaging the weighted features of all samples in the current batch:

[0157] , The weighted enhancement feature vector for domain experts is obtained by averaging the weighted features of all samples in the current batch.

[0158] ,

[0159] in, This is the set of sample indices for the training batch; This indicates the size of the training batch.

[0160] The trend analysis feature vector is compressed using a multilayer perceptron to obtain compressed trend features. The compressed semantic expert features are obtained by concatenating the compressed semantic association feature vector and the domain expert feature vector. The compressed trend features and the compressed semantic expert features are weighted and summed using the gate vector and the complement vector of the gate vector. After layer normalization, a collaborative feature vector with gate fusion is output.

[0161] In one specific embodiment, the collaborative feature vector with gated fusion is obtained by the following formula:

[0162] ,in, This represents a collaborative feature vector with gated fusion, which fuses features from three agents; The representation layer normalization operation is used to stabilize the training process; This represents a multilayer perceptron used for feature dimension compression. It contains multiple fully connected layers, such as two fully connected layers, and uses the ReLU activation function. It is a vector of all 1s with the same dimension as the gate vector; This indicates taking the complement of the gated vector, used to control the weights of the second part of the features.

[0163] In one specific embodiment, four methods—single model, ensemble learning, multi-task learning, and the heterogeneous agent cooperative network proposed in this invention—are compared. Evaluation metrics include accuracy (the proportion of correctly predicted samples), recall (the proportion of positive examples correctly predicted), and F1 score (the harmonic mean of accuracy and recall). All metrics are expressed as percentages; higher values ​​indicate better performance. The results are shown in Table 2 and... Figure 5 As shown, experimental results demonstrate that the technology of this invention significantly outperforms other comparative methods across all three evaluation metrics. The single model exhibits the most limited performance across all metrics, reflecting the inherent limitation of a single model in covering the multi-dimensional classification needs of industry intelligence. Ensemble learning methods achieve some performance improvement by combining multiple models, but the improvement is limited. Multi-task learning methods achieve further improvement by learning to transfer knowledge between different tasks through shared representation learning, but still do not reach the ideal results. The heterogeneous agent collaborative network proposed in this invention achieves feature complementarity and collaborative decision-making through the organic collaboration of trend analysis agents, semantic association agents, and domain expert agents. Therefore, it achieves the highest values ​​in accuracy, recall, and F1 score, proving the effectiveness and superiority of the multi-agent collaborative modeling mechanism.

[0164] Table 2 Performance Comparison of Different Modeling Methods

[0165] Single model Ensemble learning Multi-task learning This invention Accuracy (%) 70.7 80.9 83.6 93.8 Recall rate (%) 71.2 75.0 78.9 87.8 F1 score (%) 70.0 76.9 81.8 91.3

[0166] Figure 2This is a schematic flowchart of the method for constructing an industry intelligence analysis model based on AI intelligent agents provided in this embodiment of the invention. Specifically, it includes the following steps:

[0167] S91: Obtain industry intelligence data, the data including tags;

[0168] In one embodiment, the tags are primary and secondary category tags for industry intelligence.

[0169] S92: The industrial intelligence data is processed by the AI-based intelligent agent industrial intelligence data processing method to obtain a collaborative feature vector;

[0170] S93: Input the collaborative feature vector into the fully connected layer to obtain the predicted classification result, compare the predicted classification result with the label to generate the total loss function, optimize the total loss function, and obtain the industry intelligence analysis model.

[0171] In one embodiment, the specific steps for generating the total loss function are as follows: inputting the collaborative feature vector into the first fully connected layer, calculating the first-level category prediction probability distribution through an activation function, concatenating the collaborative feature vector with the first-level category prediction probability and inputting it into the second fully connected layer, calculating the second-level category prediction probability distribution through an activation function, jointly calculating the first-level category prediction loss and the second-level category prediction loss through a weighted cross-entropy loss function, and balancing the importance of the first-level category prediction loss and the second-level category prediction loss using hierarchical weight coefficients to obtain the total loss function; the range of the hierarchical weight coefficients is [0,1].

[0172] In one specific embodiment, the first-level category prediction probability is obtained by the following formula:

[0173] , Indicates that the sample belongs to the first Predicted probabilities for each primary category; This represents the Softmax activation function, which converts the output into a probability distribution. Indicates the first The weight matrix corresponding to each primary category is a trainable parameter; Indicates the first The bias vectors corresponding to each first-level category are trainable parameters; Indicates the first-level category index. ,like Corresponding to "technology", Corresponding to "finance"; This indicates the total number of first-level categories.

[0174] In one specific embodiment, the secondary category prediction probability is obtained by the following formula:

[0175] , Indicates that the sample belongs to the first Predicted probabilities for each secondary category; Indicates the first The weight matrix corresponding to each secondary category is a trainable parameter; Indicates that the sample belongs to the first The predicted probability of the first-level category is the... The predicted probability of the parent category corresponding to each second-level category; This represents a mapping function that returns the first... For example, if the second-level category "AI chip" maps to the first-level category "technology", then if the index of "technology" is 1, Specifically, it is obtained through a predefined mapping table; Indicates the first The bias vectors corresponding to each second-level category are trainable parameters; Indicates a second-level category index. This indicates the total number of secondary categories.

[0176] In one specific embodiment, the total number of first-level categories and the total number of second-level categories are obtained from the label hierarchy of the dataset. For example, if there are 3 first-level categories in the dataset, then... If there are 10 secondary categories, then .

[0177] In one embodiment, the formula for the total loss function is:

[0178] ,in, Represents the total loss function; This represents the hierarchical weight coefficient, used to balance the importance of the first and second-level classification losses. Its value ranges from [0,1], with the preferred value being... This represents the cross-entropy loss function, which measures the difference between the predicted probability distribution and the true distribution, encouraging the model to output a correct classification with high confidence. This represents the probability distribution of the first-level category prediction. It is the first-level category prediction probability distribution vector, and is the th element in it. One element; This represents the true label of the primary category, in one-hot format; This represents the probability distribution of second-level category predictions. It is the second-level category prediction probability distribution vector, and is the th element in it. One element; This represents the actual label for the secondary category, and it is in one-hot format.

[0179] In one embodiment, the method further includes feedback-driven online optimization, calculating the gradient of the total loss function, applying consistency constraints to the predicted classification results to update the model parameters, and updating the agent capability prototype vector to achieve online model optimization.

[0180] In a specific implementation, based on the loss gradient and the output consistency constraint gradient, the model parameters are updated by minimizing the KL divergence of the output probability distributions of the old and new models, and then combined with the learning rate and regularization strength coefficient. The formula is as follows:

[0181]

[0182] in, Indicates the first The model parameters for the next iteration, including all trainable parameters; Index for the number of iterations during model training; Indicates the first The model parameters for the next iteration, including all trainable parameters; The learning rate controls the step size for updating parameters; the optimal step size is... Indicates the loss function at The gradient at a given point represents the direction of change in the current loss.

[0183] The regularization strength coefficient adjusts the influence of the regularization term, and is preferred. ;

[0184] The gradient representing the output consistency constraint, which constrains the consistency of the outputs of the old and new models, is obtained by minimizing the KL divergence of the output probability distributions of the old and new models, and is expressed as: Ensure that the output does not change abruptly when the model is updated; This represents the KL divergence with respect to the model parameters. The gradient is used to constrain the consistency of the outputs of the old and new models. During computation, the output probability distribution is frozen. As a reference distribution, the output probability distribution is calculated through backpropagation. For model parameters The gradient is obtained using approximate calculation. Indicates about model parameters The gradient; For the first Model parameters for the next iteration The corresponding output probability distribution; For the first Model parameters The corresponding output probability distribution.

[0185] In one specific embodiment, based on the task requirement vector of the current batch of samples, the agent capability prototype vector is updated by a smoothing coefficient and a moving average of historical prototype vectors. The update formula for the agent capability prototype vector is as follows:

[0186] ,

[0187] in, This indicates an assignment operation; The smoothing coefficient controls the degree to which historical information is retained; the preferred option is... This is the set of sample indices for the training batch; Indicates the size of the training batch; Indicates the first A prototype vector of the capabilities of an agent; Indicates the first The task requirement vector for each sample.

[0188] In one specific embodiment, a training stop threshold is set during model training. When the validation loss no longer decreases significantly within a preset number of iterations (e.g., 10 iterations) (the decrease is less than the threshold, e.g., 0.001), or the validation accuracy stabilizes at the peak value (fluctuation less than 0.5%), the model is determined to have converged and training is stopped. At the same time, a maximum number of iterations (e.g., 1000 iterations) is set as a hard stopping condition to prevent overfitting or infinite loops.

[0189] In one specific embodiment, the performance of industry intelligence trend prediction is compared to evaluate the accuracy of LSTM, TCN, and the method of this invention in industry intelligence trend prediction. (See [link to documentation]). Figure 6 The horizontal axis, "Time Step," is in days (0-30 days), and the vertical axis, "Trend Strength," quantifies changes in industry indicators (normalized value -0.9 to 0.9). Red stars mark key inflection points in the true trend (days 3, 14, and 23). The true trend curve (black) exhibits non-stationary fluctuations. The conventional LSTM method (square line) shows significant lag in prediction results, with an average deviation of 0.3 units at inflection points. While the TCN method (circle line) improves the lag, it overreacts to sudden changes (e.g., overshoot of 0.25 on day 20). This technique (diamond line) perfectly matches the three key inflection points, with a maximum deviation of only 0.1. Especially at the downward inflection point on day 12, this technique accurately captures the turning point, while other methods lag by 1-2 days. Experimental results show that the dual-channel design of local fluctuations and global trends, combined with a trend-sensitive factor, effectively solves the problem of insufficient response to sudden inflection points by conventional methods.

[0190] Figure 3 This is a schematic flowchart of an industry intelligence data analysis method based on AI intelligent agents provided in an embodiment of the present invention. Specifically, it includes the following steps:

[0191] S301: Obtain industry intelligence data;

[0192] S302: Process the industry intelligence data using the AI-based intelligent agent-based industry intelligence data processing method to obtain a collaborative feature vector;

[0193] S303: Input the collaborative feature vector into the industry intelligence analysis model to obtain the predicted classification result.

[0194] This invention provides an industrial intelligence data processing device based on an AI intelligent agent, comprising: a memory and a processor, such as... Figure 7 As shown:

[0195] Memory is used to store program instructions;

[0196] The processor is used to call program instructions, which, when executed, implement an AI-based intelligent agent-based industrial intelligence data processing method, or an AI-based intelligent agent-based industrial intelligence analysis model construction method, or implement the steps of the industrial intelligence data analysis method.

[0197] The present invention provides a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements the AI-based intelligent agent-based industrial intelligence data processing method, or the AI-based intelligent agent-based industrial intelligence analysis model construction method, or the steps of the industrial intelligence data analysis method.

[0198] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0199] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0200] The modules described as separate components may or may not be physically separate. Similarly, the components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0201] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The aforementioned integrated modules can be implemented in hardware or as software functional modules.

[0202] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, which may include: read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0203] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware, and the program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0204] The computer device provided by the present invention has been described in detail above. For those skilled in the art, there will be changes in the specific implementation and application scope based on the ideas of the embodiments of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. An AI agent-based industry intelligence data processing method, characterized by, Comprise: S1: obtain industrial information data, the industrial information data comprises a DOM tree structure; S2: based on the text content and the node position information of the DOM tree in the DOM tree, generate text embedding vectors and position coding vectors; S3: concatenate the text embedding vectors and the position coding vectors to obtain semantic enhanced embedding vectors; S4: processing the semantic enhanced embedding vectors using a temporal convolution network to capture local fluctuation feature vectors and global trend feature vectors; S5: concatenate the local fluctuation feature vectors and the global trend feature vectors to obtain trend prediction vectors; S6: concatenate the semantic enhanced embedding vectors and the trend prediction vectors to output trend analysis feature vectors; input the semantic relationship graph constructed based on the semantic enhanced embedding vectors into a graph attention network to output semantic correlation feature vectors; S7: cross the trend analysis feature vectors and the semantic correlation feature vectors to output collaborative feature vectors. 2.The AI agent-based industry intelligence data processing method of claim 1, wherein, The specific steps of S2 include: S21: The DOM tree is a hierarchical structure representation, the text content of different hierarchical DOM trees is segmented to obtain text segmentation results, and the text segmentation results are obtained through an embedding matrix to obtain text embedding vectors of different levels; according to the level of the DOM tree node, the structural position information of the node in the DOM tree is captured through a position encoder to generate position coding vectors of different levels; Preferably, the specific steps of S3 include: S31: concatenate the text embedding vectors of different levels and the position coding vectors of different levels to obtain semantic enhanced embedding vectors. 3.The AI agent-based industry intelligence data processing method of claim 2, wherein, The specific steps of S31 are: statistically frequency of each word appearing in the text segmentation results and calculate the probability distribution of each word, calculate the text information entropy of each word according to the probability distribution of each word; scale the text information entropy to obtain semantic weight coefficients of different levels; use the semantic weight coefficients to weight sum the concatenated results to obtain semantic enhanced embedding vectors that fuse text semantics and structure information; Preferably, the specific steps of S4 include: S41: combine all semantic-enhanced embedding vectors within a time window of 24 hours into a recent history embedding vector sequence, and input the recent history embedding vector sequence into a first temporal convolution network for local fluctuation feature extraction, outputting a local fluctuation feature vector; combine all semantic-enhanced embedding vectors within a time window of 24 hours into a recent history embedding vector sequence, and input the recent history embedding vector sequence into a first temporal convolution network for local fluctuation feature extraction, outputting a local fluctuation feature vector; S42: The semantically enhanced embedding vector is compared with the semantically enhanced embedding vector at a time interval of... All semantically enhanced embedding vectors within a day are combined into a long-term historical embedding vector sequence, and the long-term historical embedding vector sequence is input into a second temporal convolutional network for global trend feature extraction, outputting a global trend feature vector. Preferably, said is a natural integer greater than 7; Preferred, .

4. The industrial information data processing method based on AI agent according to claim 3, characterized in that: calculating a gradient of the global trend feature vector with respect to time t, calculating a norm of the gradient using an activation function and a sensitivity coefficient mapping the norm to a trend change sensitivity factor, the sensitivity coefficient ranging from [0, 1.0] concatenate and fuse the local fluctuation feature vectors and the global trend feature vectors, weight the concatenation and fusion results element by element using a trend change sensitive factor, and output trend prediction vectors; Preferred ; Optionally, the activation function includes one or more of the following: Sigmoid function, Softmax function, tanh function, ReLU function, Leaky ReLU function, ELU function.

5. The industrial information data processing method based on AI agent according to claim 4, characterized in that: the AI agent comprises a trend analysis agent, a semantic correlation agent and a domain expert agent; optionally, the trend analysis agent is used to concatenate the semantic enhanced embedding vectors and the trend prediction vectors, capture time sequence dependence and trend features through a Transformer encoder, and output trend analysis feature vectors; Optionally, the semantic association agent is configured to input the semantic enhancement embedding vector into the semantic association agent to output a semantic association feature vector, and the specific steps are as follows: constructing a semantic relationship graph based on the semantic enhancement embedding vector, wherein the points of the semantic relationship graph are segmented words, and the edges represent semantic or grammatical relationships between the segmented words; and inputting the semantic relationship graph into a graph attention network to output the semantic association feature vector. Optionally, the domain expert agent is configured to obtain a task description, wherein the task description comprises task ID information; obtain a task embedding vector by inputting the task ID information into the embedding layer; input the task embedding vector and the semantic enhancement embedding vector into a hybrid expert layer to obtain an expert feature, wherein the hybrid expert layer comprises a gating network and a plurality of experts; route the expert feature to different domain expert networks through the gating network to output a domain expert feature vector. Optionally, the trend analysis feature vector, the semantic association feature vector, and the domain expert feature vector are multiplied element by element to realize feature cross, and a collaborative feature vector containing a domain expert feature is output through a linear projection layer. 6.The AI agent-based industry intelligence data processing method of claim 5, wherein, The method further comprises: inputting the collaborative feature vector into a dynamic task allocation mechanism, wherein the dynamic task allocation mechanism refers to extracting a sample task demand vector from the collaborative feature vector through a mapping weight matrix, calculating the KL divergence between the sample task demand vector and an AI agent capability prototype vector, and obtaining a distribution weight matrix through function and temperature coefficient normalization; the AI agent capability prototype vector refers to a trend analysis feature vector, a semantic association feature vector, and a domain expert feature vector; weighting the trend analysis feature vector, the semantic association feature vector, and the domain expert feature vector according to the distribution weight matrix to obtain a weighted enhancement feature vector; averaging and splicing the weighted enhancement feature vectors of all samples, and generating a gating vector through an activation function; using a multi-layer perception machine to compress the trend analysis feature vector to obtain a compressed trend feature, and compressing the spliced result of the semantic association feature vector and the domain expert feature vector to obtain a compressed semantic expert feature; weighting and summing the compressed trend feature and the compressed semantic expert feature using the gating vector and a complementary vector of the gating vector, and outputting a collaborative feature vector with gating fusion through layer normalization. 7.A method for constructing an AI agent-based industry information analysis model, characterized in that, comprises: S91: obtaining industrial intelligence data, wherein the data comprises a label; S92: processing the industrial intelligence data by the method of any one of claims 1-6 to obtain a collaborative feature vector; S93: inputting the collaborative feature vector into a fully connected layer to obtain a predicted classification result, comparing the predicted classification result with the label to generate a total loss function, optimizing the total loss function, and obtaining an industrial intelligence analysis model. 8.An AI agent-based industry intelligence data analysis method, characterized in that, comprises: obtaining industrial intelligence data; inputting the industrial intelligence data into the industrial intelligence analysis model in claim 7 for analysis to obtain a predicted classification result.

9. A computer device, comprising: comprises: a memory and a processor; the memory is configured to store program instructions; The processor is configured to invoke program instructions which, when executed by the processor, implement the method steps of any one of claims 1-8.

10. A computer program product comprising a computer program, characterized in that, The computer program, when executed by the processor, implements the method steps of any one of claims 1-8.

Images